[Public release 26/04] Introducing Mega MoE, FP4 Indexer and other features/fixes (#304)

* Merge with private repo

* Update README

* Update README

* Update README

* Add PyTorch requirements

* Fix sync scopes for MQA logits (#256)

* Update README

deep_gemm/include/deep_gemm/comm/barrier.cuh

@@ -0,0 +1,74 @@
+#pragma once
+
+#include <deep_gemm/ptx/ld_st.cuh>
+#include <deep_gemm/layout/sym_buffer.cuh>
+#include <deep_gemm/layout/mega_moe.cuh>
+
+namespace deep_gemm::comm {
+
+template <uint32_t kNumSMs, uint32_t kGridSyncIndex = 0, typename sync_scope_t>
+CUTLASS_DEVICE void grid_sync(const layout::Workspace& workspace,
+                              const uint32_t& sm_idx, const uint32_t& thread_idx,
+                              const sync_scope_t& sync_scope) {
+    // NOTES: the implementation idea is from `cooperative_groups::this_grid().sync()`
+    static constexpr uint32_t kFinishSumTag = 0x80000000u;
+    sync_scope();
+    if (thread_idx == 0) {
+        const auto count_ptr = workspace.get_grid_sync_count_ptr<kGridSyncIndex>();
+        const auto old_value = ptx::atomic_add_rel(
+            count_ptr, sm_idx == 0 ? (kFinishSumTag - (kNumSMs - 1)) : 1);
+        uint32_t new_value;
+        do {
+            new_value = ptx::ld_acq(count_ptr);
+        } while (((new_value ^ old_value) & kFinishSumTag) == 0);
+    }
+    sync_scope();
+}
+
+template <uint32_t kNumRanks, uint32_t kNumSMs, uint32_t kNumThreads, uint32_t kGridSyncIndex, uint32_t kTag, typename sync_scope_t>
+CUTLASS_DEVICE void nvlink_barrier(const layout::Workspace& workspace,
+                                   const layout::SymBuffer<kNumRanks>& sym_buffer,
+                                   const uint32_t& sm_idx, const uint32_t& thread_idx,
+                                   const sync_scope_t& sync_scope,
+                                   const bool& sync_prologue = true,
+                                   const bool& sync_epilogue = true) {
+    DG_STATIC_ASSERT(kNumRanks <= kNumThreads, "Insufficient threads");
+
+    // Grid sync before NVLink signaling
+    if (sync_prologue)
+        grid_sync<kNumSMs, kGridSyncIndex>(workspace, sm_idx, thread_idx, sync_scope);
+
+    // NVLink cross-rank barrier, only SM 0 participates
+    if (sm_idx == 0) {
+        auto* counter_ptr = workspace.get_nvl_barrier_counter_ptr();
+        const auto status = (*counter_ptr) & 3;
+        const auto signal_phase = status & 1, signal_sign = status >> 1;
+        auto* signal_ptr = workspace.get_nvl_barrier_signal_ptr(signal_phase);
+
+        // Send signals to remote ranks
+        if (thread_idx < kNumRanks)
+            ptx::red_add_rel_sys(sym_buffer.map(signal_ptr, thread_idx), signal_sign ? -1 : 1);
+        sync_scope();
+
+        // Update status and wait arrival (with 30s timeout, at 2 GHz)
+        constexpr int64_t kNumTimeoutCycles = 30ll * 2000000000ll;
+        if (thread_idx == 0) {
+            ptx::red_add(counter_ptr, 1);
+            const int target = signal_sign ? 0 : static_cast<int>(kNumRanks);
+            const auto start_clock = clock64();
+            while (ptx::ld_acq_sys(signal_ptr) != target) {
+                if (clock64() - start_clock >= kNumTimeoutCycles) {
+                    printf("DeepGEMM NVLink barrier timeout (30s): rank=%d, counter=%d, signal=%d, target=%d, phase=%d, sign=%d, tag=%d\n",
+                           sym_buffer.rank_idx, *counter_ptr, ptx::ld_acq_sys(signal_ptr), target, signal_phase, signal_sign, kTag);
+                    DG_DEVICE_ASSERT(false and "NVLink barrier timeout");
+                }
+            }
+        }
+    }
+
+    // Grid sync after NVLink completion
+    if (sync_epilogue)
+        grid_sync<kNumSMs, kGridSyncIndex>(workspace, sm_idx, thread_idx, sync_scope);
+}
+
+} // namespace deep_gemm::comm

deep_gemm/include/deep_gemm/common/compile.cuh

@@ -0,0 +1,18 @@
+#pragma once
+
+#include <cutlass/detail/helper_macros.hpp>
+
+#if defined(__NVCC__) or (defined(__clang__) and defined(__CUDA__)) or defined(__CUDACC_RTC__) or defined(__CLION_IDE__)
+#define DG_IN_CUDA_COMPILATION
+#endif
+
+#if defined(__NVCC__) || (defined(__clang__) and defined(__CUDA__))
+#define CUTLASS_HOST_DEVICE_NOINLINE  __device__ __host__
+#define CUTLASS_DEVICE_NOINLINE __device__
+#elif defined(__CUDACC_RTC__)
+#define CUTLASS_HOST_DEVICE_NOINLINE __device__
+#define CUTLASS_DEVICE_NOINLINE __device__
+#else
+#define CUTLASS_HOST_DEVICE_NOINLINE
+#define CUTLASS_DEVICE_NOINLINE
+#endif

deep_gemm/include/deep_gemm/common/cute_tie.cuh

@@ -1,5 +1,7 @@
 #pragma once
 
+#include <cute/int_tuple.hpp>
+
 namespace cute {
 
 struct ignore_t {

deep_gemm/include/deep_gemm/common/exception.cuh

@@ -0,0 +1,43 @@
+#pragma once
+
+#include <cuda/std/cstdint>
+#include <deep_gemm/common/compile.cuh>
+
+#ifdef __CLION_IDE__
+
+CUTLASS_HOST_DEVICE void host_device_printf(const char* format, ...) {
+    asm volatile("trap;");
+}
+
+#define printf host_device_printf
+#endif
+
+#ifndef DG_DEVICE_ASSERT
+#define DG_DEVICE_ASSERT(cond) \
+do { \
+    if (not (cond)) { \
+        printf("Assertion failed: %s:%d, condition: %s\n", __FILE__, __LINE__, #cond); \
+        asm("trap;"); \
+    } \
+} while (0)
+#endif
+
+#ifndef DG_TRAP_ONLY_DEVICE_ASSERT
+#define DG_TRAP_ONLY_DEVICE_ASSERT(cond) \
+do { \
+    if (not (cond)) \
+        asm("trap;"); \
+} while (0)
+#endif
+
+#ifndef DG_STATIC_ASSERT
+#define DG_STATIC_ASSERT(cond, ...) static_assert(cond, __VA_ARGS__)
+#endif
+
+#ifndef DG_UNIFIED_ASSERT
+#ifdef DG_IN_CUDA_COMPILATION
+#define DG_UNIFIED_ASSERT(cond) DG_DEVICE_ASSERT(cond)
+#else
+#define DG_UNIFIED_ASSERT(cond) DG_HOST_ASSERT(cond)
+#endif
+#endif

deep_gemm/include/deep_gemm/common/math.cuh

@@ -0,0 +1,149 @@
+#pragma once
+
+#include <cuda/std/cstdint>
+#include <deep_gemm/common/compile.cuh>
+#include <deep_gemm/common/exception.cuh>
+
+namespace deep_gemm::math {
+
+/// Pointer operations
+template <typename dtype_t = void>
+CUTLASS_HOST_DEVICE dtype_t* advance_ptr(void* ptr, const uint64_t num_bytes) {
+    return reinterpret_cast<dtype_t*>(static_cast<uint8_t*>(ptr) + num_bytes);
+}
+
+/// Math functions
+template <typename T>
+CUTLASS_HOST_DEVICE T ceil_div(T a, T b) {
+    return (a + b - 1) / b;
+}
+
+template <typename T>
+CUTLASS_HOST_DEVICE constexpr T constexpr_ceil_div(T a, T b) {
+    return (a + b - 1) / b;
+}
+
+template <typename T, bool kDoCeilAlignment = true>
+CUTLASS_HOST_DEVICE T align(T a, T b) {
+    return (kDoCeilAlignment ? ceil_div(a, b) : (a / b)) * b;
+}
+
+template <typename T>
+CUTLASS_HOST_DEVICE constexpr T constexpr_align(T a, T b) {
+    return constexpr_ceil_div(a, b) * b;
+}
+
+template <typename T>
+CUTLASS_HOST_DEVICE constexpr T constexpr_gcd(T a, T b) {
+    return b == 0 ? a : constexpr_gcd(b, a % b);
+}
+
+template <typename T>
+CUTLASS_HOST_DEVICE constexpr T constexpr_min(T a, T b) {
+    return a < b ? a : b;
+}
+
+template <typename T>
+CUTLASS_DEVICE void swap(T& a, T& b) {
+    T temp = a;
+    a = b;
+    b = temp;
+}
+
+#ifdef DG_IN_CUDA_COMPILATION
+CUTLASS_DEVICE float2 fma2(const float2& a, const float2& b, const float2& c) {
+#if defined(__CUDA_ARCH__) and (__CUDA_ARCH__ >= 1000)
+    return __ffma2_rn(a, b, c);
+#else
+    return make_float2(
+        __fmaf_rn(a.x, b.x, c.x),
+        __fmaf_rn(a.y, b.y, c.y)
+    );
+#endif
+}
+
+CUTLASS_HOST_DEVICE float fast_rcp(const float& x) {
+    float ret;
+    asm volatile("rcp.approx.ftz.f32 %0, %1;" : "=f"(ret) : "f"(x));
+    return ret;
+}
+
+/// Casting
+template <typename old_t>
+CUTLASS_DEVICE int cast_into_bf16_and_pack(old_t& x, old_t& y) {
+    auto bf16x2 = __float22bfloat162_rn({*reinterpret_cast<float*>(&x), *reinterpret_cast<float*>(&y)});
+    return *reinterpret_cast<int*>(&bf16x2);
+}
+
+CUTLASS_DEVICE float fast_pow2(const int& x) {
+    uint32_t bits_x = (x + 127) << 23;
+    return *reinterpret_cast<float*>(&bits_x);
+}
+
+CUTLASS_DEVICE int fast_log2_ceil(float x) {
+    const auto bits = *reinterpret_cast<uint32_t*>(&x);
+    const auto exp = bits >> 23;
+    const auto man = bits & ((1 << 23) - 1);
+    return exp - 127 + (man != 0);
+}
+
+template <bool kUseUE8M0 = true>
+CUTLASS_DEVICE void get_e4m3_sf_and_sf_inv(const float2& amax, float2& sf, float2& sf_inv) {
+    DG_STATIC_ASSERT(kUseUE8M0, "Must use UE8M0");
+    const float2 finfo_factor = {1.0 / 448.0, 1.0 / 448.0};
+    const auto scaled = __fmul2_rn(amax, finfo_factor);
+    const auto exp_x = fast_log2_ceil(scaled.x);
+    const auto exp_y = fast_log2_ceil(scaled.y);
+    sf.x = fast_pow2(exp_x), sf_inv.x = fast_pow2(-exp_x);
+    sf.y = fast_pow2(exp_y), sf_inv.y = fast_pow2(-exp_y);
+}
+
+/// Reduction
+CUTLASS_DEVICE uint32_t warp_inclusive_sum(uint32_t value, const uint32_t& lane_idx) {
+    #pragma unroll
+    for (uint32_t offset = 1; offset < 32; offset <<= 1) {
+        const uint32_t synced = __shfl_up_sync(0xffffffff, value, offset);
+        if (lane_idx >= offset)
+            value += synced;
+    }
+    return value;
+}
+
+// Operation functors
+template <typename T> struct ReduceSum { CUTLASS_DEVICE T operator()(T a, T b) const { return a + b; } };
+template <typename T> struct ReduceMax { CUTLASS_DEVICE T operator()(T a, T b) const { return a > b ? a : b; } };
+template <typename T> struct ReduceMin { CUTLASS_DEVICE T operator()(T a, T b) const { return a < b ? a : b; } };
+template <typename T> struct ReduceAnd { CUTLASS_DEVICE T operator()(T a, T b) const { return a & b; } };
+template <typename T> struct ReduceOr  { CUTLASS_DEVICE T operator()(T a, T b) const { return a | b; } };
+
+// Unified reduction function
+template <uint32_t kNumLanesPerGroup, bool kIntergroupReduce, typename T, typename Op>
+CUTLASS_DEVICE T warp_reduce(T value, Op op) {
+    DG_STATIC_ASSERT(kNumLanesPerGroup == 32 or kNumLanesPerGroup == 16 or kNumLanesPerGroup == 8 or
+                     kNumLanesPerGroup ==  4 or kNumLanesPerGroup == 2  or kNumLanesPerGroup == 1,
+                     "Invalid number of lanes");
+    constexpr uint32_t mask = 0xffffffff;
+    if constexpr (kIntergroupReduce) {
+        if constexpr (kNumLanesPerGroup <=  1) value = op(value, __shfl_xor_sync(mask, value,  1));
+        if constexpr (kNumLanesPerGroup <=  2) value = op(value, __shfl_xor_sync(mask, value,  2));
+        if constexpr (kNumLanesPerGroup <=  4) value = op(value, __shfl_xor_sync(mask, value,  4));
+        if constexpr (kNumLanesPerGroup <=  8) value = op(value, __shfl_xor_sync(mask, value,  8));
+        if constexpr (kNumLanesPerGroup <= 16) value = op(value, __shfl_xor_sync(mask, value, 16));
+    } else {
+        if constexpr (kNumLanesPerGroup >= 32) value = op(value, __shfl_xor_sync(mask, value, 16));
+        if constexpr (kNumLanesPerGroup >= 16) value = op(value, __shfl_xor_sync(mask, value,  8));
+        if constexpr (kNumLanesPerGroup >=  8) value = op(value, __shfl_xor_sync(mask, value,  4));
+        if constexpr (kNumLanesPerGroup >=  4) value = op(value, __shfl_xor_sync(mask, value,  2));
+        if constexpr (kNumLanesPerGroup >=  2) value = op(value, __shfl_xor_sync(mask, value,  1));
+    }
+    return value;
+}
+
+// Convenience aliases
+template <uint32_t kNumLanesPerGroup = 32, bool kIntergroupReduce = false, typename T>
+CUTLASS_DEVICE T warp_reduce_sum(T value) {
+    return warp_reduce<kNumLanesPerGroup, kIntergroupReduce, T>(value, ReduceSum<T>{});
+}
+#endif
+
+} // namespace deep_gemm

deep_gemm/include/deep_gemm/common/tma_copy.cuh

@@ -0,0 +1,92 @@
+#pragma once
+
+#include <cute/arch/copy_sm90_tma.hpp>
+#include <cute/arch/copy_sm100_tma.hpp>
+#include <cutlass/arch/barrier.h>
+
+#include <deep_gemm/common/exception.cuh>
+
+namespace deep_gemm::tma {
+
+template <uint32_t BLOCK_INNER, uint32_t kSwizzleMode, typename dtype_t>
+constexpr uint32_t get_inner_block_atom_size() {
+    return kSwizzleMode == 0 ? BLOCK_INNER : kSwizzleMode / sizeof(dtype_t);
+}
+
+template <uint32_t BLOCK_INNER, uint32_t BLOCK_OUTER,
+          uint32_t kSwizzleMode,
+          typename dtype_t, bool kIs3DTMA = false>
+CUTLASS_DEVICE void
+copy(void const* desc_ptr, cutlass::arch::ClusterTransactionBarrier* barrier_ptr,
+     dtype_t* smem_ptr, const uint32_t& inner_idx, const uint32_t& outer_idx,
+     const uint32_t& num_tma_multicast = 1, const uint32_t& batch_idx = 0) {
+    DG_STATIC_ASSERT(static_cast<uint64_t>(cute::TMA::CacheHintSm90::EVICT_NORMAL) ==
+                     static_cast<uint64_t>(cute::TMA::CacheHintSm100::EVICT_NORMAL), "Invalid cache hint");
+    constexpr uint32_t BLOCK_INNER_ATOM = get_inner_block_atom_size<BLOCK_INNER, kSwizzleMode, dtype_t>();
+
+    if constexpr (not kIs3DTMA) {
+        if (num_tma_multicast == 1) {
+            #pragma unroll
+            for (uint32_t i = 0; i < BLOCK_INNER / BLOCK_INNER_ATOM; ++ i) {
+                cute::SM90_TMA_LOAD_2D::copy(desc_ptr, reinterpret_cast<uint64_t*>(barrier_ptr),
+                                             static_cast<uint64_t>(cute::TMA::CacheHintSm100::EVICT_NORMAL),
+                                             smem_ptr + i * BLOCK_OUTER * BLOCK_INNER_ATOM,
+                                             inner_idx + i * BLOCK_INNER_ATOM, outer_idx);
+            }
+        } else {
+            #if (defined(__CUDA_ARCH__) and (__CUDA_ARCH__ >= 1000))
+                // 2-CTA function will send signals to the leader CTA only
+                #pragma unroll
+                for (uint32_t i = 0; i < BLOCK_INNER / BLOCK_INNER_ATOM; ++ i) {
+                    cute::SM100_TMA_2SM_LOAD_2D::copy(desc_ptr, reinterpret_cast<uint64_t*>(barrier_ptr),
+                                                      static_cast<uint64_t>(cute::TMA::CacheHintSm100::EVICT_NORMAL),
+                                                      smem_ptr + i * BLOCK_OUTER * BLOCK_INNER_ATOM,
+                                                      inner_idx + i * BLOCK_INNER_ATOM, outer_idx);
+                }
+            #elif (defined(__CUDA_ARCH__) and (__CUDA_ARCH__ >= 900))
+                if (cute::block_rank_in_cluster() == 0) {
+                    #pragma unroll
+                    for (uint32_t i = 0; i < BLOCK_INNER / BLOCK_INNER_ATOM; ++ i) {
+                        cute::SM90_TMA_LOAD_MULTICAST_2D::copy(desc_ptr, reinterpret_cast<uint64_t*>(barrier_ptr),
+                                                               (1 << num_tma_multicast) - 1, static_cast<uint64_t>(cute::TMA::CacheHintSm90::EVICT_NORMAL),
+                                                               smem_ptr + i * BLOCK_OUTER * BLOCK_INNER_ATOM,
+                                                               inner_idx + i * BLOCK_INNER_ATOM, outer_idx);
+                    }
+                }
+            #endif
+        }
+    } else {
+        if (num_tma_multicast == 1) {
+            #pragma unroll
+            for (uint32_t i = 0; i < BLOCK_INNER / BLOCK_INNER_ATOM; ++ i) {
+                cute::SM90_TMA_LOAD_3D::copy(desc_ptr, reinterpret_cast<uint64_t*>(barrier_ptr),
+                                            static_cast<uint64_t>(cute::TMA::CacheHintSm100::EVICT_NORMAL),
+                                            smem_ptr + i * BLOCK_OUTER * BLOCK_INNER_ATOM,
+                                            inner_idx + i * BLOCK_INNER_ATOM, outer_idx, batch_idx);
+            }
+        } else {
+            #if (defined(__CUDA_ARCH__) and (__CUDA_ARCH__ >= 1000))
+                // 2-CTA function will send signals to the leader CTA only
+                #pragma unroll
+                for (uint32_t i = 0; i < BLOCK_INNER / BLOCK_INNER_ATOM; ++ i) {
+                    cute::SM100_TMA_2SM_LOAD_3D::copy(desc_ptr, reinterpret_cast<uint64_t*>(barrier_ptr),
+                                                      static_cast<uint64_t>(cute::TMA::CacheHintSm100::EVICT_NORMAL),
+                                                      smem_ptr + i * BLOCK_OUTER * BLOCK_INNER_ATOM,
+                                                      inner_idx + i * BLOCK_INNER_ATOM, outer_idx, batch_idx);
+                }
+            #elif (defined(__CUDA_ARCH__) and (__CUDA_ARCH__ >= 900))
+                if (cute::block_rank_in_cluster() == 0) {
+                    #pragma unroll
+                    for (uint32_t i = 0; i < BLOCK_INNER / BLOCK_INNER_ATOM; ++ i) {
+                        cute::SM90_TMA_LOAD_MULTICAST_3D::copy(desc_ptr, reinterpret_cast<uint64_t*>(barrier_ptr),
+                                                               (1 << num_tma_multicast) - 1, static_cast<uint64_t>(cute::TMA::CacheHintSm90::EVICT_NORMAL),
+                                                               smem_ptr + i * BLOCK_OUTER * BLOCK_INNER_ATOM,
+                                                               inner_idx + i * BLOCK_INNER_ATOM, outer_idx, batch_idx);
+                    }
+                }
+            #endif
+        }
+    }
+}
+
+} // namespace deep_gemm::tma

deep_gemm/include/deep_gemm/common/types.cuh

@@ -0,0 +1,43 @@
+#pragma once
+
+#include <cute/arch/mma_sm100_desc.hpp>
+
+namespace deep_gemm {
+
+enum class MmaKind {
+    BF16        = 0,
+    MXFP8FP4    = 1,
+};
+
+constexpr CUTLASS_HOST_DEVICE int get_element_size(const MmaKind& mma_kind) {
+    switch (mma_kind) {
+        case MmaKind::BF16:     return 2;
+        case MmaKind::MXFP8FP4: return 1;
+        default: return 0;
+    }
+}
+
+enum class GemmType {
+    Normal                              = 0,
+    MGroupedContiguous                  = 1,
+    MGroupedMasked                      = 2,
+    KGroupedContiguous                  = 3,
+    Batched                             = 4,
+    MGroupedContiguousWithPsumLayout    = 5,
+};
+
+constexpr CUTLASS_HOST_DEVICE bool is_m_grouped_contiguous(const GemmType& gemm_type) {
+    switch (gemm_type) {
+        case GemmType::MGroupedContiguous:                  return true;
+        case GemmType::MGroupedContiguousWithPsumLayout:    return true;
+        default: return false;
+    }
+}
+
+enum class KernelType {
+    Kernel1D1D = 0,
+    Kernel1D2D = 1,
+    KernelNoSF = 2
+};
+
+} // namespace deep_gemm

deep_gemm/include/deep_gemm/common/utils.cuh

@@ -1,167 +1,24 @@
 #pragma once
 
-#include <cuda_bf16.h>
-#include <cuda_fp8.h>
 #include <cuda/std/cstdint>
-#include <cuda/std/utility>
-#include <cute/container/tuple.hpp>
 
-#include "cute_tie.cuh"
+#include <deep_gemm/common/exception.cuh>
 
-#ifdef __CLION_IDE__
-
-__host__ __device__ __forceinline__ void host_device_printf(const char* format, ...) {
-    asm volatile("trap;");
-}
-
-#define printf host_device_printf
-#endif
-
-#ifndef DG_DEVICE_ASSERT
-#define DG_DEVICE_ASSERT(cond) \
-do { \
-    if (not (cond)) { \
-        printf("Assertion failed: %s:%d, condition: %s\n", __FILE__, __LINE__, #cond); \
-        asm("trap;"); \
-    } \
-} while (0)
-#endif
-
-#ifndef DG_TRAP_ONLY_DEVICE_ASSERT
-#define DG_TRAP_ONLY_DEVICE_ASSERT(cond) \
-do { \
-    if (not (cond)) \
-        asm("trap;"); \
-} while (0)
-#endif
-
-#ifndef DG_STATIC_ASSERT
-#define DG_STATIC_ASSERT(cond, ...) static_assert(cond, __VA_ARGS__)
-#endif
-
-namespace deep_gemm {
+namespace deep_gemm::utils {
 
 template <typename FuncT>
 struct PatternVisitor {
     FuncT func;
 
-    __device__ __host__
+    CUTLASS_HOST_DEVICE
     explicit PatternVisitor(FuncT&& func): func(std::forward<FuncT>(func)) {}
 
-    __device__ __host__
-    auto operator [](const uint32_t& i) {
+    CUTLASS_HOST_DEVICE
+    auto operator [](const uint32_t& i) const {
         return func(i);
     }
 };
 
-template <typename T>
-__device__ __host__ T ceil_div(T a, T b) {
-    return (a + b - 1) / b;
-}
-
-template <typename T>
-__device__ __host__ constexpr T constexpr_ceil_div(T a, T b) {
-    return (a + b - 1) / b;
-}
-
-template <typename T>
-__device__ __host__ T align(T a, T b) {
-    return ceil_div(a, b) * b;
-}
-
-template <typename T>
-__device__ __host__ constexpr T constexpr_align(T a, T b) {
-    return constexpr_ceil_div(a, b) * b;
-}
-
-template <typename T>
-__device__ __host__ constexpr T constexpr_gcd(T a, T b) {
-    return b == 0 ? a : constexpr_gcd(b, a % b);
-}
-
-template<typename T>
-__forceinline__ __device__ void swap(T& a, T& b) {
-    T temp = a;
-    a = b;
-    b = temp;
-}
-
-__forceinline__ __device__ uint32_t get_sm_idx() {
-    uint32_t sm_idx;
-    asm ("mov.u32 %0, %%smid;" : "=r"(sm_idx));
-    return sm_idx;
-}
-
-__forceinline__ __device__ uint32_t get_lane_idx() {
-    uint32_t lane_id;
-    asm ("mov.u32 %0, %laneid;" : "=r"(lane_id));
-    return lane_id;
-}
-
-__device__  __forceinline__ uint32_t ld_shared(const uint32_t* ptr) {
-    uint32_t ret;
-    asm volatile("ld.shared.u32 %0, [%1];" : "=r"(ret) : "l"(__cvta_generic_to_shared(ptr)));
-    return ret;
-}
-
-__device__  __forceinline__ float2 ld_shared(const float2* ptr) {
-    float2 ret;
-    asm volatile("ld.shared.v2.f32 {%0, %1}, [%2];" : "=f"(ret.x), "=f"(ret.y) : "l"(__cvta_generic_to_shared(ptr)));
-    return ret;
-}
-
-__device__  __forceinline__ float4 ld_shared(const float4* ptr) {
-    float4 ret;
-    asm volatile("ld.shared.v4.f32 {%0, %1, %2, %3}, [%4];" : "=f"(ret.x), "=f"(ret.y), "=f"(ret.z), "=f"(ret.w) : "l"(__cvta_generic_to_shared(ptr)));
-    return ret;
-}
-
-__device__  __forceinline__ uint4 ld_shared(const uint4* ptr) {
-    uint4 ret;
-    asm volatile("ld.shared.v4.u32 {%0, %1, %2, %3}, [%4];" : "=r"(ret.x), "=r"(ret.y), "=r"(ret.z), "=r"(ret.w) : "l"(__cvta_generic_to_shared(ptr)));
-    return ret;
-}
-
-__device__  __forceinline__ float ld_shared(const float* ptr) {
-    float ret;
-    asm volatile("ld.shared.f32 %0, [%1];" : "=f"(ret) : "l"(__cvta_generic_to_shared(ptr)));
-    return ret;
-}
-
-__device__ __forceinline__ void st_shared(const float* ptr, float val) {
-    asm volatile("st.shared.f32 [%0], %1;" :: "l"(__cvta_generic_to_shared(ptr)), "f"(val));
-}
-
-__device__ __forceinline__ void st_shared(const float2* ptr, float2 val) {
-    asm volatile("st.shared.v2.f32 [%0], {%1, %2};" :: "l"(__cvta_generic_to_shared(ptr)), "f"(val.x), "f"(val.y));
-}
-
-__device__ __forceinline__ void st_shared(const uint32_t* ptr, uint32_t val) {
-    asm volatile("st.shared.u32 [%0], %1;" :: "l"(__cvta_generic_to_shared(ptr)), "r"(val));
-}
-
-__device__  __forceinline__ void st_shared(const void* ptr, uint32_t x, uint32_t y) {
-    asm volatile("st.shared.v2.u32 [%0], {%1, %2};" :: "l"(__cvta_generic_to_shared(ptr)), "r"(x), "r"(y));
-}
-
-__device__  __forceinline__ void st_shared(const void* ptr, uint32_t x, uint32_t y, uint32_t z, uint32_t w) {
-    asm volatile("st.shared.v4.u32 [%0], {%1, %2, %3, %4};" :: "l"(__cvta_generic_to_shared(ptr)), "r"(x), "r"(y), "r"(z), "r"(w));
-}
-
-__device__ __forceinline__ void st_shared(const __int128_t* ptr, __int128_t val) {
-    asm volatile("st.shared.b128 [%0], %1;" :: "l"(__cvta_generic_to_shared(ptr)), "q"(val));
-}
-
-template <typename old_t>
-__device__ __forceinline__ int cast_into_bf16_and_pack(old_t& x, old_t& y) {
-    auto bf16x2 = __float22bfloat162_rn({*reinterpret_cast<float*>(&x), *reinterpret_cast<float*>(&y)});
-    return *reinterpret_cast<int*>(&bf16x2);
-}
-
-__device__ __forceinline__ void prefetch_l1(void *ptr) {
-    asm volatile("prefetch.global.L1 [%0];" :: "l"(ptr));
-}
-
 template <uint32_t kNumBytes>
 struct Vectorized {
     static auto zeros() {
@@ -180,4 +37,14 @@ struct Vectorized {
     using vec_t = decltype(zeros());
 };
 
-} // namespace `deep_gemm`
+template <uint32_t kNumCols>
+CUTLASS_DEVICE constexpr uint32_t get_num_aligned_tmem_cols() {
+    DG_STATIC_ASSERT(kNumCols <= 512, "Too many tensor memory columns");
+    if constexpr (kNumCols <=  32) return  32;
+    if constexpr (kNumCols <=  64) return  64;
+    if constexpr (kNumCols <= 128) return 128;
+    if constexpr (kNumCols <= 256) return 256;
+    return 512;
+}
+
+} // namespace deep_gemm::utils

deep_gemm/include/deep_gemm/epilogue/sm100_store_cd.cuh

@@ -0,0 +1,137 @@
+#pragma once
+
+#include <cute/atom/copy_traits_sm100.hpp>
+
+#include <deep_gemm/common/math.cuh>
+#include <deep_gemm/common/types.cuh>
+#include <deep_gemm/common/utils.cuh>
+#include <deep_gemm/ptx/ld_st.cuh>
+#include <deep_gemm/ptx/tcgen05.cuh>
+
+namespace deep_gemm::epilogue {
+
+template <uint32_t BLOCK_M, uint32_t BLOCK_N,
+          uint32_t STORE_BLOCK_M, uint32_t STORE_BLOCK_N,
+          uint32_t kSwizzleCDMode,
+          uint32_t kNumTMAStoreStages,
+          uint32_t kNumUMMAStoreThreads,
+          GemmType kGemmType, bool kWithAccumulation,
+          typename cd_dtype_t,
+          typename epilogue_type_t,
+          typename pattern_cd_t>
+CUTLASS_DEVICE void
+sm100_store_cd(const utils::PatternVisitor<pattern_cd_t>& smem_cd, uint32_t& tma_stage_idx,
+               const uint32_t& tmem_base_addr,
+               const uint32_t& base_m_idx, const uint32_t& base_n_idx, const uint32_t& batch_idx,
+               const uint32_t& epilogue_warp_idx, const uint32_t& lane_idx,
+               const cutlass::arch::ClusterTransactionBarrier* tmem_empty_barrier,
+               const cute::TmaDescriptor& tensor_map_cd) {
+    // TMA checks
+    constexpr uint32_t kNumBankGroupBytes = 16;
+    constexpr uint32_t kNumElemsPerBankGroup = kNumBankGroupBytes / sizeof(cd_dtype_t);
+    DG_STATIC_ASSERT(kSwizzleCDMode > 0, "TMA D must be swizzled");
+    DG_STATIC_ASSERT(STORE_BLOCK_N % kNumElemsPerBankGroup == 0, "Invalid swizzling");
+    DG_STATIC_ASSERT(BLOCK_M % STORE_BLOCK_M == 0, "Invalid block sizes");
+    DG_STATIC_ASSERT(BLOCK_N % STORE_BLOCK_N == 0, "Invalid block sizes");
+
+    // Share store pipeline between blocks
+    auto advance_store_pipeline = [&]() {
+        tma_stage_idx = (tma_stage_idx + 1) % kNumTMAStoreStages;
+    };
+
+    // Iterate over M waves
+    constexpr auto kNumMWaves = BLOCK_M / STORE_BLOCK_M;
+    #pragma unroll
+    for (uint32_t w = 0; w < kNumMWaves; ++ w) {
+        // Issue every swizzled atom and pipeline STSM and TMA store
+        constexpr uint32_t kNumStores = BLOCK_N / STORE_BLOCK_N;
+        #pragma unroll
+        for (uint32_t s = 0; s < kNumStores; ++ s, advance_store_pipeline()) {
+            auto smem_base_ptr = reinterpret_cast<uint8_t*>(smem_cd[tma_stage_idx]);
+
+            // Wait shared memory to be released
+            if (epilogue_warp_idx == 0)
+                cute::tma_store_wait<kNumTMAStoreStages - 1>();
+            cutlass::arch::NamedBarrier::sync(kNumUMMAStoreThreads, 0);
+
+            // The pipeline stage
+            const auto m_idx = base_m_idx + w * STORE_BLOCK_M;
+            const auto n_idx = epilogue_type_t::apply_index_n<STORE_BLOCK_N>(base_n_idx + s * STORE_BLOCK_N);
+
+            // Store into shared memory
+            #pragma unroll
+            for (uint32_t i = 0; i < STORE_BLOCK_N / kNumElemsPerBankGroup; ++ i) {
+                // Calculate the index of the bank group to be written in the atom
+                auto bank_group_index = i + lane_idx * (kSwizzleCDMode / kNumBankGroupBytes);
+
+                // Reshape the atom in another view and swizzle
+                //  - original: `(LAYOUT_AD_M, kSwizzleCDMode / kNumBankGroupBytes)`
+                //  - new: `(LAYOUT_AD_M * kSwizzleCDMode / kNumBankGroupBytes / 8, 8)`
+                // NOTES: "8" is the number of bank groups, "16" is the swizzling pattern
+                constexpr bool kHasShortcut = (kSwizzleCDMode / kNumBankGroupBytes) == 8;
+                auto row = kHasShortcut ? (i / 8 + lane_idx) : (bank_group_index / 8);
+                auto col = kHasShortcut ? (i) : (bank_group_index % 8);
+                col ^= row % (kSwizzleCDMode / 16);
+
+                // Source and destination memory address
+                uint32_t tmem_addr = tmem_base_addr +                                       // Accumulator offset
+                                     w * BLOCK_N +                                          // Wave offset
+                                     s * STORE_BLOCK_N + i * kNumElemsPerBankGroup;         // In-block offset
+                auto smem_ptr = smem_base_ptr +                                             // Base pointer
+                                epilogue_warp_idx * 32 * kSwizzleCDMode +                   // Warp offset
+                                row * (kNumBankGroupBytes * 8) + col * kNumBankGroupBytes;  // In-atom offset
+
+                // Load from tensor memory, store into shared memory
+                uint32_t values[kNumElemsPerBankGroup];
+                if constexpr (cute::is_same_v<cd_dtype_t, float>) {
+                    // For FP32 output, read and store
+                    DG_STATIC_ASSERT(kNumElemsPerBankGroup == 4, "Invalid type");
+                    cute::SM100_TMEM_LOAD_32dp32b4x::copy(tmem_addr,
+                        values[0], values[1], values[2], values[3]);
+                    cutlass::arch::fence_view_async_tmem_load();
+                    ptx::st_shared(smem_ptr, values[0], values[1], values[2], values[3]);
+                } else {
+                    // For BF16 output, read, cast and store
+                    DG_STATIC_ASSERT(kNumElemsPerBankGroup == 8 and cute::is_same_v<cd_dtype_t, cutlass::bfloat16_t>, "Invalid type");
+                    cute::SM100_TMEM_LOAD_32dp32b8x::copy(tmem_addr,
+                        values[0], values[1], values[2], values[3],
+                        values[4], values[5], values[6], values[7]);
+                    cutlass::arch::fence_view_async_tmem_load();
+                    ptx::st_shared(
+                        smem_ptr,
+                        math::cast_into_bf16_and_pack(values[0], values[1]),
+                        math::cast_into_bf16_and_pack(values[2], values[3]),
+                        math::cast_into_bf16_and_pack(values[4], values[5]),
+                        math::cast_into_bf16_and_pack(values[6], values[7])
+                    );
+                }
+            }
+
+            // Notify tensor memory empty (only at the leader CTA) arrival ASAP
+            // NOTES: only the last stage needs to do this
+            if (w == kNumMWaves - 1 and s == BLOCK_N / STORE_BLOCK_N - 1) {
+                ptx::tcgen05_before_thread_sync();
+                tmem_empty_barrier->arrive(0u);
+            }
+
+            // Synchronize all threads and issue TMA
+            cute::tma_store_fence();
+            cutlass::arch::NamedBarrier::sync(kNumUMMAStoreThreads, 0);
+            if (epilogue_warp_idx == 0 and cute::elect_one_sync()) {
+                if constexpr (kGemmType == GemmType::Batched) {
+                    using cute_tma_t = cute::conditional_t<kWithAccumulation,
+                                            cute::SM90_TMA_REDUCE_ADD_3D, cute::SM90_TMA_STORE_3D>;
+                    cute_tma_t::copy(&tensor_map_cd, smem_base_ptr, n_idx, m_idx, batch_idx);
+                } else {
+                    using cute_tma_t = cute::conditional_t<kWithAccumulation,
+                                            cute::SM90_TMA_REDUCE_ADD_2D, cute::SM90_TMA_STORE_2D>;
+                    cute_tma_t::copy(&tensor_map_cd, smem_base_ptr, n_idx, m_idx);
+                }
+                cute::tma_store_arrive();
+            }
+            __syncwarp();
+        }
+    }
+}
+
+} // namespace deep_gemm::epilogue

deep_gemm/include/deep_gemm/epilogue/sm100_store_cd_swap_ab.cuh

@@ -0,0 +1,144 @@
+#pragma once
+
+#include <cute/atom/copy_traits_sm100.hpp>
+
+#include <deep_gemm/common/math.cuh>
+#include <deep_gemm/common/types.cuh>
+#include <deep_gemm/common/utils.cuh>
+#include <deep_gemm/ptx/ld_st.cuh>
+#include <deep_gemm/ptx/tcgen05.cuh>
+
+namespace deep_gemm::epilogue {
+
+template <uint32_t BLOCK_M, uint32_t BLOCK_N,
+          uint32_t STORE_BLOCK_M, uint32_t STORE_BLOCK_N,
+          uint32_t kSwizzleCDMode,
+          uint32_t kNumTMAStoreStages,
+          uint32_t kNumUMMAStoreThreads,
+          GemmType kGemmType, bool kWithAccumulation,
+          typename cd_dtype_t,
+          typename epilogue_type_t,
+          typename pattern_cd_t>
+CUTLASS_DEVICE void
+sm100_store_cd_swap_ab(const utils::PatternVisitor<pattern_cd_t>& smem_cd, uint32_t& tma_stage_idx,
+                       const uint32_t& tmem_base_addr,
+                       const uint32_t& base_m_idx, const uint32_t& base_n_idx, const uint32_t& batch_idx,
+                       const uint32_t& effective_m,
+                       const uint32_t& epilogue_warp_idx, const uint32_t& lane_idx,
+                       const cutlass::arch::ClusterTransactionBarrier* tmem_empty_barrier,
+                       const cute::TmaDescriptor& tensor_map_cd) {
+    // NOTES: The epilogue requires a full warpgroup to read all 128 TMEM rows,
+    //          implying STORE_BLOCK_N must be 128.
+    DG_STATIC_ASSERT(STORE_BLOCK_N == 128, "STORE_BLOCK_N must be 128 to match TMEM rows");
+
+    // TMA checks
+    constexpr uint32_t STORE_BLOCK_N_ATOM = kSwizzleCDMode / sizeof(cd_dtype_t);
+    constexpr uint32_t kNumBankGroupBytes = 16;
+    constexpr uint32_t kNumSwizzleAtomRows = 8;
+    DG_STATIC_ASSERT(kSwizzleCDMode == 128, "TMA D must be 128B swizzled");
+    DG_STATIC_ASSERT(BLOCK_M % STORE_BLOCK_M == 0, "Invalid block sizes");
+    DG_STATIC_ASSERT(BLOCK_N % STORE_BLOCK_N == 0, "Invalid block sizes");
+    DG_STATIC_ASSERT(STORE_BLOCK_M % kNumSwizzleAtomRows == 0, "Invalid swizzling");
+    DG_STATIC_ASSERT(STORE_BLOCK_N % STORE_BLOCK_N_ATOM == 0, "Invalid swizzling");
+
+    // Share store pipeline between blocks
+    auto advance_store_pipeline = [&]() {
+        tma_stage_idx = (tma_stage_idx + 1) % kNumTMAStoreStages;
+    };
+
+    // Iterate over M blocks
+    const auto num_stores = effective_m / STORE_BLOCK_M;
+    for (uint32_t s = 0; s < num_stores; ++ s, advance_store_pipeline()) {
+        // Wait shared memory to be released
+        if (epilogue_warp_idx == 0)
+            cute::tma_store_wait<kNumTMAStoreStages - 1>();
+        cutlass::arch::NamedBarrier::sync(kNumUMMAStoreThreads, 0);
+
+        // Store into shared memory
+        #pragma unroll
+        for (uint32_t i = 0; i < STORE_BLOCK_M / kNumSwizzleAtomRows; ++ i) {
+            uint32_t tmem_addr = tmem_base_addr +
+                                 s * STORE_BLOCK_M +            // Store stage offset
+                                 i * kNumSwizzleAtomRows;       // In-block offset
+            uint32_t values[kNumSwizzleAtomRows];
+
+            // Warps cooperatively write an atomic block to shared memory
+            DG_STATIC_ASSERT(STORE_BLOCK_N_ATOM % 32 == 0, "Invalid block sizes");
+            constexpr uint32_t kNumWarpsPerAtom = STORE_BLOCK_N_ATOM / 32;
+            uint32_t outer_atom_offset = (epilogue_warp_idx / kNumWarpsPerAtom) * STORE_BLOCK_M * kSwizzleCDMode;
+            uint32_t inner_atom_offset = i * kNumSwizzleAtomRows * kSwizzleCDMode;
+            auto smem_base_ptr = reinterpret_cast<uint8_t*>(smem_cd[tma_stage_idx]) + outer_atom_offset + inner_atom_offset;
+
+            if constexpr (cute::is_same_v<cd_dtype_t, float>) {
+                // NOTES: Swizzling is not required in this case, but used here for consistency with other cases
+                cute::SM100_TMEM_LOAD_32dp32b8x::copy(tmem_addr, values[0], values[1], values[2], values[3],
+                                                                 values[4], values[5], values[6], values[7]);
+                uint32_t col = lane_idx / 4;
+
+                #pragma unroll
+                for (uint32_t row = 0; row < kNumSwizzleAtomRows; ++ row) {
+                    auto smem_ptr = smem_base_ptr + row * (kNumBankGroupBytes * 8)
+                                                  + (col ^ row) * kNumBankGroupBytes
+                                                  + (lane_idx % 4) * sizeof(float);
+                    ptx::st_shared(reinterpret_cast<uint32_t*>(smem_ptr), values[row]);
+                }
+            } else {
+                // Load from TMEM using `.16x256b` shape to satisfy STSM layout requirements
+                // Start from lane index 0
+                cute::SM100_TMEM_LOAD_16dp256b1x::copy(tmem_addr,
+                                                       values[0], values[1], values[2], values[3]);
+                // Start from lane index 16
+                cute::SM100_TMEM_LOAD_16dp256b1x::copy(tmem_addr | 0x00100000,
+                                                       values[4], values[5], values[6], values[7]);
+                cutlass::arch::fence_view_async_tmem_load();
+
+                // Destination shared memory address
+                uint32_t row = lane_idx % 8;
+                uint32_t col = (epilogue_warp_idx % 2) * 4 + lane_idx / 8;
+                auto smem_ptr = smem_base_ptr + row * (kNumBankGroupBytes * 8)
+                                              + (col ^ row) * kNumBankGroupBytes;
+
+                // Store matrix with transposition
+                ptx::SM90_U32x4_STSM_T<int>::copy(math::cast_into_bf16_and_pack(values[0], values[1]),
+                                                  math::cast_into_bf16_and_pack(values[2], values[3]),
+                                                  math::cast_into_bf16_and_pack(values[4], values[5]),
+                                                  math::cast_into_bf16_and_pack(values[6], values[7]),
+                                                  smem_ptr);
+            }
+        }
+
+        // Notify tensor memory empty (only at the leader CTA) arrival ASAP
+        // NOTES: only the last stage needs to do this
+        if (s == num_stores - 1) {
+            ptx::tcgen05_before_thread_sync();
+            tmem_empty_barrier->arrive(0u);
+        }
+
+        // Synchronize all threads and issue TMA
+        cute::tma_store_fence();
+        cutlass::arch::NamedBarrier::sync(kNumUMMAStoreThreads, 0);
+        if (epilogue_warp_idx == 0 and cute::elect_one_sync()) {
+            #pragma unroll
+            for (uint32_t i = 0; i < STORE_BLOCK_N / STORE_BLOCK_N_ATOM; ++ i) {
+                auto smem_ptr = smem_cd[tma_stage_idx] + i * STORE_BLOCK_M * STORE_BLOCK_N_ATOM;
+                uint32_t m_idx = base_m_idx + s * STORE_BLOCK_M;
+                uint32_t n_idx = epilogue_type_t::apply_index_n<STORE_BLOCK_N_ATOM>(base_n_idx + i * STORE_BLOCK_N_ATOM);
+
+                // Issue 2D or 3D TMA store
+                if constexpr (kGemmType == GemmType::Batched) {
+                    using cute_tma_t = cute::conditional_t<kWithAccumulation,
+                        cute::SM90_TMA_REDUCE_ADD_3D, cute::SM90_TMA_STORE_3D>;
+                    cute_tma_t::copy(&tensor_map_cd, smem_ptr, n_idx, m_idx, batch_idx);
+                } else {
+                    using cute_tma_t = cute::conditional_t<kWithAccumulation,
+                        cute::SM90_TMA_REDUCE_ADD_2D, cute::SM90_TMA_STORE_2D>;
+                    cute_tma_t::copy(&tensor_map_cd, smem_ptr, n_idx, m_idx);
+                }
+            }
+            cute::tma_store_arrive();
+        }
+        __syncwarp();
+    }
+}
+
+} // namespace deep_gemm::epilogue

deep_gemm/include/deep_gemm/epilogue/transform.cuh

@@ -0,0 +1,24 @@
+#pragma once
+
+#include <deep_gemm/common/exception.cuh>
+
+namespace deep_gemm::epilogue::transform {
+
+struct EpilogueIdentity {
+    template <uint32_t STORE_BLOCK_N>
+    CUTLASS_DEVICE static uint32_t apply_index_n(const uint32_t& n_idx) {
+        return n_idx;
+    }
+};
+
+template <uint32_t kLeft, uint32_t kMid, uint32_t kRight>
+struct EpilogueHeadSplits: EpilogueIdentity {
+    template <uint32_t STORE_BLOCK_N>
+    CUTLASS_DEVICE static uint32_t apply_index_n(const uint32_t& n_idx) {
+        DG_STATIC_ASSERT(kLeft % STORE_BLOCK_N == 0 and kMid % STORE_BLOCK_N == 0 and
+                         kRight % STORE_BLOCK_N == 0, "Invalid head splits config");
+        return n_idx + (n_idx + kRight) / (kLeft + kRight) * kMid;
+    }
+};
+
+} // namespace deep_gemm::epilogue::transform

deep_gemm/include/deep_gemm/impls/sm100_bf16_gemm.cuh

@@ -4,14 +4,18 @@
 
 #include <cutlass/arch/barrier.h>
 
-#include <deep_gemm/common/scheduler.cuh>
-#include <deep_gemm/common/utils.cuh>
-#include <deep_gemm/common/sm100_utils.cuh>
+#include <deep_gemm/scheduler/gemm.cuh>
+#include <deep_gemm/common/math.cuh>
+#include <deep_gemm/common/tma_copy.cuh>
+#include <deep_gemm/epilogue/sm100_store_cd.cuh>
+#include <deep_gemm/epilogue/sm100_store_cd_swap_ab.cuh>
+#include <deep_gemm/epilogue/transform.cuh>
+#include <deep_gemm/mma/sm100.cuh>
+#include <deep_gemm/ptx/tcgen05.cuh>
+#include <deep_gemm/ptx/utils.cuh>
 
 namespace deep_gemm {
 
-using namespace deep_gemm::sm100;
-
 template <cute::UMMA::Major kMajorA, cute::UMMA::Major kMajorB,
           uint32_t SHAPE_M, uint32_t SHAPE_N, uint32_t SHAPE_K,
           uint32_t BLOCK_M, uint32_t BLOCK_N, uint32_t BLOCK_K_,
@@ -21,9 +25,10 @@ template <cute::UMMA::Major kMajorA, cute::UMMA::Major kMajorB,
           uint32_t kNumNonEpilogueThreads, uint32_t kNumEpilogueThreads,
           uint32_t kNumMulticast, bool kIsMulticastOnA,
           uint32_t kNumSMs,
+          bool kSwapAB,
           GemmType kGemmType, bool kWithAccumulation, typename cd_dtype_t,
           uint64_t kTensorCoreUtilControl>
-__global__ void __launch_bounds__(kNumNonEpilogueThreads + kNumEpilogueThreads, 1)
+CUTLASS_GLOBAL void __launch_bounds__(kNumNonEpilogueThreads + kNumEpilogueThreads, 1)
 sm100_bf16_gemm_impl(int* grouped_layout,
                      uint32_t shape_m, uint32_t shape_n, uint32_t shape_k,
                      const __grid_constant__ cute::TmaDescriptor tensor_map_a,
@@ -48,41 +53,31 @@ sm100_bf16_gemm_impl(int* grouped_layout,
     if constexpr (kWithAccumulation)
         DG_STATIC_ASSERT(cute::is_same_v<cd_dtype_t, float>, "Invalid C/D data dtype");
 
-    // Configs
+    // MMA Configs
     constexpr uint32_t LAYOUT_AD_M = 128;
-    constexpr uint32_t WAVE_BLOCK_M = cute::min<uint32_t>(BLOCK_M, LAYOUT_AD_M);
-    constexpr uint32_t kNumMWaves = BLOCK_M / WAVE_BLOCK_M;
-    constexpr uint32_t kNumTMAStoreStages = 2;
-    DG_STATIC_ASSERT(BLOCK_K_ == 64, "Invalid block K");
-    DG_STATIC_ASSERT(BLOCK_M % WAVE_BLOCK_M == 0 and 2 % kNumMWaves == 0, "Invalid block M");
-    DG_STATIC_ASSERT(sizeof(cutlass::bfloat16_t) * LAYOUT_AD_M % kSwizzleAMode == 0, "Invalid swizzle A mode");
-
-    // Overwrite shape constants if the compiler gives
-    shape_m = SHAPE_M != 0 ? SHAPE_M : shape_m;
-    shape_n = SHAPE_N != 0 ? SHAPE_N : shape_n;
-    shape_k = SHAPE_K != 0 ? SHAPE_K : shape_k;
-
-    // Utils
-    bool is_leader_cta = cute::block_rank_in_cluster() == 0;
-    const auto warp_idx = cutlass::canonical_warp_idx_sync();
-    const auto lane_idx = get_lane_idx();
-
-    // Align to 1024 bytes for swizzle-128B
-    extern __shared__ __align__(1024) uint8_t smem_buffer[];
-
-    // 2-CTA MMA
+    constexpr uint32_t UMMA_M = LAYOUT_AD_M * kNumMulticast;
+    constexpr uint32_t UMMA_N = kSwapAB ? BLOCK_M : BLOCK_N;
+    constexpr uint32_t UMMA_K = 16;
     constexpr uint32_t LOAD_BLOCK_M = BLOCK_M / (kIsMulticastOnA ? kNumMulticast: 1);
     constexpr uint32_t LOAD_BLOCK_N = BLOCK_N / (kIsMulticastOnA ? 1 : kNumMulticast);
-    constexpr uint32_t STORE_BLOCK_M = cute::min<uint32_t>(BLOCK_M, LAYOUT_AD_M);
-    constexpr uint32_t STORE_BLOCK_N = kSwizzleCDMode / sizeof(cd_dtype_t);
-    constexpr uint32_t kNumUMMAStoreThreads = STORE_BLOCK_M;
-    DG_STATIC_ASSERT(not kIsMulticastOnA or kNumMulticast == 1, "Invalid multicast");
-    DG_STATIC_ASSERT(LOAD_BLOCK_M == BLOCK_M, "Only support tensor memory layout A/D");
+    DG_STATIC_ASSERT(BLOCK_K_ == 64, "Invalid block K");
     DG_STATIC_ASSERT(kNumMulticast == 1 or kNumMulticast == 2, "Only support 1/2 multicast");
+    DG_STATIC_ASSERT((kSwapAB and BLOCK_N == LAYOUT_AD_M) or
+                     (not kSwapAB and (BLOCK_M == 32 or BLOCK_M == 64 or BLOCK_M == LAYOUT_AD_M)), "Invalid block size");
+
+    // Epilogue configs
+    // Always enable pipeline for better performance
+    constexpr uint32_t kNumEpilogueStages = 2;
+    constexpr uint32_t kNumTMAStoreStages = 2;
+    // NOTES: To maximize epilogue threads utilization, process an entire BLOCK_N
+    //        per store stage for swap-AB cases, and an entire BLOCK_M for non-swap cases
+    constexpr uint32_t STORE_BLOCK_M =        kSwapAB ? 16      : cute::min<uint32_t>(BLOCK_M, LAYOUT_AD_M);
+    constexpr uint32_t STORE_BLOCK_N =        kSwapAB ? BLOCK_N : kSwizzleCDMode / sizeof(cd_dtype_t);
+    constexpr uint32_t kNumUMMAStoreThreads = kSwapAB ? kNumEpilogueThreads: STORE_BLOCK_M;
     DG_STATIC_ASSERT(kNumUMMAStoreThreads % 32 == 0, "Invalid store block M");
 
     // Share memory sizes
-    constexpr uint32_t SMEM_CD_SIZE_PER_STAGE = STORE_BLOCK_M * kSwizzleCDMode;
+    constexpr uint32_t SMEM_CD_SIZE_PER_STAGE = STORE_BLOCK_M * STORE_BLOCK_N * sizeof(cd_dtype_t);
     constexpr uint32_t SMEM_CD_SIZE = SMEM_CD_SIZE_PER_STAGE * kNumTMAStoreStages;
     constexpr uint32_t SMEM_A_SIZE_PER_STAGE = LOAD_BLOCK_M * BLOCK_K * sizeof(cutlass::bfloat16_t);
     constexpr uint32_t SMEM_B_SIZE_PER_STAGE = LOAD_BLOCK_N * BLOCK_K * sizeof(cutlass::bfloat16_t);
@@ -91,50 +86,60 @@ sm100_bf16_gemm_impl(int* grouped_layout,
     DG_STATIC_ASSERT(kNumTMAStoreStages >= 1, "Invalid number of TMA stages");
 
     // NOTES: Make sure we have enough shared memory for UMMA padding
-    static constexpr uint32_t UMMA_A_SIZE_PER_STAGE = constexpr_align(LOAD_BLOCK_M, LAYOUT_AD_M) * BLOCK_K * sizeof(nv_bfloat16);
-    DG_STATIC_ASSERT(UMMA_A_SIZE_PER_STAGE <= SMEM_A_SIZE_PER_STAGE + SMEM_B_SIZE_PER_STAGE * kNumStages, "Memory Out of bound for UMMA");
-
-    // Automatically deduce the number of epilogue stages (1 or 2), according to the tensor memory size
-    // TODO: test cases of `kNumMWaves == 2 and kNumEpilogueStages == 2`
-    constexpr uint32_t kNumEpilogueStages = (2 * kNumMWaves * BLOCK_N) > 512 ? 1 : 2;
+    static constexpr uint32_t UMMA_A_SIZE_PER_STAGE = math::constexpr_align(LOAD_BLOCK_M, LAYOUT_AD_M) * BLOCK_K * sizeof(nv_bfloat16);
+    DG_STATIC_ASSERT(UMMA_A_SIZE_PER_STAGE <= SMEM_A_SIZE_PER_STAGE + SMEM_B_SIZE_PER_STAGE * kNumStages, "Memory out of bound for UMMA");
 
     // Real tensor memory size and offsets
-    constexpr uint32_t kNumAccumTmemCols = kNumEpilogueStages * kNumMWaves * BLOCK_N;
-    constexpr uint32_t kNumTmemCols = get_num_aligned_tmem_cols<kNumAccumTmemCols>();
+    constexpr uint32_t kNumAccumTmemCols = kNumEpilogueStages * UMMA_N;
+    constexpr uint32_t kNumTmemCols = utils::get_num_aligned_tmem_cols<kNumAccumTmemCols>();
+    DG_STATIC_ASSERT(32 <= kNumTmemCols and kNumTmemCols <= 512, "Invalid tensor memory columns");
+
+    // Synchronize the cluster before 2-CTA TMEM allocation
+    kNumMulticast > 1 ? cute::cluster_sync() : void();
+
+    // Utils
+    bool is_leader_cta = cute::block_rank_in_cluster() == 0;
+    const auto warp_idx = cutlass::canonical_warp_idx_sync();
+    const auto lane_idx = ptx::get_lane_idx();
 
     // Prefetch TMA descriptors at the very beginning
-    if (warp_idx == 0 and cute::elect_one_sync()) {
+    if (warp_idx == 0) {
         cute::prefetch_tma_descriptor(&tensor_map_a);
         cute::prefetch_tma_descriptor(&tensor_map_b);
         cute::prefetch_tma_descriptor(&tensor_map_cd);
     }
 
+    // Overwrite shape constants if the compiler gives
+    shape_m = SHAPE_M != 0 ? SHAPE_M : shape_m;
+    shape_n = SHAPE_N != 0 ? SHAPE_N : shape_n;
+    shape_k = SHAPE_K != 0 ? SHAPE_K : shape_k;
+
+    // Align to 1024 bytes for swizzle-128B
+    extern __shared__ __align__(1024) uint8_t smem_buffer[];
+
     // D/A/B shared memory
-    auto smem_cd = PatternVisitor([&](const uint32_t& i) {
+    auto smem_cd = utils::PatternVisitor([&](const uint32_t& i) {
         return reinterpret_cast<cd_dtype_t*>(smem_buffer + i * SMEM_CD_SIZE_PER_STAGE);
     });
-    auto smem_a  = PatternVisitor([&](const uint32_t& i) {
+    auto smem_a  = utils::PatternVisitor([&](const uint32_t& i) {
         return reinterpret_cast<cutlass::bfloat16_t*>(smem_buffer + SMEM_CD_SIZE + i * SMEM_A_SIZE_PER_STAGE);
     });
-    auto smem_b  = PatternVisitor([&](const uint32_t& i) {
+    auto smem_b  = utils::PatternVisitor([&](const uint32_t& i) {
         return reinterpret_cast<cutlass::bfloat16_t*>(smem_buffer + SMEM_CD_SIZE + kNumStages * SMEM_A_SIZE_PER_STAGE + i * SMEM_B_SIZE_PER_STAGE);
     });
 
     // Fill barriers
     auto barrier_start_ptr = reinterpret_cast<Barrier*>(smem_buffer + SMEM_CD_SIZE + kNumStages * (SMEM_A_SIZE_PER_STAGE + SMEM_B_SIZE_PER_STAGE));
-    auto full_barriers              = PatternVisitor([=](const uint32_t& i) { return barrier_start_ptr + (i); });
-    auto empty_barriers             = PatternVisitor([=](const uint32_t& i) { return barrier_start_ptr + (kNumStages + i); });
-    auto tmem_full_barriers         = PatternVisitor([=](const uint32_t& i) { return barrier_start_ptr + (kNumStages * 2 + i); });
-    auto tmem_empty_barriers        = PatternVisitor([=](const uint32_t& i) { return barrier_start_ptr + (kNumStages * 2 + kNumEpilogueStages + i); });
+    auto full_barriers              = utils::PatternVisitor([=](const uint32_t& i) { return barrier_start_ptr + (i); });
+    auto empty_barriers             = utils::PatternVisitor([=](const uint32_t& i) { return barrier_start_ptr + (kNumStages + i); });
+    auto tmem_full_barriers         = utils::PatternVisitor([=](const uint32_t& i) { return barrier_start_ptr + (kNumStages * 2 + i); });
+    auto tmem_empty_barriers        = utils::PatternVisitor([=](const uint32_t& i) { return barrier_start_ptr + (kNumStages * 2 + kNumEpilogueStages + i); });
     auto tensor_core_full_barrier   = barrier_start_ptr + kNumStages * 3 + kNumEpilogueStages * 2;
 
     // Fill the tensor memory pointer
     auto tmem_ptr_in_smem = reinterpret_cast<uint32_t*>(barrier_start_ptr + kNumStages * 3 + kNumEpilogueStages * 2 + 1);
     DG_STATIC_ASSERT(32 <= kNumTmemCols and kNumTmemCols <= 512, "Invalid tensor memory columns");
 
-    if (kNumMulticast > 1)
-        cute::cluster_sync();
-
     // Initialize barriers
     if (warp_idx == 1 and cute::elect_one_sync()) {
         #pragma unroll
@@ -162,9 +167,13 @@ sm100_bf16_gemm_impl(int* grouped_layout,
     }
     kNumMulticast > 1 ? cute::cluster_sync() : __syncthreads();
 
+    // Wait for primary kernel completion
+    cudaGridDependencySynchronize();
+
     // Block scheduler
     uint32_t m_block_idx, n_block_idx;
-    auto scheduler = Scheduler<kGemmType, BLOCK_M, BLOCK_N, kNumGroups, kNumMulticast, kIsMulticastOnA, kNumSMs>(shape_m, shape_n, shape_k, grouped_layout);
+    auto scheduler = sched::Scheduler<kGemmType, BLOCK_M, BLOCK_N, kNumGroups, kNumMulticast, kIsMulticastOnA, kNumSMs>(
+        shape_m, shape_n, shape_k, grouped_layout);
 
     // Pipeline and TMA phases
     uint32_t stage_idx = 0, phase = 0, tensor_core_phase = 0;
@@ -181,16 +190,20 @@ sm100_bf16_gemm_impl(int* grouped_layout,
         // TMA load warp
         // Persistently schedule over blocks
         while (scheduler.get_next_block(m_block_idx, n_block_idx)) {
-            const auto& num_total_k_blocks = ceil_div(scheduler.current_shape_k, BLOCK_K);
+            // Use dynamic load block M, when swap-AB is enabled
+            const auto load_block_m = kSwapAB ? scheduler.get_aligned_effective_m_in_block(m_block_idx) / kNumMulticast : LOAD_BLOCK_M;
+
+            // For k-grouped layout, the number of block K is variable
+            const auto num_total_k_blocks = math::ceil_div(scheduler.current_shape_k, BLOCK_K);
             for (uint32_t k_block_idx = 0; k_block_idx < num_total_k_blocks; advance_pipeline(k_block_idx)) {
                 // Wait consumer release
                 empty_barriers[stage_idx]->wait(phase ^ 1);
 
                 // Compute offsets
                 // NOTES: the group is always concatenated with the outer dimension
-                uint32_t m_idx = scheduler.template get_global_idx<(kGemmType == GemmType::MGroupedMasked), IndexType::MN> (
+                uint32_t m_idx = scheduler.template get_global_idx<(kGemmType == GemmType::MGroupedMasked), sched::IndexType::MN> (
                     shape_m, BLOCK_M, m_block_idx);
-                uint32_t n_idx = scheduler.template get_global_idx<(kMajorB == cute::UMMA::Major::K), IndexType::MN> (
+                uint32_t n_idx = scheduler.template get_global_idx<(kMajorB == cute::UMMA::Major::K), sched::IndexType::MN> (
                     shape_n, BLOCK_N, n_block_idx, m_block_idx);
 
                 // NOTES: `k_idx` is actually the k index default for K-major, while `k_b_idx` may be MN-major
@@ -198,14 +211,14 @@ sm100_bf16_gemm_impl(int* grouped_layout,
                 DG_STATIC_ASSERT(kGemmType == GemmType::Normal or kGemmType == GemmType::KGroupedContiguous or kGemmType == GemmType::Batched or
                                  kMajorA == cute::UMMA::Major::K, "Invalid major");
                 uint32_t k_idx = k_block_idx * BLOCK_K;
-                uint32_t k_a_idx = scheduler.template get_global_idx<(kMajorA == cute::UMMA::Major::MN), IndexType::K> (
+                uint32_t k_a_idx = scheduler.template get_global_idx<(kMajorA == cute::UMMA::Major::MN), sched::IndexType::K> (
                     shape_k, BLOCK_K, k_block_idx, m_block_idx);
-                uint32_t k_b_idx = scheduler.template get_global_idx<(kMajorB == cute::UMMA::Major::MN), IndexType::K> (
+                uint32_t k_b_idx = scheduler.template get_global_idx<(kMajorB == cute::UMMA::Major::MN), sched::IndexType::K> (
                     shape_k, BLOCK_K, k_block_idx, m_block_idx);
 
                 // Add 2 CTA offsets
                 if constexpr (kNumMulticast > 1) {
-                    m_idx += kIsMulticastOnA ? (cute::block_rank_in_cluster() * LOAD_BLOCK_M) : 0;
+                    m_idx += kIsMulticastOnA ? (cute::block_rank_in_cluster() * load_block_m) : 0;
                     n_idx += kIsMulticastOnA ? 0 : (cute::block_rank_in_cluster() * LOAD_BLOCK_N);
                 }
 
@@ -213,16 +226,16 @@ sm100_bf16_gemm_impl(int* grouped_layout,
                 constexpr bool kIsBatchedMM = (kGemmType == GemmType::Batched);
                 const uint32_t batch_idx = (kIsBatchedMM ? scheduler.current_group_idx : 0);
                 if constexpr (kMajorA == cute::UMMA::Major::K)
-                    tma_copy<BLOCK_K, LOAD_BLOCK_M, kSwizzleAMode, cutlass::bfloat16_t, kIsBatchedMM>(
+                    tma::copy<BLOCK_K, LOAD_BLOCK_M, kSwizzleAMode, cutlass::bfloat16_t, kIsBatchedMM>(
                         &tensor_map_a, full_barriers[stage_idx], smem_a[stage_idx], k_a_idx, m_idx, kNumMulticast, batch_idx);
                 if constexpr (kMajorA == cute::UMMA::Major::MN)
-                    tma_copy<LOAD_BLOCK_M, BLOCK_K, kSwizzleAMode, cutlass::bfloat16_t, kIsBatchedMM>(
+                    tma::copy<LOAD_BLOCK_M, BLOCK_K, kSwizzleAMode, cutlass::bfloat16_t, kIsBatchedMM>(
                         &tensor_map_a, full_barriers[stage_idx], smem_a[stage_idx], m_idx, k_a_idx, kNumMulticast, batch_idx);
                 if constexpr (kMajorB == cute::UMMA::Major::K)
-                    tma_copy<BLOCK_K, LOAD_BLOCK_N, kSwizzleBMode, cutlass::bfloat16_t, kIsBatchedMM>(
+                    tma::copy<BLOCK_K, LOAD_BLOCK_N, kSwizzleBMode, cutlass::bfloat16_t, kIsBatchedMM>(
                         &tensor_map_b, full_barriers[stage_idx], smem_b[stage_idx], k_b_idx, n_idx, kNumMulticast, batch_idx);
                 if constexpr (kMajorB == cute::UMMA::Major::MN)
-                    tma_copy<LOAD_BLOCK_N, BLOCK_K, kSwizzleBMode, cutlass::bfloat16_t, kIsBatchedMM>(
+                    tma::copy<LOAD_BLOCK_N, BLOCK_K, kSwizzleBMode, cutlass::bfloat16_t, kIsBatchedMM>(
                         &tensor_map_b, full_barriers[stage_idx], smem_b[stage_idx], n_idx, k_b_idx, kNumMulticast, batch_idx);
 
                 // Arrive at full barriers
@@ -238,17 +251,16 @@ sm100_bf16_gemm_impl(int* grouped_layout,
         // MMA issue warp
         // NOTES: only the leader CTA will do this
         // Make instruction descriptor
-        // TODO: refactor `UMMA_M` calculation
-        constexpr uint32_t UMMA_M = LAYOUT_AD_M * (kIsMulticastOnA ? 1 : kNumMulticast);
-        constexpr uint32_t UMMA_N = BLOCK_N * (kIsMulticastOnA ? kNumMulticast : 1);
-        constexpr uint32_t UMMA_K = 32 / sizeof(cutlass::bfloat16_t);
-        auto instr_desc = cute::UMMA::make_instr_desc<cutlass::bfloat16_t, cutlass::bfloat16_t, float, UMMA_M, UMMA_N, kMajorA, kMajorB>();
+        auto instr_desc = kSwapAB ? cute::UMMA::make_instr_desc<cutlass::bfloat16_t, cutlass::bfloat16_t, float,
+                                                                UMMA_M, UMMA_N, kMajorB, kMajorA>()
+                                  : cute::UMMA::make_instr_desc<cutlass::bfloat16_t, cutlass::bfloat16_t, float,
+                                                                UMMA_M, UMMA_N, kMajorA, kMajorB>();
 
         DG_STATIC_ASSERT(kNumStages <= 32, "Too many stages");
         // Merged stages only happens in NT normal GEMM cases
         constexpr uint32_t BLOCK_ATOM_K = BLOCK_K / kNumStagesPerMerge;
-        auto a_desc = make_umma_desc<kMajorA, LOAD_BLOCK_M, BLOCK_ATOM_K, kSwizzleAMode>(smem_a[0], 0, 0);
-        auto b_desc = make_umma_desc<kMajorB, LOAD_BLOCK_N, BLOCK_ATOM_K, kSwizzleBMode>(smem_b[0], 0, 0);
+        auto a_desc = mma::sm100::make_umma_desc<kMajorA, LOAD_BLOCK_M, BLOCK_ATOM_K, kSwizzleAMode>(smem_a[0], 0, 0);
+        auto b_desc = mma::sm100::make_umma_desc<kMajorB, LOAD_BLOCK_N, BLOCK_ATOM_K, kSwizzleBMode>(smem_b[0], 0, 0);
         uint32_t a_desc_lo = lane_idx < kNumStages ? a_desc.lo + lane_idx * SMEM_A_SIZE_PER_STAGE / 16 : 0u;
         uint32_t b_desc_lo = lane_idx < kNumStages ? b_desc.lo + lane_idx * SMEM_B_SIZE_PER_STAGE / 16 : 0u;
 
@@ -265,7 +277,7 @@ sm100_bf16_gemm_impl(int* grouped_layout,
             auto accum_stage_idx = scheduler.current_iter % kNumEpilogueStages;
             auto accum_phase_idx = (scheduler.current_iter / kNumEpilogueStages) & 1;
             tmem_empty_barriers[accum_stage_idx]->wait(accum_phase_idx ^ 1);
-            tcgen05_after_thread_sync();
+            ptx::tcgen05_after_thread_sync();
 
             // UMMA and empty barrier arrival alias
             auto umma_arrive = [](const uint64_t* barrier) {
@@ -282,36 +294,45 @@ sm100_bf16_gemm_impl(int* grouped_layout,
                 // NOTES: the tensor memory accumulator pipeline has nothing to do with multicasting
                 if (do_tmem_full_arrive)
                     umma_arrive(reinterpret_cast<uint64_t*>(tmem_full_barriers[accum_stage_idx]));
+                __syncwarp();
             };
 
+            // Dynamic update of UMMA N based on effective M, when swap-AB is enabled
+            if constexpr (kSwapAB) {
+                uint32_t umma_n = scheduler.get_aligned_effective_m_in_block(m_block_idx);
+                mma::sm100::update_instr_desc_with_umma_n(instr_desc, umma_n);
+            }
+
             // Launch MMAs
-            const auto& num_total_k_blocks = ceil_div(scheduler.current_shape_k, BLOCK_K);
+            const auto num_total_k_blocks = math::ceil_div(scheduler.current_shape_k, BLOCK_K);
             for (uint32_t k_block_idx = 0; k_block_idx < num_total_k_blocks; advance_pipeline(k_block_idx)) {
                 // Wait TMA arrival
                 full_barriers[stage_idx]->wait(phase);
-                tcgen05_after_thread_sync();
+                ptx::tcgen05_after_thread_sync();
 
                 // Issue UMMA in the leader CTA
-                using mma_t = cute::conditional_t<kNumMulticast == 1, SM100_MMA_F16BF16_SS, SM100_MMA_F16BF16_2x1SM_SS>;
-                const auto& runtime_instr_desc = cute::UMMA::make_runtime_instr_desc(instr_desc);
-                const auto& a_desc_base_lo = __shfl_sync(0xffffffff, a_desc_lo, static_cast<int>(stage_idx));
-                const auto& b_desc_base_lo = __shfl_sync(0xffffffff, b_desc_lo, static_cast<int>(stage_idx));
+                using mma_t = cute::conditional_t<kNumMulticast == 1, ptx::SM100_MMA_F16BF16_SS, ptx::SM100_MMA_F16BF16_2x1SM_SS>;
+                const auto runtime_instr_desc = cute::UMMA::make_runtime_instr_desc(instr_desc);
+                const auto a_desc_base_lo = __shfl_sync(0xffffffff, a_desc_lo, static_cast<int>(stage_idx));
+                const auto b_desc_base_lo = __shfl_sync(0xffffffff, b_desc_lo, static_cast<int>(stage_idx));
                 if (cute::elect_one_sync()) {
                     #pragma unroll
                     for (uint32_t k = 0; k < BLOCK_K / UMMA_K; ++ k) {
                         uint32_t atom_k_idx = k * UMMA_K / BLOCK_ATOM_K;
-                        b_desc.lo = advance_umma_desc_lo<kMajorB, LOAD_BLOCK_N, kSwizzleBMode, cutlass::bfloat16_t>(b_desc_base_lo, atom_k_idx * LOAD_BLOCK_N * BLOCK_ATOM_K, k * UMMA_K % BLOCK_ATOM_K);
-                        #pragma unroll
-                        for (uint32_t w = 0; w < kNumMWaves; ++ w) {
-                            DG_STATIC_ASSERT((WAVE_BLOCK_M * BLOCK_K) % 128 == 0, "Invalid swizzling offset");
-                            a_desc.lo = advance_umma_desc_lo<kMajorA, LOAD_BLOCK_M, kSwizzleAMode, cutlass::bfloat16_t>(a_desc_base_lo, atom_k_idx * LOAD_BLOCK_M * BLOCK_ATOM_K + w * WAVE_BLOCK_M * BLOCK_ATOM_K, k * UMMA_K % BLOCK_ATOM_K);
-                            mma_t::fma(a_desc, b_desc,
-                                       accum_stage_idx * kNumMWaves * BLOCK_N + w * BLOCK_N,
-                                       k_block_idx > 0 or k > 0,
-                                       runtime_instr_desc);
+                        a_desc.lo = mma::sm100::advance_umma_desc_lo<kMajorA, LOAD_BLOCK_M, kSwizzleAMode, cutlass::bfloat16_t>(
+                                        a_desc_base_lo, atom_k_idx * LOAD_BLOCK_M * BLOCK_ATOM_K, k * UMMA_K % BLOCK_ATOM_K);
+                        b_desc.lo = mma::sm100::advance_umma_desc_lo<kMajorB, LOAD_BLOCK_N, kSwizzleBMode, cutlass::bfloat16_t>(
+                                        b_desc_base_lo, atom_k_idx * LOAD_BLOCK_N * BLOCK_ATOM_K, k * UMMA_K % BLOCK_ATOM_K);
+                        if (kSwapAB) {
+                            mma_t::fma(b_desc, a_desc, accum_stage_idx * UMMA_N,
+                                       k_block_idx > 0 or k > 0, runtime_instr_desc);
+                        } else {
+                            mma_t::fma(a_desc, b_desc, accum_stage_idx * UMMA_N,
+                                       k_block_idx > 0 or k > 0, runtime_instr_desc);
                         }
                     }
                 }
+                __syncwarp();
 
                 // Commit to the mbarrier object
                 // No explicit `tcgen05.fence::before_thread_sync` is needed, as this is implicitly performed by `tcgen05.commit`
@@ -322,15 +343,16 @@ sm100_bf16_gemm_impl(int* grouped_layout,
                 if constexpr (kTensorCoreUtilControl < 100) {
                     // For utilization control
                     umma_arrive(reinterpret_cast<uint64_t*>(tensor_core_full_barrier));
+                    __syncwarp();
 
                     // Wait for last UMMA to be done
                     tensor_core_full_barrier->wait(tensor_core_phase);
                     tensor_core_phase ^= 1;
 
                     // Sleep for certain cycles
-                    constexpr static uint64_t kNumUMMACycles = (2ull * LAYOUT_AD_M * kNumMWaves * BLOCK_N * BLOCK_K) / 8192ull;
+                    constexpr static uint64_t kNumUMMACycles = (2ull * UMMA_M * UMMA_N * BLOCK_K) / 8192ull;
                     constexpr static uint64_t kNumDummyCycles = (100ull - kTensorCoreUtilControl) * kNumUMMACycles / kTensorCoreUtilControl;
-                    const auto& start_clock = clock64();
+                    const auto start_clock = clock64();
                     if (cute::elect_one_sync())
                         while (clock64() - start_clock < kNumDummyCycles) {}
                     __syncwarp();
@@ -339,9 +361,9 @@ sm100_bf16_gemm_impl(int* grouped_layout,
         }
 
         // To safely deconstruct barriers, we need another round of waits
-        const auto& iter_idx = scheduler.current_iter - 1;
+        const auto iter_idx = scheduler.current_iter - 1;
         if (kNumMulticast > 1 and iter_idx >= 0) {
-            const auto& accum_phase_idx = (iter_idx / kNumEpilogueStages) & 1;
+            const auto accum_phase_idx = (iter_idx / kNumEpilogueStages) & 1;
             tmem_empty_barriers[iter_idx % kNumEpilogueStages]->wait(accum_phase_idx);
         }
     } else if (warp_idx >= kNumNonEpilogueThreads / 32 and warp_idx < (kNumNonEpilogueThreads + kNumUMMAStoreThreads) / 32) {
@@ -351,19 +373,10 @@ sm100_bf16_gemm_impl(int* grouped_layout,
         // NOTES: tensor memory addresses are simplified, as the hardware will ignore the warp index bits,
         // i.e., no need for `tmem_ptr |= (epilogue_warp_idx * 32) << 16`.
         // NOTES: we also forbid two CTAs to share the same SM and its tensor memory
-        DG_TRAP_ONLY_DEVICE_ASSERT(ld_shared(tmem_ptr_in_smem) == 0);
-
-        // TMA checks
-        constexpr uint32_t kNumBankGroupBytes = 16;
-        constexpr uint32_t kNumElemsPerBankGroup = kNumBankGroupBytes / sizeof(cd_dtype_t);
-        DG_STATIC_ASSERT(kSwizzleCDMode > 0, "TMA D must be swizzled");
-        DG_STATIC_ASSERT(STORE_BLOCK_N % kNumElemsPerBankGroup == 0, "Invalid swizzling");
+        DG_TRAP_ONLY_DEVICE_ASSERT(ptx::ld_shared(tmem_ptr_in_smem) == 0);
 
         // Share store pipeline between blocks
         uint32_t tma_stage_idx = 0;
-        auto advance_store_pipeline = [&]() {
-            tma_stage_idx = (tma_stage_idx + 1) % kNumTMAStoreStages;
-        };
 
         // Persistently schedule over blocks
         while (scheduler.get_next_block(m_block_idx, n_block_idx)) {
@@ -372,106 +385,44 @@ sm100_bf16_gemm_impl(int* grouped_layout,
 
             // Wait UMMA arrival
             tmem_full_barriers[accum_stage_idx]->wait(accum_phase_idx);
-            tcgen05_after_thread_sync();
+            ptx::tcgen05_after_thread_sync();
 
             // Load from tensor memory into registers, and write shared memory with STSM
-            DG_STATIC_ASSERT(kNumEpilogueThreads == 128, "Epilogue threads not enough");
-            DG_STATIC_ASSERT(BLOCK_N % STORE_BLOCK_N == 0, "Invalid block sizes");
+            const auto tmem_base_addr = accum_stage_idx * UMMA_N;
+            const auto base_m_idx = scheduler.template get_global_idx<
+                (not is_m_grouped_contiguous(kGemmType)), sched::IndexType::MN>(shape_m, BLOCK_M, m_block_idx);
+            const auto base_n_idx = n_block_idx * BLOCK_N;
 
-            // Iterate over M waves
-            #pragma unroll
-            for (uint32_t w = 0; w < kNumMWaves; ++ w) {
-                // Issue every swizzled atom and pipeline STSM and TMA store
-                constexpr uint32_t kNumStores = BLOCK_N / STORE_BLOCK_N;
-                #pragma unroll
-                for (uint32_t s = 0; s < kNumStores; ++ s, advance_store_pipeline()) {
-                    // Wait shared memory to be released
-                    if (epilogue_warp_idx == 0)
-                        cute::tma_store_wait<kNumTMAStoreStages - 1>();
-                    cutlass::arch::NamedBarrier::sync(kNumUMMAStoreThreads, 0);
-
-                    // The pipeline stage
-                    const auto m_idx = scheduler.template get_global_idx<(not is_m_grouped_contiguous(kGemmType)), IndexType::MN>(shape_m, BLOCK_M, m_block_idx) + w * WAVE_BLOCK_M;
-                    const auto n_idx = n_block_idx * BLOCK_N + s * STORE_BLOCK_N;
-
-                    // Store into shared memory
-                    #pragma unroll
-                    for (uint32_t i = 0; i < STORE_BLOCK_N / kNumElemsPerBankGroup; ++ i) {
-                        // Calculate the index of the bank group to be written in the atom
-                        auto bank_group_index = i + lane_idx * (kSwizzleCDMode / kNumBankGroupBytes);
-
-                        // Reshape the atom in another view and swizzle
-                        //  - original: `(LAYOUT_AD_M, kSwizzleCDMode / kNumBankGroupBytes)`
-                        //  - new: `(LAYOUT_AD_M * kSwizzleCDMode / kNumBankGroupBytes / 8, 8)`
-                        // NOTES: "8" is the number of bank groups, "16" is the swizzling pattern
-                        constexpr bool kHasShortcut = (kSwizzleCDMode / kNumBankGroupBytes) == 8;
-                        auto row = kHasShortcut ? (i / 8 + lane_idx) : (bank_group_index / 8);
-                        auto col = kHasShortcut ? (i) : (bank_group_index % 8);
-                        col ^= row % (kSwizzleCDMode / 16);
-
-                        // Source and destination memory address
-                        uint32_t tmem_addr = accum_stage_idx * kNumMWaves * BLOCK_N +               // Accumulator offset
-                                             w * BLOCK_N +                                          // Wave offset
-                                             s * STORE_BLOCK_N + i * kNumElemsPerBankGroup;         // In-block offset
-                        auto smem_ptr = reinterpret_cast<uint8_t*>(smem_cd[tma_stage_idx]) +        // Base pointer
-                                        epilogue_warp_idx * 32 * kSwizzleCDMode +                   // Warp offset
-                                        row * (kNumBankGroupBytes * 8) + col * kNumBankGroupBytes;  // In-atom offset
-
-                        // Load from tensor memory, store into shared memory
-                        uint32_t values[kNumElemsPerBankGroup];
-                        if constexpr (cute::is_same_v<cd_dtype_t, float>) {
-                            // For FP32 output, read and store
-                            DG_STATIC_ASSERT(kNumElemsPerBankGroup == 4, "Invalid type");
-                            cute::SM100_TMEM_LOAD_32dp32b4x::copy(tmem_addr,
-                                values[0], values[1], values[2], values[3]);
-                            cutlass::arch::fence_view_async_tmem_load();
-                            st_shared(smem_ptr, values[0], values[1], values[2], values[3]);
-                        } else {
-                            // For BF16 output, read, cast and store
-                            DG_STATIC_ASSERT(kNumElemsPerBankGroup == 8 and cute::is_same_v<cd_dtype_t, cutlass::bfloat16_t>, "Invalid type");
-                            cute::SM100_TMEM_LOAD_32dp32b8x::copy(tmem_addr,
-                                values[0], values[1], values[2], values[3],
-                                values[4], values[5], values[6], values[7]);
-                            cutlass::arch::fence_view_async_tmem_load();
-                            st_shared(smem_ptr,
-                                      cast_into_bf16_and_pack(values[0], values[1]),
-                                      cast_into_bf16_and_pack(values[2], values[3]),
-                                      cast_into_bf16_and_pack(values[4], values[5]),
-                                      cast_into_bf16_and_pack(values[6], values[7]));
-                        }
-                    }
-
-                    // Notify tensor memory empty (only at the leader CTA) arrival ASAP
-                    // NOTES: only the last stage needs to do this
-                    if (w == kNumMWaves - 1 and s == BLOCK_N / STORE_BLOCK_N - 1) {
-                        tcgen05_before_thread_sync();
-                        tmem_empty_barriers[accum_stage_idx]->arrive(0u);
-                    }
-                    __syncwarp();
-
-                    // Synchronize all threads and issue TMA
-                    cute::tma_store_fence();
-                    cutlass::arch::NamedBarrier::sync(kNumUMMAStoreThreads, 0);
-                    if (epilogue_warp_idx == 0 and cute::elect_one_sync()) {
-                        if constexpr (kGemmType == GemmType::Batched) {
-                            using cute_tma_t = cute::conditional_t<kWithAccumulation,
-                                cute::SM90_TMA_REDUCE_ADD_3D, cute::SM90_TMA_STORE_3D>;
-                            cute_tma_t::copy(&tensor_map_cd, smem_cd[tma_stage_idx],
-                                             n_idx, m_idx, scheduler.current_group_idx);
-                        } else {
-                            using cute_tma_t = cute::conditional_t<kWithAccumulation,
-                                cute::SM90_TMA_REDUCE_ADD_2D, cute::SM90_TMA_STORE_2D>;
-                            cute_tma_t::copy(&tensor_map_cd, smem_cd[tma_stage_idx], n_idx, m_idx);
-                        }
-                        cute::tma_store_arrive();
-                    }
-                }
+            if constexpr (kSwapAB) {
+                const auto effective_m = scheduler.get_aligned_effective_m_in_block(m_block_idx);
+                epilogue::sm100_store_cd_swap_ab<BLOCK_M, BLOCK_N, STORE_BLOCK_M, STORE_BLOCK_N,
+                    kSwizzleCDMode, kNumTMAStoreStages, kNumUMMAStoreThreads,
+                    kGemmType, kWithAccumulation,
+                    cd_dtype_t, epilogue::transform::EpilogueIdentity>
+                (smem_cd, tma_stage_idx, tmem_base_addr,
+                 base_m_idx, base_n_idx, scheduler.current_group_idx,
+                 effective_m,
+                 epilogue_warp_idx, lane_idx,
+                 tmem_empty_barriers[accum_stage_idx],
+                 tensor_map_cd);
+            } else {
+                epilogue::sm100_store_cd<BLOCK_M, BLOCK_N, STORE_BLOCK_M, STORE_BLOCK_N,
+                    kSwizzleCDMode, kNumTMAStoreStages, kNumUMMAStoreThreads,
+                    kGemmType, kWithAccumulation,
+                    cd_dtype_t, epilogue::transform::EpilogueIdentity>
+                (smem_cd, tma_stage_idx, tmem_base_addr,
+                 base_m_idx, base_n_idx, scheduler.current_group_idx,
+                 epilogue_warp_idx, lane_idx,
+                 tmem_empty_barriers[accum_stage_idx],
+                 tensor_map_cd);
             }
         }
     }
 
-    // Deallocate tensor memory
+    // TODO: Remove redundant synchronization
     kNumMulticast > 1 ? cute::cluster_sync() : __syncthreads();
+
+    // Deallocate tensor memory
     if (warp_idx == 0)
         Allocator().free(0, kNumTmemCols);
 

deep_gemm/include/deep_gemm/impls/sm100_bmk_bnk_mn.cuh

@@ -5,18 +5,19 @@
 #include <cutlass/arch/barrier.h>
 
 #include <deep_gemm/common/utils.cuh>
-#include <deep_gemm/common/sm100_utils.cuh>
+#include <deep_gemm/mma/sm100.cuh>
+#include <deep_gemm/ptx/ld_st.cuh>
+#include <deep_gemm/ptx/tcgen05.cuh>
+#include <deep_gemm/ptx/utils.cuh>
 
 namespace deep_gemm {
 
-using namespace deep_gemm::sm100;
-
 template <uint32_t SHAPE_M, uint32_t SHAPE_N, uint32_t SHAPE_K,
           uint32_t BLOCK_M, uint32_t BLOCK_N, uint32_t BLOCK_K,
           uint32_t kSplitFactor,
           uint32_t kSwizzleABMode, uint32_t kSwizzleCDMode,
           uint32_t kNumStages, uint32_t kNumThreads>
-__global__ void __launch_bounds__(kNumThreads, 1)
+CUTLASS_GLOBAL void __launch_bounds__(kNumThreads, 1)
 sm100_bmn_bnk_mn_gemm_impl(uint32_t shape_s,
                            const __grid_constant__ cute::TmaDescriptor tensor_map_a,
                            const __grid_constant__ cute::TmaDescriptor tensor_map_b,
@@ -30,7 +31,7 @@ sm100_bmn_bnk_mn_gemm_impl(uint32_t shape_s,
 
     // Utils
     const auto warp_idx = cutlass::canonical_warp_idx_sync();
-    const auto lane_idx = get_lane_idx();
+    const auto lane_idx = ptx::get_lane_idx();
     DG_STATIC_ASSERT(BLOCK_M == LAYOUT_AD_M and BLOCK_N == 128 and BLOCK_K == 64, "Invalid block size");
     DG_STATIC_ASSERT(kSwizzleABMode == 128 and kSwizzleCDMode == 128, "Invalid swizzle mode");
 
@@ -51,24 +52,24 @@ sm100_bmn_bnk_mn_gemm_impl(uint32_t shape_s,
     }
 
     // Real tensor memory size and offsets
-    constexpr uint32_t kNumTmemCols = get_num_aligned_tmem_cols<BLOCK_N>();
+    constexpr uint32_t kNumTmemCols = utils::get_num_aligned_tmem_cols<BLOCK_N>();
 
     // Fill D/A/B
-    auto smem_cd = PatternVisitor([&](const uint32_t& i) {
+    auto smem_cd = utils::PatternVisitor([&](const uint32_t& i) {
         return reinterpret_cast<float*>(smem_buffer + (i * SMEM_CD_SIZE_PER_STAGE));
     });
-    auto smem_a  = PatternVisitor([&](const uint32_t& i) {
+    auto smem_a  = utils::PatternVisitor([&](const uint32_t& i) {
         return reinterpret_cast<cutlass::bfloat16_t*>(smem_buffer + (SMEM_CD_SIZE + i * SMEM_A_SIZE_PER_STAGE));
     });
-    auto smem_b  = PatternVisitor([&](const uint32_t& i) {
+    auto smem_b  = utils::PatternVisitor([&](const uint32_t& i) {
         return reinterpret_cast<cutlass::bfloat16_t*>(smem_buffer + (SMEM_CD_SIZE + kNumStages * SMEM_A_SIZE_PER_STAGE + i * SMEM_B_SIZE_PER_STAGE));
     });
 
     // Fill barriers
     auto barrier_start_ptr = reinterpret_cast<Barrier*>(smem_buffer + SMEM_CD_SIZE +
             kNumStages * (SMEM_A_SIZE_PER_STAGE + SMEM_B_SIZE_PER_STAGE));
-    auto full_barriers     = PatternVisitor([=](const uint32_t& i) { return barrier_start_ptr + (i); });
-    auto empty_barriers    = PatternVisitor([=](const uint32_t& i) { return barrier_start_ptr + (kNumStages + i); });
+    auto full_barriers     = utils::PatternVisitor([=](const uint32_t& i) { return barrier_start_ptr + (i); });
+    auto empty_barriers    = utils::PatternVisitor([=](const uint32_t& i) { return barrier_start_ptr + (kNumStages + i); });
     auto tmem_full_barrier = barrier_start_ptr + (kNumStages * 2);
 
     // Fill the tensor memory pointer
@@ -93,14 +94,17 @@ sm100_bmn_bnk_mn_gemm_impl(uint32_t shape_s,
     __syncthreads();
 
     // Block indices
-    const uint32_t num_n_blocks = ceil_div(SHAPE_N, BLOCK_N);
-    const uint32_t num_mn_blocks = num_n_blocks * ceil_div(SHAPE_M, BLOCK_M);
+    const uint32_t num_n_blocks = math::ceil_div(SHAPE_N, BLOCK_N);
+    const uint32_t num_mn_blocks = num_n_blocks * math::ceil_div(SHAPE_M, BLOCK_M);
     const uint32_t mn_block_idx = blockIdx.x % num_mn_blocks;
     const uint32_t sk_block_idx = blockIdx.x / num_mn_blocks;
     const uint32_t n_block_idx = mn_block_idx % num_n_blocks;
     const uint32_t m_block_idx = mn_block_idx / num_n_blocks;
     const uint32_t num_total_stages = cute::min(kSplitFactor, shape_s * (SHAPE_K / BLOCK_K) - sk_block_idx * kSplitFactor);
 
+    // Wait for primary kernel completion
+    cudaGridDependencySynchronize();
+
     if (warp_idx == 0) {
         // TMA load warp
         for (uint32_t s = 0; s < num_total_stages; ++ s) {
@@ -115,8 +119,8 @@ sm100_bmn_bnk_mn_gemm_impl(uint32_t shape_s,
 
             // Issue TMAs
             if (cute::elect_one_sync()) {
-                tma_copy<BLOCK_K, BLOCK_M, kSwizzleABMode>(&tensor_map_a, full_barriers[stage_idx], smem_a[stage_idx], k_idx, m_idx + s_idx * SHAPE_M);
-                tma_copy<BLOCK_K, BLOCK_N, kSwizzleABMode>(&tensor_map_b, full_barriers[stage_idx], smem_b[stage_idx], k_idx, n_idx + s_idx * SHAPE_N);
+                tma::copy<BLOCK_K, BLOCK_M, kSwizzleABMode>(&tensor_map_a, full_barriers[stage_idx], smem_a[stage_idx], k_idx, m_idx + s_idx * SHAPE_M);
+                tma::copy<BLOCK_K, BLOCK_N, kSwizzleABMode>(&tensor_map_b, full_barriers[stage_idx], smem_b[stage_idx], k_idx, n_idx + s_idx * SHAPE_N);
             }
 
             // Arrive at full barriers
@@ -134,8 +138,8 @@ sm100_bmn_bnk_mn_gemm_impl(uint32_t shape_s,
         auto instr_desc = cute::UMMA::make_instr_desc<cutlass::bfloat16_t, cutlass::bfloat16_t, float, UMMA_M, UMMA_N, cute::UMMA::Major::K, cute::UMMA::Major::K>();
 
         DG_STATIC_ASSERT(kNumStages <= 32, "Too many stages");
-        auto a_desc = make_umma_desc<cute::UMMA::Major::K, BLOCK_M, BLOCK_K, kSwizzleABMode>(smem_a[0], 0, 0);
-        auto b_desc = make_umma_desc<cute::UMMA::Major::K, BLOCK_N, BLOCK_K, kSwizzleABMode>(smem_b[0], 0, 0);
+        auto a_desc = mma::sm100::make_umma_desc<cute::UMMA::Major::K, BLOCK_M, BLOCK_K, kSwizzleABMode>(smem_a[0], 0, 0);
+        auto b_desc = mma::sm100::make_umma_desc<cute::UMMA::Major::K, BLOCK_N, BLOCK_K, kSwizzleABMode>(smem_b[0], 0, 0);
         uint32_t a_desc_lo = lane_idx < kNumStages ? a_desc.lo + lane_idx * SMEM_A_SIZE_PER_STAGE / 16 : 0u;
         uint32_t b_desc_lo = lane_idx < kNumStages ? b_desc.lo + lane_idx * SMEM_B_SIZE_PER_STAGE / 16 : 0u;
 
@@ -147,14 +151,14 @@ sm100_bmn_bnk_mn_gemm_impl(uint32_t shape_s,
                          "Invalid MMA instruction shape");
 
         // Wait tensor memory empty barrier arrival
-        tcgen05_after_thread_sync();
+        ptx::tcgen05_after_thread_sync();
 
         // Launch MMAs
         for (uint32_t s = 0; s < num_total_stages; ++ s) {
             // Wait TMA arrival
             const auto& stage_idx = s % kNumStages;
             full_barriers[stage_idx]->wait((s / kNumStages) & 1);
-            tcgen05_after_thread_sync();
+            ptx::tcgen05_after_thread_sync();
 
             // Issue UMMA in the leader CTA
             const auto& runtime_instr_desc = cute::UMMA::make_runtime_instr_desc(instr_desc);
@@ -163,9 +167,11 @@ sm100_bmn_bnk_mn_gemm_impl(uint32_t shape_s,
             if (cute::elect_one_sync()) {
                 #pragma unroll
                 for (uint32_t k = 0; k < BLOCK_K / UMMA_K; ++ k) {
-                    a_desc.lo = advance_umma_desc_lo<cute::UMMA::Major::K, BLOCK_M, kSwizzleABMode, cutlass::bfloat16_t>(a_desc_base_lo, 0, k * UMMA_K);
-                    b_desc.lo = advance_umma_desc_lo<cute::UMMA::Major::K, BLOCK_N, kSwizzleABMode, cutlass::bfloat16_t>(b_desc_base_lo, 0, k * UMMA_K);
-                    SM100_MMA_F16BF16_SS::fma(a_desc, b_desc, 0, s > 0 or k > 0, runtime_instr_desc);
+                    a_desc.lo = mma::sm100::advance_umma_desc_lo<cute::UMMA::Major::K, BLOCK_M, kSwizzleABMode, cutlass::bfloat16_t>(
+                        a_desc_base_lo, 0, k * UMMA_K);
+                    b_desc.lo = mma::sm100::advance_umma_desc_lo<cute::UMMA::Major::K, BLOCK_N, kSwizzleABMode, cutlass::bfloat16_t>(
+                        b_desc_base_lo, 0, k * UMMA_K);
+                    ptx::SM100_MMA_F16BF16_SS::fma(a_desc, b_desc, 0, s > 0 or k > 0, runtime_instr_desc);
                 }
             }
 
@@ -180,7 +186,7 @@ sm100_bmn_bnk_mn_gemm_impl(uint32_t shape_s,
     // i.e., no need for `tmem_ptr |= (warp_idx * 32) << 16`.
     // NOTES: we also forbid two CTAs to share the same SM and its tensor memory
     if (warp_idx == 2)
-        DG_TRAP_ONLY_DEVICE_ASSERT(ld_shared(tmem_ptr_in_smem) == 0);
+        DG_TRAP_ONLY_DEVICE_ASSERT(ptx::ld_shared(tmem_ptr_in_smem) == 0);
 
     // TMA checks
     constexpr uint32_t kNumBankGroupBytes = 16;
@@ -191,7 +197,7 @@ sm100_bmn_bnk_mn_gemm_impl(uint32_t shape_s,
 
     // Wait UMMA arrival
     tmem_full_barrier->wait(0);
-    tcgen05_after_thread_sync();
+    ptx::tcgen05_after_thread_sync();
 
     // Load from tensor memory into registers, and write shared memory with STSM
     DG_STATIC_ASSERT(BLOCK_N % STORE_BLOCK_N == 0, "Invalid block sizes");
@@ -239,7 +245,7 @@ sm100_bmn_bnk_mn_gemm_impl(uint32_t shape_s,
             cute::SM100_TMEM_LOAD_32dp32b4x::copy(tmem_addr,
                 values[0], values[1], values[2], values[3]);
             cutlass::arch::fence_view_async_tmem_load();
-            st_shared(smem_ptr, values[0], values[1], values[2], values[3]);
+            ptx::st_shared(smem_ptr, values[0], values[1], values[2], values[3]);
         }
 
         // Synchronize all threads and issue TMA
@@ -251,7 +257,6 @@ sm100_bmn_bnk_mn_gemm_impl(uint32_t shape_s,
         }
     }
 
-    __syncthreads();
     // Deallocate tensor memory by warp 1
     // NOTES: warp 0 is doing TMA stores
     if (warp_idx == 1)

deep_gemm/include/deep_gemm/impls/sm100_fp4_mqa_logits.cuh

@@ -0,0 +1,456 @@
+#pragma once
+
+#include <cutlass/arch/barrier.h>
+#include <cutlass/arch/reg_reconfig.h>
+
+#include <cute/arch/cluster_sm90.hpp>
+#include <cute/arch/copy_sm90_desc.hpp>
+
+#include <deep_gemm/common/cute_tie.cuh>
+#include <deep_gemm/common/utils.cuh>
+#include <deep_gemm/mma/sm100.cuh>
+#include <deep_gemm/ptx/ld_st.cuh>
+#include <deep_gemm/ptx/tcgen05.cuh>
+#include <deep_gemm/ptx/utils.cuh>
+
+namespace deep_gemm {
+
+template <uint32_t kNumHeads, uint32_t kHeadDim,
+          bool kIsCompressedLogits,
+          uint32_t BLOCK_Q, uint32_t BLOCK_KV,
+          uint32_t kNumQStages, uint32_t kNumKVStages,
+          uint32_t kNumSMs,
+          uint32_t kNumSpecializedThreads, uint32_t kNumMathThreads,
+          typename logits_dtype_t,
+          uint32_t kNumMathWarpGroups = kNumMathThreads / 128>
+CUTLASS_GLOBAL __launch_bounds__(kNumSpecializedThreads + kNumMathThreads, 1)
+void sm100_fp4_mqa_logits(const uint32_t seq_len, const uint32_t seq_len_kv,
+                          const uint32_t max_seqlen_k,
+                          const uint32_t logits_stride,
+                          const uint32_t* cu_seq_len_k_start,
+                          const uint32_t* cu_seq_len_k_end,
+                          logits_dtype_t* logits,
+                          const __grid_constant__ cute::TmaDescriptor tensor_map_q,
+                          const __grid_constant__ cute::TmaDescriptor tensor_map_sf_q,
+                          const __grid_constant__ cute::TmaDescriptor tensor_map_kv,
+                          const __grid_constant__ cute::TmaDescriptor tensor_map_sf_kv,
+                          const __grid_constant__ cute::TmaDescriptor tensor_map_weights) {
+    using Barrier = cutlass::arch::ClusterTransactionBarrier;
+
+    // Utils
+    const auto sm_idx = blockIdx.x;
+    const auto warp_idx = cutlass::canonical_warp_idx_sync();
+    const auto warpgroup_idx = warp_idx / 4;
+    const auto lane_idx = ptx::get_lane_idx();
+    constexpr uint32_t kSpecWarpStart = kNumMathWarpGroups * 4;
+
+    // Prefetch TMA descriptors
+    if (warp_idx == kSpecWarpStart) {
+        cute::prefetch_tma_descriptor(&tensor_map_q);
+        cute::prefetch_tma_descriptor(&tensor_map_sf_q);
+        cute::prefetch_tma_descriptor(&tensor_map_weights);
+        cute::prefetch_tma_descriptor(&tensor_map_kv);
+        cute::prefetch_tma_descriptor(&tensor_map_sf_kv);
+    }
+
+    // UMMA configs
+    static constexpr uint32_t kNumTmemStages = 3;
+    static constexpr uint32_t kNumUTCCPAlignedElems = 128;
+    static constexpr uint32_t UMMA_M = 128;
+    static constexpr uint32_t UMMA_N = BLOCK_Q * kNumHeads;
+    static constexpr uint32_t UMMA_K = 64;
+    static constexpr uint32_t kNumSFQ  = math::constexpr_align(BLOCK_Q * kNumHeads, kNumUTCCPAlignedElems);
+    static constexpr uint32_t kNumSFKV = math::constexpr_align(BLOCK_KV, kNumUTCCPAlignedElems);
+    static constexpr uint32_t kRealNumSFQ = BLOCK_Q * kNumHeads;
+    DG_STATIC_ASSERT(kNumSpecializedThreads == 128 and kNumMathThreads % 128 == 0, "Invalid threads");
+    DG_STATIC_ASSERT(BLOCK_KV == kNumMathWarpGroups * UMMA_M and BLOCK_KV % kNumUTCCPAlignedElems == 0, "Invalid `BLOCK_KV`");
+
+    // Shared memory configs
+    static constexpr uint32_t kSwizzleAlignment = 8 * (kHeadDim / 2);
+    static constexpr uint32_t SMEM_Q_SIZE_PER_STAGE      = BLOCK_Q * kNumHeads * (kHeadDim / 2);
+    static constexpr uint32_t SMEM_SF_Q_SIZE_PER_STAGE   = kNumSFQ * sizeof(int);
+    static constexpr uint32_t SMEM_KV_SIZE_PER_STAGE     = BLOCK_KV * (kHeadDim / 2);
+    static constexpr uint32_t SMEM_SF_KV_SIZE_PER_STAGE  = kNumSFKV * sizeof(int);
+    static constexpr uint32_t SMEM_WEIGHT_SIZE_PER_STAGE = BLOCK_Q * kNumHeads * sizeof(float);
+
+    // Align to swizzling alignment bytes
+    extern __shared__ __align__(kSwizzleAlignment) uint8_t smem_buffer[];
+    DG_STATIC_ASSERT(SMEM_Q_SIZE_PER_STAGE  % kSwizzleAlignment == 0, "Unaligned TMA swizzling");
+    DG_STATIC_ASSERT(SMEM_KV_SIZE_PER_STAGE % kSwizzleAlignment == 0, "Unaligned TMA swizzling");
+
+    // Q and KV data on shared memory
+    auto smem_q = utils::PatternVisitor([&](const uint32_t& i) {
+        return smem_buffer + SMEM_Q_SIZE_PER_STAGE * i;
+    });
+    auto smem_kv = utils::PatternVisitor([&](const uint32_t& i) {
+        return smem_buffer + SMEM_Q_SIZE_PER_STAGE * kNumQStages + SMEM_KV_SIZE_PER_STAGE * i;
+    });
+    const auto smem_sf_ptr = smem_buffer + (SMEM_Q_SIZE_PER_STAGE * kNumQStages + SMEM_KV_SIZE_PER_STAGE * kNumKVStages);
+    auto smem_sf_q = utils::PatternVisitor([&](const uint32_t& i) {
+        return reinterpret_cast<uint32_t*>(smem_sf_ptr + SMEM_SF_Q_SIZE_PER_STAGE * i);
+    });
+    auto smem_sf_kv = utils::PatternVisitor([&](const uint32_t& i) {
+        return reinterpret_cast<uint32_t*>(smem_sf_ptr + SMEM_SF_Q_SIZE_PER_STAGE * kNumQStages + SMEM_SF_KV_SIZE_PER_STAGE * i);
+    });
+    auto smem_weights = utils::PatternVisitor([&](const uint32_t& i) {
+        return reinterpret_cast<float*>(smem_sf_ptr + SMEM_SF_Q_SIZE_PER_STAGE * kNumQStages + SMEM_SF_KV_SIZE_PER_STAGE * kNumKVStages
+                                                    + SMEM_WEIGHT_SIZE_PER_STAGE * i);
+    });
+
+    // Barriers and TMEM pointer on shared memory
+    const auto barrier_ptr = reinterpret_cast<Barrier*>(smem_weights[kNumQStages]);
+    auto full_q_barriers     = utils::PatternVisitor([&](const uint32_t& i) { return barrier_ptr + i; });
+    auto empty_q_barriers    = utils::PatternVisitor([&](const uint32_t& i) { return barrier_ptr + kNumQStages + i; });
+    auto full_kv_barriers    = utils::PatternVisitor([&](const uint32_t& i) { return barrier_ptr + kNumQStages * 2 + i; });
+    auto empty_kv_barriers   = utils::PatternVisitor([&](const uint32_t& i) { return barrier_ptr + kNumQStages * 2 + kNumKVStages + i; });
+    const auto tmem_barrier_ptr = barrier_ptr + kNumQStages * 2 + kNumKVStages * 2;
+    auto full_tmem_barriers  = utils::PatternVisitor([&](const uint32_t& i) { return tmem_barrier_ptr + i; });
+    auto empty_tmem_barriers = utils::PatternVisitor([&](const uint32_t& i) { return tmem_barrier_ptr + kNumTmemStages + i; });
+    auto tmem_ptr_in_smem    = reinterpret_cast<uint32_t*>(tmem_barrier_ptr + kNumTmemStages * 2);
+
+    // Tensor memory configs
+    constexpr uint32_t kNumAccumTmemCols = BLOCK_Q * kNumHeads * kNumTmemStages;
+    constexpr uint32_t kNumTmemCols = utils::get_num_aligned_tmem_cols<kNumAccumTmemCols + kNumSFQ / 32 + kNumSFKV / 32>();
+    constexpr uint32_t kTmemStartColOfSFQ = kNumAccumTmemCols;
+    constexpr uint32_t kTmemStartColOfSFKV = kNumAccumTmemCols + kNumSFQ / 32;
+    DG_STATIC_ASSERT(kNumTmemCols <= 512, "Too many tensor memory");
+
+    // Initialize barriers
+    if (warp_idx == kSpecWarpStart + 1 and cute::elect_one_sync()) {
+        #pragma unroll
+        for (uint32_t i = 0; i < kNumQStages; ++ i) {
+            full_q_barriers[i]->init(1);
+            empty_q_barriers[i]->init(kNumMathThreads + 32);
+        }
+        #pragma unroll
+        for (uint32_t i = 0; i < kNumKVStages; ++ i) {
+            full_kv_barriers[i]->init(1);
+            empty_kv_barriers[i]->init(1);
+        }
+        #pragma unroll
+        for (uint32_t i = 0; i < kNumTmemStages; ++i) {
+            full_tmem_barriers[i]->init(1);
+            empty_tmem_barriers[i]->init(128);
+        }
+        cutlass::arch::fence_barrier_init();
+    }
+
+    // Allocate tensor memory
+    if (warp_idx == kSpecWarpStart + 2)
+        cute::TMEM::Allocator1Sm().allocate(kNumTmemCols, tmem_ptr_in_smem);
+    __syncthreads();
+
+    // Scheduler
+    const uint32_t num_q_blocks = math::ceil_div(seq_len, BLOCK_Q);
+    uint32_t seq_k_start[BLOCK_Q], seq_k_end[BLOCK_Q];
+    auto load_schedule = [&](const uint32_t& q_idx) -> cute::tuple<uint32_t, uint32_t> {
+        uint32_t start = cute::numeric_limits<uint32_t>::max();
+        uint32_t end = cute::numeric_limits<uint32_t>::min();
+        #pragma unroll
+        for (uint32_t i = 0; i < BLOCK_Q; ++ i) {
+            const auto row_idx = cute::min(q_idx * BLOCK_Q + i, seq_len - 1);
+            seq_k_start[i] = cute::min(cu_seq_len_k_start[row_idx], seq_len_kv);
+            seq_k_end[i] = cute::min(cu_seq_len_k_end[row_idx], seq_len_kv);
+            start = cute::min(start, seq_k_start[i]);
+            end = cute::max(end, seq_k_end[i]);
+        }
+        // TMA alignment requirements for SF KV
+        start = start / 4 * 4;
+        return {start, math::ceil_div(end - start, BLOCK_KV)};
+    };
+
+    // Make Q, KV and TMEM pipeline
+    auto make_pipeline = [](const uint32_t& num_stages) {
+        // Return current stage and phase, and advance pipeline by steps
+        return [iter_idx = 0u, num_stages](const uint32_t& step = 1) mutable -> cute::tuple<uint32_t, uint32_t> {
+            uint32_t current_idx = iter_idx;
+            iter_idx += step;
+            return {current_idx % num_stages, (current_idx / num_stages) & 1};
+        };
+    };
+    auto advance_q_pipeline    = make_pipeline(kNumQStages);
+    auto advance_kv_pipeline   = make_pipeline(kNumKVStages);
+    auto advance_tmem_pipeline = make_pipeline(kNumTmemStages);
+
+    // Register reconfigurations
+    constexpr uint32_t kNumSpecializedRegisters = 56;
+    constexpr uint32_t kNumMathRegisters = 224;
+
+    // Wait for primary kernel completion
+    cudaGridDependencySynchronize();
+
+    if (warp_idx == kSpecWarpStart) {
+        // TMA warp for loading Q
+        cutlass::arch::warpgroup_reg_dealloc<kNumSpecializedRegisters>();
+
+        // Enumerate Q blocks
+        if (cute::elect_one_sync()) {
+            for (uint32_t q_idx = sm_idx; q_idx < num_q_blocks; q_idx += kNumSMs) {
+                // Wait Q consumer release
+                CUTE_TIE_DECL(advance_q_pipeline(), q_stage_idx, q_phase);
+                empty_q_barriers[q_stage_idx]->wait(q_phase ^ 1);
+
+                // Issue TMA Q
+                cute::SM90_TMA_LOAD_2D::copy(&tensor_map_q, reinterpret_cast<uint64_t*>(full_q_barriers[q_stage_idx]),
+                                            static_cast<uint64_t>(cute::TMA::CacheHintSm100::EVICT_NORMAL),
+                                            smem_q[q_stage_idx], 0, q_idx * BLOCK_Q * kNumHeads);
+                tma::copy<BLOCK_Q * kNumHeads, 1, 0>(&tensor_map_sf_q, full_q_barriers[q_stage_idx], smem_sf_q[q_stage_idx], 0, q_idx * BLOCK_Q);
+                tma::copy<kNumHeads, BLOCK_Q, 0>(&tensor_map_weights, full_q_barriers[q_stage_idx], smem_weights[q_stage_idx], 0, q_idx * BLOCK_Q);
+                full_q_barriers[q_stage_idx]->arrive_and_expect_tx(SMEM_Q_SIZE_PER_STAGE + kRealNumSFQ * sizeof(int) + SMEM_WEIGHT_SIZE_PER_STAGE);
+            }
+        }
+        __syncwarp();
+    } else if (warp_idx == kSpecWarpStart + 1) {
+        // TMA warp for loading KV cache
+        cutlass::arch::warpgroup_reg_dealloc<kNumSpecializedRegisters>();
+
+        if (cute::elect_one_sync()) {
+            // Enumerate Q blocks
+            for (uint32_t q_idx = sm_idx; q_idx < num_q_blocks; q_idx += kNumSMs) {
+                // Load KV block ranges
+                CUTE_TIE_DECL(load_schedule(q_idx), kv_start, num_kv_blocks);
+
+                // Enumerate KV blocks
+                for (uint32_t kv_idx = 0; kv_idx < num_kv_blocks; ++ kv_idx) {
+                    // Wait KV consumer release
+                    CUTE_TIE_DECL(advance_kv_pipeline(), kv_stage_idx, kv_phase);
+                    empty_kv_barriers[kv_stage_idx]->wait(kv_phase ^ 1);
+
+                    // Issue TMA KV
+                    cute::SM90_TMA_LOAD_2D::copy(&tensor_map_kv, reinterpret_cast<uint64_t*>(full_kv_barriers[kv_stage_idx]),
+                                                 static_cast<uint64_t>(cute::TMA::CacheHintSm100::EVICT_NORMAL),
+                                                 smem_kv[kv_stage_idx], 0, kv_start + kv_idx * BLOCK_KV);
+                    tma::copy<BLOCK_KV, 1, 0>(&tensor_map_sf_kv, full_kv_barriers[kv_stage_idx],
+                                              smem_sf_kv[kv_stage_idx],
+                                              kv_start + kv_idx * BLOCK_KV, 0);
+                    full_kv_barriers[kv_stage_idx]->arrive_and_expect_tx(SMEM_KV_SIZE_PER_STAGE + SMEM_SF_KV_SIZE_PER_STAGE);
+                }
+            }
+        }
+    } else if (warp_idx == kSpecWarpStart + 2) {
+        // UMMA warp
+        cutlass::arch::warpgroup_reg_dealloc<kNumSpecializedRegisters>();
+        DG_TRAP_ONLY_DEVICE_ASSERT(ptx::ld_shared(tmem_ptr_in_smem) == 0);
+
+        // UTCCP transposer
+        auto utccp_required_smem_warp_transpose = [&](const uint32_t* smem_ptr) {
+            DG_STATIC_ASSERT(kNumUTCCPAlignedElems == 128, "Invalid aligned elements");
+            uint32_t values[4];
+            #pragma unroll
+            for (uint32_t i = 0; i < 4; ++ i)
+                values[i] = ptx::ld_shared(smem_ptr + (i ^ (lane_idx >> 3)) * 32 + lane_idx);
+            __syncwarp();
+            #pragma unroll
+            for (uint32_t i = 0; i < 4; ++ i)
+                ptx::st_shared(smem_ptr + lane_idx * 4 + (i ^ (lane_idx >> 3)), values[i]);
+        };
+
+        // Make UMMA desc
+        auto instr_desc = cute::UMMA::make_instr_desc_block_scaled<cutlass::float_e2m1_t, cutlass::float_e2m1_t, float, cutlass::float_ue8m0_t,
+                                                                   UMMA_M, UMMA_N, cute::UMMA::Major::K, cute::UMMA::Major::K>();
+        auto sf_desc = mma::sm100::make_sf_desc(nullptr);
+
+        // Enumerate Q blocks
+        for (uint32_t q_idx = sm_idx; q_idx < num_q_blocks; q_idx += kNumSMs) {
+            // Load KV block ranges
+            CUTE_TIE_DECL(load_schedule(q_idx), kv_start, num_kv_blocks);
+
+            // Wait TMA Q arrivals
+            CUTE_TIE_DECL(advance_q_pipeline(), q_stage_idx, q_phase);
+            full_q_barriers[q_stage_idx]->wait(q_phase);
+
+            // Transpose and copy SF Q
+            #pragma unroll
+            for (uint32_t i = 0; i < kNumSFQ / kNumUTCCPAlignedElems; ++ i) {
+                auto smem_ptr = smem_sf_q[q_stage_idx] + i * kNumUTCCPAlignedElems;
+                utccp_required_smem_warp_transpose(smem_ptr);
+                cutlass::arch::fence_view_async_shared();
+                mma::sm100::replace_smem_desc_addr(sf_desc, smem_ptr);
+                if (cute::elect_one_sync())
+                    cute::SM100_UTCCP_4x32dp128bit_1cta::copy(sf_desc, kTmemStartColOfSFQ + i * 4);
+                __syncwarp();
+            }
+
+            // Enumerate KV blocks
+            for (uint32_t kv_idx = 0; kv_idx < num_kv_blocks; ++ kv_idx) {
+                // Wait TMA KV arrivals
+                CUTE_TIE_DECL(advance_kv_pipeline(), kv_stage_idx, kv_phase);
+                full_kv_barriers[kv_stage_idx]->wait(kv_phase);
+
+                // Transpose
+                #pragma unroll
+                for (uint32_t i = 0; i < kNumSFKV / kNumUTCCPAlignedElems; ++ i) {
+                    auto smem_ptr = smem_sf_kv[kv_stage_idx] + i * kNumUTCCPAlignedElems;
+                    utccp_required_smem_warp_transpose(smem_ptr);
+                    cutlass::arch::fence_view_async_shared();
+                }
+
+                // UMMA with SF
+                if (cute::elect_one_sync()) {
+                    // Copy SF KV
+                    #pragma unroll
+                    for (uint32_t i = 0; i < kNumSFKV / kNumUTCCPAlignedElems; ++ i) {
+                        auto smem_ptr = smem_sf_kv[kv_stage_idx] + i * kNumUTCCPAlignedElems;
+                        mma::sm100::replace_smem_desc_addr(sf_desc, smem_ptr);
+                        cute::SM100_UTCCP_4x32dp128bit_1cta::copy(sf_desc, kTmemStartColOfSFKV + i * 4);
+                    }
+
+                    #pragma unroll
+                    for (uint32_t i = 0; i < kNumMathWarpGroups; ++ i) {
+                        // Wait TMEM release
+                        CUTE_TIE_DECL(advance_tmem_pipeline(), tmem_stage_idx, tmem_phase);
+                        uint32_t tmem_addr = tmem_stage_idx * UMMA_N;
+
+                        empty_tmem_barriers[tmem_stage_idx]->wait(tmem_phase ^ 1);
+                        ptx::tcgen05_after_thread_sync();
+
+                        // Issue UMMA with SF
+                        #pragma unroll
+                        for (uint32_t k = 0; k < kHeadDim / UMMA_K; ++ k) {
+                            auto runtime_instr_desc = mma::sm100::make_runtime_instr_desc_with_sf_id(instr_desc, k * 2, k * 2);
+                            // TODO: generalize umma desc
+                            DG_STATIC_ASSERT(kHeadDim == 128, "Invalid head dim");
+                            auto a_desc = mma::sm100::make_smem_desc(
+                                cute::UMMA::LayoutType::SWIZZLE_64B,
+                                smem_kv[kv_stage_idx] + i * UMMA_M * (kHeadDim / 2) + k * UMMA_K / 2,
+                                8 * (kHeadDim / 2), 0);
+                            auto b_desc = mma::sm100::make_smem_desc(
+                                cute::UMMA::LayoutType::SWIZZLE_64B,
+                                smem_q[q_stage_idx] + k * UMMA_K / 2,
+                                8 * (kHeadDim / 2), 0);
+                            ptx::SM100_MMA_MXF4_SS::fma(
+                                a_desc, b_desc, tmem_addr, k, runtime_instr_desc,
+                                kTmemStartColOfSFKV + i * 4, kTmemStartColOfSFQ);
+                        }
+                        // TODO: move this into `deep_gemm/ptx/tcgen05.cuh`
+                        asm volatile("tcgen05.commit.cta_group::1.mbarrier::arrive::one.shared::cluster.b64 [%0];"
+                                     ::"r"(cute::cast_smem_ptr_to_uint(full_tmem_barriers[tmem_stage_idx])));
+                    }
+                }
+                cutlass::arch::umma_arrive(reinterpret_cast<uint64_t*>(empty_kv_barriers[kv_stage_idx]));
+            }
+
+            // UMMA warp must also arrive on empty_q to prevent running ahead
+            // of math warps in the Q pipeline. Without this, UMMA can consume
+            // kNumQStages Q blocks before math warps release any, causing a
+            // circular dependency: UMMA waits full_q -> TMA_Q waits empty_q
+            // -> Math waits full_tmem -> UMMA (already moved on).
+            empty_q_barriers[q_stage_idx]->arrive();
+        }
+    } else if (warp_idx == kSpecWarpStart + 3) {
+        cutlass::arch::warpgroup_reg_dealloc<kNumSpecializedRegisters>();
+    } else if (warp_idx < kSpecWarpStart) {
+        // Math warpgroups for reduce
+        cutlass::arch::warpgroup_reg_alloc<kNumMathRegisters>();
+
+        const auto math_warpgroup_idx = warpgroup_idx;
+        const auto math_thread_idx = threadIdx.x;
+
+        // Helper lambda for loading tensor memory
+        auto tmem_load = [](auto num_elems_c, const uint32_t& tmem_addr, float* accum) {
+            constexpr uint32_t N = decltype(num_elems_c)::value;
+            DG_STATIC_ASSERT(N == 32 or N == 64, "Unsupported TMEM load size");
+            using Loader = cute::conditional_t<N == 32,
+                cute::SM100_TMEM_LOAD_32dp32b32x,
+                cute::SM100_TMEM_LOAD_32dp32b64x>;
+            [&]<size_t... Is>(cute::index_sequence<Is...>) {
+                Loader::copy(tmem_addr, reinterpret_cast<uint32_t*>(accum)[Is]...);
+            }(cute::make_index_sequence<N>{});
+            cutlass::arch::fence_view_async_tmem_load();
+        };
+
+        // Math warpgroups process TMEM stages alternately
+        // Advance pipeline to align with the assigned stage
+        advance_tmem_pipeline(math_warpgroup_idx);
+
+        // Local register buffers
+        float accum[kNumHeads];
+        float weights[BLOCK_Q][kNumHeads];
+
+        // Enumerate Q blocks
+        for (uint32_t q_idx = sm_idx; q_idx < num_q_blocks; q_idx += kNumSMs) {
+            // Load KV block ranges
+            CUTE_TIE_DECL(load_schedule(q_idx), kv_start, num_kv_blocks);
+
+            // Wait TMA Q arrivals
+            CUTE_TIE_DECL(advance_q_pipeline(), q_stage_idx, q_phase);
+            full_q_barriers[q_stage_idx]->wait(q_phase);
+
+            // Read weights
+            // TODO: optimize bank conflicts
+            #pragma unroll
+            for (uint32_t i = 0; i < BLOCK_Q; ++ i) {
+                #pragma unroll
+                for (uint32_t j = 0; j < kNumHeads; ++ j)
+                    weights[i][j] = ptx::ld_shared(smem_weights[q_stage_idx] + i * kNumHeads + j);
+            }
+
+            // Enumerate KV blocks
+            for (uint32_t kv_idx = 0; kv_idx < num_kv_blocks; ++ kv_idx) {
+                // Calculate KV offset in advance
+                auto kv_offset = kv_start + kv_idx * BLOCK_KV + math_thread_idx;
+
+                // Advance pipeline by `kNumMathWarpGroups` steps
+                // Wait UMMA arrival
+                CUTE_TIE_DECL(advance_tmem_pipeline(kNumMathWarpGroups), tmem_stage_idx, tmem_phase);
+                full_tmem_barriers[tmem_stage_idx]->wait(tmem_phase);
+                ptx::tcgen05_after_thread_sync();
+
+                // Reduce over the head dim and store
+                #pragma unroll
+                for (uint32_t i = 0; i < BLOCK_Q; ++ i) {
+                    // Load accumulator from TMEM
+                    uint32_t tmem_addr = tmem_stage_idx * UMMA_N + i * kNumHeads;
+                    tmem_load(cute::Int<kNumHeads>{}, tmem_addr, accum);
+
+                    // Release TMEM empty
+                    if (i == BLOCK_Q - 1) {
+                        ptx::tcgen05_before_thread_sync();
+                        empty_tmem_barriers[tmem_stage_idx]->arrive();
+                    }
+
+                    // Accumulate weighted ReLU in parallel
+                    auto sum_0 = make_float2(0, 0);
+                    auto sum_1 = make_float2(0, 0);
+
+                    const auto transform = [&](const uint32_t& j, const float2& sum) {
+                        auto a = make_float2(fmaxf(accum[j], 0), fmaxf(accum[j + 1], 0));
+                        auto b = make_float2(weights[i][j], weights[i][j + 1]);
+                        return __ffma2_rn(a, b, sum);
+                    };
+
+                    #pragma unroll
+                    for (uint32_t j = 0; j < kNumHeads; j += 4) {
+                        sum_0 = transform(j, sum_0);
+                        sum_1 = transform(j + 2, sum_1);
+                    }
+
+                    auto sum = __fadd2_rn(sum_0, sum_1);
+                    auto result = static_cast<logits_dtype_t>(sum.x + sum.y);
+
+                    // Store into the global memory
+                    // NOTES: we have redundant writes here, consider more carefully
+                    // TODO: optimize performance
+                    const auto q_offset = (q_idx * BLOCK_Q + i) * static_cast<uint64_t>(logits_stride);
+                    if constexpr (kIsCompressedLogits) {
+                        if (seq_k_start[i] <= kv_offset and kv_offset < seq_k_end[i])
+                            logits[q_offset + kv_offset - seq_k_start[i]] = result;
+                    } else {
+                        logits[q_offset + kv_offset] = result;
+                    }
+                    __syncwarp();
+                }
+            }
+
+            // Release last Q empty
+            empty_q_barriers[q_stage_idx]->arrive();
+        }
+
+        // Free tensor memory
+        cutlass::arch::NamedBarrier(kNumMathThreads, 0).sync();
+        if (warp_idx == 0)
+            cute::TMEM::Allocator1Sm().free(0, kNumTmemCols);
+    }
+}
+
+} // namespace deep_gemm

deep_gemm/include/deep_gemm/impls/sm100_fp4_paged_mqa_logits.cuh

@@ -0,0 +1,496 @@
+#pragma once
+
+#include <cutlass/arch/barrier.h>
+#include <cutlass/arch/reg_reconfig.h>
+
+#include <cute/arch/cluster_sm90.hpp>
+#include <cute/arch/copy_sm90_desc.hpp>
+
+#include <deep_gemm/common/cute_tie.cuh>
+#include <deep_gemm/common/math.cuh>
+#include <deep_gemm/common/tma_copy.cuh>
+#include <deep_gemm/common/utils.cuh>
+#include <deep_gemm/mma/sm100.cuh>
+#include <deep_gemm/ptx/ld_st.cuh>
+#include <deep_gemm/ptx/tcgen05.cuh>
+#include <deep_gemm/ptx/utils.cuh>
+#include <deep_gemm/scheduler/paged_mqa_logits.cuh>
+
+namespace deep_gemm {
+
+template <uint32_t kNextN, uint32_t kNumHeads,
+          uint32_t kHeadDim, uint32_t BLOCK_KV,
+          bool kIsContextLens2D,
+          uint32_t kNumQStages, uint32_t kNumKVStages,
+          uint32_t SPLIT_KV,
+          uint32_t kNumSpecializedThreads, uint32_t kNumMathThreads,
+          typename logits_dtype_t,
+          uint32_t kNumMathWarpGroups = kNumMathThreads / 128>
+CUTLASS_GLOBAL __launch_bounds__(kNumSpecializedThreads + kNumMathThreads, 1)
+void sm100_fp4_paged_mqa_logits(const uint32_t batch_size,
+                                const uint32_t logits_stride, const uint32_t block_table_stride,
+                                const uint32_t* context_lens, logits_dtype_t* logits,
+                                const uint32_t* block_table, const uint32_t* schedule_meta,
+                                const __grid_constant__ cute::TmaDescriptor tensor_map_q,
+                                const __grid_constant__ cute::TmaDescriptor tensor_map_sf_q,
+                                const __grid_constant__ cute::TmaDescriptor tensor_map_kv,
+                                const __grid_constant__ cute::TmaDescriptor tensor_map_sf_kv,
+                                const __grid_constant__ cute::TmaDescriptor tensor_map_weights) {
+    using Barrier = cutlass::arch::ClusterTransactionBarrier;
+
+    // Utils
+    const auto sm_idx = blockIdx.x;
+    const auto warp_idx = cutlass::canonical_warp_idx_sync();
+    const auto warpgroup_idx = warp_idx / 4;
+    const auto lane_idx = ptx::get_lane_idx();
+    constexpr uint32_t kSpecWarpStart = kNumMathWarpGroups * 4;
+
+    // Prefetch TMA descriptors
+    if (warp_idx == kSpecWarpStart) {
+        cute::prefetch_tma_descriptor(&tensor_map_q);
+        cute::prefetch_tma_descriptor(&tensor_map_sf_q);
+        cute::prefetch_tma_descriptor(&tensor_map_weights);
+        cute::prefetch_tma_descriptor(&tensor_map_kv);
+        cute::prefetch_tma_descriptor(&tensor_map_sf_kv);
+    }
+
+    // Next-N atom configs
+    static constexpr uint32_t kNextNAtom = (kNextN % 2 == 0) ? 2 : 1;
+    static constexpr uint32_t kNumNextNAtoms = kNextN / kNextNAtom;
+    static constexpr bool kSingleAtom = (kNumNextNAtoms == 1);
+
+    // UMMA configs
+    static constexpr uint32_t kNumTmemStages = 3;
+    static constexpr uint32_t kNumUTCCPAlignedElems = 128;
+    static constexpr uint32_t UMMA_M = 128;
+    static constexpr uint32_t UMMA_N = kNextNAtom * kNumHeads;
+    static constexpr uint32_t UMMA_K = 64;
+    static constexpr uint32_t kNumSFQAtom  = math::constexpr_align(kNextNAtom * kNumHeads, kNumUTCCPAlignedElems);
+    static constexpr uint32_t kNumSFKV = math::constexpr_align(SPLIT_KV, kNumUTCCPAlignedElems);
+    static constexpr uint32_t kRealNumSFQAtom = kNextNAtom * kNumHeads;
+    DG_STATIC_ASSERT(kNumSpecializedThreads == 128 and kNumMathThreads % 128 == 0, "Invalid threads");
+    DG_STATIC_ASSERT(SPLIT_KV == kNumMathWarpGroups * UMMA_M and SPLIT_KV % kNumUTCCPAlignedElems == 0, "Invalid `SPLIT_KV`");
+
+    // Shared memory configs
+    static constexpr uint32_t kSwizzleAlignment = 8 * (kHeadDim / 2);
+    static constexpr uint32_t SMEM_Q_SIZE_PER_STAGE      = kNextNAtom * kNumHeads * (kHeadDim / 2);
+    static constexpr uint32_t SMEM_SF_Q_SIZE_PER_STAGE   = kNumSFQAtom * sizeof(int);
+    static constexpr uint32_t SMEM_KV_SIZE_PER_STAGE     = SPLIT_KV * (kHeadDim / 2);
+    static constexpr uint32_t SMEM_SF_KV_SIZE_PER_STAGE  = kNumSFKV * sizeof(int);
+    static constexpr uint32_t SMEM_WEIGHT_SIZE_PER_STAGE = kNextNAtom * kNumHeads * sizeof(float);
+
+    // Align to swizzling alignment bytes
+    extern __shared__ __align__(kSwizzleAlignment) uint8_t smem_buffer[];
+    DG_STATIC_ASSERT(SMEM_Q_SIZE_PER_STAGE  % kSwizzleAlignment == 0, "Unaligned TMA swizzling");
+    DG_STATIC_ASSERT(SMEM_KV_SIZE_PER_STAGE % kSwizzleAlignment == 0, "Unaligned TMA swizzling");
+
+    // Q and KV data on shared memory
+    auto smem_q = utils::PatternVisitor([&](const uint32_t& i) {
+        return smem_buffer + SMEM_Q_SIZE_PER_STAGE * i;
+    });
+    auto smem_kv = utils::PatternVisitor([&](const uint32_t& i) {
+        return smem_buffer + SMEM_Q_SIZE_PER_STAGE * kNumQStages + SMEM_KV_SIZE_PER_STAGE * i;
+    });
+    const auto smem_sf_ptr = smem_buffer + (SMEM_Q_SIZE_PER_STAGE * kNumQStages + SMEM_KV_SIZE_PER_STAGE * kNumKVStages);
+    auto smem_sf_q = utils::PatternVisitor([&](const uint32_t& i) {
+        return reinterpret_cast<uint32_t*>(smem_sf_ptr + SMEM_SF_Q_SIZE_PER_STAGE * i);
+    });
+    auto smem_sf_kv = utils::PatternVisitor([&](const uint32_t& i) {
+        return reinterpret_cast<uint32_t*>(smem_sf_ptr + SMEM_SF_Q_SIZE_PER_STAGE * kNumQStages + SMEM_SF_KV_SIZE_PER_STAGE * i);
+    });
+    auto smem_weights = utils::PatternVisitor([&](const uint32_t& i) {
+        return reinterpret_cast<float*>(smem_sf_ptr + SMEM_SF_Q_SIZE_PER_STAGE * kNumQStages + SMEM_SF_KV_SIZE_PER_STAGE * kNumKVStages
+                                                    + SMEM_WEIGHT_SIZE_PER_STAGE * i);
+    });
+
+    // Barriers and TMEM pointer on shared memory
+    const auto barrier_ptr = reinterpret_cast<Barrier*>(smem_weights[kNumQStages]);
+    auto full_q_barriers     = utils::PatternVisitor([&](const uint32_t& i) { return barrier_ptr + i; });
+    auto empty_q_barriers    = utils::PatternVisitor([&](const uint32_t& i) { return barrier_ptr + kNumQStages + i; });
+    auto full_kv_barriers    = utils::PatternVisitor([&](const uint32_t& i) { return barrier_ptr + kNumQStages * 2 + i; });
+    auto empty_kv_barriers   = utils::PatternVisitor([&](const uint32_t& i) { return barrier_ptr + kNumQStages * 2 + kNumKVStages + i; });
+    const auto tmem_barrier_ptr = barrier_ptr + kNumQStages * 2 + kNumKVStages * 2;
+    auto full_tmem_barriers  = utils::PatternVisitor([&](const uint32_t& i) { return tmem_barrier_ptr + i; });
+    auto empty_tmem_barriers = utils::PatternVisitor([&](const uint32_t& i) { return tmem_barrier_ptr + kNumTmemStages + i; });
+    auto tmem_ptr_in_smem    = reinterpret_cast<uint32_t*>(tmem_barrier_ptr + kNumTmemStages * 2);
+
+    // Tensor memory configs
+    constexpr uint32_t kNumAccumTmemCols = kNextNAtom * kNumHeads * kNumTmemStages;
+    constexpr uint32_t kNumTmemCols = utils::get_num_aligned_tmem_cols<kNumAccumTmemCols + kNumSFQAtom / 32 + kNumSFKV / 32>();
+    constexpr uint32_t kTmemStartColOfSFQ = kNumAccumTmemCols;
+    constexpr uint32_t kTmemStartColOfSFKV = kNumAccumTmemCols + kNumSFQAtom / 32;
+    DG_STATIC_ASSERT(kNumTmemCols <= 512, "Too many tensor memory");
+
+    // Initialize barriers
+    if (warp_idx == kSpecWarpStart and cute::elect_one_sync()) {
+        #pragma unroll
+        for (uint32_t i = 0; i < kNumQStages; ++ i) {
+            full_q_barriers[i]->init(1);
+            empty_q_barriers[i]->init(kNumMathThreads + 32);
+        }
+        cutlass::arch::fence_barrier_init();
+    }
+    if (warp_idx == kSpecWarpStart + 1 and cute::elect_one_sync()) {
+        #pragma unroll
+        for (uint32_t i = 0; i < kNumKVStages; ++ i) {
+            full_kv_barriers[i]->init(1);
+            empty_kv_barriers[i]->init(1);
+        }
+        cutlass::arch::fence_barrier_init();
+    }
+    if (warp_idx == kSpecWarpStart + 2) {
+        if (cute::elect_one_sync()) {
+            #pragma unroll
+            for (uint32_t i = 0; i < kNumTmemStages; ++i) {
+                full_tmem_barriers[i]->init(1);
+                empty_tmem_barriers[i]->init(128);
+            }
+            cutlass::arch::fence_barrier_init();
+        }
+        // Allocate tensor memory
+        cute::TMEM::Allocator1Sm().allocate(kNumTmemCols, tmem_ptr_in_smem);
+    }
+    __syncthreads();
+
+    // Wait for primary kernel completion
+    cudaGridDependencySynchronize();
+
+    // Scheduler
+    constexpr uint32_t kNumBlocksPerSplit = SPLIT_KV / BLOCK_KV;
+    using Scheduler = sched::PagedMQALogitsScheduler<kNextN, kIsContextLens2D, BLOCK_KV, kNumBlocksPerSplit, kNumNextNAtoms>;
+    DG_STATIC_ASSERT(SPLIT_KV == BLOCK_KV * kNumBlocksPerSplit, "Invalid `SPLIT_KV`");
+
+    // Make Q, KV and TMEM pipeline
+    auto make_pipeline = [](const uint32_t& num_stages) {
+        // Return current stage and phase, and advance pipeline by steps
+        return [iter_idx = 0u, num_stages](const uint32_t& step = 1) mutable -> cute::tuple<uint32_t, uint32_t> {
+            uint32_t current_idx = iter_idx;
+            iter_idx += step;
+            return {current_idx % num_stages, (current_idx / num_stages) & 1};
+        };
+    };
+    auto advance_q_pipeline    = make_pipeline(kNumQStages);
+    auto advance_kv_pipeline   = make_pipeline(kNumKVStages);
+    auto advance_tmem_pipeline = make_pipeline(kNumTmemStages);
+
+    // Register reconfigurations
+    constexpr uint32_t kNumSpecializedRegisters = 56;
+    constexpr uint32_t kNumMathRegisters = 224;
+
+    if (warp_idx == kSpecWarpStart) {
+        // TMA warp for loading Q
+        cutlass::arch::warpgroup_reg_dealloc<kNumSpecializedRegisters>();
+
+        if (cute::elect_one_sync()) {
+            auto scheduler = Scheduler(sm_idx, context_lens, schedule_meta);
+
+            // Persistently schedule over blocks
+            // Initialize outside valid range to indicate no previous task
+            uint32_t last_q_atom_idx = batch_size * kNumNextNAtoms;
+            uint32_t q_atom_idx, _, __;
+            while (scheduler.fetch_next_task(q_atom_idx, _, __)) {
+                // Issue TMA Q when (q_idx, atom_idx) changes
+                if (q_atom_idx != last_q_atom_idx) {
+                    // Wait Q consumer release
+                    CUTE_TIE_DECL(advance_q_pipeline(), q_stage_idx, q_phase);
+                    empty_q_barriers[q_stage_idx]->wait(q_phase ^ 1);
+
+                    // Issue TMA Q
+                    cute::SM90_TMA_LOAD_2D::copy(&tensor_map_q, reinterpret_cast<uint64_t*>(full_q_barriers[q_stage_idx]),
+                                                 static_cast<uint64_t>(cute::TMA::CacheHintSm100::EVICT_NORMAL),
+                                                 smem_q[q_stage_idx], 0, q_atom_idx * kNextNAtom * kNumHeads);
+                    tma::copy<kNextNAtom * kNumHeads, 1, 0>(&tensor_map_sf_q, full_q_barriers[q_stage_idx], smem_sf_q[q_stage_idx], 0, q_atom_idx * kNextNAtom);
+                    tma::copy<kNumHeads, kNextNAtom, 0>(&tensor_map_weights, full_q_barriers[q_stage_idx], smem_weights[q_stage_idx], 0, q_atom_idx * kNextNAtom);
+                    full_q_barriers[q_stage_idx]->arrive_and_expect_tx(SMEM_Q_SIZE_PER_STAGE + kRealNumSFQAtom * sizeof(int) + SMEM_WEIGHT_SIZE_PER_STAGE);
+                }
+                last_q_atom_idx = q_atom_idx;
+            }
+        }
+        __syncwarp();
+    } else if (warp_idx == kSpecWarpStart + 1) {
+        // TMA warp for loading KV cache
+        cutlass::arch::warpgroup_reg_dealloc<kNumSpecializedRegisters>();
+        auto scheduler = Scheduler(sm_idx, context_lens, schedule_meta);
+
+        // Persistently schedule over blocks
+        uint32_t kv_block_idx_ptr = 32, kv_block_idx_storage;
+        uint32_t last_q_atom_idx = batch_size * kNumNextNAtoms;
+        uint32_t q_atom_idx, kv_idx, num_kv;
+        while (scheduler.fetch_next_task(q_atom_idx, kv_idx, num_kv)) {
+            // Reset block table cache on kv restart
+            if (q_atom_idx != last_q_atom_idx)
+                kv_block_idx_ptr = 32;
+            last_q_atom_idx = q_atom_idx;
+
+            // Coalesced load of block table
+            if (kv_block_idx_ptr == 32) {
+                kv_block_idx_ptr = 0;
+                const auto block_table_offset = (q_atom_idx / kNumNextNAtoms) * static_cast<uint64_t>(block_table_stride);
+                kv_block_idx_storage = (kv_idx + lane_idx < num_kv)
+                    ? block_table[block_table_offset + kv_idx + lane_idx] : 0;
+            }
+
+            // Broadcast KV block indices 
+            int kv_block_idx[kNumBlocksPerSplit];
+            #pragma unroll
+            for (int i = 0; i < kNumBlocksPerSplit; ++ i)
+                kv_block_idx[i] = __shfl_sync(0xffffffff, kv_block_idx_storage, kv_block_idx_ptr + i);
+            kv_block_idx_ptr += kNumBlocksPerSplit;
+            DG_STATIC_ASSERT(32 % kNumBlocksPerSplit == 0, "Invalid `SPLIT_KV`");
+
+            // Wait KV consumer release
+            CUTE_TIE_DECL(advance_kv_pipeline(), kv_stage_idx, kv_phase);
+            
+            // Issue TMA KV
+            if (cute::elect_one_sync()) {
+                empty_kv_barriers[kv_stage_idx]->wait(kv_phase ^ 1);
+                #pragma unroll
+                for (int i = 0; i < kNumBlocksPerSplit; ++ i) {
+                    cute::SM90_TMA_LOAD_3D::copy(&tensor_map_kv, reinterpret_cast<uint64_t*>(full_kv_barriers[kv_stage_idx]),
+                                                 static_cast<uint64_t>(cute::TMA::CacheHintSm100::EVICT_NORMAL),
+                                                 smem_kv[kv_stage_idx] + (BLOCK_KV * kHeadDim / 2) * i,
+                                                 0, 0, kv_block_idx[i]);
+                    tma::copy<BLOCK_KV, 1, 0>(&tensor_map_sf_kv, full_kv_barriers[kv_stage_idx],
+                                              smem_sf_kv[kv_stage_idx] + BLOCK_KV * i,
+                                              0, kv_block_idx[i]);
+                }
+                full_kv_barriers[kv_stage_idx]->arrive_and_expect_tx(SMEM_KV_SIZE_PER_STAGE + SMEM_SF_KV_SIZE_PER_STAGE);
+            }
+        }
+    } else if (warp_idx == kSpecWarpStart + 2) {
+        // UMMA warp
+        cutlass::arch::warpgroup_reg_dealloc<kNumSpecializedRegisters>();
+        auto scheduler = Scheduler(sm_idx, context_lens, schedule_meta);
+        DG_TRAP_ONLY_DEVICE_ASSERT(ptx::ld_shared(tmem_ptr_in_smem) == 0);
+
+        // UTCCP transposer
+        auto utccp_required_smem_warp_transpose = [&](const uint32_t* smem_ptr) {
+            DG_STATIC_ASSERT(kNumUTCCPAlignedElems == 128, "Invalid aligned elements");
+            uint32_t values[4];
+            #pragma unroll
+            for (uint32_t i = 0; i < 4; ++ i)
+                values[i] = ptx::ld_shared(smem_ptr + (i ^ (lane_idx >> 3)) * 32 + lane_idx);
+            __syncwarp();
+            #pragma unroll
+            for (uint32_t i = 0; i < 4; ++ i)
+                ptx::st_shared(smem_ptr + lane_idx * 4 + (i ^ (lane_idx >> 3)), values[i]);
+        };
+
+        // Make UMMA desc
+        auto instr_desc = cute::UMMA::make_instr_desc_block_scaled<cutlass::float_e2m1_t, cutlass::float_e2m1_t, float, cutlass::float_ue8m0_t,
+                                                                   UMMA_M, UMMA_N, cute::UMMA::Major::K, cute::UMMA::Major::K>();
+        auto sf_desc = mma::sm100::make_sf_desc(nullptr);
+
+        // Persistently schedule over blocks
+        uint32_t last_q_atom_idx = batch_size * kNumNextNAtoms;
+        uint32_t q_atom_idx, kv_idx, _;
+        while (scheduler.fetch_next_task(q_atom_idx, kv_idx, _)) {
+            // Wait TMA Q arrivals
+            uint32_t q_stage_idx, q_phase;
+            if (q_atom_idx != last_q_atom_idx) {
+                CUTE_TIE(advance_q_pipeline(), q_stage_idx, q_phase);
+
+                // Release previous Q empty (UMMA warp must participate to prevent
+                // running ahead of math warps in the Q pipeline)
+                if (last_q_atom_idx != batch_size * kNumNextNAtoms)
+                    empty_q_barriers[(q_stage_idx + kNumQStages - 1) % kNumQStages]->arrive();
+
+                full_q_barriers[q_stage_idx]->wait(q_phase);
+
+                // Transpose and copy SF Q
+                #pragma unroll
+                for (uint32_t i = 0; i < kNumSFQAtom / kNumUTCCPAlignedElems; ++ i) {
+                    auto smem_ptr = smem_sf_q[q_stage_idx] + i * kNumUTCCPAlignedElems;
+                    utccp_required_smem_warp_transpose(smem_ptr);
+                    cutlass::arch::fence_view_async_shared();
+                    mma::sm100::replace_smem_desc_addr(sf_desc, smem_ptr);
+                    if (cute::elect_one_sync())
+                        cute::SM100_UTCCP_4x32dp128bit_1cta::copy(sf_desc, kTmemStartColOfSFQ + i * 4);
+                    __syncwarp();
+                }
+            }
+            last_q_atom_idx = q_atom_idx;
+
+            // Wait TMA KV arrivals
+            CUTE_TIE_DECL(advance_kv_pipeline(), kv_stage_idx, kv_phase);
+            full_kv_barriers[kv_stage_idx]->wait(kv_phase);
+
+            // Transpose
+            #pragma unroll
+            for (uint32_t i = 0; i < kNumSFKV / kNumUTCCPAlignedElems; ++ i) {
+                auto smem_ptr = smem_sf_kv[kv_stage_idx] + i * kNumUTCCPAlignedElems;
+                utccp_required_smem_warp_transpose(smem_ptr);
+                cutlass::arch::fence_view_async_shared();
+            }
+
+            // UMMA with SF
+            if (cute::elect_one_sync()) {
+                // Copy SF KV
+                #pragma unroll
+                for (uint32_t i = 0; i < kNumSFKV / kNumUTCCPAlignedElems; ++ i) {
+                    auto smem_ptr = smem_sf_kv[kv_stage_idx] + i * kNumUTCCPAlignedElems;
+                    mma::sm100::replace_smem_desc_addr(sf_desc, smem_ptr);
+                    cute::SM100_UTCCP_4x32dp128bit_1cta::copy(sf_desc, kTmemStartColOfSFKV + i * 4);
+                }
+
+                #pragma unroll
+                for (uint32_t i = 0; i < kNumMathWarpGroups; ++ i) {
+                    // Wait TMEM release
+                    CUTE_TIE_DECL(advance_tmem_pipeline(), tmem_stage_idx, tmem_phase);
+                    uint32_t tmem_addr = tmem_stage_idx * UMMA_N;
+
+                    empty_tmem_barriers[tmem_stage_idx]->wait(tmem_phase ^ 1);
+                    ptx::tcgen05_after_thread_sync();
+
+                    // Issue UMMA with SF
+                    #pragma unroll
+                    for (uint32_t k = 0; k < kHeadDim / UMMA_K; ++ k) {
+                        auto runtime_instr_desc = mma::sm100::make_runtime_instr_desc_with_sf_id(instr_desc, k * 2, k * 2);
+                        // TODO: generalize UMMA desc
+                        DG_STATIC_ASSERT(kHeadDim == 128, "Invalid head dim");
+                        auto a_desc = mma::sm100::make_smem_desc(
+                            cute::UMMA::LayoutType::SWIZZLE_64B,
+                            smem_kv[kv_stage_idx] + i * UMMA_M * (kHeadDim / 2) + k * UMMA_K / 2,
+                            8 * (kHeadDim / 2), 0);
+                        auto b_desc = mma::sm100::make_smem_desc(
+                            cute::UMMA::LayoutType::SWIZZLE_64B,
+                            smem_q[q_stage_idx] + k * UMMA_K / 2,
+                            8 * (kHeadDim / 2), 0);
+                        ptx::SM100_MMA_MXF4_SS::fma(a_desc, b_desc, tmem_addr, k, runtime_instr_desc,
+                                                    kTmemStartColOfSFKV + i * 4, kTmemStartColOfSFQ);
+                    }
+                    // TODO: move this PTX into headers
+                    asm volatile("tcgen05.commit.cta_group::1.mbarrier::arrive::one.shared::cluster.b64 [%0];"
+                                 ::"r"(cute::cast_smem_ptr_to_uint(full_tmem_barriers[tmem_stage_idx])));
+                }
+            }
+            cutlass::arch::umma_arrive(reinterpret_cast<uint64_t*>(empty_kv_barriers[kv_stage_idx]));
+        }
+    } else if (warp_idx == kSpecWarpStart + 3) {
+        cutlass::arch::warpgroup_reg_dealloc<kNumSpecializedRegisters>();
+    } else if (warp_idx < kSpecWarpStart) {
+        // Math warpgroups for reduce
+        cutlass::arch::warpgroup_reg_alloc<kNumMathRegisters>();
+        auto scheduler = Scheduler(sm_idx, context_lens, schedule_meta);
+
+        const auto math_warpgroup_idx = warpgroup_idx;
+        const auto math_thread_idx = warp_idx * 32 + lane_idx;
+
+        // Helper lambda for loading tensor memory
+        auto tmem_load = [](auto num_elems_c, const uint32_t& tmem_addr, float* accum) {
+            constexpr int N = decltype(num_elems_c)::value;
+            DG_STATIC_ASSERT(N == 32 or N == 64, "Unsupported TMEM load size");
+            using Loader = cute::conditional_t<N == 32,
+                cute::SM100_TMEM_LOAD_32dp32b32x,
+                cute::SM100_TMEM_LOAD_32dp32b64x>;
+            [&]<size_t... Is>(cute::index_sequence<Is...>) {
+                Loader::copy(tmem_addr, reinterpret_cast<uint32_t*>(accum)[Is]...);
+            }(cute::make_index_sequence<N>{});
+            cutlass::arch::fence_view_async_tmem_load();
+        };
+
+        // Math warpgroups process TMEM stages alternately
+        // Advance pipeline to align with the assigned stage
+        advance_tmem_pipeline(math_warpgroup_idx);
+
+        // Local register buffers
+        float accum[kNumHeads];
+        float weights[kNextNAtom][kNumHeads];
+
+        // Persistently schedule over blocks
+        uint32_t last_q_atom_idx = batch_size * kNumNextNAtoms;
+        uint32_t q_atom_idx, kv_idx, _;
+        while (scheduler.fetch_next_task(q_atom_idx, kv_idx, _)) {
+            if (q_atom_idx != last_q_atom_idx) {
+                CUTE_TIE_DECL(advance_q_pipeline(), q_stage_idx, q_phase);
+
+                // Release last Q empty
+                if (last_q_atom_idx != batch_size * kNumNextNAtoms)
+                    empty_q_barriers[(q_stage_idx + kNumQStages - 1) % kNumQStages]->arrive();
+
+                // Wait TMA Q arrivals
+                full_q_barriers[q_stage_idx]->wait(q_phase);
+
+                // Read weights
+                #pragma unroll
+                for (uint32_t i = 0; i < kNextNAtom; ++ i) {
+                    #pragma unroll
+                    for (uint32_t j = 0; j < kNumHeads; j += 4) {
+                        float4 raw = ptx::ld_shared((float4*)(smem_weights[q_stage_idx] + i * kNumHeads + j));
+                        weights[i][j + 0] = raw.x;
+                        weights[i][j + 1] = raw.y;
+                        weights[i][j + 2] = raw.z;
+                        weights[i][j + 3] = raw.w;
+                    }
+                }
+            }
+            last_q_atom_idx = q_atom_idx;
+
+            // Calculate KV offset in advance
+            auto kv_offset = q_atom_idx * kNextNAtom * static_cast<uint64_t>(logits_stride) + kv_idx * BLOCK_KV + math_thread_idx;
+
+            // Advance pipeline by `kNumMathWarpGroups` steps
+            // Wait UMMA arrival
+            CUTE_TIE_DECL(advance_tmem_pipeline(kNumMathWarpGroups), tmem_stage_idx, tmem_phase);
+            full_tmem_barriers[tmem_stage_idx]->wait(tmem_phase);
+            ptx::tcgen05_after_thread_sync();
+
+            // Reduce over the head dim and store
+            #pragma unroll
+            for (uint32_t i = 0; i < kNextNAtom; ++ i) {
+                // Load accumulator from TMEM
+                uint32_t tmem_addr = tmem_stage_idx * UMMA_N + i * kNumHeads;
+                tmem_load(cute::Int<kNumHeads>{}, tmem_addr, accum);
+
+                // Release TMEM empty
+                if (i == kNextNAtom - 1) {
+                    ptx::tcgen05_before_thread_sync();
+                    empty_tmem_barriers[tmem_stage_idx]->arrive();
+                }
+
+                // Accumulate weighted ReLU in parallel
+                auto sum_0 = make_float2(0, 0);
+                auto sum_1 = make_float2(0, 0);
+
+                const auto transform = [&](const uint32_t& j, const float2& sum) {
+                    auto a = make_float2(fmaxf(accum[j], 0), fmaxf(accum[j + 1], 0));
+                    auto b = make_float2(weights[i][j], weights[i][j + 1]);
+                    return __ffma2_rn(a, b, sum);
+                };
+
+                #pragma unroll
+                for (uint32_t j = 0; j < kNumHeads; j += 4) {
+                    sum_0 = transform(j, sum_0);
+                    sum_1 = transform(j + 2, sum_1);
+                }
+
+                auto sum = __fadd2_rn(sum_0, sum_1);
+                auto result = static_cast<logits_dtype_t>(sum.x + sum.y);
+
+                // Store into the global memory
+                const auto dst_offset = kv_offset + i * static_cast<uint64_t>(logits_stride);
+                if constexpr(sizeof(logits_dtype_t) == 2) {
+                    // Pack two adjacent bf16 lanes into uint32 for wider store
+                    uint16_t my_bits = *reinterpret_cast<const uint16_t*>(&result);
+                    uint16_t neighbor_bits = __shfl_down_sync(0xffffffff, my_bits, 1);
+                    uint32_t packed;
+                    asm volatile("mov.b32 %0, {%1, %2};" : "=r"(packed) : "h"(my_bits), "h"(neighbor_bits));
+                    if (lane_idx % 2 == 0)
+                        *reinterpret_cast<uint32_t*>(logits + dst_offset) = packed;
+                } else {
+                    logits[dst_offset] = result;
+                }
+                // this sync warp prevent the next load tmem from reordering
+                // nvcc may reorder it to overlap with the current tmem load, lead to large register usage
+                __syncwarp();
+            }
+        }
+
+        // Free tensor memory
+        cutlass::arch::NamedBarrier(kNumMathThreads, 0).sync();
+        if (warp_idx == 0)
+            cute::TMEM::Allocator1Sm().free(0, kNumTmemCols);
+    }
+}
+
+} // namespace deep_gemm

deep_gemm/include/deep_gemm/impls/sm100_fp8_fp4_gemm_1d1d.cuh

@@ -0,0 +1,514 @@
+#pragma once
+#pragma clang diagnostic push
+#pragma clang diagnostic ignored "-Wunknown-attributes"
+
+#include <cutlass/arch/barrier.h>
+
+#include <deep_gemm/common/math.cuh>
+#include <deep_gemm/common/tma_copy.cuh>
+#include <deep_gemm/epilogue/transform.cuh>
+#include <deep_gemm/epilogue/sm100_store_cd.cuh>
+#include <deep_gemm/epilogue/sm100_store_cd_swap_ab.cuh>
+#include <deep_gemm/mma/sm100.cuh>
+#include <deep_gemm/scheduler/gemm.cuh>
+#include <deep_gemm/ptx/utils.cuh>
+
+namespace deep_gemm {
+
+template <cute::UMMA::Major kMajorA, cute::UMMA::Major kMajorB,
+          uint32_t kGranKA, uint32_t kGranKB,
+          uint32_t SHAPE_M, uint32_t SHAPE_N, uint32_t SHAPE_K,
+          uint32_t BLOCK_M, uint32_t BLOCK_N, uint32_t BLOCK_K,
+          uint32_t kNumGroups,
+          uint32_t kSwizzleAMode, uint32_t kSwizzleBMode, uint32_t kSwizzleCDMode,
+          uint32_t kNumStages,
+          uint32_t kNumNonEpilogueThreads, uint32_t kNumEpilogueThreads,
+          uint32_t kNumMulticast, bool kIsMulticastOnA,
+          uint32_t kNumSMs,
+          bool kSwapAB,
+          GemmType kGemmType, bool kWithAccumulation,
+          typename a_dtype_t, typename b_dtype_t, typename cd_dtype_t,
+          typename epilogue_type_t>
+CUTLASS_GLOBAL void __launch_bounds__(kNumNonEpilogueThreads + kNumEpilogueThreads, 1)
+sm100_fp8_fp4_gemm_1d1d_impl(int* grouped_layout,
+                             uint32_t shape_m, uint32_t shape_n, uint32_t shape_k,
+                             const __grid_constant__ cute::TmaDescriptor tensor_map_a,
+                             const __grid_constant__ cute::TmaDescriptor tensor_map_b,
+                             const __grid_constant__ cute::TmaDescriptor tensor_map_sfa,
+                             const __grid_constant__ cute::TmaDescriptor tensor_map_sfb,
+                             const __grid_constant__ cute::TmaDescriptor tensor_map_cd) {
+#if (defined(__CUDA_ARCH__) and (__CUDA_ARCH__ >= 1000)) or defined(__CLION_IDE__)
+    using Barrier = cutlass::arch::ClusterTransactionBarrier;
+    using Allocator = cute::conditional_t<kNumMulticast == 1, cute::TMEM::Allocator1Sm, cute::TMEM::Allocator2Sm>;
+
+    // GEMM with accumulation must have FP32 output
+    if constexpr (kWithAccumulation)
+        DG_STATIC_ASSERT(cute::is_same_v<cd_dtype_t, float>, "Invalid C/D data dtype");
+
+    // MMA Configs
+    constexpr uint32_t LAYOUT_AD_M = 128;
+    constexpr uint32_t UMMA_M = LAYOUT_AD_M * kNumMulticast;
+    constexpr uint32_t UMMA_N = kSwapAB ? BLOCK_M : BLOCK_N;
+    constexpr uint32_t UMMA_K = 32;
+    constexpr uint32_t LOAD_BLOCK_M = BLOCK_M / (kIsMulticastOnA ? kNumMulticast: 1);
+    constexpr uint32_t LOAD_BLOCK_N = BLOCK_N / (kIsMulticastOnA ? 1 : kNumMulticast);
+    DG_STATIC_ASSERT(BLOCK_K == 128, "Invalid block K");
+    DG_STATIC_ASSERT(kNumMulticast == 1 or kNumMulticast == 2, "Only support 1/2 multicast");
+    DG_STATIC_ASSERT((kSwapAB and BLOCK_N == LAYOUT_AD_M) or
+                     (not kSwapAB and (BLOCK_M == 32 or BLOCK_M == 64 or BLOCK_M == LAYOUT_AD_M)), "Invalid block size");
+
+    // SF configs
+    constexpr uint32_t kNumUTCCPAlignedElems = 128;
+    constexpr uint32_t SF_BLOCK_M = math::constexpr_align(BLOCK_M, kNumUTCCPAlignedElems);
+    constexpr uint32_t SF_BLOCK_N = math::constexpr_align(BLOCK_N, kNumUTCCPAlignedElems);
+    constexpr uint32_t kNumSFAStagesPerLoad = kGranKA == 32 ? 1 : 4;
+    constexpr uint32_t kNumSFBStagesPerLoad = kGranKB == 32 ? 1 : 4;
+    DG_STATIC_ASSERT(kGranKA == 32 or kGranKA == 128, "Invalid granularity K for A");
+    DG_STATIC_ASSERT(kGranKB == 32 or kGranKB == 128, "Invalid granularity K for B");
+    DG_STATIC_ASSERT((kGemmType != GemmType::KGroupedContiguous) or kGranKA == kGranKB, "K-grouped SF requires kGranKA == kGranKB");
+
+    // Epilogue configs
+    // Always enable pipeline for better performance
+    constexpr uint32_t kNumEpilogueStages = 2;
+    constexpr uint32_t kNumTMAStoreStages = 2;
+    // NOTES: To maximize epilogue threads utilization, process an entire BLOCK_N
+    //        per store stage for swap-AB cases, and an entire BLOCK_M for non-swap cases
+    constexpr uint32_t STORE_BLOCK_M =        kSwapAB ? 16      : cute::min<uint32_t>(BLOCK_M, LAYOUT_AD_M);
+    constexpr uint32_t STORE_BLOCK_N =        kSwapAB ? BLOCK_N : kSwizzleCDMode / sizeof(cd_dtype_t);
+    constexpr uint32_t kNumUMMAStoreThreads = kSwapAB ? kNumEpilogueThreads: STORE_BLOCK_M;
+    DG_STATIC_ASSERT(kNumUMMAStoreThreads % 32 == 0, "Invalid store block M");
+
+    // Share memory sizes
+    constexpr uint32_t SMEM_CD_SIZE_PER_STAGE = STORE_BLOCK_M * STORE_BLOCK_N * sizeof(cd_dtype_t);
+    constexpr uint32_t SMEM_CD_SIZE = SMEM_CD_SIZE_PER_STAGE * kNumTMAStoreStages;
+    constexpr uint32_t SMEM_A_SIZE_PER_STAGE = LOAD_BLOCK_M * BLOCK_K * sizeof(a_dtype_t);
+    constexpr uint32_t SMEM_B_SIZE_PER_STAGE = LOAD_BLOCK_N * BLOCK_K * sizeof(b_dtype_t);
+    constexpr uint32_t SMEM_SFA_SIZE_PER_STAGE = SF_BLOCK_M * sizeof(uint32_t);
+    constexpr uint32_t SMEM_SFB_SIZE_PER_STAGE = SF_BLOCK_N * sizeof(uint32_t);
+    DG_STATIC_ASSERT(SMEM_CD_SIZE % 1024 == 0 and SMEM_A_SIZE_PER_STAGE % 1024 == 0 and SMEM_B_SIZE_PER_STAGE % 1024 == 0, 
+                     "Shared memory of A/B must be aligned to 1024 bytes");
+    // NOTES: Make sure we have enough shared memory for UMMA padding
+    constexpr uint32_t UMMA_A_SIZE_PER_STAGE = math::constexpr_align(LOAD_BLOCK_M, LAYOUT_AD_M) * BLOCK_K * sizeof(a_dtype_t);
+    DG_STATIC_ASSERT(UMMA_A_SIZE_PER_STAGE <= SMEM_A_SIZE_PER_STAGE + SMEM_B_SIZE_PER_STAGE * kNumStages, "Memory Out of bound for UMMA");
+
+    // Tensor memory size and offsets
+    constexpr uint32_t kNumAccumTmemCols = UMMA_N * kNumEpilogueStages;
+    constexpr uint32_t kNumSFATmemCols = SF_BLOCK_M / 32;
+    constexpr uint32_t kNumSFBTmemCols = SF_BLOCK_N / 32;
+    constexpr uint32_t kNumTmemCols = utils::get_num_aligned_tmem_cols<kNumAccumTmemCols + kNumSFATmemCols + kNumSFBTmemCols>();
+    constexpr uint32_t kTmemStartColOfSFA = kNumAccumTmemCols;
+    constexpr uint32_t kTmemStartColOfSFB = kNumAccumTmemCols + kNumSFATmemCols;
+    DG_STATIC_ASSERT(32 <= kNumTmemCols and kNumTmemCols <= 512, "Invalid tensor memory columns");
+
+    // Synchronize the cluster before 2-CTA TMEM allocation
+    kNumMulticast > 1 ? cute::cluster_sync() : void();
+
+    // Utils
+    const bool is_leader_cta = cute::block_rank_in_cluster() == 0;
+    const auto warp_idx = cutlass::canonical_warp_idx_sync();
+    const auto lane_idx = ptx::get_lane_idx();
+
+    // Prefetch TMA descriptors at the very beginning
+    if (warp_idx == 0) {
+        cute::prefetch_tma_descriptor(&tensor_map_a);
+        cute::prefetch_tma_descriptor(&tensor_map_b);
+        cute::prefetch_tma_descriptor(&tensor_map_sfa);
+        cute::prefetch_tma_descriptor(&tensor_map_sfb);
+        cute::prefetch_tma_descriptor(&tensor_map_cd);
+    }
+
+    // Overwrite shape constants if the compiler gives
+    shape_m = SHAPE_M != 0 ? SHAPE_M : shape_m;
+    shape_n = SHAPE_N != 0 ? SHAPE_N : shape_n;
+    shape_k = SHAPE_K != 0 ? SHAPE_K : shape_k;
+    const auto shape_sfa_k = math::ceil_div(shape_k, kGranKA * 4);
+    const auto shape_sfb_k = math::ceil_div(shape_k, kGranKB * 4);
+
+    // Align to 1024 bytes for swizzle-128B
+    extern __shared__ __align__(1024) uint8_t smem_buffer[];
+
+    // D/A/B shared memory
+    auto smem_cd = utils::PatternVisitor([&](const uint32_t& i) {
+        return reinterpret_cast<cd_dtype_t*>(smem_buffer + i * SMEM_CD_SIZE_PER_STAGE); 
+    });
+    auto smem_a  = utils::PatternVisitor([&](const uint32_t& i) {
+        return reinterpret_cast<a_dtype_t*>(smem_buffer + SMEM_CD_SIZE + i * SMEM_A_SIZE_PER_STAGE);
+    });
+    auto smem_b  = utils::PatternVisitor([&](const uint32_t& i) {
+        return reinterpret_cast<b_dtype_t*>(smem_buffer + SMEM_CD_SIZE + kNumStages * SMEM_A_SIZE_PER_STAGE + i * SMEM_B_SIZE_PER_STAGE);
+    });
+
+    // SFA/SFB shared memory
+    auto sf_start_ptr = reinterpret_cast<uint8_t*>(smem_b[kNumStages]);
+    auto smem_sfa = utils::PatternVisitor([=](const uint32_t& i) {
+        return reinterpret_cast<uint32_t*>(sf_start_ptr + i * SMEM_SFA_SIZE_PER_STAGE);
+    });
+    auto smem_sfb = utils::PatternVisitor([=](const uint32_t& i) {
+        return reinterpret_cast<uint32_t*>(sf_start_ptr + kNumStages * SMEM_SFA_SIZE_PER_STAGE + i * SMEM_SFB_SIZE_PER_STAGE);
+    });
+
+    // Barriers and tensor memory pointer
+    auto barrier_start_ptr = reinterpret_cast<Barrier*>(smem_sfb[kNumStages]);;
+    auto full_barriers          = utils::PatternVisitor([=](const uint32_t& i) { return barrier_start_ptr + (i); });
+    auto empty_barriers         = utils::PatternVisitor([=](const uint32_t& i) { return barrier_start_ptr + (kNumStages + i); });
+    auto with_sf_full_barriers  = utils::PatternVisitor([=](const uint32_t& i) { return barrier_start_ptr + (kNumStages * 2 + i); });
+    auto tmem_full_barriers     = utils::PatternVisitor([=](const uint32_t& i) { return barrier_start_ptr + (kNumStages * 3 + i); });
+    auto tmem_empty_barriers    = utils::PatternVisitor([=](const uint32_t& i) { return barrier_start_ptr + (kNumStages * 3 + kNumEpilogueStages + i); });
+    auto tmem_ptr_in_smem  = reinterpret_cast<uint32_t*>(barrier_start_ptr + kNumStages * 3 + kNumEpilogueStages * 2);
+
+    // Initialize barriers
+    if (warp_idx == 1 and cute::elect_one_sync()) {
+        #pragma unroll
+        for (uint32_t i = 0; i < kNumStages; ++ i) {
+            // Arrive at all CTAs
+            full_barriers[i]->init(1);
+            empty_barriers[i]->init(1);
+            // Arrive only at the leader CTA
+            with_sf_full_barriers[i]->init(kNumMulticast * 32);
+        }
+        #pragma unroll
+        for (uint32_t i = 0; i < kNumEpilogueStages; ++ i) {
+            // Arrive at all CTAs
+            tmem_full_barriers[i]->init(1);
+            // Arrive only at the leader CTA
+            tmem_empty_barriers[i]->init(kNumMulticast * kNumUMMAStoreThreads);
+        }
+
+        // Make initialized barrier visible in async proxy
+        cutlass::arch::fence_barrier_init();
+    } else if (warp_idx == 2) {
+        // Allocate tensor memory
+        Allocator().allocate(kNumTmemCols, tmem_ptr_in_smem);
+    }
+    kNumMulticast > 1 ? cute::cluster_sync() : __syncthreads();
+
+    // Wait for primary kernel completion
+    cudaGridDependencySynchronize();
+
+    // Block scheduler
+    uint32_t m_block_idx, n_block_idx;
+    auto scheduler = sched::Scheduler<kGemmType, BLOCK_M, BLOCK_N, kNumGroups, kNumMulticast, kIsMulticastOnA, kNumSMs, kGranKA * 4>(
+        shape_m, shape_n, shape_k, grouped_layout);
+
+    // Pipeline and TMA phases
+    uint32_t stage_idx = 0, phase = 0;
+    auto advance_pipeline = [&](uint32_t& k_block_idx) {
+        ++ k_block_idx;
+
+        // Flip phases only if reach the next first stage
+        stage_idx = stage_idx == kNumStages - 1 ? 0 : stage_idx + 1;
+        phase ^= stage_idx == 0;
+    };
+
+    // Dispatch warps into different roles
+    if (warp_idx == 0 and cute::elect_one_sync()) {
+        // TMA load warp
+        // Persistently schedule over blocks
+        while (scheduler.get_next_block(m_block_idx, n_block_idx)) {
+            // Use dynamic load block M, when swap-AB is enabled
+            const auto load_block_m = kSwapAB ? scheduler.get_aligned_effective_m_in_block(m_block_idx) / kNumMulticast : LOAD_BLOCK_M;
+
+            // For k-grouped layout, the number of block K is variable
+            const auto num_total_k_blocks = math::ceil_div(scheduler.current_shape_k, BLOCK_K);
+            for (uint32_t k_block_idx = 0; k_block_idx < num_total_k_blocks; advance_pipeline(k_block_idx)) {
+                // Wait consumer release
+                empty_barriers[stage_idx]->wait(phase ^ 1);
+
+                // Compute offsets
+                // NOTES: the group is always concatenated with the outer dimension
+                uint32_t m_idx = scheduler.template get_global_idx<(kGemmType == GemmType::MGroupedMasked), sched::IndexType::MN> (
+                    shape_m, BLOCK_M, m_block_idx);
+                uint32_t n_idx = scheduler.template get_global_idx<(kMajorB == cute::UMMA::Major::K), sched::IndexType::MN> (
+                    shape_n, BLOCK_N, n_block_idx, m_block_idx);
+
+                // NOTES: `k_idx` is actually the k index default for K-major, while `k_b_idx` may be MN-major
+                // And for all m-grouped GEMMs, A must be K-majored
+                DG_STATIC_ASSERT(kGemmType == GemmType::Normal or kGemmType == GemmType::KGroupedContiguous or kGemmType == GemmType::Batched or
+                                 kMajorA == cute::UMMA::Major::K, "Invalid major");
+                uint32_t k_idx = k_block_idx * BLOCK_K;
+                uint32_t k_a_idx = scheduler.template get_global_idx<(kMajorA == cute::UMMA::Major::MN), sched::IndexType::K> (
+                    shape_k, BLOCK_K, k_block_idx, m_block_idx);
+                uint32_t k_b_idx = scheduler.template get_global_idx<(kMajorB == cute::UMMA::Major::MN), sched::IndexType::K> (
+                    shape_k, BLOCK_K, k_block_idx, m_block_idx);
+
+                // Add 2 CTA offsets
+                if constexpr (kNumMulticast > 1) {
+                    m_idx += kIsMulticastOnA ? (cute::block_rank_in_cluster() * load_block_m) : 0;
+                    n_idx += kIsMulticastOnA ? 0 : (cute::block_rank_in_cluster() * LOAD_BLOCK_N);
+                }
+
+                // Issue TMAs
+                constexpr bool kIsBatchedMM = (kGemmType == GemmType::Batched);
+                const uint32_t batch_idx = (kIsBatchedMM ? scheduler.current_group_idx : 0);
+                if constexpr (kMajorA == cute::UMMA::Major::K)
+                    tma::copy<BLOCK_K, LOAD_BLOCK_M, kSwizzleAMode, a_dtype_t, kIsBatchedMM>(
+                        &tensor_map_a, full_barriers[stage_idx], smem_a[stage_idx], k_a_idx, m_idx, 1, batch_idx);
+                if constexpr (kMajorA == cute::UMMA::Major::MN)
+                    tma::copy<LOAD_BLOCK_M, BLOCK_K, kSwizzleAMode, a_dtype_t, kIsBatchedMM>(
+                        &tensor_map_a, full_barriers[stage_idx], smem_a[stage_idx], m_idx, k_a_idx, 1, batch_idx);
+                if constexpr (kMajorB == cute::UMMA::Major::K)
+                    tma::copy<BLOCK_K, LOAD_BLOCK_N, kSwizzleBMode, b_dtype_t, kIsBatchedMM>(
+                        &tensor_map_b, full_barriers[stage_idx], smem_b[stage_idx], k_b_idx, n_idx, 1, batch_idx);
+                if constexpr (kMajorB == cute::UMMA::Major::MN)
+                    tma::copy<LOAD_BLOCK_N, BLOCK_K, kSwizzleBMode, b_dtype_t, kIsBatchedMM>(
+                        &tensor_map_b, full_barriers[stage_idx], smem_b[stage_idx], n_idx, k_b_idx, 1, batch_idx);
+                auto num_arrival_bytes = SMEM_A_SIZE_PER_STAGE / (std::is_same_v<a_dtype_t, cutlass::float_e4m3_t> ? 1 : 2) +
+                                         SMEM_B_SIZE_PER_STAGE / (std::is_same_v<b_dtype_t, cutlass::float_e4m3_t> ? 1 : 2);
+
+                // Issue SFA and SFB TMAs at certain stages
+                // No swizzling, so one TMA for one SF is enough
+                if (k_block_idx % kNumSFAStagesPerLoad == 0) {
+                    uint32_t sfa_m_idx = m_block_idx * BLOCK_M;
+                    uint32_t sfa_k_idx = scheduler.template get_global_idx<(not is_m_grouped_contiguous(kGemmType)), sched::IndexType::SF_K>(
+                        shape_sfa_k, 1, math::ceil_div(k_idx, BLOCK_K * kNumSFAStagesPerLoad));
+                    tma::copy<BLOCK_M, 1, 0>(&tensor_map_sfa, full_barriers[stage_idx], smem_sfa[stage_idx], sfa_m_idx, sfa_k_idx);
+                    num_arrival_bytes += BLOCK_M * sizeof(uint32_t);
+                }
+                if (k_block_idx % kNumSFBStagesPerLoad == 0) {
+                    uint32_t sfb_n_idx = n_block_idx * BLOCK_N;
+                    uint32_t sfb_k_idx = scheduler.template get_global_idx<true, sched::IndexType::SF_K>(
+                        shape_sfb_k, 1, math::ceil_div(k_idx, BLOCK_K * kNumSFBStagesPerLoad), m_block_idx);
+                    tma::copy<BLOCK_N, 1, 0>(&tensor_map_sfb, full_barriers[stage_idx], smem_sfb[stage_idx], sfb_n_idx, sfb_k_idx);
+                    num_arrival_bytes += BLOCK_N * sizeof(uint32_t);
+                }
+
+                // Arrive at full barriers
+                full_barriers[stage_idx]->arrive_and_expect_tx(num_arrival_bytes);
+            }
+        }
+    } else if (warp_idx == 1 and is_leader_cta) {
+        // MMA issue warp
+        // NOTES: only the leader CTA will do this
+        // Make instruction descriptor
+        auto instr_desc = kSwapAB ? cute::UMMA::make_instr_desc_block_scaled<b_dtype_t, a_dtype_t, float, cutlass::float_ue8m0_t,
+                                                                             UMMA_M, UMMA_N, kMajorB, kMajorA>()
+                                  : cute::UMMA::make_instr_desc_block_scaled<a_dtype_t, b_dtype_t, float, cutlass::float_ue8m0_t,
+                                                                             UMMA_M, UMMA_N, kMajorA, kMajorB>();
+        auto sf_desc = mma::sm100::make_sf_desc(nullptr);
+
+        DG_STATIC_ASSERT(kNumStages <= 32, "Too many stages");
+        auto a_desc = mma::sm100::make_umma_desc<kMajorA, LOAD_BLOCK_M, BLOCK_K, kSwizzleAMode>(smem_a[0], 0, 0);
+        auto b_desc = mma::sm100::make_umma_desc<kMajorB, LOAD_BLOCK_N, BLOCK_K, kSwizzleBMode>(smem_b[0], 0, 0);
+        uint32_t a_desc_lo = lane_idx < kNumStages ? a_desc.lo + lane_idx * SMEM_A_SIZE_PER_STAGE / 16 : 0u;
+        uint32_t b_desc_lo = lane_idx < kNumStages ? b_desc.lo + lane_idx * SMEM_B_SIZE_PER_STAGE / 16 : 0u;
+
+        // Checks for MMA instructions
+        // NOTES: CUTLASS does not have such checks except the MMA traits, but we are not using these traits
+        DG_STATIC_ASSERT((UMMA_M == 64  and UMMA_N %  8 == 0 and  8 <= UMMA_N and UMMA_N <= 256) or
+                         (UMMA_M == 128 and UMMA_N % 16 == 0 and 16 <= UMMA_N and UMMA_N <= 256) or
+                         (UMMA_M == 256 and UMMA_N % 16 == 0 and 16 <= UMMA_N and UMMA_N <= 256),
+                         "Invalid MMA instruction shape");
+
+        // Persistently schedule over blocks
+        while (scheduler.get_next_block(m_block_idx, n_block_idx)) {
+            // Wait tensor memory empty barrier arrival
+            auto accum_stage_idx = scheduler.current_iter % kNumEpilogueStages;
+            auto accum_phase_idx = (scheduler.current_iter / kNumEpilogueStages) & 1;
+            tmem_empty_barriers[accum_stage_idx]->wait(accum_phase_idx ^ 1);
+            ptx::tcgen05_after_thread_sync();
+
+            // Empty barrier arrival
+            auto empty_barrier_arrive = [&](const bool& do_tmem_full_arrive) {
+                auto umma_arrive = [](const uint64_t* barrier) {
+                    if constexpr (kNumMulticast == 1) {
+                        cutlass::arch::umma_arrive(barrier);
+                    } else {
+                        constexpr uint16_t kCTAMask = (1 << kNumMulticast) - 1;
+                        cutlass::arch::umma_arrive_multicast_2x1SM(barrier, kCTAMask);
+                    }
+                };
+                umma_arrive(reinterpret_cast<uint64_t*>(empty_barriers[stage_idx]));
+
+                // NOTES: the tensor memory accumulator pipeline has nothing to do with multicasting
+                if (do_tmem_full_arrive)
+                    umma_arrive(reinterpret_cast<uint64_t*>(tmem_full_barriers[accum_stage_idx]));
+                __syncwarp();
+            };
+
+            // Dynamic update of UMMA N based on effective M, when swap-AB is enabled
+            if constexpr (kSwapAB) {
+                uint32_t umma_n = scheduler.get_aligned_effective_m_in_block(m_block_idx);
+                mma::sm100::update_instr_desc_with_umma_n(instr_desc, umma_n);
+            }
+
+            // Launch MMAs
+            const auto num_total_k_blocks = math::ceil_div(scheduler.current_shape_k, BLOCK_K);
+            #pragma unroll 4
+            for (uint32_t k_block_idx = 0; k_block_idx < num_total_k_blocks; advance_pipeline(k_block_idx)) {
+                // Wait TMA and SF-transpose arrival
+                with_sf_full_barriers[stage_idx]->wait(phase);
+                ptx::tcgen05_after_thread_sync();
+
+                const auto a_desc_base_lo = ptx::exchange(a_desc_lo, stage_idx);
+                const auto b_desc_base_lo = ptx::exchange(b_desc_lo, stage_idx);
+                if (cute::elect_one_sync()) {
+                    // Do SF copy at certain stages
+                    // TODO: process shared memory descriptor by addition
+                    using cute_utccp_t = cute::conditional_t<kNumMulticast == 1,
+                        cute::SM100_UTCCP_4x32dp128bit_1cta, cute::SM100_UTCCP_4x32dp128bit_2cta>;
+                    const uint32_t sfa_stage_in_group_idx = k_block_idx % kNumSFAStagesPerLoad;
+                    if (sfa_stage_in_group_idx == 0) {
+                        #pragma unroll
+                        for (uint32_t i = 0; i < SF_BLOCK_M / kNumUTCCPAlignedElems; ++ i) {
+                            auto smem_ptr = smem_sfa[stage_idx] + i * kNumUTCCPAlignedElems;
+                            mma::sm100::replace_smem_desc_addr(sf_desc, smem_ptr);
+                            cute_utccp_t::copy(sf_desc, kTmemStartColOfSFA + i * 4);
+                        }
+                    }
+                    const uint32_t sfb_stage_in_group_idx = k_block_idx % kNumSFBStagesPerLoad;
+                    if (sfb_stage_in_group_idx == 0) {
+                        #pragma unroll
+                        for (uint32_t i = 0; i < SF_BLOCK_N / kNumUTCCPAlignedElems; ++ i) {
+                            auto smem_ptr = smem_sfb[stage_idx] + i * kNumUTCCPAlignedElems;
+                            mma::sm100::replace_smem_desc_addr(sf_desc, smem_ptr);
+                            cute_utccp_t::copy(sf_desc, kTmemStartColOfSFB + i * 4);
+                        }
+                    }
+
+                    // Issue UMMA
+                    using mma_t = cute::conditional_t<
+                        kNumMulticast == 1, ptx::SM100_MMA_MXF8F6F4_SS, ptx::SM100_MMA_MXF8F6F4_2x1SM_SS>;
+                    #pragma unroll
+                    for (uint32_t k = 0; k < BLOCK_K / UMMA_K; ++ k) {
+                        const uint32_t sfa_id = (kGranKA == 32 ? k : sfa_stage_in_group_idx);
+                        const uint32_t sfb_id = (kGranKB == 32 ? k : sfb_stage_in_group_idx);
+                        const auto runtime_instr_desc = kSwapAB ?
+                            mma::sm100::make_runtime_instr_desc_with_sf_id(instr_desc, sfb_id, sfa_id):
+                            mma::sm100::make_runtime_instr_desc_with_sf_id(instr_desc, sfa_id, sfb_id);
+
+                        a_desc.lo = mma::sm100::advance_umma_desc_lo<kMajorA, LOAD_BLOCK_M, kSwizzleAMode, a_dtype_t>(a_desc_base_lo, 0, k * UMMA_K);
+                        b_desc.lo = mma::sm100::advance_umma_desc_lo<kMajorB, LOAD_BLOCK_N, kSwizzleBMode, b_dtype_t>(b_desc_base_lo, 0, k * UMMA_K);
+                        if constexpr (kSwapAB) {
+                            mma_t::fma(b_desc, a_desc, accum_stage_idx * UMMA_N,
+                                       k_block_idx > 0 or k > 0, runtime_instr_desc,
+                                       kTmemStartColOfSFB, kTmemStartColOfSFA);
+                        } else {
+                            mma_t::fma(a_desc, b_desc, accum_stage_idx * UMMA_N,
+                                       k_block_idx > 0 or k > 0, runtime_instr_desc,
+                                       kTmemStartColOfSFA, kTmemStartColOfSFB);
+                        }
+                    }
+                }
+                __syncwarp();
+
+                // Commit to the mbarrier object
+                // No explicit `tcgen05.fence::before_thread_sync` is needed, as this is implicitly performed by `tcgen05.commit`
+                empty_barrier_arrive(k_block_idx == num_total_k_blocks - 1);
+            }
+        }
+
+        // To safely deconstruct barriers, we need another round of waits
+        const auto iter_idx = scheduler.current_iter - 1;
+        if (kNumMulticast > 1 and iter_idx >= 0) {
+            const auto accum_phase_idx = (iter_idx / kNumEpilogueStages) & 1;
+            tmem_empty_barriers[iter_idx % kNumEpilogueStages]->wait(accum_phase_idx);
+        }
+    } else if (warp_idx == 2) {
+        // UTCCP transposer
+        auto utccp_required_smem_warp_transpose = [&](const uint32_t* smem_ptr) {
+            DG_STATIC_ASSERT(kNumUTCCPAlignedElems == 128, "Invalid aligned elements");
+            uint32_t values[4];
+            #pragma unroll
+            for (uint32_t i = 0; i < 4; ++ i)
+                values[i] = ptx::ld_shared(smem_ptr + (i ^ (lane_idx >> 3)) * 32 + lane_idx);
+            __syncwarp();
+            #pragma unroll
+            for (uint32_t i = 0; i < 4; ++ i)
+                ptx::st_shared(smem_ptr + lane_idx * 4 + (i ^ (lane_idx >> 3)), values[i]);
+        };
+
+        while (scheduler.get_next_block(m_block_idx, n_block_idx)) {
+            const auto num_total_k_blocks = math::ceil_div(scheduler.current_shape_k, BLOCK_K);
+            for (uint32_t k_block_idx = 0; k_block_idx < num_total_k_blocks; advance_pipeline(k_block_idx)) {
+                // Wait TMA arrival
+                full_barriers[stage_idx]->wait(phase);
+
+                // Transpose for UTCCP at certain stages
+                if (k_block_idx % kNumSFAStagesPerLoad == 0) {
+                    #pragma unroll
+                    for (uint32_t i = 0; i < SF_BLOCK_M / kNumUTCCPAlignedElems; ++ i)
+                        utccp_required_smem_warp_transpose(smem_sfa[stage_idx] + i * kNumUTCCPAlignedElems);
+                    // TODO: figure out whether the proxy fence is valid for 2-CTA cases
+                    cutlass::arch::fence_view_async_shared();
+                }
+                if (k_block_idx % kNumSFBStagesPerLoad == 0) {
+                    #pragma unroll
+                    for (uint32_t i = 0; i < SF_BLOCK_N / kNumUTCCPAlignedElems; ++ i)
+                        utccp_required_smem_warp_transpose(smem_sfb[stage_idx] + i * kNumUTCCPAlignedElems);
+                    // TODO: figure out whether the proxy fence is valid for 2-CTA cases
+                    cutlass::arch::fence_view_async_shared();
+                }
+
+                // Arrive
+                with_sf_full_barriers[stage_idx]->arrive(0u);
+            }
+        }
+    } else if (warp_idx >= kNumNonEpilogueThreads / 32 and warp_idx < (kNumNonEpilogueThreads + kNumUMMAStoreThreads) / 32) {
+        // Epilogue warp groups
+        const auto epilogue_warp_idx = warp_idx - (kNumNonEpilogueThreads / 32);
+
+        // NOTES: tensor memory addresses are simplified, as the hardware will ignore the warp index bits,
+        // i.e., no need for `tmem_ptr |= (epilogue_warp_idx * 32) << 16`.
+        // NOTES: we also forbid two CTAs to share the same SM and its tensor memory
+        DG_TRAP_ONLY_DEVICE_ASSERT(ptx::ld_shared(tmem_ptr_in_smem) == 0);
+
+        // Share store pipeline between blocks
+        uint32_t tma_stage_idx = 0;
+
+        // Persistently schedule over blocks
+        while (scheduler.get_next_block(m_block_idx, n_block_idx)) {
+            auto accum_stage_idx = scheduler.current_iter % kNumEpilogueStages;
+            auto accum_phase_idx = (scheduler.current_iter / kNumEpilogueStages) & 1;
+
+            // Wait UMMA arrival
+            tmem_full_barriers[accum_stage_idx]->wait(accum_phase_idx);
+            ptx::tcgen05_after_thread_sync();
+
+            const auto tmem_base_addr = accum_stage_idx * UMMA_N;
+            const auto base_m_idx = scheduler.template get_global_idx<(not is_m_grouped_contiguous(kGemmType)), sched::IndexType::MN>(shape_m, BLOCK_M, m_block_idx);
+            const auto base_n_idx = n_block_idx * BLOCK_N;
+
+            if constexpr (kSwapAB) {
+                const auto effective_m = scheduler.get_aligned_effective_m_in_block(m_block_idx);
+                epilogue::sm100_store_cd_swap_ab<
+                    BLOCK_M, BLOCK_N, STORE_BLOCK_M, STORE_BLOCK_N,
+                    kSwizzleCDMode, kNumTMAStoreStages, kNumUMMAStoreThreads,
+                    kGemmType, kWithAccumulation,
+                    cd_dtype_t, epilogue_type_t>
+                (smem_cd, tma_stage_idx, tmem_base_addr,
+                 base_m_idx, base_n_idx, scheduler.current_group_idx,
+                 effective_m,
+                 epilogue_warp_idx, lane_idx,
+                 tmem_empty_barriers[accum_stage_idx],
+                 tensor_map_cd);
+            } else {
+                epilogue::sm100_store_cd<
+                    BLOCK_M, BLOCK_N, STORE_BLOCK_M, STORE_BLOCK_N,
+                    kSwizzleCDMode, kNumTMAStoreStages, kNumUMMAStoreThreads,
+                    kGemmType, kWithAccumulation,
+                    cd_dtype_t, epilogue_type_t>
+                (smem_cd, tma_stage_idx, tmem_base_addr,
+                 base_m_idx, base_n_idx, scheduler.current_group_idx,
+                 epilogue_warp_idx, lane_idx,
+                 tmem_empty_barriers[accum_stage_idx],
+                 tensor_map_cd);
+            }
+        }
+    }
+
+    // TODO: Remove redundant synchronization
+    kNumMulticast > 1 ? cute::cluster_sync() : __syncthreads();
+
+    // Deallocate tensor memory
+    if (warp_idx == 0)
+        Allocator().free(0, kNumTmemCols);
+
+#else
+    if (blockIdx.x == 0 and threadIdx.x == 0)
+        DG_DEVICE_ASSERT(false and "This kernel only support sm_100f");
+#endif
+}
+
+};  // namespace deep_gemm
+
+#pragma clang diagnostic pop

deep_gemm/include/deep_gemm/impls/sm100_fp8_mqa_logits.cuh

@@ -6,27 +6,31 @@
 #include <cute/arch/cluster_sm90.hpp>
 #include <cute/arch/copy_sm90_desc.hpp>
 
+#include <deep_gemm/common/cute_tie.cuh>
+#include <deep_gemm/common/math.cuh>
+#include <deep_gemm/common/tma_copy.cuh>
 #include <deep_gemm/common/utils.cuh>
-#include <deep_gemm/common/sm90_utils.cuh>
-#include <deep_gemm/common/sm100_utils.cuh>
+#include <deep_gemm/mma/sm100.cuh>
+#include <deep_gemm/ptx/ld_st.cuh>
+#include <deep_gemm/ptx/tcgen05.cuh>
+#include <deep_gemm/ptx/utils.cuh>
 
 namespace deep_gemm {
 
-using namespace deep_gemm::sm90;
-using namespace deep_gemm::sm100;
-
 template <uint32_t kNumHeads, uint32_t kHeadDim,
           bool kIsCompressedLogits,
           uint32_t BLOCK_Q, uint32_t BLOCK_KV,
           uint32_t kNumQStages, uint32_t kNumKVStages,
+          uint32_t kNumSMs,
           uint32_t kNumSpecializedThreads, uint32_t kNumMathThreads,
+          typename logits_dtype_t,
           uint32_t kNumMathWarpGroups = kNumMathThreads / 128>
-__global__ __launch_bounds__(kNumSpecializedThreads + kNumMathThreads, 1)
+CUTLASS_GLOBAL __launch_bounds__(kNumSpecializedThreads + kNumMathThreads, 1)
 void sm100_fp8_mqa_logits(const uint32_t seq_len, const uint32_t seq_len_kv,
-                          const uint32_t max_seqlen_k, const uint64_t stride_logits,
+                          const uint32_t max_seqlen_k, const uint32_t stride_logits,
                           uint32_t* cu_seq_len_k_start,
                           uint32_t* cu_seq_len_k_end,
-                          float* logits,
+                          logits_dtype_t* logits,
                           const __grid_constant__ cute::TmaDescriptor tensor_map_q,
                           const __grid_constant__ cute::TmaDescriptor tensor_map_kv,
                           const __grid_constant__ cute::TmaDescriptor tensor_map_kv_scales,
@@ -35,26 +39,26 @@ void sm100_fp8_mqa_logits(const uint32_t seq_len, const uint32_t seq_len_kv,
     // Normally, `h (kNumHeads) == 32` and `d (kHeadDim) == 64`
     // For one block, we process `[q_start:q_end, h, d] @ [kv_start:kv_end, d] -> [q_start:q_end, kv_start:kv_end]`
     // Q should be load only at once for a block
-    const auto& num_q_blocks = ceil_div(seq_len, BLOCK_Q);
+    const auto num_q_blocks = math::ceil_div(seq_len, BLOCK_Q);
 
     // Types
     using Barrier = cutlass::arch::ClusterTransactionBarrier;
 
-    // NOTES: use `__shfl_sync` to encourage NVCC to use unified registers
-    const auto& warp_idx = __shfl_sync(0xffffffff, threadIdx.x / 32, 0);
-    const auto& warp_in_group_idx = warp_idx % 4;
-    const auto& warpgroup_idx = warp_idx / 4;
-    const auto& lane_idx = get_lane_idx();
+    // Utils
+    const auto sm_idx = blockIdx.x;
+    const auto warp_idx = cutlass::canonical_warp_idx_sync();
+    const auto warpgroup_idx = warp_idx / 4;
+    const auto lane_idx = ptx::get_lane_idx();
+    constexpr uint32_t kSpecWarpStart = kNumMathWarpGroups * 4;
 
     // Prefetch TMA descriptors
     DG_STATIC_ASSERT(kNumSpecializedThreads == 128 and kNumMathThreads % 128 == 0, "Invalid threads");
-    if (warp_idx == kNumMathThreads / 32 and cute::elect_one_sync()) {
+    if (warp_idx == kSpecWarpStart) {
         cute::prefetch_tma_descriptor(&tensor_map_q);
         cute::prefetch_tma_descriptor(&tensor_map_kv);
         cute::prefetch_tma_descriptor(&tensor_map_kv_scales);
         cute::prefetch_tma_descriptor(&tensor_map_weights);
     }
-    __syncwarp();
 
     // Shared memory configs
     // NOTES: weight may be unaligned
@@ -62,7 +66,7 @@ void sm100_fp8_mqa_logits(const uint32_t seq_len, const uint32_t seq_len_kv,
     static constexpr uint32_t SMEM_WEIGHT_SIZE_PER_STAGE = BLOCK_Q * kNumHeads * sizeof(float);
     static constexpr uint32_t SMEM_KV_SIZE_PER_STAGE = BLOCK_KV * kHeadDim * sizeof(__nv_fp8_e4m3);
     static constexpr uint32_t SMEM_KV_SCALE_SIZE_PER_STAGE = BLOCK_KV * sizeof(float);
-    static constexpr uint32_t ALIGNED_SMEM_KV_SCALE_SIZE_PER_STAGE = constexpr_align(SMEM_KV_SCALE_SIZE_PER_STAGE, 512u);
+    static constexpr uint32_t ALIGNED_SMEM_KV_SCALE_SIZE_PER_STAGE = math::constexpr_align(SMEM_KV_SCALE_SIZE_PER_STAGE, 512u);
 
     // Align to 512 bytes for swizzle-64B
     extern __shared__ __align__(512) uint8_t smem_buffer[];
@@ -75,19 +79,19 @@ void sm100_fp8_mqa_logits(const uint32_t seq_len, const uint32_t seq_len_kv,
     DG_STATIC_ASSERT(kNumTmemCols <= 512, "Too many tensor memory");
 
     // Data on shared memory
-    auto smem_q = PatternVisitor([&](const uint32_t& i) {
+    auto smem_q = utils::PatternVisitor([&](const uint32_t& i) {
         return reinterpret_cast<__nv_fp8_e4m3*>(smem_buffer +
             SMEM_Q_SIZE_PER_STAGE * i);
     });
-    auto smem_weights = PatternVisitor([&](const uint32_t& i) {
+    auto smem_weights = utils::PatternVisitor([&](const uint32_t& i) {
         return reinterpret_cast<float*>(smem_buffer +
             SMEM_Q_SIZE_PER_STAGE * kNumQStages + SMEM_WEIGHT_SIZE_PER_STAGE * i);
     });
-    auto smem_kv = PatternVisitor([&](const uint32_t& i) {
+    auto smem_kv = utils::PatternVisitor([&](const uint32_t& i) {
         return reinterpret_cast<__nv_fp8_e4m3*>(smem_buffer + (
             SMEM_Q_SIZE_PER_STAGE * kNumQStages + SMEM_WEIGHT_SIZE_PER_STAGE * kNumQStages + SMEM_KV_SIZE_PER_STAGE * i));
     });
-    auto smem_kv_scales =  PatternVisitor([&](const uint32_t& i) {
+    auto smem_kv_scales = utils::PatternVisitor([&](const uint32_t& i) {
         return reinterpret_cast<float*>(smem_buffer +
             SMEM_Q_SIZE_PER_STAGE * kNumQStages + SMEM_WEIGHT_SIZE_PER_STAGE * kNumQStages +
             SMEM_KV_SIZE_PER_STAGE * kNumKVStages + ALIGNED_SMEM_KV_SCALE_SIZE_PER_STAGE * i);
@@ -95,76 +99,77 @@ void sm100_fp8_mqa_logits(const uint32_t seq_len, const uint32_t seq_len_kv,
 
     // TMA barriers
     auto barrier_ptr = reinterpret_cast<Barrier*>(smem_kv_scales[kNumKVStages]);
-    auto full_q_barriers     = PatternVisitor([&](const uint32_t& i) { return barrier_ptr + i; });
-    auto empty_q_barriers    = PatternVisitor([&](const uint32_t& i) { return barrier_ptr + (kNumQStages + i); });
-    auto full_kv_barriers    = PatternVisitor([&](const uint32_t& i) { return barrier_ptr + (kNumQStages * 2 + i); });
-    auto empty_kv_barriers   = PatternVisitor([&](const uint32_t& i) { return barrier_ptr + (kNumQStages * 2 + kNumKVStages + i); });
-    auto full_umma_barriers  = PatternVisitor([&](const uint32_t& i) { return barrier_ptr + (kNumQStages * 2 + kNumKVStages * 2 + i); });
-    auto empty_umma_barriers = PatternVisitor([&](const uint32_t& i) { return barrier_ptr + (kNumQStages * 2 + kNumKVStages * 2 + kNumMathWarpGroups + i); });
+    auto full_q_barriers     = utils::PatternVisitor([&](const uint32_t& i) { return barrier_ptr + i; });
+    auto empty_q_barriers    = utils::PatternVisitor([&](const uint32_t& i) { return barrier_ptr + (kNumQStages + i); });
+    auto full_kv_barriers    = utils::PatternVisitor([&](const uint32_t& i) { return barrier_ptr + (kNumQStages * 2 + i); });
+    auto empty_kv_barriers   = utils::PatternVisitor([&](const uint32_t& i) { return barrier_ptr + (kNumQStages * 2 + kNumKVStages + i); });
+    auto full_umma_barriers  = utils::PatternVisitor([&](const uint32_t& i) { return barrier_ptr + (kNumQStages * 2 + kNumKVStages * 2 + i); });
+    auto empty_umma_barriers = utils::PatternVisitor([&](const uint32_t& i) { return barrier_ptr + (kNumQStages * 2 + kNumKVStages * 2 + kNumMathWarpGroups + i); });
 
     // Tensor memory allocation
     auto tmem_ptr_in_smem = reinterpret_cast<uint32_t*>(barrier_ptr + kNumQStages * 2 + kNumKVStages * 2 + kNumMathWarpGroups * 2);
 
     // Initialize barriers
     DG_STATIC_ASSERT(kNumSpecializedThreads % 128 == 0 and kNumSpecializedThreads >= 64, "Invalid threads");
-    const bool& is_tma_load_warp = (warp_idx == (kNumMathThreads / 32));
-    const bool& is_umma_warp = (warp_idx == (kNumMathThreads / 32 + 1));
-    if (is_tma_load_warp and cute::elect_one_sync()) {
+    if (warp_idx == kSpecWarpStart and cute::elect_one_sync()) {
         #pragma unroll
         for (uint32_t i = 0; i < kNumQStages; ++ i) {
             full_q_barriers[i]->init(1);
-            empty_q_barriers[i]->init(kNumMathThreads);
+            empty_q_barriers[i]->init(kNumMathThreads + 32);
         }
         #pragma unroll
         for (uint32_t i = 0; i < kNumKVStages; ++ i) {
             full_kv_barriers[i]->init(1);
             empty_kv_barriers[i]->init(kNumMathThreads);
         }
-        #pragma unroll
-        for (uint32_t i = 0; i < kNumMathWarpGroups; ++ i) {
-            full_umma_barriers[i]->init(1);
-            empty_umma_barriers[i]->init(128);
-        }
-
-        // Make initialized barrier visible in async proxy
         cutlass::arch::fence_barrier_init();
-    } else if (is_umma_warp) {
+    }
+    if (warp_idx == kSpecWarpStart + 1) {
+        if (cute::elect_one_sync()) {
+            #pragma unroll
+            for (uint32_t i = 0; i < kNumMathWarpGroups; ++ i) {
+                full_umma_barriers[i]->init(1);
+                empty_umma_barriers[i]->init(128);
+            }
+            cutlass::arch::fence_barrier_init();
+        }
         // Allocate tensor memory
         cute::TMEM::Allocator1Sm().allocate(kNumTmemCols, tmem_ptr_in_smem);
     }
     __syncthreads();
 
     // Register reconfigurations
-    constexpr uint32_t kNumSpecializedRegisters = 24;
-    constexpr uint32_t kNumMathRegisters = 240;
+    constexpr uint32_t kNumSpecializedRegisters = 40;
+    constexpr uint32_t kNumMathRegisters = 232;
 
     // Block scheduler
-    uint32_t block_q_idx = blockIdx.x, q_iter_idx = 0;
-    const auto& get_next_block_q_idx = [&]() -> cute::tuple<uint32_t, uint32_t> {
-        return {block_q_idx + gridDim.x, q_iter_idx + 1};
+    uint32_t block_q_idx = sm_idx, q_iter_idx = 0;
+    const auto get_next_block_q_idx = [&]() -> cute::tuple<uint32_t, uint32_t> {
+        return {block_q_idx + kNumSMs, q_iter_idx + 1};
     };
     uint32_t seq_k_start[BLOCK_Q], seq_k_end[BLOCK_Q];
-    const auto& load_schedule = [&](const uint32_t& q_iter_offset = 0) -> cute::tuple<uint32_t, uint32_t, uint32_t, uint32_t> {
+    const auto load_schedule = [&](const uint32_t& q_iter_offset = 0) -> cute::tuple<uint32_t, uint32_t, uint32_t, uint32_t> {
         uint32_t start = cute::numeric_limits<uint32_t>::max();
         uint32_t end = cute::numeric_limits<uint32_t>::min();
 
         #pragma unroll
         for (uint32_t i = 0; i < BLOCK_Q; ++ i) {
-            const auto& q_idx = min(block_q_idx * BLOCK_Q + i, seq_len - 1);
-            seq_k_start[i] = __ldg(cu_seq_len_k_start + q_idx);
-            seq_k_end[i] = __ldg(cu_seq_len_k_end + q_idx);
+            const auto q_idx = min(block_q_idx * BLOCK_Q + i, seq_len - 1);
+            seq_k_start[i] = cu_seq_len_k_start[q_idx];
+            seq_k_end[i] = cu_seq_len_k_end[q_idx];
             start = min(start, min(seq_k_start[i], seq_len_kv));
             end = max(end, min(seq_k_end[i], seq_len_kv));
         }
+        // TMA alignment requirements for SF KV
         start = start / 4 * 4;
         return {(q_iter_idx + q_iter_offset) % kNumQStages,       // Q pipeline stage
                 ((q_iter_idx + q_iter_offset) / kNumQStages) & 1, // Q pipeline phase
-                start, ceil_div(end - start, BLOCK_KV)};          // Task info
+                start, math::ceil_div(end - start, BLOCK_KV)};          // Task info
     };
 
     // KV pipeline
     uint32_t num_total_kv_blocks = 0;
-    const auto& get_kv_pipeline = [&](const uint32_t& kv_block_idx) -> cute::tuple<uint32_t, uint32_t> {
+    const auto get_kv_pipeline = [&](const uint32_t& kv_block_idx) -> cute::tuple<uint32_t, uint32_t> {
         return {
             (num_total_kv_blocks + kv_block_idx) % kNumKVStages,         // KV pipeline stage
             ((num_total_kv_blocks + kv_block_idx) / kNumKVStages) & 1    // KV pipeline phase
@@ -177,13 +182,16 @@ void sm100_fp8_mqa_logits(const uint32_t seq_len, const uint32_t seq_len_kv,
     constexpr uint32_t UMMA_K = 32 / sizeof(cutlass::float_e4m3_t);
     constexpr uint32_t UMMA_N = BLOCK_Q * kNumHeads;
 
-    if (is_tma_load_warp) {
+    // Wait for primary kernel completion
+    cudaGridDependencySynchronize();
+
+    if (warp_idx == kSpecWarpStart) {
         cutlass::arch::warpgroup_reg_dealloc<kNumSpecializedRegisters>();
 
         // Prefetch
-        const auto& issue_tma_q = [&](const uint32_t& stage_idx, const auto& block_idx) {
-            tma_copy<kHeadDim, BLOCK_Q * kNumHeads, kHeadDim>(&tensor_map_q, full_q_barriers[stage_idx], smem_q[stage_idx], 0, block_idx * BLOCK_Q * kNumHeads);
-            tma_copy<kNumHeads, BLOCK_Q, 0>(&tensor_map_weights, full_q_barriers[stage_idx], smem_weights[stage_idx], 0, block_idx * BLOCK_Q);
+        const auto issue_tma_q = [&](const uint32_t& stage_idx, const auto& block_idx) {
+            tma::copy<kHeadDim, BLOCK_Q * kNumHeads, kHeadDim>(&tensor_map_q, full_q_barriers[stage_idx], smem_q[stage_idx], 0, block_idx * BLOCK_Q * kNumHeads);
+            tma::copy<kNumHeads, BLOCK_Q, 0>(&tensor_map_weights, full_q_barriers[stage_idx], smem_weights[stage_idx], 0, block_idx * BLOCK_Q);
             full_q_barriers[stage_idx]->arrive_and_expect_tx(SMEM_Q_SIZE_PER_STAGE + SMEM_WEIGHT_SIZE_PER_STAGE);
         };
         if (cute::elect_one_sync() and block_q_idx < num_q_blocks)
@@ -209,10 +217,10 @@ void sm100_fp8_mqa_logits(const uint32_t seq_len, const uint32_t seq_len_kv,
                     empty_kv_barriers[kv_stage_idx]->wait(kv_phase ^ 1);
 
                     // Issue TMA KV
-                    tma_copy<kHeadDim, BLOCK_KV, kHeadDim>(&tensor_map_kv, full_kv_barriers[kv_stage_idx],
-                                                           smem_kv[kv_stage_idx], 0, kv_start + kv_block_idx * BLOCK_KV);
-                    tma_copy<BLOCK_KV, 1, 0>(&tensor_map_kv_scales, full_kv_barriers[kv_stage_idx],
-                                             smem_kv_scales[kv_stage_idx], kv_start + kv_block_idx * BLOCK_KV, 0);
+                    tma::copy<kHeadDim, BLOCK_KV, kHeadDim>(&tensor_map_kv, full_kv_barriers[kv_stage_idx],
+                                                            smem_kv[kv_stage_idx], 0, kv_start + kv_block_idx * BLOCK_KV);
+                    tma::copy<BLOCK_KV, 1, 0>(&tensor_map_kv_scales, full_kv_barriers[kv_stage_idx],
+                                              smem_kv_scales[kv_stage_idx], kv_start + kv_block_idx * BLOCK_KV, 0);
                     full_kv_barriers[kv_stage_idx]->arrive_and_expect_tx(SMEM_KV_SIZE_PER_STAGE + SMEM_KV_SCALE_SIZE_PER_STAGE);
                 }
                 num_total_kv_blocks += num_kv_blocks;
@@ -221,11 +229,11 @@ void sm100_fp8_mqa_logits(const uint32_t seq_len, const uint32_t seq_len_kv,
                 CUTE_TIE(get_next_block_q_idx(), block_q_idx, q_iter_idx);
             }
         }
-    } else if (is_umma_warp) {
+    } else if (warp_idx == kSpecWarpStart + 1) {
         cutlass::arch::warpgroup_reg_dealloc<kNumSpecializedRegisters>();
 
         // Require full allocation
-        DG_TRAP_ONLY_DEVICE_ASSERT(ld_shared(tmem_ptr_in_smem) == 0);
+        DG_TRAP_ONLY_DEVICE_ASSERT(ptx::ld_shared(tmem_ptr_in_smem) == 0);
 
         // Make UMMA desc
         auto instr_desc = cute::UMMA::make_instr_desc<cutlass::float_e4m3_t, cutlass::float_e4m3_t, float,
@@ -252,12 +260,12 @@ void sm100_fp8_mqa_logits(const uint32_t seq_len, const uint32_t seq_len_kv,
                 #pragma unroll
                 for (uint32_t i = 0; i < kNumMathWarpGroups; ++ i) {
                     empty_umma_barriers[i]->wait(((num_total_kv_blocks + kv_block_idx) & 1) ^ 1);
-                    tcgen05_after_thread_sync();
+                    ptx::tcgen05_after_thread_sync();
                     #pragma unroll
                     for (uint32_t k = 0; k < kHeadDim / UMMA_K; ++ k) {
-                        auto a_desc = make_umma_desc<cute::UMMA::Major::K, 0, kHeadDim, kHeadDim>(
+                        auto a_desc = mma::sm100::make_umma_desc<cute::UMMA::Major::K, 0, kHeadDim, kHeadDim>(
                             smem_kv[kv_stage_idx], i * UMMA_M, k * UMMA_K);
-                        auto b_desc = make_umma_desc<cute::UMMA::Major::K, 0, kHeadDim, kHeadDim>(
+                        auto b_desc = mma::sm100::make_umma_desc<cute::UMMA::Major::K, 0, kHeadDim, kHeadDim>(
                             smem_q[q_stage_idx], 0, k * UMMA_K);
                         cute::SM100_MMA_F8F6F4_SS::fma(a_desc, b_desc, i * UMMA_N, k, runtime_instr_desc);
                     }
@@ -266,23 +274,37 @@ void sm100_fp8_mqa_logits(const uint32_t seq_len, const uint32_t seq_len_kv,
             }
             num_total_kv_blocks += num_kv_blocks;
 
+            // UMMA warp must also arrive on empty_q to prevent running ahead
+            // of math warps in the Q pipeline
+            empty_q_barriers[q_stage_idx]->arrive();
+
             // Jump to the next block
             CUTE_TIE(get_next_block_q_idx(), block_q_idx, q_iter_idx);
         }
-    } else if (warp_idx >= kNumMathThreads / 32) {
+    } else if (warp_idx == kSpecWarpStart + 2 or warp_idx == kSpecWarpStart + 3) {
         cutlass::arch::warpgroup_reg_dealloc<kNumSpecializedRegisters>();
-    } else if (warp_idx < kNumMathThreads / 32) {
+    } else if (warp_idx < kSpecWarpStart) {
         cutlass::arch::warpgroup_reg_alloc<kNumMathRegisters>();
 
         // Offsets
-        const auto& tmem_start = __shfl_sync(0xffffffff, warpgroup_idx * UMMA_N, 0);
-        const auto& warp_offset = warp_idx * 32;
-        const auto& v_offset = lane_idx;
+        const auto tmem_start = warpgroup_idx * UMMA_N;
+        const auto math_thread_idx = warp_idx * 32 + lane_idx;
 
-        // Preload weights
-        constexpr uint32_t kNumWeightsInReg = cute::min(52, kNumHeads);
-        float weights[BLOCK_Q][kNumWeightsInReg];
-        DG_STATIC_ASSERT(kNumWeightsInReg % 4 == 0, "Invalid number of weights in registers");
+        // Helper lambda for loading tensor memory
+        auto tmem_load = [](auto num_elems_c, const uint32_t& tmem_addr, float* accum) {
+            constexpr int N = decltype(num_elems_c)::value;
+            DG_STATIC_ASSERT(N == 32 or N == 64, "Unsupported TMEM load size");
+            using Loader = cute::conditional_t<N == 32,
+                cute::SM100_TMEM_LOAD_32dp32b32x,
+                cute::SM100_TMEM_LOAD_32dp32b64x>;
+            [&]<size_t... Is>(cute::index_sequence<Is...>) {
+                Loader::copy(tmem_addr, reinterpret_cast<uint32_t*>(accum)[Is]...);
+            }(cute::make_index_sequence<N>{});
+            cutlass::arch::fence_view_async_tmem_load();
+        };
+
+        // Local register buffers
+        float weights[BLOCK_Q][kNumHeads];
 
         while (block_q_idx < num_q_blocks) {
             CUTE_TIE_DECL(load_schedule(), q_stage_idx, q_phase, kv_start, num_kv_blocks);
@@ -293,9 +315,9 @@ void sm100_fp8_mqa_logits(const uint32_t seq_len, const uint32_t seq_len_kv,
             // Read weights
             #pragma unroll
             for (uint32_t i = 0; i < BLOCK_Q; ++ i) {
-                for (uint32_t j = 0; j < kNumWeightsInReg; ++ j) {
-                    weights[i][j] = ld_shared(smem_weights[q_stage_idx] + i * kNumHeads + j);
-                }
+                #pragma unroll
+                for (uint32_t j = 0; j < kNumHeads; ++ j)
+                    weights[i][j] = ptx::ld_shared(smem_weights[q_stage_idx] + i * kNumHeads + j);
             }
 
             // Compute over KV blocks
@@ -307,82 +329,59 @@ void sm100_fp8_mqa_logits(const uint32_t seq_len, const uint32_t seq_len_kv,
                 full_kv_barriers[kv_stage_idx]->wait(kv_phase);
 
                 // Read per-KV scales
-                float scale_kv = ld_shared(smem_kv_scales[kv_stage_idx] + warp_offset + v_offset);
+                float scale_kv = ptx::ld_shared(smem_kv_scales[kv_stage_idx] + math_thread_idx);
 
                 // Wait UMMA arrival
                 full_umma_barriers[warpgroup_idx]->wait((num_total_kv_blocks + kv_block_idx) & 1);
-                tcgen05_after_thread_sync();
+                ptx::tcgen05_after_thread_sync();
 
                 // Release KV empty
                 empty_kv_barriers[kv_stage_idx]->arrive();
 
                 // Reduce over the head dim and store
-                const auto& kv_offset = kv_start + kv_block_idx * BLOCK_KV + warp_offset;
-                static constexpr uint32_t kNumAccumPerReduce = kNumHeads / 2;
+                const auto kv_offset = kv_start + kv_block_idx * BLOCK_KV + math_thread_idx;
                 DG_STATIC_ASSERT(kNumHeads % 8 == 0, "Invalid head");
 
-                constexpr uint32_t kNumLDTMElems = kNumHeads * BLOCK_Q;
-                DG_STATIC_ASSERT(kNumLDTMElems == 32 or kNumLDTMElems == 64 or kNumLDTMElems == 128, "Invalid kNumLDTMElems");
-                uint32_t shifted_accum[kNumLDTMElems];
-                auto tmem_load = [&](auto... Is) {
-                    if constexpr (kNumLDTMElems == 32) {
-                        cute::SM100_TMEM_LOAD_32dp32b32x::copy(tmem_start, shifted_accum[Is]...);
-                    } else if constexpr (kNumLDTMElems == 64) {
-                        cute::SM100_TMEM_LOAD_32dp32b64x::copy(tmem_start, shifted_accum[Is]...);
-                    } else if constexpr (kNumLDTMElems == 128) {
-                        cute::SM100_TMEM_LOAD_32dp32b128x::copy(tmem_start, shifted_accum[Is]...);
-                    }
-                };
-                [&]<size_t... Is>(cute::index_sequence<Is...>) { tmem_load(Is...); }(cute::make_index_sequence<kNumLDTMElems>{});
-                cutlass::arch::fence_view_async_tmem_load();
-
-                tcgen05_before_thread_sync();
-                empty_umma_barriers[warpgroup_idx]->arrive();
-                
                 #pragma unroll
                 for (uint32_t i = 0; i < BLOCK_Q; ++ i) {
-                    auto accum = reinterpret_cast<float*>(shifted_accum + i * kNumHeads);
+                    // Load accumulator from TMEM
+                    float accum[kNumHeads];
+                    tmem_load(cute::Int<kNumHeads>{}, tmem_start + i * kNumHeads, accum);
 
+                    // Release TMEM empty
+                    if (i == BLOCK_Q - 1) {
+                        ptx::tcgen05_before_thread_sync();
+                        empty_umma_barriers[warpgroup_idx]->arrive();
+                    }
+
+                    // Accumulate weighted ReLU in parallel
                     auto sum_0 = make_float2(0, 0);
                     auto sum_1 = make_float2(0, 0);
 
-                    const auto& transform_reg = [&](const uint32_t& j, const float2& sum) {
+                    const auto transform = [&](const uint32_t& j, const float2& sum) {
                         auto a = make_float2(fmaxf(accum[j], 0), fmaxf(accum[j + 1], 0));
                         auto b = make_float2(weights[i][j], weights[i][j + 1]);
                         return __ffma2_rn(a, b, sum);
                     };
 
                     #pragma unroll
-                    for (int j = 0; j < kNumWeightsInReg; j += 4) {
-                        sum_0 = transform_reg(j, sum_0);
-                        sum_1 = transform_reg(j + 2, sum_1);
-                    }
-
-                    const auto& transform_smem = [&](const uint32_t& j, const float2& sum) {
-                        auto a = make_float2(fmaxf(accum[j], 0), fmaxf(accum[j + 1], 0));
-                        auto b = make_float2(ld_shared(smem_weights[q_stage_idx] + i * kNumHeads + j),
-                                             ld_shared(smem_weights[q_stage_idx] + i * kNumHeads + j + 1));
-                        return __ffma2_rn(a, b, sum);
-                    };
-
-                    #pragma unroll
-                    for (int j = kNumWeightsInReg; j < kNumHeads; j += 4) {
-                        sum_0 = transform_smem(j, sum_0);
-                        sum_1 = transform_smem(j + 2, sum_1);
+                    for (uint32_t j = 0; j < kNumHeads; j += 4) {
+                        sum_0 = transform(j, sum_0);
+                        sum_1 = transform(j + 2, sum_1);
                     }
 
                     auto sum = __fadd2_rn(sum_0, sum_1);
-                    float result = scale_kv * (sum.x + sum.y);
+                    auto result = static_cast<logits_dtype_t>(scale_kv * (sum.x + sum.y));
 
                     // Store into the global memory
-                    // NOTES: we have redundant writes here, consider more carefully
-                    const uint32_t& q_idx = block_q_idx * BLOCK_Q + i;
+                    const auto q_offset = (block_q_idx * BLOCK_Q + i) * static_cast<uint64_t>(stride_logits);
                     if constexpr (kIsCompressedLogits) {
-                        if (seq_k_start[i] <= kv_offset + v_offset and kv_offset + v_offset < seq_k_end[i])
-                            logits[q_idx * stride_logits + kv_offset + v_offset - seq_k_start[i]] = result;
+                        if (seq_k_start[i] <= kv_offset and kv_offset < seq_k_end[i])
+                            logits[q_offset + kv_offset - seq_k_start[i]] = result;
                     } else {
-                        logits[q_idx * stride_logits + kv_offset + v_offset] = result;
+                        logits[q_offset + kv_offset] = result;
                     }
+                    __syncwarp();
                 }
             }
             num_total_kv_blocks += num_kv_blocks;
@@ -393,12 +392,12 @@ void sm100_fp8_mqa_logits(const uint32_t seq_len, const uint32_t seq_len_kv,
             // Jump to the next block
             CUTE_TIE(get_next_block_q_idx(), block_q_idx, q_iter_idx);
         }
-    }
 
-    // Free tensor memory
-    __syncthreads();
-    if (is_tma_load_warp)
-        cute::TMEM::Allocator1Sm().free(0, kNumTmemCols);
+        // Free tensor memory
+        cutlass::arch::NamedBarrier(kNumMathThreads, 0).sync();
+        if (warp_idx == 0)
+            cute::TMEM::Allocator1Sm().free(0, kNumTmemCols);
+    }
 }
 
 } // namespace deep_gemm

deep_gemm/include/deep_gemm/impls/sm100_fp8_paged_mqa_logits.cuh

@@ -6,28 +6,30 @@
 #include <cute/arch/cluster_sm90.hpp>
 #include <cute/arch/copy_sm90_desc.hpp>
 
+#include <deep_gemm/common/cute_tie.cuh>
+#include <deep_gemm/common/math.cuh>
+#include <deep_gemm/common/tma_copy.cuh>
 #include <deep_gemm/common/utils.cuh>
-#include <deep_gemm/common/sm90_utils.cuh>
-#include <deep_gemm/common/sm100_utils.cuh>
-
-#include <deep_gemm/impls/sm90_fp8_paged_mqa_logits.cuh>
+#include <deep_gemm/mma/sm100.cuh>
+#include <deep_gemm/ptx/ld_st.cuh>
+#include <deep_gemm/ptx/tcgen05.cuh>
+#include <deep_gemm/ptx/utils.cuh>
+#include <deep_gemm/scheduler/paged_mqa_logits.cuh>
 
 namespace deep_gemm {
 
-using namespace deep_gemm::sm90;
-using namespace deep_gemm::sm100;
-
 template <uint32_t kNextN, uint32_t kNumHeads,
           uint32_t kHeadDim, uint32_t BLOCK_KV,
           bool kIsContextLens2D,
           uint32_t kNumQStages, uint32_t kNumKVStages,
           uint32_t SPLIT_KV,
           uint32_t kNumSpecializedThreads, uint32_t kNumMathThreads,
+          typename logits_dtype_t,
           uint32_t kNumMathWarpGroups = kNumMathThreads / 128>
-__global__ __launch_bounds__(kNumSpecializedThreads + kNumMathThreads, 1)
+CUTLASS_GLOBAL __launch_bounds__(kNumSpecializedThreads + kNumMathThreads, 1)
 void sm100_fp8_paged_mqa_logits(const uint32_t batch_size,
-                                const uint64_t logits_stride, const uint64_t block_table_stride,
-                                const uint32_t* context_lens, float* logits,
+                                const uint32_t logits_stride, const uint32_t block_table_stride,
+                                const uint32_t* context_lens, logits_dtype_t* logits,
                                 const uint32_t* block_table, const uint32_t* schedule_meta,
                                 const __grid_constant__ cute::TmaDescriptor tensor_map_q,
                                 const __grid_constant__ cute::TmaDescriptor tensor_map_kv,
@@ -35,27 +37,33 @@ void sm100_fp8_paged_mqa_logits(const uint32_t batch_size,
                                 const __grid_constant__ cute::TmaDescriptor tensor_map_weights) {
     using Barrier = cutlass::arch::ClusterTransactionBarrier;
 
-    // NOTES: use `__shfl_sync` to encourage NVCC to use unified registers
-    const auto& warp_idx = __shfl_sync(0xffffffff, threadIdx.x / 32, 0);
-    const auto& warpgroup_idx = warp_idx / 4;
-    const auto& lane_idx = get_lane_idx();
+    // Utils
+    const auto sm_idx = blockIdx.x;
+    const auto warp_idx = cutlass::canonical_warp_idx_sync();
+    const auto warpgroup_idx = warp_idx / 4;
+    const auto lane_idx = ptx::get_lane_idx();
+    constexpr uint32_t kSpecWarpStart = kNumMathWarpGroups * 4;
 
     // Prefetch TMA descriptors
     DG_STATIC_ASSERT(kNumSpecializedThreads == 128 and kNumMathThreads % 128 == 0, "Invalid threads");
-    if (warp_idx == kNumMathThreads / 32 and cute::elect_one_sync()) {
+    if (warp_idx == kSpecWarpStart) {
         cute::prefetch_tma_descriptor(&tensor_map_q);
         cute::prefetch_tma_descriptor(&tensor_map_kv);
         cute::prefetch_tma_descriptor(&tensor_map_kv_scales);
         cute::prefetch_tma_descriptor(&tensor_map_weights);
     }
-    __syncwarp();
+
+    // Next-N atom configs
+    static constexpr uint32_t kNextNAtom = (kNextN % 2 == 0) ? 2 : 1;
+    static constexpr uint32_t kNumNextNAtoms = kNextN / kNextNAtom;
+    static constexpr bool kSingleAtom = (kNumNextNAtoms == 1);
 
     // Shared memory configs
     static constexpr uint32_t kSwizzleAlignment = kHeadDim * 8;
-    static constexpr uint32_t SMEM_Q_SIZE_PER_STAGE = kNextN * kNumHeads * kHeadDim * sizeof(__nv_fp8_e4m3);
+    static constexpr uint32_t SMEM_Q_SIZE_PER_STAGE = kNextNAtom * kNumHeads * kHeadDim * sizeof(__nv_fp8_e4m3);
     static constexpr uint32_t SMEM_KV_SIZE_PER_STAGE = SPLIT_KV * kHeadDim * sizeof(__nv_fp8_e4m3);
     static constexpr uint32_t SMEM_KV_SCALE_SIZE_PER_STAGE = SPLIT_KV * sizeof(float);
-    static constexpr uint32_t SMEM_WEIGHT_SIZE_PER_STAGE = kNextN * kNumHeads * sizeof(float);
+    static constexpr uint32_t SMEM_WEIGHT_SIZE_PER_STAGE = kNextNAtom * kNumHeads * sizeof(float);
 
     // Align to swizzling alignment bytes
     extern __shared__ __align__(kSwizzleAlignment) uint8_t smem_buffer[];
@@ -63,43 +71,40 @@ void sm100_fp8_paged_mqa_logits(const uint32_t batch_size,
     DG_STATIC_ASSERT(SMEM_KV_SIZE_PER_STAGE % kSwizzleAlignment == 0, "Unaligned TMA swizzling");
 
     // Q and KV data on shared memory
-    auto smem_q = PatternVisitor([&](const uint32_t& i) {
+    auto smem_q = utils::PatternVisitor([&](const uint32_t& i) {
         return reinterpret_cast<__nv_fp8_e4m3*>(smem_buffer + SMEM_Q_SIZE_PER_STAGE * i);
     });
-    auto smem_kv = PatternVisitor([&](const uint32_t& i) {
+    auto smem_kv = utils::PatternVisitor([&](const uint32_t& i) {
         return reinterpret_cast<__nv_fp8_e4m3*>(smem_buffer + SMEM_Q_SIZE_PER_STAGE * kNumQStages + SMEM_KV_SIZE_PER_STAGE * i);
     });
     constexpr auto smem_offset = SMEM_Q_SIZE_PER_STAGE * kNumQStages + SMEM_KV_SIZE_PER_STAGE * kNumKVStages;
-    auto smem_kv_scales = PatternVisitor([&](const uint32_t& i) {
+    auto smem_kv_scales = utils::PatternVisitor([&](const uint32_t& i) {
         return reinterpret_cast<float*>(smem_buffer + smem_offset + SMEM_KV_SCALE_SIZE_PER_STAGE * i);
     });
-    auto smem_weights = PatternVisitor([&](const uint32_t& i) {
+    auto smem_weights = utils::PatternVisitor([&](const uint32_t& i) {
         return reinterpret_cast<float*>(smem_buffer + smem_offset + SMEM_KV_SCALE_SIZE_PER_STAGE * kNumKVStages + SMEM_WEIGHT_SIZE_PER_STAGE * i);
     });
 
     // Barriers and TMEM pointer on shared memory
     const auto barrier_ptr = reinterpret_cast<Barrier*>(smem_weights[kNumQStages]);
-    auto full_q_barriers     = PatternVisitor([&](const uint32_t& i) { return barrier_ptr + i; });
-    auto empty_q_barriers    = PatternVisitor([&](const uint32_t& i) { return barrier_ptr + kNumQStages + i; });
-    auto full_kv_barriers    = PatternVisitor([&](const uint32_t& i) { return barrier_ptr + kNumQStages * 2 + i; });
-    auto empty_kv_barriers   = PatternVisitor([&](const uint32_t& i) { return barrier_ptr + kNumQStages * 2 + kNumKVStages + i; });
+    auto full_q_barriers     = utils::PatternVisitor([&](const uint32_t& i) { return barrier_ptr + i; });
+    auto empty_q_barriers    = utils::PatternVisitor([&](const uint32_t& i) { return barrier_ptr + kNumQStages + i; });
+    auto full_kv_barriers    = utils::PatternVisitor([&](const uint32_t& i) { return barrier_ptr + kNumQStages * 2 + i; });
+    auto empty_kv_barriers   = utils::PatternVisitor([&](const uint32_t& i) { return barrier_ptr + kNumQStages * 2 + kNumKVStages + i; });
     const auto umma_barrier_ptr = barrier_ptr + kNumQStages * 2 + kNumKVStages * 2;
-    auto full_umma_barriers  = PatternVisitor([&](const uint32_t& i) { return umma_barrier_ptr + i; });
-    auto empty_umma_barriers = PatternVisitor([&](const uint32_t& i) { return umma_barrier_ptr + kNumMathWarpGroups + i; });
+    auto full_umma_barriers  = utils::PatternVisitor([&](const uint32_t& i) { return umma_barrier_ptr + i; });
+    auto empty_umma_barriers = utils::PatternVisitor([&](const uint32_t& i) { return umma_barrier_ptr + kNumMathWarpGroups + i; });
     auto tmem_ptr_in_smem    = reinterpret_cast<uint32_t*>(umma_barrier_ptr + kNumMathWarpGroups * 2);
 
-    constexpr uint32_t kNumTmemCols = kNextN * kNumHeads * kNumMathWarpGroups;
+    constexpr uint32_t kNumTmemCols = kNextNAtom * kNumHeads * kNumMathWarpGroups;
     DG_STATIC_ASSERT(kNumTmemCols <= 512, "Too many tensor memory");
-    const bool& is_math_warp = (warp_idx < kNumMathWarpGroups * 4);
-    const bool& is_tma_load_warp = (warp_idx == kNumMathWarpGroups * 4);
-    const bool& is_umma_warp = (warp_idx == kNumMathWarpGroups * 4 + 1);
 
     // Initialize barriers
-    if (is_tma_load_warp and cute::elect_one_sync()) {
+    if (warp_idx == kSpecWarpStart and cute::elect_one_sync()) {
         #pragma unroll
         for (uint32_t i = 0; i < kNumQStages; ++ i) {
             full_q_barriers[i]->init(1);
-            empty_q_barriers[i]->init(kNumMathThreads);
+            empty_q_barriers[i]->init(kNumMathThreads + 32);
         }
         #pragma unroll
         for (uint32_t i = 0; i < kNumKVStages; ++ i) {
@@ -108,7 +113,7 @@ void sm100_fp8_paged_mqa_logits(const uint32_t batch_size,
         }
         cutlass::arch::fence_barrier_init();
     }
-    if (is_umma_warp) {
+    if (warp_idx == kSpecWarpStart + 1) {
         if (cute::elect_one_sync()) {
             #pragma unroll
             for (uint32_t i = 0; i < kNumMathWarpGroups; ++i) {
@@ -123,66 +128,76 @@ void sm100_fp8_paged_mqa_logits(const uint32_t batch_size,
     __syncthreads();
 
     // Register reconfigurations
-    constexpr uint32_t kNumSpecializedRegisters = 40;
-    constexpr uint32_t kNumMathRegisters = 232;
+    constexpr uint32_t kNumSpecializedRegisters = 56;
+    constexpr uint32_t kNumMathRegisters = 224;
+
+    // Wait for primary kernel completion
+    cudaGridDependencySynchronize();
 
     // Scheduler
     constexpr uint32_t kNumBlocksPerSplit = SPLIT_KV / BLOCK_KV;
-    auto scheduler = PagedMQALogitsScheduler<kNextN, kIsContextLens2D, BLOCK_KV, kNumBlocksPerSplit>(batch_size, blockIdx.x, context_lens, schedule_meta);
+    using Scheduler = sched::PagedMQALogitsScheduler<kNextN, kIsContextLens2D, BLOCK_KV, kNumBlocksPerSplit, kNumNextNAtoms>;
     DG_STATIC_ASSERT(SPLIT_KV == BLOCK_KV * kNumBlocksPerSplit, "Invalid `SPLIT_KV`");
 
     // Q and KV pipeline
-    const auto& get_q_pipeline = [=](const uint32_t& q_iter_idx) -> cute::tuple<uint32_t, uint32_t> {
+    const auto get_q_pipeline = [=](const uint32_t& q_iter_idx) -> cute::tuple<uint32_t, uint32_t> {
         return {q_iter_idx % kNumQStages, (q_iter_idx / kNumQStages) & 1}; // Q pipeline stage and phase
     };
-    const auto& get_kv_pipeline = [=](const uint32_t& kv_iter_idx) -> cute::tuple<uint32_t, uint32_t> {
+    const auto get_kv_pipeline = [=](const uint32_t& kv_iter_idx) -> cute::tuple<uint32_t, uint32_t> {
         return {kv_iter_idx % kNumKVStages, (kv_iter_idx / kNumKVStages) & 1}; // KV pipeline stage and phase
     };
-    uint32_t q_iter_idx = 0, kv_iter_idx = 0;
 
     // UMMA settings
     // Construct instruction with layout D
     constexpr uint32_t UMMA_M = 128;
     constexpr uint32_t UMMA_K = 32 / sizeof(cutlass::float_e4m3_t);
-    constexpr uint32_t UMMA_N = kNextN * kNumHeads;
+    constexpr uint32_t UMMA_N = kNextNAtom * kNumHeads;
     DG_STATIC_ASSERT(SPLIT_KV == UMMA_M * kNumMathWarpGroups, "Invalid `SPLIT_KV`");
 
-    if (is_tma_load_warp) {
-        // TMA warp-group for loading data
+    if (warp_idx == kSpecWarpStart) {
+        // TMA warp for loading data
         cutlass::arch::warpgroup_reg_dealloc<kNumSpecializedRegisters>();
+        auto scheduler = Scheduler(sm_idx, context_lens, schedule_meta);
+        uint32_t q_iter_idx = 0, kv_iter_idx = 0;
 
-        const auto& issue_tma_q = [&](const uint32_t& stage_idx, const uint32_t& q_idx) {
+        const auto issue_tma_q = [&](const uint32_t& stage_idx, const uint32_t& q_atom_idx) {
             if (cute::elect_one_sync()) {
-                tma_copy<kHeadDim, kNextN * kNumHeads, kHeadDim>(&tensor_map_q, full_q_barriers[stage_idx], smem_q[stage_idx], 0, q_idx * kNextN * kNumHeads);
-                tma_copy<kNextN * kNumHeads, 1, 0>(&tensor_map_weights, full_q_barriers[stage_idx], smem_weights[stage_idx], 0, q_idx);
+                tma::copy<kHeadDim, kNextNAtom * kNumHeads, kHeadDim>(&tensor_map_q, full_q_barriers[stage_idx], smem_q[stage_idx], 0, q_atom_idx * kNextNAtom * kNumHeads);
+                tma::copy<kNextNAtom * kNumHeads, 1, 0>(&tensor_map_weights, full_q_barriers[stage_idx], smem_weights[stage_idx], 0, q_atom_idx * kNextNAtom);
                 full_q_barriers[stage_idx]->arrive_and_expect_tx(SMEM_Q_SIZE_PER_STAGE + SMEM_WEIGHT_SIZE_PER_STAGE);
             }
         };
 
-        // Initialize `q_idx` outside `[0, batch_size)` to indicate it was none
-        uint32_t q_idx = batch_size, kv_idx, num_kv;
-        uint32_t next_q_idx, next_kv_idx, next_num_kv;
+        // Initialize outside valid range to indicate no previous task
+        uint32_t q_atom_idx = batch_size * kNumNextNAtoms, kv_idx, num_kv;
+        uint32_t next_q_atom_idx, next_kv_idx, next_num_kv;
         bool fetched_next_task;
 
         // Prefetch the first Q
-        if ((fetched_next_task = scheduler.fetch_next_task(next_q_idx, next_kv_idx, next_num_kv)))
-            issue_tma_q(0, next_q_idx), q_iter_idx = 1;
+        if ((fetched_next_task = scheduler.fetch_next_task(next_q_atom_idx, next_kv_idx, next_num_kv)))
+            issue_tma_q(0, next_q_atom_idx), q_iter_idx = 1;
 
-        int kv_block_idx_ptr = 32;
+        uint32_t kv_block_idx_ptr = 32;
         uint32_t kv_block_idx_storage;
 
         while (fetched_next_task) {
-            // Prefetch next Q when current Q changes
-            bool prefetch_q = (q_idx != next_q_idx and scheduler.exist_q_idx(next_q_idx + 1));
-            q_idx = next_q_idx;
+            // Prefetch next Q when (q, atom) changes
+            bool prefetch_q = (q_atom_idx != next_q_atom_idx) and scheduler.exist_q_atom_idx(next_q_atom_idx + 1);
+
+            if (q_atom_idx != next_q_atom_idx)
+                kv_block_idx_ptr = 32;
+
+            q_atom_idx = next_q_atom_idx;
             kv_idx = next_kv_idx;
             num_kv = next_num_kv;
 
             // Read KV block index
-            // TODO: deal with `-1`?
-            if (kv_idx == 0 or kv_block_idx_ptr == 32) {
+            // TODO(xuzhean): consider -1
+            if (kv_block_idx_ptr == 32) {
                 kv_block_idx_ptr = 0;
-                kv_block_idx_storage = (kv_idx + lane_idx < num_kv ? __ldg(block_table + q_idx * block_table_stride + (kv_idx + lane_idx)) : 0);
+                const auto block_table_offset = (q_atom_idx / kNumNextNAtoms) * static_cast<uint64_t>(block_table_stride);
+                kv_block_idx_storage = (kv_idx + lane_idx < num_kv)
+                    ? block_table[block_table_offset + kv_idx + lane_idx] : 0;
             }
             DG_STATIC_ASSERT(32 % kNumBlocksPerSplit == 0, "Invalid `UMMA_M`");
 
@@ -190,12 +205,12 @@ void sm100_fp8_paged_mqa_logits(const uint32_t batch_size,
             if (prefetch_q) {
                 CUTE_TIE_DECL(get_q_pipeline(q_iter_idx ++), q_stage_idx, q_phase);
                 empty_q_barriers[q_stage_idx]->wait(q_phase ^ 1);
-                issue_tma_q(q_stage_idx, q_idx + 1);
+                issue_tma_q(q_stage_idx, q_atom_idx + 1);
             }
 
-            int kv_block_idx[kNumBlocksPerSplit];
+            uint32_t kv_block_idx[kNumBlocksPerSplit];
             #pragma unroll
-            for (int i = 0; i < kNumBlocksPerSplit; ++ i)
+            for (uint32_t i = 0; i < kNumBlocksPerSplit; ++ i)
                 kv_block_idx[i] = __shfl_sync(0xffffffff, kv_block_idx_storage, kv_block_idx_ptr + i);
             kv_block_idx_ptr += kNumBlocksPerSplit;
 
@@ -205,45 +220,53 @@ void sm100_fp8_paged_mqa_logits(const uint32_t batch_size,
 
             if (cute::elect_one_sync()) {
                 #pragma unroll
-                for (int i = 0; i < kNumBlocksPerSplit; ++ i) {
-                    tma_copy<kHeadDim, BLOCK_KV, 0, __nv_fp8_e4m3, true>(&tensor_map_kv, full_kv_barriers[kv_stage_idx],
-                                                                         smem_kv[kv_stage_idx] + (BLOCK_KV * kHeadDim) * i,
-                                                                         0, 0, 1, kv_block_idx[i]);
-                    tma_copy<BLOCK_KV, 1, 0>(&tensor_map_kv_scales, full_kv_barriers[kv_stage_idx],
-                                             smem_kv_scales[kv_stage_idx] + BLOCK_KV * i,
-                                             0, kv_block_idx[i]);
+                for (uint32_t i = 0; i < kNumBlocksPerSplit; ++ i) {
+                    tma::copy<kHeadDim, BLOCK_KV, 0, __nv_fp8_e4m3, true>(&tensor_map_kv, full_kv_barriers[kv_stage_idx],
+                                                                          smem_kv[kv_stage_idx] + (BLOCK_KV * kHeadDim) * i,
+                                                                          0, 0, 1, kv_block_idx[i]);
+                    tma::copy<BLOCK_KV, 1, 0>(&tensor_map_kv_scales, full_kv_barriers[kv_stage_idx],
+                                              smem_kv_scales[kv_stage_idx] + BLOCK_KV * i,
+                                              0, kv_block_idx[i]);
                 }
                 full_kv_barriers[kv_stage_idx]->arrive_and_expect_tx(SMEM_KV_SIZE_PER_STAGE + SMEM_KV_SCALE_SIZE_PER_STAGE);
             }
 
             // Fetch next task
-            fetched_next_task = scheduler.fetch_next_task(next_q_idx, next_kv_idx, next_num_kv);
+            fetched_next_task = scheduler.fetch_next_task(next_q_atom_idx, next_kv_idx, next_num_kv);
         }
-    } else if (is_umma_warp) {
+    } else if (warp_idx == kSpecWarpStart + 1) {
         cutlass::arch::warpgroup_reg_dealloc<kNumSpecializedRegisters>();
+        auto scheduler = Scheduler(sm_idx, context_lens, schedule_meta);
+        uint32_t q_iter_idx = 0, kv_iter_idx = 0;
 
         // Require full allocation
-        DG_TRAP_ONLY_DEVICE_ASSERT(ld_shared(tmem_ptr_in_smem) == 0);
+        DG_TRAP_ONLY_DEVICE_ASSERT(ptx::ld_shared(tmem_ptr_in_smem) == 0);
 
         // Make UMMA desc
         auto instr_desc = cute::UMMA::make_instr_desc<cutlass::float_e4m3_t, cutlass::float_e4m3_t, float,
                                                       UMMA_M, UMMA_N, cute::UMMA::Major::K, cute::UMMA::Major::K>();
         auto runtime_instr_desc = cute::UMMA::make_runtime_instr_desc(instr_desc);
 
-        uint32_t q_idx = batch_size, kv_idx;
-        uint32_t next_q_idx, next_kv_idx, next_num_kv;
+        uint32_t q_atom_idx = batch_size * kNumNextNAtoms, kv_idx;
+        uint32_t next_q_atom_idx, next_kv_idx, next_num_kv;
         uint32_t q_stage_idx, q_phase;
         uint32_t umma_phase = 1;
 
-        while (scheduler.fetch_next_task(next_q_idx, next_kv_idx, next_num_kv)) {
-            if (q_idx != next_q_idx) {
+        while (scheduler.fetch_next_task(next_q_atom_idx, next_kv_idx, next_num_kv)) {
+            if (q_atom_idx != next_q_atom_idx) {
+                // Release previous Q empty (UMMA warp must participate to prevent
+                // running ahead of math warps in the Q pipeline)
+                if (q_iter_idx > 0)
+                    empty_q_barriers[(q_iter_idx - 1) % kNumQStages]->arrive();
+
                 CUTE_TIE(get_q_pipeline(q_iter_idx ++), q_stage_idx, q_phase);
                 full_q_barriers[q_stage_idx]->wait(q_phase);
             }
 
-            q_idx = next_q_idx;
+            q_atom_idx = next_q_atom_idx;
             kv_idx = next_kv_idx;
 
+            // Wait KV arrival
             CUTE_TIE_DECL(get_kv_pipeline(kv_iter_idx ++), kv_stage_idx, kv_phase);
             full_kv_barriers[kv_stage_idx]->wait(kv_phase);
 
@@ -251,12 +274,12 @@ void sm100_fp8_paged_mqa_logits(const uint32_t batch_size,
             #pragma unroll
             for (uint32_t i = 0; i < kNumMathWarpGroups; ++ i) {
                 empty_umma_barriers[i]->wait(umma_phase);    
-                tcgen05_after_thread_sync();
+                ptx::tcgen05_after_thread_sync();
                 #pragma unroll
                 for (uint32_t k = 0; k < kHeadDim / UMMA_K; ++ k) {
-                    auto a_desc = make_umma_desc<cute::UMMA::Major::K, 0, kHeadDim, kHeadDim>(
+                    auto a_desc = mma::sm100::make_umma_desc<cute::UMMA::Major::K, 0, kHeadDim, kHeadDim>(
                         smem_kv[kv_stage_idx], i * UMMA_M, k * UMMA_K);
-                    auto b_desc = make_umma_desc<cute::UMMA::Major::K, 0, kHeadDim, kHeadDim>(
+                    auto b_desc = mma::sm100::make_umma_desc<cute::UMMA::Major::K, 0, kHeadDim, kHeadDim>(
                         smem_q[q_stage_idx], 0, k * UMMA_K);
                     cute::SM100_MMA_F8F6F4_SS::fma(a_desc, b_desc, i * UMMA_N, k, runtime_instr_desc);
                 }
@@ -264,29 +287,45 @@ void sm100_fp8_paged_mqa_logits(const uint32_t batch_size,
             }
             umma_phase ^= 1;
         }
-    } else if (is_math_warp) {
-        // Math warp-groups for WGMMA
+    } else if (warp_idx == kSpecWarpStart + 2 or warp_idx == kSpecWarpStart + 3) {
+        cutlass::arch::warpgroup_reg_dealloc<kNumSpecializedRegisters>();
+    } else if (warp_idx < kSpecWarpStart) {
+        // Math warpgroups for reduce
         cutlass::arch::warpgroup_reg_alloc<kNumMathRegisters>();
+        auto scheduler = Scheduler(sm_idx, context_lens, schedule_meta);
+        uint32_t q_iter_idx = 0, kv_iter_idx = 0;
 
         // Offsets
-        const auto& tmem_start = __shfl_sync(0xffffffff, warpgroup_idx * UMMA_N, 0);
-        const uint32_t thread_idx = threadIdx.x;
+        const auto math_warpgroup_idx = warpgroup_idx;
+        const auto tmem_start = math_warpgroup_idx * UMMA_N;
+        const auto math_thread_idx = warp_idx * 32 + lane_idx;
 
-        // Weights
-        constexpr uint32_t kNumWeightsInReg = (kNextN == 1 ? kNumHeads : cute::min(48, kNumHeads));
-        float weights[kNextN][kNumWeightsInReg];
-        DG_STATIC_ASSERT(kNumWeightsInReg % 4 == 0, "Invalid number of weights in registers");
+        // Helper lambda for loading tensor memory
+        auto tmem_load = [](auto num_elems_c, const uint32_t& tmem_addr, float* accum) {
+            constexpr int N = decltype(num_elems_c)::value;
+            DG_STATIC_ASSERT(N == 32 or N == 64, "Unsupported TMEM load size");
+            using Loader = cute::conditional_t<N == 32,
+                cute::SM100_TMEM_LOAD_32dp32b32x,
+                cute::SM100_TMEM_LOAD_32dp32b64x>;
+            [&]<size_t... Is>(cute::index_sequence<Is...>) {
+                Loader::copy(tmem_addr, reinterpret_cast<uint32_t*>(accum)[Is]...);
+            }(cute::make_index_sequence<N>{});
+            cutlass::arch::fence_view_async_tmem_load();
+        };
 
-        // Initialize `q_idx` outside `[0, batch_size)` to indicate it was none
-        uint32_t q_idx = batch_size, kv_idx;
-        uint32_t next_q_idx, next_kv_idx, next_num_kv;
+        // Local register buffers
+        float weights[kNextNAtom][kNumHeads];
+
+        // Initialize outside valid range to indicate no previous task
+        uint32_t q_atom_idx = batch_size * kNumNextNAtoms, kv_idx;
+        uint32_t next_q_atom_idx, next_kv_idx, next_num_kv;
         uint32_t q_stage_idx, q_phase;
         uint32_t umma_phase = 0;
 
-        while (scheduler.fetch_next_task(next_q_idx, next_kv_idx, next_num_kv)) {
-            // Current Q changes
-            if (q_idx != next_q_idx) {
-                // Release Last Q empty
+        while (scheduler.fetch_next_task(next_q_atom_idx, next_kv_idx, next_num_kv)) {
+            // Q or atom changes
+            if (q_atom_idx != next_q_atom_idx) {
+                // Release last Q empty
                 if (q_iter_idx > 0)
                     empty_q_barriers[(q_iter_idx - 1) % kNumQStages]->arrive();
 
@@ -296,30 +335,30 @@ void sm100_fp8_paged_mqa_logits(const uint32_t batch_size,
 
                 // Read weights
                 #pragma unroll
-                for (uint32_t i = 0; i < kNextN; ++ i) {
-                    for (uint32_t j = 0; j < kNumWeightsInReg; ++ j)
-                        weights[i][j] = ld_shared(smem_weights[q_stage_idx] + i * kNumHeads + j);
+                for (uint32_t i = 0; i < kNextNAtom; ++ i) {
+                    #pragma unroll
+                    for (uint32_t j = 0; j < kNumHeads; ++ j)
+                        weights[i][j] = ptx::ld_shared(smem_weights[q_stage_idx] + i * kNumHeads + j);
                 }
             }
 
-            // Get current Q and KV index
-            q_idx = next_q_idx;
+            // Get current task indices
+            q_atom_idx = next_q_atom_idx;
             kv_idx = next_kv_idx;
 
             // Calculate KV offset in advance
-            auto kv_offset = q_idx * kNextN * logits_stride + kv_idx * BLOCK_KV;
+            auto kv_offset = q_atom_idx * kNextNAtom * static_cast<uint64_t>(logits_stride) + kv_idx * BLOCK_KV;
 
-            // Compute `[kNextN * kNumHeads, kHeadDim] @ [SPLIT_KV, kHeadDim] -> [kNextN, SPLIT_KV]`
             // Wait TMA KV arrival
             CUTE_TIE_DECL(get_kv_pipeline(kv_iter_idx ++), kv_stage_idx, kv_phase);
             full_kv_barriers[kv_stage_idx]->wait(kv_phase);
 
             // Read per-KV scales
-            float scale_kv = ld_shared(smem_kv_scales[kv_stage_idx] + thread_idx);
+            float scale_kv = ptx::ld_shared(smem_kv_scales[kv_stage_idx] + math_thread_idx);
 
             // Wait UMMA arrival
-            full_umma_barriers[warpgroup_idx]->wait(umma_phase);
-            tcgen05_after_thread_sync();
+            full_umma_barriers[math_warpgroup_idx]->wait(umma_phase);
+            ptx::tcgen05_after_thread_sync();
             umma_phase ^= 1;
 
             // Release KV empty
@@ -327,72 +366,49 @@ void sm100_fp8_paged_mqa_logits(const uint32_t batch_size,
 
             // Reduce over the head dim and store
             DG_STATIC_ASSERT(kNumHeads % 8 == 0, "Invalid head");
-            constexpr uint32_t kNumLDTMElems = kNumHeads * kNextN;
-            uint32_t shifted_accum[kNumLDTMElems];
-            DG_STATIC_ASSERT(kNumLDTMElems == 32 or kNumLDTMElems == 64 or kNumLDTMElems == 128, "Invalid LDTM");
-            auto tmem_load = [&](auto... Is) {
-                if constexpr (kNumLDTMElems == 32) {
-                    cute::SM100_TMEM_LOAD_32dp32b32x::copy(tmem_start, shifted_accum[Is]...);
-                } else if constexpr (kNumLDTMElems == 64) {
-                    cute::SM100_TMEM_LOAD_32dp32b64x::copy(tmem_start, shifted_accum[Is]...);
-                } else if constexpr (kNumLDTMElems == 128) {
-                    cute::SM100_TMEM_LOAD_32dp32b128x::copy(tmem_start, shifted_accum[Is]...);
-                }
-            };
-            [&]<size_t... Is>(cute::index_sequence<Is...>) { tmem_load(Is...); }(cute::make_index_sequence<kNumLDTMElems>{});
-            cutlass::arch::fence_view_async_tmem_load();
-
-            tcgen05_before_thread_sync();
-            empty_umma_barriers[warpgroup_idx]->arrive();
 
             #pragma unroll
-            for (uint32_t i = 0; i < kNextN; ++ i) {
-                auto accum = reinterpret_cast<float*>(shifted_accum + i * kNumHeads);
+            for (uint32_t i = 0; i < kNextNAtom; ++ i) {
+                // Load accumulator from TMEM
+                float accum[kNumHeads];
+                tmem_load(cute::Int<kNumHeads>{}, tmem_start + i * kNumHeads, accum);
 
+                // Release TMEM empty
+                if (i == kNextNAtom - 1) {
+                    ptx::tcgen05_before_thread_sync();
+                    empty_umma_barriers[math_warpgroup_idx]->arrive();
+                }
+
+                // Accumulate weighted ReLU in parallel
                 auto sum_0 = make_float2(0, 0);
                 auto sum_1 = make_float2(0, 0);
 
-                const auto& transform_reg = [&](const uint32_t& j, const float2& sum) {
+                const auto transform = [&](const uint32_t& j, const float2& sum) {
                     auto a = make_float2(fmaxf(accum[j], 0), fmaxf(accum[j + 1], 0));
                     auto b = make_float2(weights[i][j], weights[i][j + 1]);
                     return __ffma2_rn(a, b, sum);
                 };
 
                 #pragma unroll
-                for (int j = 0; j < kNumWeightsInReg; j += 4) {
-                    sum_0 = transform_reg(j, sum_0);
-                    sum_1 = transform_reg(j + 2, sum_1);
-                }
-
-                const auto& transform_smem = [&](const uint32_t& j, const float2& sum) {
-                    auto a = make_float2(fmaxf(accum[j], 0), fmaxf(accum[j + 1], 0));
-                    auto b = make_float2(ld_shared(smem_weights[q_stage_idx] + i * kNumHeads + j),
-                                         ld_shared(smem_weights[q_stage_idx] + i * kNumHeads + j + 1));
-                    return __ffma2_rn(a, b, sum);
-                };
-
-                #pragma unroll
-                for (int j = kNumWeightsInReg; j < kNumHeads; j += 4) {
-                    sum_0 = transform_smem(j, sum_0);
-                    sum_1 = transform_smem(j + 2, sum_1);
+                for (uint32_t j = 0; j < kNumHeads; j += 4) {
+                    sum_0 = transform(j, sum_0);
+                    sum_1 = transform(j + 2, sum_1);
                 }
 
                 auto sum = __fadd2_rn(sum_0, sum_1);
-                float result = scale_kv * (sum.x + sum.y);
+                auto result = static_cast<logits_dtype_t>(scale_kv * (sum.x + sum.y));
 
                 // Store into the global memory
-                // NOTES: we have redundant writes here, consider more carefully
-                logits[kv_offset + i * logits_stride + thread_idx] = result;
+                logits[kv_offset + i * static_cast<uint64_t>(logits_stride) + math_thread_idx] = result;
+                __syncwarp();
             }
         }
-    } else {
-        cutlass::arch::warpgroup_reg_dealloc<kNumSpecializedRegisters>();
-    }
 
-    // Free tensor memory
-    __syncthreads();
-    if (is_umma_warp)
-        cute::TMEM::Allocator1Sm().free(0, kNumTmemCols);
+        // Free tensor memory
+        cutlass::arch::NamedBarrier(kNumMathThreads, 0).sync();
+        if (warp_idx == 0)
+            cute::TMEM::Allocator1Sm().free(0, kNumTmemCols);
+    }
 }
 
 } // namespace deep_gemm

deep_gemm/include/deep_gemm/impls/sm100_tf32_hc_prenorm_gemm.cuh

@@ -4,20 +4,22 @@
 
 #include <cutlass/arch/barrier.h>
 
-#include <deep_gemm/common/reduction.cuh>
+#include <deep_gemm/common/cute_tie.cuh>
+#include <deep_gemm/common/math.cuh>
+#include <deep_gemm/common/tma_copy.cuh>
 #include <deep_gemm/common/utils.cuh>
-#include <deep_gemm/common/sm90_utils.cuh>
-#include <deep_gemm/common/sm100_utils.cuh>
+#include <deep_gemm/mma/sm100.cuh>
+#include <deep_gemm/ptx/ld_st.cuh>
+#include <deep_gemm/ptx/tcgen05.cuh>
+#include <deep_gemm/ptx/utils.cuh>
 
 namespace deep_gemm {
 
-using namespace deep_gemm::sm100;
-
 template <uint32_t kSwizzleMode, uint32_t kSwizzleBase = 16>
-__device__ __forceinline__
+CUTLASS_DEVICE
 uint32_t get_swizzled_smem_offset(const uint32_t& offset, const uint32_t& lane_idx) {
     // Calculate the index of the bank group to be written in the atom
-    const auto& bank_group_idx = offset + lane_idx * (kSwizzleMode / kSwizzleBase);
+    const auto bank_group_idx = offset + lane_idx * (kSwizzleMode / kSwizzleBase);
 
     // Reshape the atom in another view and swizzle
     //  - original: `(BLOCK_N, kSwizzleMode / kSwizzleBase)`
@@ -37,7 +39,7 @@ template <uint32_t SHAPE_N, uint32_t SHAPE_K,
           uint32_t kSwizzleCDMode,
           uint32_t kNumStages,
           uint32_t kNumMMAThreads, uint32_t kNumCastAndReduceThreads>
-__global__ void __launch_bounds__(kNumMMAThreads + kNumCastAndReduceThreads, 1)
+CUTLASS_GLOBAL void __launch_bounds__(kNumMMAThreads + kNumCastAndReduceThreads, 1)
 sm100_tf32_hc_prenorm_gemm_impl(const uint32_t shape_m,
                                 const __grid_constant__ cute::TmaDescriptor tensor_map_a,
                                 const __grid_constant__ cute::TmaDescriptor tensor_map_b,
@@ -58,7 +60,7 @@ sm100_tf32_hc_prenorm_gemm_impl(const uint32_t shape_m,
 
     // Utils
     const auto warp_idx = cutlass::canonical_warp_idx_sync();
-    const auto lane_idx = get_lane_idx();
+    const auto lane_idx = ptx::get_lane_idx();
 
     // Align to 1024 bytes for swizzle-128B
     extern __shared__ __align__(1024) uint8_t smem_buffer[];
@@ -70,7 +72,7 @@ sm100_tf32_hc_prenorm_gemm_impl(const uint32_t shape_m,
     DG_STATIC_ASSERT(SMEM_CD_SIZE % 1024 == 0, "Shared memory of A/B must be aligned to 1024 bytes");
 
     // Real tensor memory size and offsets
-    constexpr uint32_t kNumTmemCols = get_num_aligned_tmem_cols<BLOCK_K * kNumCastStages + BLOCK_N>();
+    constexpr uint32_t kNumTmemCols = utils::get_num_aligned_tmem_cols<BLOCK_K * kNumCastStages + BLOCK_N>();
 
     // Prefetch TMA descriptors at the very beginning
     if (warp_idx == 0 and cute::elect_one_sync()) {
@@ -82,20 +84,20 @@ sm100_tf32_hc_prenorm_gemm_impl(const uint32_t shape_m,
     // Data on shared memory (layout as ordered below)
     // Fill D/A/B pointers
     auto smem_cd = reinterpret_cast<float*>(smem_buffer);
-    auto smem_a = PatternVisitor([&](const uint32_t& i) {
+    auto smem_a = utils::PatternVisitor([&](const uint32_t& i) {
         return reinterpret_cast<nv_bfloat16*>(smem_buffer + (SMEM_CD_SIZE + i * SMEM_A_SIZE_PER_STAGE));
     });
-    auto smem_b = PatternVisitor([&](const uint32_t& i) {
+    auto smem_b = utils::PatternVisitor([&](const uint32_t& i) {
         return reinterpret_cast<float*>(smem_buffer + (SMEM_CD_SIZE + kNumStages * SMEM_A_SIZE_PER_STAGE + i * SMEM_B_SIZE_PER_STAGE));
     });
 
     // Fill barriers
     auto barrier_start_ptr = reinterpret_cast<Barrier*>(smem_buffer + SMEM_CD_SIZE +
         kNumStages * (SMEM_A_SIZE_PER_STAGE + SMEM_B_SIZE_PER_STAGE));
-    auto full_barriers           = PatternVisitor([=](const uint32_t& i) { return barrier_start_ptr + (i); });
-    auto full_cast_barriers      = PatternVisitor([=](const uint32_t& i) { return barrier_start_ptr + (kNumStages + i); });
-    auto empty_barriers          = PatternVisitor([=](const uint32_t& i) { return barrier_start_ptr + (kNumStages * 2 + i); });
-    auto empty_cast_barriers     = PatternVisitor([=](const uint32_t& i) { return barrier_start_ptr + (kNumStages * 3 + i); });
+    auto full_barriers           = utils::PatternVisitor([=](const uint32_t& i) { return barrier_start_ptr + (i); });
+    auto full_cast_barriers      = utils::PatternVisitor([=](const uint32_t& i) { return barrier_start_ptr + (kNumStages + i); });
+    auto empty_barriers          = utils::PatternVisitor([=](const uint32_t& i) { return barrier_start_ptr + (kNumStages * 2 + i); });
+    auto empty_cast_barriers     = utils::PatternVisitor([=](const uint32_t& i) { return barrier_start_ptr + (kNumStages * 3 + i); });
     auto tmem_full_barrier       = barrier_start_ptr + kNumStages * 4;
 
     // Fill the tensor memory pointer
@@ -121,7 +123,7 @@ sm100_tf32_hc_prenorm_gemm_impl(const uint32_t shape_m,
     }
     __syncthreads();
 
-    constexpr uint32_t kNumKBlocks = constexpr_ceil_div(SHAPE_K, BLOCK_K);
+    constexpr uint32_t kNumKBlocks = math::constexpr_ceil_div(SHAPE_K, BLOCK_K);
     constexpr uint32_t kNumKBlocksPerSplit = kNumKBlocks / kNumSplits;
     constexpr uint32_t kRemainKBlocks = kNumKBlocks % kNumSplits;
     const uint32_t block_idx = __shfl_sync(0xffffffff, blockIdx.x, 0);
@@ -131,6 +133,9 @@ sm100_tf32_hc_prenorm_gemm_impl(const uint32_t shape_m,
     const uint32_t m_offset = shape_m * k_split_idx;
     const uint32_t num_total_stages = kNumKBlocksPerSplit + (k_split_idx < kRemainKBlocks);
 
+    // Wait for primary kernel completion
+    cudaGridDependencySynchronize();
+
     // Dispatch warps into different roles
     if (warp_idx < kNumMMAThreads / 32) {
         // TMA load warp
@@ -145,8 +150,8 @@ sm100_tf32_hc_prenorm_gemm_impl(const uint32_t shape_m,
                 uint32_t k_idx = k_offset + s * BLOCK_K;
 
                 // Issue TMAs
-                tma_copy<BLOCK_K, BLOCK_M, kSwizzleAMode>(&tensor_map_a, full_barriers[stage_idx], smem_a[stage_idx], k_idx, m_idx);
-                tma_copy<BLOCK_K, BLOCK_N, kSwizzleBMode>(&tensor_map_b, full_barriers[stage_idx], smem_b[stage_idx], k_idx, 0);
+                tma::copy<BLOCK_K, BLOCK_M, kSwizzleAMode>(&tensor_map_a, full_barriers[stage_idx], smem_a[stage_idx], k_idx, m_idx);
+                tma::copy<BLOCK_K, BLOCK_N, kSwizzleBMode>(&tensor_map_b, full_barriers[stage_idx], smem_b[stage_idx], k_idx, 0);
 
                 // Arrive at full barriers
                 constexpr uint32_t kNumArrivalBytes = SMEM_A_SIZE_PER_STAGE + SMEM_B_SIZE_PER_STAGE;
@@ -168,7 +173,7 @@ sm100_tf32_hc_prenorm_gemm_impl(const uint32_t shape_m,
             const auto& runtime_instr_desc = cute::UMMA::make_runtime_instr_desc(instr_desc);
 
             DG_STATIC_ASSERT(kNumStages <= 32, "Too many stages");
-            auto b_desc = make_umma_desc<kMajorB, BLOCK_N, BLOCK_SWIZZLED_BK, kSwizzleBMode>(smem_b[0], 0, 0);
+            auto b_desc = mma::sm100::make_umma_desc<kMajorB, BLOCK_N, BLOCK_SWIZZLED_BK, kSwizzleBMode>(smem_b[0], 0, 0);
             const uint32_t& b_desc_lo = lane_idx < kNumStages ? b_desc.lo + lane_idx * SMEM_B_SIZE_PER_STAGE / 16 : 0u;
 
             // Checks for MMA instructions
@@ -185,7 +190,7 @@ sm100_tf32_hc_prenorm_gemm_impl(const uint32_t shape_m,
                 const auto& stage_idx = s % kNumStages;
                 const auto& cast_stage_idx = s % kNumCastStages;
                 full_cast_barriers[cast_stage_idx]->wait((s / kNumCastStages) & 1);
-                tcgen05_after_thread_sync();
+                ptx::tcgen05_after_thread_sync();
 
                 // Issue UMMA
                 const auto& b_desc_base_lo = __shfl_sync(0xffffffff, b_desc_lo, static_cast<int>(stage_idx));
@@ -194,7 +199,7 @@ sm100_tf32_hc_prenorm_gemm_impl(const uint32_t shape_m,
                     const uint32_t& atom_idx = (k * UMMA_K) / BLOCK_SWIZZLED_BK;
                     const uint32_t& in_atom_idx = (k * UMMA_K) % BLOCK_SWIZZLED_BK;
                     const uint32_t& offset = atom_idx * BLOCK_N * BLOCK_SWIZZLED_BK;
-                    b_desc.lo = advance_umma_desc_lo<kMajorB, BLOCK_N, kSwizzleBMode, float>(b_desc_base_lo, offset, in_atom_idx);
+                    b_desc.lo = mma::sm100::advance_umma_desc_lo<kMajorB, BLOCK_N, kSwizzleBMode, float>(b_desc_base_lo, offset, in_atom_idx);
                     umma_t::fma(BLOCK_K * cast_stage_idx + k * UMMA_K, b_desc, BLOCK_K * kNumCastStages, s > 0 or k > 0, runtime_instr_desc);
                 }
 
@@ -218,7 +223,7 @@ sm100_tf32_hc_prenorm_gemm_impl(const uint32_t shape_m,
 
         // Wait UMMA arrival
         tmem_full_barrier->wait(0);
-        tcgen05_after_thread_sync();
+        ptx::tcgen05_after_thread_sync();
 
         // Load from tensor memory into registers, and write shared memory with STSM
         DG_STATIC_ASSERT(kNumMMAThreads == 128, "Epilogue threads not enough");
@@ -239,7 +244,7 @@ sm100_tf32_hc_prenorm_gemm_impl(const uint32_t shape_m,
                 values[0], values[1], values[2], values[3]);
             cutlass::arch::fence_view_async_tmem_load();
             if (BLOCK_M == 128 or (BLOCK_M == 64 and lane_idx < 16))
-                st_shared(smem_ptr, values[0], values[1], values[2], values[3]);
+                ptx::st_shared(smem_ptr, values[0], values[1], values[2], values[3]);
             if constexpr (BLOCK_M == 64)
                 __syncwarp();
         }
@@ -290,9 +295,9 @@ sm100_tf32_hc_prenorm_gemm_impl(const uint32_t shape_m,
             #pragma unroll
             for (uint32_t i = 0; i < kNumLoads; i += 2) {
                 auto smem_ptr = smem_base_ptr + get_swizzled_smem_offset<kSwizzleAMode>(i + lane_idx / 16, lane_idx % 16);
-                sm90::SM90_U32x4_LDSM_N::copy(uint32_values[0][i + 0], uint32_values[1][i + 0],
-                                              uint32_values[0][i + 1], uint32_values[1][i + 1],
-                                              smem_ptr);
+                ptx::SM90_U32x4_LDSM_N::copy(uint32_values[0][i + 0], uint32_values[1][i + 0],
+                                             uint32_values[0][i + 1], uint32_values[1][i + 1],
+                                             smem_ptr);
             }
 
             // Wait tensor memory empty
@@ -321,15 +326,15 @@ sm100_tf32_hc_prenorm_gemm_impl(const uint32_t shape_m,
             cutlass::arch::fence_view_async_tmem_store();
 
             // Arrive for issuing MMAs
-            tcgen05_before_thread_sync();
+            ptx::tcgen05_before_thread_sync();
             full_cast_barriers[cast_stage_idx]->arrive();
         }
 
         // Intra-warp reduction and write back
         #pragma unroll
         for (uint32_t u = 0; u < 2; ++ u) {
-            const auto& reduced_sum = warp_reduce_sum<4>(sum[u].x + sum[u].y);
-            const auto& m_idx = m_block_idx * BLOCK_M + sub_warp_idx * BLOCK_M_PER_WARP + lane_idx / 4 + u * 8;
+            const auto reduced_sum = math::warp_reduce_sum<4>(sum[u].x + sum[u].y);
+            const auto m_idx = m_block_idx * BLOCK_M + sub_warp_idx * BLOCK_M_PER_WARP + lane_idx / 4 + u * 8;
             if (lane_idx % 4 == 0 and m_idx < shape_m)
                 sqr_sum[m_offset + m_idx] = reduced_sum;
         }

deep_gemm/include/deep_gemm/impls/sm90_bf16_gemm.cuh

@@ -11,14 +11,19 @@
 #include <cute/arch/copy_sm90_tma.hpp>
 #include <cute/arch/mma_sm100_desc.hpp>
 
+#include <deep_gemm/common/math.cuh>
 #include <deep_gemm/common/utils.cuh>
-#include <deep_gemm/common/scheduler.cuh>
-#include <deep_gemm/common/sm90_utils.cuh>
+#include <deep_gemm/common/tma_copy.cuh>
+#include <deep_gemm/common/types.cuh>
+#include <deep_gemm/mma/sm90.cuh>
+#include <deep_gemm/epilogue/transform.cuh>
+#include <deep_gemm/ptx/ld_st.cuh>
+#include <deep_gemm/ptx/utils.cuh>
+#include <deep_gemm/ptx/wgmma.cuh>
+#include <deep_gemm/scheduler/gemm.cuh>
 
 namespace deep_gemm {
 
-using namespace deep_gemm::sm90;
-
 template <cute::UMMA::Major kMajorA, cute::UMMA::Major kMajorB,
           uint32_t SHAPE_M, uint32_t SHAPE_N, uint32_t SHAPE_K,
           uint32_t kNumGroups,
@@ -30,7 +35,7 @@ template <cute::UMMA::Major kMajorA, cute::UMMA::Major kMajorB,
           uint32_t kNumSMs,
           GemmType kGemmType, bool kWithAccumulation,
           typename cd_dtype_t>
-__global__ __launch_bounds__(kNumTMAThreads + kNumMathThreads, 1) void
+CUTLASS_GLOBAL __launch_bounds__(kNumTMAThreads + kNumMathThreads, 1) void
 sm90_bf16_gemm_impl(int* grouped_layout,
                     uint32_t shape_m, uint32_t shape_n, uint32_t shape_k,
                     const __grid_constant__ cute::TmaDescriptor tensor_map_a,
@@ -51,7 +56,7 @@ sm90_bf16_gemm_impl(int* grouped_layout,
     constexpr uint32_t kNumStages = kNumStages_ / kNumStagesPerMerge;
 
     // Types
-    using WGMMA = typename BF16MMASelector<BLOCK_N, kMajorA, kMajorB>::type;
+    using WGMMA = typename mma::sm90::BF16MMASelector<BLOCK_N, kMajorA, kMajorB>::type;
     using Barrier = cutlass::arch::ClusterTransactionBarrier;
     DG_STATIC_ASSERT(BLOCK_M % WGMMA::M == 0 or BLOCK_M < WGMMA::M, "Invalid block size");
 
@@ -61,7 +66,7 @@ sm90_bf16_gemm_impl(int* grouped_layout,
     shape_k = SHAPE_K != 0 ? SHAPE_K : shape_k;
 
     // Shared memory
-    static constexpr uint32_t SMEM_D_SIZE = constexpr_align(BLOCK_M * BLOCK_N * static_cast<uint32_t>(sizeof(cd_dtype_t)), 1024u);
+    static constexpr uint32_t SMEM_D_SIZE = math::constexpr_align(BLOCK_M * BLOCK_N * static_cast<uint32_t>(sizeof(cd_dtype_t)), 1024u);
     static constexpr uint32_t SMEM_A_SIZE_PER_STAGE = BLOCK_M * BLOCK_K * sizeof(__nv_bfloat16);
     static constexpr uint32_t SMEM_B_SIZE_PER_STAGE = BLOCK_N * BLOCK_K * sizeof(__nv_bfloat16);
 
@@ -71,7 +76,7 @@ sm90_bf16_gemm_impl(int* grouped_layout,
 
     // Configs
     const uint32_t warp_idx = __shfl_sync(0xffffffff, threadIdx.x / 32, 0);
-    const uint32_t lane_idx = get_lane_idx();
+    const uint32_t lane_idx = ptx::get_lane_idx();
 
     // Prefetch TMA descriptors at the very beginning
     if (warp_idx == kNumMathThreads / 32 and cute::elect_one_sync()) {
@@ -88,17 +93,17 @@ sm90_bf16_gemm_impl(int* grouped_layout,
 
     // D/A/B shared memory
     auto smem_d = reinterpret_cast<cd_dtype_t*>(smem_buffer);
-    auto smem_a = PatternVisitor([&](const uint32_t& i) {
+    auto smem_a = utils::PatternVisitor([&](const uint32_t& i) {
         return reinterpret_cast<cutlass::bfloat16_t*>(smem_buffer + SMEM_D_SIZE + i * SMEM_A_SIZE_PER_STAGE);
     });
-    auto smem_b = PatternVisitor([&](const uint32_t& i) {
+    auto smem_b = utils::PatternVisitor([&](const uint32_t& i) {
         return reinterpret_cast<cutlass::bfloat16_t*>(smem_buffer + SMEM_D_SIZE + kNumStages * SMEM_A_SIZE_PER_STAGE + i * SMEM_B_SIZE_PER_STAGE);
     });
 
     // Fill barriers
     auto barrier_start_ptr = reinterpret_cast<Barrier*>(smem_buffer + SMEM_D_SIZE + kNumStages * (SMEM_A_SIZE_PER_STAGE + SMEM_B_SIZE_PER_STAGE));
-    auto full_barriers  = PatternVisitor([=](const uint32_t& i) { return barrier_start_ptr + (i); });
-    auto empty_barriers = PatternVisitor([=](const uint32_t& i) { return barrier_start_ptr + (kNumStages + i); });
+    auto full_barriers  = utils::PatternVisitor([=](const uint32_t& i) { return barrier_start_ptr + (i); });
+    auto empty_barriers = utils::PatternVisitor([=](const uint32_t& i) { return barrier_start_ptr + (kNumStages + i); });
 
     // Initialize barriers
     if (warp_idx == kNumMathThreads / 32 + 1 and cute::elect_one_sync()) {
@@ -119,9 +124,12 @@ sm90_bf16_gemm_impl(int* grouped_layout,
     constexpr uint32_t kNumTMARegisters = 48;
     constexpr uint32_t kNumMathRegisters = kNumMathThreads == 128 ? 248 : 224;
 
+    // Wait for primary kernel completion
+    cudaGridDependencySynchronize();
+
     // Block scheduler
     uint32_t m_block_idx, n_block_idx;
-    auto scheduler = Scheduler<kGemmType, BLOCK_M, BLOCK_N, kNumGroups, kNumTMAMulticast, kIsTMAMulticastOnA, kNumSMs>(shape_m, shape_n, shape_k, grouped_layout);
+    auto scheduler = sched::Scheduler<kGemmType, BLOCK_M, BLOCK_N, kNumGroups, kNumTMAMulticast, kIsTMAMulticastOnA, kNumSMs>(shape_m, shape_n, shape_k, grouped_layout);
 
     // Pipeline and TMA phases
     uint32_t stage_idx = 0, phase = 0;
@@ -151,7 +159,7 @@ sm90_bf16_gemm_impl(int* grouped_layout,
                 const uint32_t num_tma_multicast_b = (not kIsTMAMulticastOnA and is_tma_multicast_valid) ? kNumTMAMulticast : 1;
                 DG_STATIC_ASSERT(kNumTMAMulticast <= 2, "Scheduler does not support > 2 TMA multicast");
 
-                const auto& num_total_k_blocks = ceil_div(scheduler.current_shape_k, BLOCK_K);
+                const auto num_total_k_blocks = math::ceil_div(scheduler.current_shape_k, BLOCK_K);
                 for (uint32_t k_block_idx = 0; k_block_idx < num_total_k_blocks; advance_pipeline(k_block_idx)) {
                     // Wait consumer release
                     empty_barriers[stage_idx]->wait(phase ^ 1);
@@ -159,31 +167,30 @@ sm90_bf16_gemm_impl(int* grouped_layout,
                     constexpr bool kWithGroupOffsetA = kGemmType == GemmType::MGroupedMasked;
                     auto& full_barrier = *full_barriers[stage_idx];
 
-                    const auto m_idx = scheduler.template get_global_idx<kWithGroupOffsetA, IndexType::MN>(shape_m, BLOCK_M, m_block_idx);
-                    const auto n_idx = scheduler.template get_global_idx<(kMajorB == cute::UMMA::Major::K), IndexType::MN>(shape_n, BLOCK_N, n_block_idx, m_block_idx);
+                    const auto m_idx = scheduler.template get_global_idx<kWithGroupOffsetA, sched::IndexType::MN>(shape_m, BLOCK_M, m_block_idx);
+                    const auto n_idx = scheduler.template get_global_idx<(kMajorB == cute::UMMA::Major::K), sched::IndexType::MN>(shape_n, BLOCK_N, n_block_idx, m_block_idx);
 
                     DG_STATIC_ASSERT(kGemmType == GemmType::Normal or kGemmType == GemmType::KGroupedContiguous or kMajorA == cute::UMMA::Major::K, "Invalid major");
-                    uint32_t k_a_idx = scheduler.template get_global_idx<(kMajorA == cute::UMMA::Major::MN), IndexType::K> (
+                    uint32_t k_a_idx = scheduler.template get_global_idx<(kMajorA == cute::UMMA::Major::MN), sched::IndexType::K> (
                         shape_k, BLOCK_K, k_block_idx, m_block_idx);
-                    uint32_t k_b_idx = scheduler.template get_global_idx<(kMajorB == cute::UMMA::Major::MN), IndexType::K> (
+                    uint32_t k_b_idx = scheduler.template get_global_idx<(kMajorB == cute::UMMA::Major::MN), sched::IndexType::K> (
                         shape_k, BLOCK_K, k_block_idx, m_block_idx);
 
                     // Issue TMAs
                     constexpr bool kIsBatchedMM = (kGemmType == GemmType::Batched);
                     const uint32_t batch_idx = (kIsBatchedMM ? scheduler.current_group_idx : 0);
                     if constexpr (kMajorA == cute::UMMA::Major::K)
-                        tma_copy<BLOCK_K, BLOCK_M, kSwizzleAMode, cutlass::bfloat16_t, kIsBatchedMM>(
+                        tma::copy<BLOCK_K, BLOCK_M, kSwizzleAMode, cutlass::bfloat16_t, kIsBatchedMM>(
                             &tensor_map_a, &full_barrier, smem_a[stage_idx], k_a_idx, m_idx, num_tma_multicast_a, batch_idx);
                     if constexpr (kMajorA == cute::UMMA::Major::MN)
-                        tma_copy<BLOCK_M, BLOCK_K, kSwizzleAMode, cutlass::bfloat16_t, kIsBatchedMM>(
+                        tma::copy<BLOCK_M, BLOCK_K, kSwizzleAMode, cutlass::bfloat16_t, kIsBatchedMM>(
                             &tensor_map_a, &full_barrier, smem_a[stage_idx], m_idx, k_a_idx, num_tma_multicast_a, batch_idx);
                     if constexpr (kMajorB == cute::UMMA::Major::K)
-                        tma_copy<BLOCK_K, BLOCK_N, kSwizzleBMode, cutlass::bfloat16_t, kIsBatchedMM>(
+                        tma::copy<BLOCK_K, BLOCK_N, kSwizzleBMode, cutlass::bfloat16_t, kIsBatchedMM>(
                             &tensor_map_b, &full_barrier, smem_b[stage_idx], k_b_idx, n_idx, num_tma_multicast_b, batch_idx);
                     if constexpr (kMajorB == cute::UMMA::Major::MN)
-                        tma_copy<BLOCK_N, BLOCK_K, kSwizzleBMode, cutlass::bfloat16_t, kIsBatchedMM>(
+                        tma::copy<BLOCK_N, BLOCK_K, kSwizzleBMode, cutlass::bfloat16_t, kIsBatchedMM>(
                             &tensor_map_b, &full_barrier, smem_b[stage_idx], n_idx, k_b_idx, num_tma_multicast_b, batch_idx);
-
                     full_barrier.arrive_and_expect_tx(SMEM_A_SIZE_PER_STAGE + SMEM_B_SIZE_PER_STAGE);
                 }
             }
@@ -203,8 +210,8 @@ sm90_bf16_gemm_impl(int* grouped_layout,
         
         // Merged stages only happens in NT normal GEMM cases
         constexpr uint32_t BLOCK_ATOM_K = BLOCK_K / kNumStagesPerMerge;
-        auto a_desc = make_gmma_desc<kMajorA, BLOCK_M, BLOCK_ATOM_K, kSwizzleAMode>(smem_a[0], math_wg_idx * WGMMA::M, 0);
-        auto b_desc = make_gmma_desc<kMajorB, BLOCK_N, BLOCK_ATOM_K, kSwizzleBMode>(smem_b[0], 0, 0);
+        auto a_desc = mma::sm90::make_gmma_desc<kMajorA, BLOCK_M, BLOCK_ATOM_K, kSwizzleAMode>(smem_a[0], math_wg_idx * WGMMA::M, 0);
+        auto b_desc = mma::sm90::make_gmma_desc<kMajorB, BLOCK_N, BLOCK_ATOM_K, kSwizzleBMode>(smem_b[0], 0, 0);
         const uint32_t a_desc_lo = __shfl_sync(0xffffffff, a_desc.reg32_[0], 0);
         const uint32_t b_desc_lo = __shfl_sync(0xffffffff, b_desc.reg32_[0], 0);
 
@@ -229,10 +236,10 @@ sm90_bf16_gemm_impl(int* grouped_layout,
             };
 
             // TODO: remove some useless computation for unaligned Ms
-            const auto& num_total_k_blocks = ceil_div(scheduler.current_shape_k, BLOCK_K);
+            const auto num_total_k_blocks = math::ceil_div(scheduler.current_shape_k, BLOCK_K);
             for (uint32_t k_block_idx = 0; k_block_idx < num_total_k_blocks; advance_pipeline(k_block_idx)) {
-                const auto& a_desc_base_lo = a_desc_lo + stage_idx * (SMEM_A_SIZE_PER_STAGE / 16);
-                const auto& b_desc_base_lo = b_desc_lo + stage_idx * (SMEM_B_SIZE_PER_STAGE / 16);
+                const auto a_desc_base_lo = a_desc_lo + stage_idx * (SMEM_A_SIZE_PER_STAGE / 16);
+                const auto b_desc_base_lo = b_desc_lo + stage_idx * (SMEM_B_SIZE_PER_STAGE / 16);
 
                 // Wait TMA arrivals
                 full_barriers[stage_idx]->wait(phase);
@@ -240,26 +247,26 @@ sm90_bf16_gemm_impl(int* grouped_layout,
                 // Commit WGMMA instructions
                 #pragma unroll
                 for (uint32_t i = 0; i < WGMMA::kNumAccum * (BLOCK_M / WAVE_BLOCK_M); ++ i)
-                    warpgroup_fence_operand(accum[i]);
-                warpgroup_arrive();
+                    ptx::warpgroup_fence_operand(accum[i]);
+                ptx::warpgroup_arrive();
                 #pragma unroll
                 for (uint32_t local_idx = 0; local_idx < BLOCK_M / WAVE_BLOCK_M; ++ local_idx) {
                     auto shifted_accum = accum + WGMMA::kNumAccum * local_idx;
                     #pragma unroll
                     for (uint32_t k = 0; k < BLOCK_K / WGMMA::K; ++ k) {
-                        const uint32_t& atom_k_idx = k * WGMMA::K / BLOCK_ATOM_K;
-                        a_desc.reg32_[0] = advance_gmma_desc_lo<kMajorA, BLOCK_M, BLOCK_ATOM_K, kSwizzleAMode, nv_bfloat16>(
+                        const uint32_t atom_k_idx = k * WGMMA::K / BLOCK_ATOM_K;
+                        a_desc.reg32_[0] = mma::sm90::advance_gmma_desc_lo<kMajorA, BLOCK_M, BLOCK_ATOM_K, kSwizzleAMode, nv_bfloat16>(
                             a_desc_base_lo, local_idx * WAVE_BLOCK_M, (k * WGMMA::K) % BLOCK_ATOM_K, atom_k_idx * BLOCK_M * BLOCK_ATOM_K);
-                        b_desc.reg32_[0] = advance_gmma_desc_lo<kMajorB, BLOCK_N, BLOCK_ATOM_K, kSwizzleBMode, nv_bfloat16>(
+                        b_desc.reg32_[0] = mma::sm90::advance_gmma_desc_lo<kMajorB, BLOCK_N, BLOCK_ATOM_K, kSwizzleBMode, nv_bfloat16>(
                             b_desc_base_lo, 0, (k * WGMMA::K) % BLOCK_ATOM_K, atom_k_idx * BLOCK_N * BLOCK_ATOM_K);
                         WGMMA::wgmma(a_desc, b_desc, shifted_accum, 1);
                     }
                 }
-                warpgroup_commit_batch();
+                ptx::warpgroup_commit_batch();
                 #pragma unroll
                 for (uint32_t i = 0; i < WGMMA::kNumAccum * (BLOCK_M / WAVE_BLOCK_M); ++ i)
-                    warpgroup_fence_operand(accum[i]);
-                warpgroup_wait<0>();
+                    ptx::warpgroup_fence_operand(accum[i]);
+                ptx::warpgroup_wait<0>();
 
                 // Notify barrier arrival
                 empty_barrier_arrive(stage_idx);
@@ -324,7 +331,7 @@ sm90_bf16_gemm_impl(int* grouped_layout,
                         }
 
                         // NOTES: only 16 lanes' addresses are used
-                        SM90_U32x2_STSM_N<nv_bfloat162>::copy(
+                        ptx::SM90_U32x2_STSM_N<nv_bfloat162>::copy(
                             __float22bfloat162_rn({shifted_accum[i * 4 + 0], shifted_accum[i * 4 + 1]}),
                             __float22bfloat162_rn({shifted_accum[i * 4 + 2], shifted_accum[i * 4 + 3]}),
                             smem_ptr
@@ -341,8 +348,8 @@ sm90_bf16_gemm_impl(int* grouped_layout,
                     auto smem_d_1 = reinterpret_cast<float2*>(smem_d + (m_offset + warp_idx * WGMMA_M_PER_WARP + lane_idx / 4 + 8) * BLOCK_N + (lane_idx % 4) * 2);
                     #pragma unroll
                     for (uint32_t i = 0; i < WGMMA::kNumAccum / 4; ++ i) {
-                        st_shared(smem_d_0 + i * 4, make_float2(shifted_accum[i * 4 + 0], shifted_accum[i * 4 + 1]));
-                        st_shared(smem_d_1 + i * 4, make_float2(shifted_accum[i * 4 + 2], shifted_accum[i * 4 + 3]));
+                        ptx::st_shared(smem_d_0 + i * 4, make_float2(shifted_accum[i * 4 + 0], shifted_accum[i * 4 + 1]));
+                        ptx::st_shared(smem_d_1 + i * 4, make_float2(shifted_accum[i * 4 + 2], shifted_accum[i * 4 + 3]));
                     }
                 }
             }
@@ -350,7 +357,7 @@ sm90_bf16_gemm_impl(int* grouped_layout,
             cutlass::arch::NamedBarrier::sync(kNumWGMMAStoreThreads, 0);
 
             // Use TMA store to write back to global memory
-            const auto m_idx = scheduler.template get_global_idx<(not is_m_grouped_contiguous(kGemmType)), IndexType::MN>(shape_m, BLOCK_M, m_block_idx);
+            const auto m_idx = scheduler.template get_global_idx<(not is_m_grouped_contiguous(kGemmType)), sched::IndexType::MN>(shape_m, BLOCK_M, m_block_idx);
             DG_STATIC_ASSERT(kNumWGMMAStoreThreads >= BLOCK_N / TMA_D_BLOCK_N, "Too many TMA blocks");
             if (threadIdx.x < BLOCK_N / TMA_D_BLOCK_N) {
                 auto in_block_n_offset = threadIdx.x * TMA_D_BLOCK_N;

deep_gemm/include/deep_gemm/impls/sm90_bmk_bnk_mn.cuh

@@ -4,26 +4,32 @@
 #include <cutlass/arch/barrier.h>
 #include <cutlass/arch/reg_reconfig.h>
 
+#include <deep_gemm/common/math.cuh>
 #include <deep_gemm/common/utils.cuh>
-#include <deep_gemm/common/sm90_utils.cuh>
+#include <deep_gemm/common/tma_copy.cuh>
+#include <deep_gemm/common/types.cuh>
+#include <deep_gemm/mma/sm90.cuh>
+#include <deep_gemm/epilogue/transform.cuh>
+#include <deep_gemm/ptx/ld_st.cuh>
+#include <deep_gemm/ptx/utils.cuh>
+#include <deep_gemm/ptx/wgmma.cuh>
+#include <deep_gemm/scheduler/gemm.cuh>
 
 namespace deep_gemm {
 
-using namespace deep_gemm::sm90;
-
 template <uint32_t SHAPE_M, uint32_t SHAPE_N, uint32_t SHAPE_K,
           uint32_t BLOCK_M, uint32_t BLOCK_N, uint32_t BLOCK_K,
           uint32_t kSplitFactor,
           uint32_t kNumStages,
           uint32_t kNumTMAThreads, uint32_t kNumMathThreads>
-__global__ __launch_bounds__(kNumTMAThreads + kNumMathThreads, 1) void
+CUTLASS_GLOBAL __launch_bounds__(kNumTMAThreads + kNumMathThreads, 1) void
 sm90_bmn_bnk_mn_gemm_impl(const uint32_t shape_s,
                           const __grid_constant__ cute::TmaDescriptor tensor_map_a,
                           const __grid_constant__ cute::TmaDescriptor tensor_map_b,
                           float *d) {
 #if (defined(__CUDA_ARCH__) and (__CUDA_ARCH__ >= 900)) or defined(__CLION_IDE__)
     // Types
-    using WGMMA = typename BF16MMASelector<BLOCK_N>::type;
+    using WGMMA = typename mma::sm90::BF16MMASelector<BLOCK_N>::type;
     using Barrier = cutlass::arch::ClusterTransactionBarrier;
     DG_STATIC_ASSERT(BLOCK_M % WGMMA::M == 0, "Invalid block size");
 
@@ -33,7 +39,7 @@ sm90_bmn_bnk_mn_gemm_impl(const uint32_t shape_s,
 
     // Configs
     const uint32_t warp_idx = __shfl_sync(0xffffffff, threadIdx.x / 32, 0);
-    const uint32_t lane_idx = get_lane_idx();
+    const uint32_t lane_idx = ptx::get_lane_idx();
     DG_STATIC_ASSERT(BLOCK_M == 128, "Invalid block M");
     DG_STATIC_ASSERT(kNumTMAThreads == 128, "Invalid number of TMA threads");
     DG_STATIC_ASSERT(kNumMathThreads == 256, "Invalid number of math threads");
@@ -48,17 +54,17 @@ sm90_bmn_bnk_mn_gemm_impl(const uint32_t shape_s,
     // Align to 1024 bytes for swizzle-128B
     // Fill shared memory pointers
     extern __shared__ __align__(1024) uint8_t smem_buffer[];
-    auto smem_a = PatternVisitor([&](const uint32_t& i) {
+    auto smem_a = utils::PatternVisitor([&](const uint32_t& i) {
         return reinterpret_cast<__nv_bfloat16*>(smem_buffer + (i * SMEM_A_SIZE_PER_STAGE));
     });
-    auto smem_b = PatternVisitor([&](const uint32_t& i) {
+    auto smem_b = utils::PatternVisitor([&](const uint32_t& i) {
         return reinterpret_cast<__nv_bfloat16*>(smem_buffer + (kNumStages * SMEM_A_SIZE_PER_STAGE + i * SMEM_B_SIZE_PER_STAGE));
     });
 
     // Fill barriers
     auto barrier_start_ptr = reinterpret_cast<Barrier*>(smem_buffer + kNumStages * (SMEM_A_SIZE_PER_STAGE + SMEM_B_SIZE_PER_STAGE));
-    auto full_barriers     = PatternVisitor([=](const uint32_t& i) { return barrier_start_ptr + (i); });
-    auto empty_barriers    = PatternVisitor([=](const uint32_t& i) { return barrier_start_ptr + (kNumStages + i); });
+    auto full_barriers     = utils::PatternVisitor([=](const uint32_t& i) { return barrier_start_ptr + (i); });
+    auto empty_barriers    = utils::PatternVisitor([=](const uint32_t& i) { return barrier_start_ptr + (kNumStages + i); });
 
     // Initialize barriers
     if (warp_idx == 1 and cute::elect_one_sync()) {
@@ -80,14 +86,17 @@ sm90_bmn_bnk_mn_gemm_impl(const uint32_t shape_s,
     constexpr uint32_t kNumMathRegisters = 232;
 
    // Block indices
-    const uint32_t num_n_blocks = ceil_div(SHAPE_N, BLOCK_N);
-    const uint32_t num_mn_blocks = num_n_blocks * ceil_div(SHAPE_M, BLOCK_M);
+    const uint32_t num_n_blocks = math::ceil_div(SHAPE_N, BLOCK_N);
+    const uint32_t num_mn_blocks = num_n_blocks * math::ceil_div(SHAPE_M, BLOCK_M);
     const uint32_t mn_block_idx = blockIdx.x % num_mn_blocks;
     const uint32_t sk_block_idx = blockIdx.x / num_mn_blocks;
     const uint32_t n_block_idx = mn_block_idx % num_n_blocks;
     const uint32_t m_block_idx = mn_block_idx / num_n_blocks;
     const uint32_t num_total_stages = cute::min(kSplitFactor, shape_s * (SHAPE_K / BLOCK_K) - sk_block_idx * kSplitFactor);
 
+    // Wait for primary kernel completion
+    cudaGridDependencySynchronize();
+
     if (warp_idx >= kNumMathThreads / 32) {
         // TMA warp-group for loading data
         cutlass::arch::warpgroup_reg_dealloc<kNumTMARegisters>();
@@ -98,18 +107,18 @@ sm90_bmn_bnk_mn_gemm_impl(const uint32_t shape_s,
             #pragma unroll
             for (uint32_t s = 0; s < num_total_stages; ++ s) {
                 // Wait consumer release
-                const auto& stage_idx = s % kNumStages;
+                const auto stage_idx = s % kNumStages;
                 empty_barriers[stage_idx]->wait((s / kNumStages + 1) & 1);
 
                 auto& full_barrier = *full_barriers[stage_idx];
-                const uint32_t& sk_idx = (sk_block_idx * kSplitFactor + s) * BLOCK_K;
-                const uint32_t& k_idx = sk_idx % SHAPE_K;
-                const uint32_t& s_idx = sk_idx / SHAPE_K;
+                const uint32_t sk_idx = (sk_block_idx * kSplitFactor + s) * BLOCK_K;
+                const uint32_t k_idx = sk_idx % SHAPE_K;
+                const uint32_t s_idx = sk_idx / SHAPE_K;
 
                 constexpr uint32_t kSwizzle = BLOCK_K * sizeof(nv_bfloat16);
-                tma_copy<BLOCK_K, BLOCK_M, kSwizzle>(
+                tma::copy<BLOCK_K, BLOCK_M, kSwizzle>(
                     &tensor_map_a, &full_barrier, smem_a[stage_idx], k_idx, m_block_idx * BLOCK_M + s_idx * SHAPE_M, 1);
-                tma_copy<BLOCK_K, BLOCK_N, kSwizzle>(
+                tma::copy<BLOCK_K, BLOCK_N, kSwizzle>(
                     &tensor_map_b, &full_barrier, smem_b[stage_idx], k_idx, n_block_idx * BLOCK_N + s_idx * SHAPE_N, 1);
                 full_barrier.arrive_and_expect_tx(SMEM_A_SIZE_PER_STAGE + SMEM_B_SIZE_PER_STAGE);
             }
@@ -125,32 +134,32 @@ sm90_bmn_bnk_mn_gemm_impl(const uint32_t shape_s,
         // Launch MMAs
         for (uint32_t s = 0; s < num_total_stages; ++ s) {
             // Wait TMA arrivals
-            const auto& stage_idx = s % kNumStages;
+            const auto stage_idx = s % kNumStages;
             full_barriers[stage_idx]->wait((s / kNumStages) & 1);
 
             // Commit WGMMA instructions
             #pragma unroll
             for (uint32_t i = 0; i < WGMMA::kNumAccum; ++ i)
-                warpgroup_fence_operand(accum[i]);
-            warpgroup_arrive();
+                ptx::warpgroup_fence_operand(accum[i]);
+            ptx::warpgroup_arrive();
             #pragma unroll
             for (uint32_t k = 0; k < BLOCK_K / WGMMA::K; ++ k) {
-                auto desc_a = make_smem_desc(smem_a[stage_idx] + (math_wg_idx * WGMMA::M) * BLOCK_K + k * WGMMA::K, 1);
-                auto desc_b = make_smem_desc(smem_b[stage_idx] + k * WGMMA::K, 1);
+                auto desc_a = mma::sm90::make_smem_desc(smem_a[stage_idx] + (math_wg_idx * WGMMA::M) * BLOCK_K + k * WGMMA::K, 1);
+                auto desc_b = mma::sm90::make_smem_desc(smem_b[stage_idx] + k * WGMMA::K, 1);
                 WGMMA::wgmma(desc_a, desc_b, accum, 1);
             }
-            warpgroup_commit_batch();
+            ptx::warpgroup_commit_batch();
             #pragma unroll
             for (uint32_t i = 0; i < WGMMA::kNumAccum; ++ i)
-                warpgroup_fence_operand(accum[i]);
-            warpgroup_wait<0>();
+                ptx::warpgroup_fence_operand(accum[i]);
+            ptx::warpgroup_wait<0>();
 
             // Notify barrier arrival at the last warpgroup wave
             empty_barriers[stage_idx]->arrive();
         }
 
-        const auto& row = m_block_idx * BLOCK_M + warp_idx * 16 + lane_idx / 4;
-        const auto& col = n_block_idx * BLOCK_N + (lane_idx % 4) * 2;
+        const auto row = m_block_idx * BLOCK_M + warp_idx * 16 + lane_idx / 4;
+        const auto col = n_block_idx * BLOCK_N + (lane_idx % 4) * 2;
         #pragma unroll
         for (uint32_t i = 0; i < WGMMA::kNumAccum / 4; ++ i) {
             if (col + i * 8 >= SHAPE_N)

deep_gemm/include/deep_gemm/impls/sm90_fp8_gemm_1d1d.cuh

@@ -6,18 +6,26 @@
 #include <cutlass/arch/barrier.h>
 #include <cutlass/arch/reg_reconfig.h>
 
+#include <cute/int_tuple.hpp>
 #include <cute/arch/cluster_sm90.hpp>
 #include <cute/arch/copy_sm90_desc.hpp>
 #include <cute/arch/copy_sm90_tma.hpp>
 
+#include <deep_gemm/common/cute_tie.cuh>
+#include <deep_gemm/common/math.cuh>
 #include <deep_gemm/common/utils.cuh>
-#include <deep_gemm/common/scheduler.cuh>
-#include <deep_gemm/common/sm90_utils.cuh>
+#include <deep_gemm/common/tma_copy.cuh>
+#include <deep_gemm/common/types.cuh>
+#include <deep_gemm/mma/sm90.cuh>
+#include <deep_gemm/epilogue/transform.cuh>
+#include <deep_gemm/ptx/ld_st.cuh>
+#include <deep_gemm/ptx/tma.cuh>
+#include <deep_gemm/ptx/utils.cuh>
+#include <deep_gemm/ptx/wgmma.cuh>
+#include <deep_gemm/scheduler/gemm.cuh>
 
 namespace deep_gemm {
 
-using namespace deep_gemm::sm90;
-
 template <uint32_t SHAPE_M, uint32_t SHAPE_N, uint32_t SHAPE_K,
           uint32_t kNumGroups,
           uint32_t BLOCK_M, uint32_t BLOCK_N, uint32_t BLOCK_K,
@@ -27,7 +35,7 @@ template <uint32_t SHAPE_M, uint32_t SHAPE_N, uint32_t SHAPE_K,
           uint32_t kNumTMAMulticast, bool kIsTMAMulticastOnA,
           uint32_t kNumSMs,
           GemmType kGemmType, typename cd_dtype_t>
-__global__ __launch_bounds__(kNumTMAThreads + kNumMathThreads, 1) void
+CUTLASS_GLOBAL __launch_bounds__(kNumTMAThreads + kNumMathThreads, 1) void
 sm90_fp8_gemm_1d1d_impl(__nv_fp8_e4m3* gmem_a_ptr, __nv_fp8_e4m3* gmem_b_ptr,
                         int* grouped_layout,
                         cute::TmaDescriptor* tensor_map_buffer,
@@ -45,7 +53,7 @@ sm90_fp8_gemm_1d1d_impl(__nv_fp8_e4m3* gmem_a_ptr, __nv_fp8_e4m3* gmem_b_ptr,
     DG_STATIC_ASSERT(kGemmType == GemmType::Normal or kGemmType == GemmType::KGroupedContiguous, "Invalid GEMM type");
 
     // Types
-    using WGMMA = typename FP8MMASelector<BLOCK_N>::type;
+    using WGMMA = typename mma::sm90::FP8MMASelector<BLOCK_N>::type;
     using Barrier = cutlass::arch::ClusterTransactionBarrier;
     DG_STATIC_ASSERT(BLOCK_M % WGMMA::M == 0, "Invalid block size");
 
@@ -55,13 +63,13 @@ sm90_fp8_gemm_1d1d_impl(__nv_fp8_e4m3* gmem_a_ptr, __nv_fp8_e4m3* gmem_b_ptr,
     shape_k = SHAPE_K != 0 ? SHAPE_K : shape_k;
 
     // Shared memory
-    static constexpr uint32_t SMEM_TENSOR_MAP_SIZE = (kGemmType == GemmType::KGroupedContiguous ? sizeof(cute::TmaDescriptor) * 4 : 0);
+    static constexpr uint32_t SMEM_TENSOR_MAP_SIZE = (kGemmType == GemmType::KGroupedContiguous ? sizeof(cute::TmaDescriptor) * 2 : 0);
     static constexpr uint32_t SMEM_D_SIZE = BLOCK_M * BLOCK_N * sizeof(float);
     static constexpr uint32_t SMEM_A_SIZE_PER_STAGE = BLOCK_M * BLOCK_K * sizeof(__nv_fp8_e4m3);
     static constexpr uint32_t SMEM_B_SIZE_PER_STAGE = BLOCK_N * BLOCK_K * sizeof(__nv_fp8_e4m3);
     static constexpr uint32_t SMEM_SFA_SIZE_PER_STAGE = BLOCK_M * sizeof(float);
     static constexpr uint32_t SMEM_SFB_SIZE_PER_STAGE = BLOCK_N * sizeof(float);
-    static constexpr uint32_t ALIGNED_SMEM_SFB_SIZE_PER_STAGE = constexpr_align(SMEM_SFB_SIZE_PER_STAGE, 128u);
+    static constexpr uint32_t ALIGNED_SMEM_SFB_SIZE_PER_STAGE = math::constexpr_align(SMEM_SFB_SIZE_PER_STAGE, 128u);
     DG_STATIC_ASSERT(SMEM_SFA_SIZE_PER_STAGE % 128 == 0, "Invalid TMA alignment");
 
     // Configs
@@ -83,47 +91,41 @@ sm90_fp8_gemm_1d1d_impl(__nv_fp8_e4m3* gmem_a_ptr, __nv_fp8_e4m3* gmem_b_ptr,
     DG_STATIC_ASSERT(SMEM_D_SIZE % 1024 == 0, "Shared memory of A/B must be aligned to 1024 bytes");
 
     // Tensor maps on shared and global memory
-    auto smem_tensor_map_a = PatternVisitor([&](const uint32_t& i) {
-        return reinterpret_cast<cute::TmaDescriptor*>(smem_buffer + static_cast<uint32_t>(sizeof(cute::TmaDescriptor)) * i);
-    });
-    auto smem_tensor_map_b = PatternVisitor([&](const uint32_t& i) {
-        return reinterpret_cast<cute::TmaDescriptor*>(smem_buffer + static_cast<uint32_t>(sizeof(cute::TmaDescriptor)) * (2 + i));
-    });
-    auto gmem_tensor_map_a = PatternVisitor([=](const uint32_t& i) { return tensor_map_buffer + blockIdx.x * 4 + i; });
-    auto gmem_tensor_map_b = PatternVisitor([=](const uint32_t& i) { return tensor_map_buffer + blockIdx.x * 4 + 2 + i; });
+    auto smem_tensor_map_a = reinterpret_cast<cute::TmaDescriptor*>(smem_buffer);
+    auto smem_tensor_map_b = smem_tensor_map_a + 1;
+    auto gmem_tensor_map_a = tensor_map_buffer + blockIdx.x * 2;
+    auto gmem_tensor_map_b = gmem_tensor_map_a + 1;
 
     // Data on shared memory
     auto smem_d = reinterpret_cast<float*>(smem_buffer + SMEM_TENSOR_MAP_SIZE);
-    auto smem_a = PatternVisitor([&](const uint32_t& i) {
+    auto smem_a = utils::PatternVisitor([&](const uint32_t& i) {
         return reinterpret_cast<__nv_fp8_e4m3*>(smem_buffer + (SMEM_TENSOR_MAP_SIZE + SMEM_D_SIZE + i * SMEM_A_SIZE_PER_STAGE)); 
     });
-    auto smem_b = PatternVisitor([&](const uint32_t& i) {
+    auto smem_b = utils::PatternVisitor([&](const uint32_t& i) {
         return reinterpret_cast<__nv_fp8_e4m3*>(smem_buffer + (SMEM_TENSOR_MAP_SIZE + SMEM_D_SIZE + kNumStages * SMEM_A_SIZE_PER_STAGE + i * SMEM_B_SIZE_PER_STAGE));
     });
     constexpr auto SMEM_SF_OFFSET = SMEM_TENSOR_MAP_SIZE + SMEM_D_SIZE + kNumStages * (SMEM_A_SIZE_PER_STAGE + SMEM_B_SIZE_PER_STAGE);
-    auto smem_sfa = PatternVisitor([&](const uint32_t& i) {
+    auto smem_sfa = utils::PatternVisitor([&](const uint32_t& i) {
         return reinterpret_cast<float*>(smem_buffer + (SMEM_SF_OFFSET + i * SMEM_SFA_SIZE_PER_STAGE));
     });
-    auto smem_sfb = PatternVisitor([&](const uint32_t& i) {
+    auto smem_sfb = utils::PatternVisitor([&](const uint32_t& i) {
         return reinterpret_cast<float*>(smem_buffer + (SMEM_SF_OFFSET + kNumStages * SMEM_SFA_SIZE_PER_STAGE + i * ALIGNED_SMEM_SFB_SIZE_PER_STAGE));
     });
 
     // Barriers on shared memory
     constexpr auto SMEM_BARRIER_OFFSET = SMEM_SF_OFFSET + kNumStages * (SMEM_SFA_SIZE_PER_STAGE + ALIGNED_SMEM_SFB_SIZE_PER_STAGE);
-    auto full_barriers = PatternVisitor([&](const uint32_t& i) {
+    auto full_barriers = utils::PatternVisitor([&](const uint32_t& i) {
         return reinterpret_cast<Barrier*>(smem_buffer + (SMEM_BARRIER_OFFSET + i * static_cast<uint32_t>(sizeof(Barrier))));
     });
-    auto empty_barriers = PatternVisitor([&](const uint32_t& i) {
+    auto empty_barriers = utils::PatternVisitor([&](const uint32_t& i) {
         return reinterpret_cast<Barrier*>(smem_buffer + (SMEM_BARRIER_OFFSET + (kNumStages + i) * static_cast<uint32_t>(sizeof(Barrier))));
     });
 
     if (warp_idx == kNumMathThreads / 32 + 1 and cute::elect_one_sync()) {
         // Load tensormap A/B to shared memory
         if constexpr (kGemmType == GemmType::KGroupedContiguous) {
-            *smem_tensor_map_a[0] = tensor_map_a_base;
-            *smem_tensor_map_a[1] = tensor_map_a_base;
-            *smem_tensor_map_b[0] = tensor_map_b_base;
-            *smem_tensor_map_b[1] = tensor_map_b_base;
+            *smem_tensor_map_a = tensor_map_a_base;
+            *smem_tensor_map_b = tensor_map_b_base;
         }
 
         // Initialize barriers
@@ -149,12 +151,15 @@ sm90_fp8_gemm_1d1d_impl(__nv_fp8_e4m3* gmem_a_ptr, __nv_fp8_e4m3* gmem_b_ptr,
     constexpr uint32_t kNumTMARegisters = (kNumPipelineUnrolls == 0 ? 40 : 24);
     constexpr uint32_t kNumMathRegisters = (kNumPipelineUnrolls == 0 ? 232 : 240);
 
+    // Wait for primary kernel completion
+    cudaGridDependencySynchronize();
+    
     // Block scheduler
     uint32_t m_block_idx, n_block_idx;
-    auto scheduler = Scheduler<kGemmType, BLOCK_M, BLOCK_N, kNumGroups, kNumTMAMulticast, kIsTMAMulticastOnA, kNumSMs, 128u>(shape_m, shape_n, shape_k, grouped_layout);
+    auto scheduler = sched::Scheduler<kGemmType, BLOCK_M, BLOCK_N, kNumGroups, kNumTMAMulticast, kIsTMAMulticastOnA, kNumSMs, 128u>(shape_m, shape_n, shape_k, grouped_layout);
 
     // TMA and MMA pipeline
-    const auto& get_pipeline = [=](const uint32_t& iter_idx) -> cute::tuple<uint32_t, uint32_t> {
+    const auto get_pipeline = [=](const uint32_t& iter_idx) -> cute::tuple<uint32_t, uint32_t> {
         return {iter_idx % kNumStages, (iter_idx / kNumStages) & 1}; // Pipeline stage and phase
     };
     uint32_t iter_idx = 0;
@@ -165,10 +170,7 @@ sm90_fp8_gemm_1d1d_impl(__nv_fp8_e4m3* gmem_a_ptr, __nv_fp8_e4m3* gmem_b_ptr,
 
         // NOTES: only one thread (or warp) will be used
         if (warp_idx == kNumMathThreads / 32 and cute::elect_one_sync()) {
-            const cute::TmaDescriptor* current_tensor_map_a = &tensor_map_a_base;
-            const cute::TmaDescriptor* current_tensor_map_b = &tensor_map_b_base;
             uint32_t last_group_idx = kNumGroups;
-            uint32_t prefetched_next_group_idx = kNumGroups;  // Track which group was prefetched
 
             // Persistently schedule over blocks
             while (scheduler.get_next_block(m_block_idx, n_block_idx)) {
@@ -179,63 +181,26 @@ sm90_fp8_gemm_1d1d_impl(__nv_fp8_e4m3* gmem_a_ptr, __nv_fp8_e4m3* gmem_b_ptr,
                 const uint32_t num_tma_multicast_b = (not kIsTMAMulticastOnA and is_tma_multicast_valid) ? kNumTMAMulticast : 1;
                 DG_STATIC_ASSERT(kNumTMAMulticast <= 2, "Scheduler does not support > 2 TMA multicast");
                 
-                const uint32_t& num_k_blocks = ceil_div(scheduler.current_shape_k, BLOCK_K);
-                const uint32_t& m_idx = m_block_idx * BLOCK_M;
-                const uint32_t& n_idx = n_block_idx * BLOCK_N;
+                const uint32_t num_k_blocks = math::ceil_div(scheduler.current_shape_k, BLOCK_K);
+                const uint32_t m_idx = m_block_idx * BLOCK_M;
+                const uint32_t n_idx = n_block_idx * BLOCK_N;
 
-                if (kGemmType == GemmType::KGroupedContiguous and last_group_idx != scheduler.current_group_idx) {
-                    const uint32_t& stage_idx = scheduler.current_num_valid_groups & 1;
-                    const uint32_t& next_stage_idx = stage_idx ^ 1;
-                    last_group_idx = scheduler.current_group_idx;
+                if (kGemmType == GemmType::KGroupedContiguous && last_group_idx != scheduler.current_group_idx) {  
+                    last_group_idx = scheduler.current_group_idx;  
 
-                    // Check if the current group matches the prefetched group
-                    // If not, we need to prepare the correct tensor map for the current group
-                    if (scheduler.current_num_valid_groups > 0 &&
-                        scheduler.current_group_idx != prefetched_next_group_idx) {
-                        // The prefetched tensor map doesn't match current group
-                        // This happens when block count is small (< num_SMs) and scheduler skips groups
-                        // Need to prepare the correct tensor map for current group
-                        // Use scheduler.current_k_cumsum which correctly tracks k offset even when groups are skipped
-                        const uint64_t current_k_offset = scheduler.current_k_cumsum;
-                        tensor_map_replace_global_addr_in_smem(smem_tensor_map_a[stage_idx],
-                            gmem_a_ptr + current_k_offset * shape_m);
-                        tensor_map_replace_global_addr_in_smem(smem_tensor_map_b[stage_idx],
-                            gmem_b_ptr + current_k_offset * shape_n);
-                        tensor_map_replace_global_inner_dim_stride_in_smem(smem_tensor_map_a[stage_idx],
-                            scheduler.current_shape_k, scheduler.current_shape_k);
-                        tensor_map_replace_global_inner_dim_stride_in_smem(smem_tensor_map_b[stage_idx],
-                            scheduler.current_shape_k, scheduler.current_shape_k);
-                        *(gmem_tensor_map_a[stage_idx]) = *(smem_tensor_map_a[stage_idx]);
-                        *(gmem_tensor_map_b[stage_idx]) = *(smem_tensor_map_b[stage_idx]);
-                        // NOTE: Don't call tensor_map_release_cta() here!
-                        // We're preparing the current tensor map, not the next one.
-                        // It will be acquired immediately in the "Get current tensor map" section below.
-                    }
+                    // Directly update current tensor map
+                    const uint64_t current_k_offset = scheduler.current_k_cumsum;
+                    ptx::tensor_map_replace_global_addr_in_smem(smem_tensor_map_a, gmem_a_ptr + current_k_offset * shape_m);
+                    ptx::tensor_map_replace_global_addr_in_smem(smem_tensor_map_b, gmem_b_ptr + current_k_offset * shape_n);
+                    ptx::tensor_map_replace_global_inner_dim_stride_in_smem(smem_tensor_map_a, scheduler.current_shape_k, scheduler.current_shape_k);
+                    ptx::tensor_map_replace_global_inner_dim_stride_in_smem(smem_tensor_map_b, scheduler.current_shape_k, scheduler.current_shape_k);
+                    *(gmem_tensor_map_a) = *(smem_tensor_map_a);  
+                    *(gmem_tensor_map_b) = *(smem_tensor_map_b);  
+                    ptx::tensor_map_release_gpu();
 
-                    // Prepare next tensor map (prefetch for next group)
-                    if (scheduler.next_group_idx < kNumGroups) {
-                        // Calculate next group's k offset using scheduler-provided information
-                        // This ensures consistency even when groups are skipped
-                        const uint64_t next_k_offset = static_cast<uint64_t>(scheduler.current_k_cumsum) + scheduler.current_shape_k;
-                        tensor_map_replace_global_addr_in_smem(smem_tensor_map_a[next_stage_idx], gmem_a_ptr + next_k_offset * shape_m);
-                        tensor_map_replace_global_addr_in_smem(smem_tensor_map_b[next_stage_idx], gmem_b_ptr + next_k_offset * shape_n);
-                        tensor_map_replace_global_inner_dim_stride_in_smem(smem_tensor_map_a[next_stage_idx], scheduler.next_shape_k, scheduler.next_shape_k);
-                        tensor_map_replace_global_inner_dim_stride_in_smem(smem_tensor_map_b[next_stage_idx], scheduler.next_shape_k, scheduler.next_shape_k);
-                        *(gmem_tensor_map_a[next_stage_idx]) = *(smem_tensor_map_a[next_stage_idx]);
-                        *(gmem_tensor_map_b[next_stage_idx]) = *(smem_tensor_map_b[next_stage_idx]);
-                        tensor_map_release_cta();
-                        prefetched_next_group_idx = scheduler.next_group_idx;  // Record which group was prefetched
-                    } else {
-                        prefetched_next_group_idx = kNumGroups;  // No more groups to prefetch
-                    }
-
-                    // Get current tensor map
-                    if (scheduler.current_num_valid_groups > 0) {
-                        tensor_map_acquire_cta(gmem_tensor_map_a[stage_idx]);
-                        tensor_map_acquire_cta(gmem_tensor_map_b[stage_idx]);
-                        current_tensor_map_a = gmem_tensor_map_a[stage_idx];
-                        current_tensor_map_b = gmem_tensor_map_b[stage_idx];
-                    }
+                    // Immediately acquire current tensor map
+                    ptx::tensor_map_acquire_gpu(gmem_tensor_map_a);
+                    ptx::tensor_map_acquire_gpu(gmem_tensor_map_b);
                 }
 
                 #pragma unroll kNumPipelineUnrolls
@@ -246,12 +211,14 @@ sm90_fp8_gemm_1d1d_impl(__nv_fp8_e4m3* gmem_a_ptr, __nv_fp8_e4m3* gmem_b_ptr,
 
                     // Issue TMA
                     auto& full_barrier = *full_barriers[stage_idx];
-                    const uint32_t& k_idx = k_block_idx * BLOCK_K;
-                    const uint32_t& sf_k_idx = scheduler.current_sf_k_cumsum + k_block_idx;
-                    tma_copy<BLOCK_M, BLOCK_K, 0>(&tensor_map_sfa, &full_barrier, smem_sfa[stage_idx], m_idx, sf_k_idx, num_tma_multicast_a);
-                    tma_copy<BLOCK_N, BLOCK_K, 0>(&tensor_map_sfb, &full_barrier, smem_sfb[stage_idx], n_idx, sf_k_idx, num_tma_multicast_b);
-                    tma_copy<BLOCK_K, BLOCK_M, kSwizzleAMode>(current_tensor_map_a, &full_barrier, smem_a[stage_idx], k_idx, m_idx, num_tma_multicast_a);
-                    tma_copy<BLOCK_K, BLOCK_N, kSwizzleBMode>(current_tensor_map_b, &full_barrier, smem_b[stage_idx], k_idx, n_idx, num_tma_multicast_b);
+                    const uint32_t k_idx = k_block_idx * BLOCK_K;
+                    const uint32_t sf_k_idx = scheduler.current_sf_k_cumsum + k_block_idx;
+                    const auto tensor_map_a_ptr = (kGemmType == GemmType::KGroupedContiguous ? gmem_tensor_map_a : &tensor_map_a_base);
+                    const auto tensor_map_b_ptr = (kGemmType == GemmType::KGroupedContiguous ? gmem_tensor_map_b : &tensor_map_b_base);
+                    tma::copy<BLOCK_M, BLOCK_K, 0>(&tensor_map_sfa, &full_barrier, smem_sfa[stage_idx], m_idx, sf_k_idx, num_tma_multicast_a);
+                    tma::copy<BLOCK_N, BLOCK_K, 0>(&tensor_map_sfb, &full_barrier, smem_sfb[stage_idx], n_idx, sf_k_idx, num_tma_multicast_b);
+                    tma::copy<BLOCK_K, BLOCK_M, kSwizzleAMode>(tensor_map_a_ptr, &full_barrier, smem_a[stage_idx], k_idx, m_idx, num_tma_multicast_a);
+                    tma::copy<BLOCK_K, BLOCK_N, kSwizzleBMode>(tensor_map_b_ptr, &full_barrier, smem_b[stage_idx], k_idx, n_idx, num_tma_multicast_b);
                     full_barrier.arrive_and_expect_tx(SMEM_A_SIZE_PER_STAGE + SMEM_B_SIZE_PER_STAGE + SMEM_SFA_SIZE_PER_STAGE + SMEM_SFB_SIZE_PER_STAGE);
                 }
             }
@@ -278,9 +245,9 @@ sm90_fp8_gemm_1d1d_impl(__nv_fp8_e4m3* gmem_a_ptr, __nv_fp8_e4m3* gmem_b_ptr,
         while (scheduler.get_next_block(m_block_idx, n_block_idx)) {
             // Accumulation for WGMMA or CUDA promotion
             DG_STATIC_ASSERT(BLOCK_M == WGMMA::M * (BLOCK_M <= 64 ? 1 : 2), "Invalid block sizes");
-            const uint32_t& current_shape_k = (kGemmType == GemmType::KGroupedContiguous ? scheduler.current_shape_k : shape_k);
-            const uint32_t& current_group_idx = (kGemmType == GemmType::KGroupedContiguous ? scheduler.current_group_idx : 0);
-            const uint32_t& num_k_blocks = ceil_div(current_shape_k, BLOCK_K);
+            const uint32_t current_shape_k = (kGemmType == GemmType::KGroupedContiguous ? scheduler.current_shape_k : shape_k);
+            const uint32_t current_group_idx = (kGemmType == GemmType::KGroupedContiguous ? scheduler.current_group_idx : 0);
+            const uint32_t num_k_blocks = math::ceil_div(current_shape_k, BLOCK_K);
             float accum[WGMMA::kNumAccum], final_accum[WGMMA::kNumAccum] = {0};
             float2 scales_b[WGMMA::kNumAccum / 4];
 
@@ -302,30 +269,30 @@ sm90_fp8_gemm_1d1d_impl(__nv_fp8_e4m3* gmem_a_ptr, __nv_fp8_e4m3* gmem_b_ptr,
 
                 // Read A scales
                 // NOTES: all shared memory read must be prior to `warpgroup_arrive` to avoid next scheduled block polluting the results
-                auto scale_a_0 = ld_shared(smem_sfa[stage_idx] + r_0);
-                auto scale_a_1 = ld_shared(smem_sfa[stage_idx] + r_1);
+                auto scale_a_0 = ptx::ld_shared(smem_sfa[stage_idx] + r_0);
+                auto scale_a_1 = ptx::ld_shared(smem_sfa[stage_idx] + r_1);
 
                 // Read B scales
                 #pragma unroll
                 for (int i = 0; i < WGMMA::kNumAccum / 4; ++i)
-                    scales_b[i] = ld_shared(reinterpret_cast<float2*>(smem_sfb[stage_idx] + i * 8 + col_idx * 2));
+                    scales_b[i] = ptx::ld_shared(reinterpret_cast<float2*>(smem_sfb[stage_idx] + i * 8 + col_idx * 2));
 
                 // Commit WGMMA instructions
                 #pragma unroll
                 for (uint32_t i = 0; i < WGMMA::kNumAccum; ++ i)
-                    warpgroup_fence_operand(accum[i]);
-                warpgroup_arrive();
+                    ptx::warpgroup_fence_operand(accum[i]);
+                ptx::warpgroup_arrive();
                 #pragma unroll
                 for (uint32_t k = 0; k < BLOCK_K / WGMMA::K; ++ k) {
-                    auto desc_a = make_smem_desc(smem_a[stage_idx] + math_wg_idx * WGMMA::M * BLOCK_K + k * WGMMA::K, 1);
-                    auto desc_b = make_smem_desc(smem_b[stage_idx] + k * WGMMA::K, 1);
+                    auto desc_a = mma::sm90::make_smem_desc(smem_a[stage_idx] + math_wg_idx * WGMMA::M * BLOCK_K + k * WGMMA::K, 1);
+                    auto desc_b = mma::sm90::make_smem_desc(smem_b[stage_idx] + k * WGMMA::K, 1);
                     WGMMA::wgmma(desc_a, desc_b, accum, k);
                 }
-                warpgroup_commit_batch();
+                ptx::warpgroup_commit_batch();
                 #pragma unroll
                 for (uint32_t i = 0; i < WGMMA::kNumAccum; ++ i)
-                    warpgroup_fence_operand(accum[i]);
-                warpgroup_wait<0>();
+                    ptx::warpgroup_fence_operand(accum[i]);
+                ptx::warpgroup_wait<0>();
 
                 // Notify barrier arrival
                 empty_barrier_arrive(stage_idx);
@@ -348,12 +315,12 @@ sm90_fp8_gemm_1d1d_impl(__nv_fp8_e4m3* gmem_a_ptr, __nv_fp8_e4m3* gmem_b_ptr,
             cutlass::arch::NamedBarrier::sync(128, math_wg_idx);
 
             // Store to D shared memory
-            const auto& smem_d_0 = reinterpret_cast<float2*>(smem_d + r_0 * BLOCK_N + col_idx * 2);
-            const auto& smem_d_1 = reinterpret_cast<float2*>(smem_d + r_1 * BLOCK_N + col_idx * 2);
+            const auto smem_d_0 = reinterpret_cast<float2*>(smem_d + r_0 * BLOCK_N + col_idx * 2);
+            const auto smem_d_1 = reinterpret_cast<float2*>(smem_d + r_1 * BLOCK_N + col_idx * 2);
             #pragma unroll
             for (auto i = 0; i < WGMMA::kNumAccum / 4; ++ i) {
-                st_shared(smem_d_0 + i * 4, {final_accum[i * 4 + 0], final_accum[i * 4 + 1]});
-                st_shared(smem_d_1 + i * 4, {final_accum[i * 4 + 2], final_accum[i * 4 + 3]});
+                ptx::st_shared(smem_d_0 + i * 4, {final_accum[i * 4 + 0], final_accum[i * 4 + 1]});
+                ptx::st_shared(smem_d_1 + i * 4, {final_accum[i * 4 + 2], final_accum[i * 4 + 3]});
             }
             cute::tma_store_fence();
             cutlass::arch::NamedBarrier::sync(128, math_wg_idx);

deep_gemm/include/deep_gemm/impls/sm90_fp8_gemm_1d2d.cuh

@@ -10,17 +10,21 @@
 #include <cute/arch/copy_sm90_desc.hpp>
 #include <cute/arch/copy_sm90_tma.hpp>
 
-#include <deep_gemm/common/epilogue_utils.cuh>
+#include <deep_gemm/common/math.cuh>
 #include <deep_gemm/common/utils.cuh>
-#include <deep_gemm/common/scheduler.cuh>
-#include <deep_gemm/common/sm90_utils.cuh>
+#include <deep_gemm/common/tma_copy.cuh>
+#include <deep_gemm/common/types.cuh>
+#include <deep_gemm/mma/sm90.cuh>
+#include <deep_gemm/epilogue/transform.cuh>
+#include <deep_gemm/ptx/ld_st.cuh>
+#include <deep_gemm/ptx/utils.cuh>
+#include <deep_gemm/ptx/wgmma.cuh>
+#include <deep_gemm/scheduler/gemm.cuh>
 
 namespace deep_gemm {
 
-using namespace deep_gemm::sm90;
-
 template <uint32_t kNumFormerIters, uint32_t kGap, uint32_t kEnd, typename func_t>
-__device__ void dispatch_num_former_iters(uint32_t num_former_iters, const func_t& func) {
+CUTLASS_DEVICE void dispatch_num_former_iters(uint32_t num_former_iters, const func_t& func) {
     if (num_former_iters == kNumFormerIters) {
         func(cute::Int<kNumFormerIters>{});
         return;
@@ -35,12 +39,12 @@ template <cute::UMMA::Major kMajorSFB,
           uint32_t kNumGroups,
           uint32_t BLOCK_M, uint32_t BLOCK_N, uint32_t BLOCK_K,
           uint32_t kSwizzleAMode, uint32_t kSwizzleBMode, uint32_t kSwizzleDMode,
-          uint32_t kNumStages, uint32_t kNumLastStages,
+          uint32_t kNumStages,
           uint32_t kNumTMAThreads, uint32_t kNumMathThreads,
           uint32_t kNumTMAMulticast, bool kIsTMAMulticastOnA,
           uint32_t kNumSMs, GemmType kGemmType,
           typename epilogue_type_t>
-__global__ __launch_bounds__(kNumTMAThreads + kNumMathThreads, 1) void
+CUTLASS_GLOBAL __launch_bounds__(kNumTMAThreads + kNumMathThreads, 1) void
 sm90_fp8_gemm_1d2d_impl(float* sfb, int* grouped_layout,
                         uint32_t shape_m, uint32_t shape_n, uint32_t shape_k,
                         const __grid_constant__ cute::TmaDescriptor tensor_map_a,
@@ -50,10 +54,12 @@ sm90_fp8_gemm_1d2d_impl(float* sfb, int* grouped_layout,
 #if (defined(__CUDA_ARCH__) and (__CUDA_ARCH__ >= 900)) or defined(__CLION_IDE__)
     // Scaling checks
     DG_STATIC_ASSERT(BLOCK_K == 128, "Only support per-128-channel FP8 scaling");
-    DG_STATIC_ASSERT(constexpr_ceil_div(BLOCK_N, BLOCK_K) == 1 or (constexpr_gcd(BLOCK_N, BLOCK_K) == BLOCK_N - BLOCK_K), "Too much B scales in a single block");
+    DG_STATIC_ASSERT(
+        math::constexpr_ceil_div(BLOCK_N, BLOCK_K) == 1 or
+        (math::constexpr_gcd(BLOCK_N, BLOCK_K) == BLOCK_N - BLOCK_K), "Too much B scales in a single block");
 
     // Types
-    using WGMMA = typename FP8MMASelector<BLOCK_N>::type;
+    using WGMMA = typename mma::sm90::FP8MMASelector<BLOCK_N>::type;
     using Barrier = cutlass::arch::ClusterTransactionBarrier;
     DG_STATIC_ASSERT(BLOCK_M % WGMMA::M == 0 or BLOCK_M < WGMMA::M, "Invalid block size");
 
@@ -64,23 +70,23 @@ sm90_fp8_gemm_1d2d_impl(float* sfb, int* grouped_layout,
 
     // Shared memory
     static constexpr bool kMustUseUniformedScaleB = (BLOCK_K % BLOCK_N == 0);
-    static constexpr uint32_t SMEM_D_SIZE = constexpr_align(BLOCK_M * BLOCK_N * static_cast<uint32_t>(sizeof(__nv_bfloat16)), 1024u);
+    static constexpr uint32_t SMEM_D_SIZE = math::constexpr_align(BLOCK_M * BLOCK_N * static_cast<uint32_t>(sizeof(__nv_bfloat16)), 1024u);
     static constexpr uint32_t SMEM_A_SIZE_PER_STAGE = BLOCK_M * BLOCK_K * sizeof(__nv_fp8_e4m3);
     static constexpr uint32_t SMEM_B_SIZE_PER_STAGE = BLOCK_N * BLOCK_K * sizeof(__nv_fp8_e4m3);
     static constexpr uint32_t SMEM_SFA_SIZE_PER_STAGE = BLOCK_M * sizeof(float);
-    static constexpr uint32_t ALIGNED_SMEM_SFA_SIZE_PER_STAGE = constexpr_align(SMEM_SFA_SIZE_PER_STAGE, 128u);
-    const uint32_t& shape_k_scales = ceil_div(shape_k, BLOCK_K);
-    const uint32_t& shape_n_sfb = ceil_div(shape_n, BLOCK_K);
-    const uint32_t& smem_sfb_size = align<uint32_t>(shape_k_scales * (kMustUseUniformedScaleB ? 1 : 2) * sizeof(float), sizeof(Barrier));
+    static constexpr uint32_t ALIGNED_SMEM_SFA_SIZE_PER_STAGE = math::constexpr_align(SMEM_SFA_SIZE_PER_STAGE, 128u);
+    const uint32_t shape_k_scales = math::ceil_div(shape_k, BLOCK_K);
+    const uint32_t shape_n_sfb = math::ceil_div(shape_n, BLOCK_K);
+    const uint32_t smem_sfb_size = math::align<uint32_t>(shape_k_scales * (kMustUseUniformedScaleB ? 1 : 2) * sizeof(float), sizeof(Barrier));
 
     // NOTES: Make sure we have enough shared memory for WGMMA padding
     static constexpr uint32_t WGMMA_A_SIZE_PER_STAGE = WGMMA::M * BLOCK_K * sizeof(__nv_fp8_e4m3);
     DG_STATIC_ASSERT(WGMMA_A_SIZE_PER_STAGE <= SMEM_A_SIZE_PER_STAGE + SMEM_B_SIZE_PER_STAGE * kNumStages, "Memory Out of bound for WGMMA");
 
     // Configs
-    const uint32_t num_total_k_blocks = ceil_div(shape_k, BLOCK_K);
+    const uint32_t num_total_k_blocks = math::ceil_div(shape_k, BLOCK_K);
     const uint32_t warp_idx = __shfl_sync(0xffffffff, threadIdx.x / 32, 0);
-    const uint32_t lane_idx = get_lane_idx();
+    const uint32_t lane_idx = ptx::get_lane_idx();
 
     // Prefetch TMA descriptors at the very beginning
     if (warp_idx == kNumMathThreads / 32 and cute::elect_one_sync()) {
@@ -97,22 +103,22 @@ sm90_fp8_gemm_1d2d_impl(float* sfb, int* grouped_layout,
 
     // Data on shared memory
     auto smem_d = reinterpret_cast<__nv_bfloat16*>(smem_buffer);
-    auto smem_a = PatternVisitor([&](const uint32_t& i) {
+    auto smem_a = utils::PatternVisitor([&](const uint32_t& i) {
         return reinterpret_cast<__nv_fp8_e4m3*>(smem_buffer + SMEM_D_SIZE + i * SMEM_A_SIZE_PER_STAGE);
     });
-    auto smem_b = PatternVisitor([&](const uint32_t& i) {
+    auto smem_b = utils::PatternVisitor([&](const uint32_t& i) {
         return reinterpret_cast<__nv_fp8_e4m3*>(smem_buffer + SMEM_D_SIZE + kNumStages * SMEM_A_SIZE_PER_STAGE + i * SMEM_B_SIZE_PER_STAGE);
     });
     constexpr uint32_t SMEM_SF_OFFSET = SMEM_D_SIZE + kNumStages * (SMEM_A_SIZE_PER_STAGE + SMEM_B_SIZE_PER_STAGE);
-    auto smem_sfa = PatternVisitor([&](const uint32_t& i) {
+    auto smem_sfa = utils::PatternVisitor([&](const uint32_t& i) {
         return reinterpret_cast<float*>(smem_buffer + SMEM_SF_OFFSET + i * ALIGNED_SMEM_SFA_SIZE_PER_STAGE);
     });
     auto smem_sfb = reinterpret_cast<float*>(smem_buffer + SMEM_SF_OFFSET + kNumStages * ALIGNED_SMEM_SFA_SIZE_PER_STAGE);
 
     // Fill barriers
     auto barrier_start_ptr = reinterpret_cast<Barrier*>(reinterpret_cast<uint8_t*>(smem_sfb) + smem_sfb_size);
-    auto full_barriers     = PatternVisitor([&](const uint32_t& i) { return barrier_start_ptr + i; });
-    auto empty_barriers    = PatternVisitor([&](const uint32_t& i) { return barrier_start_ptr + kNumStages + i; });
+    auto full_barriers     = utils::PatternVisitor([&](const uint32_t& i) { return barrier_start_ptr + i; });
+    auto empty_barriers    = utils::PatternVisitor([&](const uint32_t& i) { return barrier_start_ptr + kNumStages + i; });
 
     // Initialize barriers
     DG_STATIC_ASSERT(kNumTMAMulticast <= 32, "Too many TMA multicast");
@@ -136,9 +142,12 @@ sm90_fp8_gemm_1d2d_impl(float* sfb, int* grouped_layout,
     constexpr uint32_t kNumTMARegisters = 40;
     constexpr uint32_t kNumMathRegisters = kNumMathThreads == 128 ? 248 : 232;
 
+    // Wait for primary kernel completion
+    cudaGridDependencySynchronize();
+
     // Block scheduler
     uint32_t m_block_idx, n_block_idx;
-    auto scheduler = Scheduler<kGemmType, BLOCK_M, BLOCK_N, kNumGroups, kNumTMAMulticast, kIsTMAMulticastOnA, kNumSMs>(shape_m, shape_n, shape_k, grouped_layout);
+    auto scheduler = sched::Scheduler<kGemmType, BLOCK_M, BLOCK_N, kNumGroups, kNumTMAMulticast, kIsTMAMulticastOnA, kNumSMs>(shape_m, shape_n, shape_k, grouped_layout);
 
     // Pipeline and TMA phases
     uint32_t stage_idx = 0, phase = 0;
@@ -177,15 +186,15 @@ sm90_fp8_gemm_1d2d_impl(float* sfb, int* grouped_layout,
                     constexpr bool kWithGroupOffsetA = kGemmType == GemmType::MGroupedMasked;
                     auto& full_barrier = *full_barriers[stage_idx];
                     const uint32_t k_idx = k_block_idx * BLOCK_K;
-                    tma_copy<BLOCK_K, BLOCK_M, kSwizzleAMode, __nv_fp8_e4m3, kIsBatchedMM>(&tensor_map_a, &full_barrier,
+                    tma::copy<BLOCK_K, BLOCK_M, kSwizzleAMode, __nv_fp8_e4m3, kIsBatchedMM>(&tensor_map_a, &full_barrier,
                              smem_a[stage_idx], k_idx, scheduler.get_global_idx<kWithGroupOffsetA>(shape_m, BLOCK_M, m_block_idx),
                              num_tma_multicast_a, batch_idx);
-                    tma_copy<BLOCK_M, BLOCK_K, 0>(&tensor_map_sfa, &full_barrier,
-                             smem_sfa[stage_idx], m_block_idx * BLOCK_M, scheduler.template get_global_idx<kWithGroupOffsetA, IndexType::SF_K>(shape_k_scales, 1, k_block_idx),
+                    tma::copy<BLOCK_M, BLOCK_K, 0>(&tensor_map_sfa, &full_barrier,
+                             smem_sfa[stage_idx], m_block_idx * BLOCK_M, scheduler.template get_global_idx<kWithGroupOffsetA, sched::IndexType::SF_K>(shape_k_scales, 1, k_block_idx),
                              num_tma_multicast_a);
 
                     // Issue TMA B
-                    tma_copy<BLOCK_K, BLOCK_N, kSwizzleBMode, __nv_fp8_e4m3, kIsBatchedMM>(&tensor_map_b, &full_barrier,
+                    tma::copy<BLOCK_K, BLOCK_N, kSwizzleBMode, __nv_fp8_e4m3, kIsBatchedMM>(&tensor_map_b, &full_barrier,
                              smem_b[stage_idx], k_idx, scheduler.get_global_idx<true>(shape_n, BLOCK_N, n_block_idx, m_block_idx),
                              num_tma_multicast_b, batch_idx);
                     full_barrier.arrive_and_expect_tx(SMEM_A_SIZE_PER_STAGE + SMEM_B_SIZE_PER_STAGE + SMEM_SFA_SIZE_PER_STAGE);
@@ -206,8 +215,8 @@ sm90_fp8_gemm_1d2d_impl(float* sfb, int* grouped_layout,
         const auto math_wg_idx = __shfl_sync(0xffffffff, threadIdx.x / 128, 0);
         const auto r_0 = warp_idx * 16 + lane_idx / 4, r_1 = r_0 + 8;
 
-        auto a_desc = make_smem_desc(smem_a[0] + math_wg_idx * WGMMA::M * BLOCK_K, 1);
-        auto b_desc = make_smem_desc(smem_b[0], 1);
+        auto a_desc = mma::sm90::make_smem_desc(smem_a[0] + math_wg_idx * WGMMA::M * BLOCK_K, 1);
+        auto b_desc = mma::sm90::make_smem_desc(smem_b[0], 1);
         const uint32_t a_desc_lo = __shfl_sync(0xffffffff, a_desc.reg32_[0], 0);
         const uint32_t b_desc_lo = __shfl_sync(0xffffffff, b_desc.reg32_[0], 0);
 
@@ -225,14 +234,14 @@ sm90_fp8_gemm_1d2d_impl(float* sfb, int* grouped_layout,
             // Load B scales with math warp-groups
             // NOTES: except the first warp, we want to overlap loading B scales with TMA stores between tasks
             if (threadIdx.x >= 32) {
-                auto previous_group_offset = scheduler.template get_global_idx<true, IndexType::SF_K>(shape_n_sfb * shape_k_scales, 0, 0, m_block_idx);
+                auto previous_group_offset = scheduler.template get_global_idx<true, sched::IndexType::SF_K>(shape_n_sfb * shape_k_scales, 0, 0, m_block_idx);
                 const uint32_t stride_n_sfb = kMajorSFB == cute::UMMA::Major::MN ? 1 : shape_k_scales;
                 const uint32_t stride_k_sfb = kMajorSFB == cute::UMMA::Major::MN ? shape_n_sfb : 1;
                 auto local_sfb = sfb + previous_group_offset + ((n_block_idx * BLOCK_N) / BLOCK_K) * stride_n_sfb;
 
                 #pragma unroll
                 for (uint32_t i = threadIdx.x - 32; i < num_sfb; i += kNumMathThreads - 32)
-                    st_shared(smem_sfb + i, __ldg(i < shape_k_scales ? local_sfb + i * stride_k_sfb : local_sfb + (i - shape_k_scales) * stride_k_sfb + stride_n_sfb));
+                    ptx::st_shared(smem_sfb + i, i < shape_k_scales ? local_sfb[i * stride_k_sfb] : local_sfb[(i - shape_k_scales) * stride_k_sfb + stride_n_sfb]);
             }
             cutlass::arch::NamedBarrier::sync(kNumMathThreads, 0);
 
@@ -259,22 +268,22 @@ sm90_fp8_gemm_1d2d_impl(float* sfb, int* grouped_layout,
             // Skip useless computations
             if (scheduler.is_computation_valid(m_block_idx, math_wg_idx * WGMMA::M)) {
                 // The compiler must know the dynamic variable `num_former_iters`'s real value
-                constexpr bool kShouldOptimize = BLOCK_K / constexpr_gcd(BLOCK_K, BLOCK_N) <= 4 and not kMustUseUniformedScaleB;
-                constexpr uint32_t kGap = constexpr_gcd(BLOCK_K, BLOCK_N) / 8;
+                constexpr bool kShouldOptimize = BLOCK_K / math::constexpr_gcd(BLOCK_K, BLOCK_N) <= 4 and not kMustUseUniformedScaleB;
+                constexpr uint32_t kGap = math::constexpr_gcd(BLOCK_K, BLOCK_N) / 8;
                 constexpr uint32_t kEnd = kShouldOptimize ? BLOCK_K / 8 : 0;
 
                 // Dispatch `num_former_iters` and launch MMAs
                 dispatch_num_former_iters<0, kGap, kEnd>(kShouldOptimize ? num_former_iters : 0, [&](auto _) {
                     #pragma unroll 8
                     for (uint32_t k_block_idx = 0; k_block_idx < num_total_k_blocks; advance_pipeline(k_block_idx)) {
-                        const auto& a_desc_base_lo = a_desc_lo + stage_idx * (SMEM_A_SIZE_PER_STAGE / 16);
-                        const auto& b_desc_base_lo = b_desc_lo + stage_idx * (SMEM_B_SIZE_PER_STAGE / 16);
+                        const auto a_desc_base_lo = a_desc_lo + stage_idx * (SMEM_A_SIZE_PER_STAGE / 16);
+                        const auto b_desc_base_lo = b_desc_lo + stage_idx * (SMEM_B_SIZE_PER_STAGE / 16);
 
                         // Read B scales
-                        float scale_b_0 = ld_shared(smem_sfb + k_block_idx), scale_b_1;
+                        float scale_b_0 = ptx::ld_shared(smem_sfb + k_block_idx), scale_b_1;
                         // NOTES: even some blocks do not need to read the second row, but we still load one to align with other blocks
                         if constexpr (not kMustUseUniformedScaleB)
-                            scale_b_1 = ld_shared(smem_sfb + k_block_idx + shape_k_scales);
+                            scale_b_1 = ptx::ld_shared(smem_sfb + k_block_idx + shape_k_scales);
 
                         // Wait TMA arrivals
                         full_barriers[stage_idx]->wait(phase);
@@ -286,25 +295,25 @@ sm90_fp8_gemm_1d2d_impl(float* sfb, int* grouped_layout,
 
                             // Read A scales
                             // NOTES: all shared memory read must be prior to `warpgroup_arrive` to avoid next scheduled block polluting the results
-                            auto scale_a_0 = do_wgmma_store ? ld_shared(smem_sfa[stage_idx] + r_0 + m_offset) : 0;
-                            auto scale_a_1 = do_wgmma_store ? ld_shared(smem_sfa[stage_idx] + r_1 + m_offset) : 0;
+                            auto scale_a_0 = do_wgmma_store ? ptx::ld_shared(smem_sfa[stage_idx] + r_0 + m_offset) : 0;
+                            auto scale_a_1 = do_wgmma_store ? ptx::ld_shared(smem_sfa[stage_idx] + r_1 + m_offset) : 0;
 
                             // Commit WGMMA instructions
                             #pragma unroll
                             for (uint32_t i = 0; i < WGMMA::kNumAccum; ++ i)
-                                warpgroup_fence_operand(accum[i]);
-                            warpgroup_arrive();
+                                ptx::warpgroup_fence_operand(accum[i]);
+                            ptx::warpgroup_arrive();
                             #pragma unroll
                             for (uint32_t k = 0; k < BLOCK_K / WGMMA::K; ++ k) {
                                 a_desc.reg32_[0] = a_desc_base_lo + (m_offset * BLOCK_K + k * WGMMA::K) / 16;
                                 b_desc.reg32_[0] = b_desc_base_lo + k * WGMMA::K / 16;
                                 WGMMA::wgmma(a_desc, b_desc, accum, k);
                             }
-                            warpgroup_commit_batch();
+                            ptx::warpgroup_commit_batch();
                             #pragma unroll
                             for (uint32_t i = 0; i < WGMMA::kNumAccum; ++ i)
-                                warpgroup_fence_operand(accum[i]);
-                            warpgroup_wait<0>();
+                                ptx::warpgroup_fence_operand(accum[i]);
+                            ptx::warpgroup_wait<0>();
 
                             // Notify barrier arrival at the last warpgroup wave
                             if (local_idx == BLOCK_M / WAVE_BLOCK_M - 1)
@@ -325,7 +334,7 @@ sm90_fp8_gemm_1d2d_impl(float* sfb, int* grouped_layout,
                             #pragma unroll
                             for (uint32_t i = 0; i < WGMMA::kNumAccum / 4; ++ i) {
                                 // NOTES: for unrolled `num_former_iters` cases, we expect the compiler to automatically make it a constant
-                                const bool& predicate = kMustUseUniformedScaleB or i < num_former_iters;
+                                const bool predicate = kMustUseUniformedScaleB or i < num_former_iters;
                                 shifted_accum[i * 4 + 0] += (predicate ? scale_0_0 : scale_0_1) * accum[i * 4 + 0];
                                 shifted_accum[i * 4 + 1] += (predicate ? scale_0_0 : scale_0_1) * accum[i * 4 + 1];
                                 shifted_accum[i * 4 + 2] += (predicate ? scale_1_0 : scale_1_1) * accum[i * 4 + 2];
@@ -399,7 +408,7 @@ sm90_fp8_gemm_1d2d_impl(float* sfb, int* grouped_layout,
                     }
 
                     // NOTES: only 16 lanes' addresses are used
-                    SM90_U32x2_STSM_N<nv_bfloat162>::copy(
+                    ptx::SM90_U32x2_STSM_N<nv_bfloat162>::copy(
                         __float22bfloat162_rn({shifted_accum[i * 4 + 0], shifted_accum[i * 4 + 1]}),
                         __float22bfloat162_rn({shifted_accum[i * 4 + 2], shifted_accum[i * 4 + 3]}),
                         smem_ptr

deep_gemm/include/deep_gemm/impls/sm90_fp8_mqa_logits.cuh

@@ -7,36 +7,31 @@
 #include <cute/arch/copy_sm90_desc.hpp>
 #include <cute/arch/mma_sm90_desc.hpp>
 
+#include <deep_gemm/common/cute_tie.cuh>
+#include <deep_gemm/common/math.cuh>
 #include <deep_gemm/common/utils.cuh>
-#include <deep_gemm/common/sm90_utils.cuh>
+#include <deep_gemm/common/tma_copy.cuh>
+#include <deep_gemm/common/types.cuh>
+#include <deep_gemm/mma/sm90.cuh>
+#include <deep_gemm/ptx/ld_st.cuh>
+#include <deep_gemm/ptx/utils.cuh>
+#include <deep_gemm/ptx/wgmma.cuh>
 
 namespace deep_gemm {
 
-using namespace deep_gemm::sm90;
-
-// ReSharper disable once CppNotAllPathsReturnValue
-template <uint32_t kHeadDim>
-static constexpr int to_swizzle_cute_type() {
-    DG_STATIC_ASSERT(kHeadDim == 32 or kHeadDim == 64 or kHeadDim == 128, "Invalid swizzling");
-    if constexpr (kHeadDim == 32)
-        return static_cast<int>(cute::SM90::GMMA::LayoutType::B32);
-    if constexpr (kHeadDim == 64)
-        return static_cast<int>(cute::SM90::GMMA::LayoutType::B64);
-    if constexpr (kHeadDim == 128)
-        return static_cast<int>(cute::SM90::GMMA::LayoutType::B128);
-}
-
 template <uint32_t kNumHeads, uint32_t kHeadDim,
           bool kIsCompressedLogits,
           uint32_t BLOCK_Q, uint32_t BLOCK_KV,
           uint32_t kNumQStages, uint32_t kNumKVStages,
-          uint32_t kNumTMAThreads, uint32_t kNumMathThreads>
-__global__ __launch_bounds__(kNumTMAThreads + kNumMathThreads, 1)
+          uint32_t kNumSMs,
+          uint32_t kNumTMAThreads, uint32_t kNumMathThreads,
+          typename logits_dtype_t>
+CUTLASS_GLOBAL __launch_bounds__(kNumTMAThreads + kNumMathThreads, 1)
 void sm90_fp8_mqa_logits(const uint32_t seq_len, const uint32_t seq_len_kv,
-                         const uint32_t max_seqlen_k, const uint64_t stride_logits,
+                         const uint32_t max_seqlen_k, const uint32_t stride_logits,
                          uint32_t* cu_seq_len_k_start,
                          uint32_t* cu_seq_len_k_end,
-                         float* logits,
+                         logits_dtype_t* logits,
                          const __grid_constant__ cute::TmaDescriptor tensor_map_q,
                          const __grid_constant__ cute::TmaDescriptor tensor_map_kv,
                          const __grid_constant__ cute::TmaDescriptor tensor_map_kv_scales,
@@ -44,10 +39,10 @@ void sm90_fp8_mqa_logits(const uint32_t seq_len, const uint32_t seq_len_kv,
     // TODO: consider TMA multicast
     // For one block, we process `[q_start:q_end, h, d] @ [kv_start:kv_end, d] -> [q_start:q_end, kv_start:kv_end]`
     // Q should be load only at once for a block
-    const auto& num_q_blocks = ceil_div(seq_len, BLOCK_Q);
+    const auto num_q_blocks = math::ceil_div(seq_len, BLOCK_Q);
 
     // Types
-    using WGMMA = typename FP8MMASelector<BLOCK_Q * kNumHeads>::type;
+    using WGMMA = typename mma::sm90::FP8MMASelector<BLOCK_Q * kNumHeads>::type;
     using Barrier = cutlass::arch::ClusterTransactionBarrier;
 
     // Prefetch TMA descriptors
@@ -74,19 +69,19 @@ void sm90_fp8_mqa_logits(const uint32_t seq_len, const uint32_t seq_len_kv,
     DG_STATIC_ASSERT(SMEM_KV_SIZE_PER_STAGE % kSwizzleAlignment == 0, "Unaligned TMA swizzling");
 
     // Data on shared memory
-    auto smem_q = PatternVisitor([&](const uint32_t& i) {
+    auto smem_q = utils::PatternVisitor([&](const uint32_t& i) {
         return reinterpret_cast<__nv_fp8_e4m3*>(smem_buffer +
             SMEM_Q_SIZE_PER_STAGE * i);
     });
-    auto smem_kv = PatternVisitor([&](const uint32_t& i) {
+    auto smem_kv = utils::PatternVisitor([&](const uint32_t& i) {
         return reinterpret_cast<__nv_fp8_e4m3*>(smem_buffer + (
             SMEM_Q_SIZE_PER_STAGE * kNumQStages + SMEM_KV_SIZE_PER_STAGE * i));
     });
-    auto smem_weights = PatternVisitor([&](const uint32_t& i) {
+    auto smem_weights = utils::PatternVisitor([&](const uint32_t& i) {
         return reinterpret_cast<float*>(smem_buffer +
             SMEM_Q_SIZE_PER_STAGE * kNumQStages + SMEM_KV_SIZE_PER_STAGE * kNumKVStages + SMEM_WEIGHT_SIZE_PER_STAGE * i);
     });
-    auto smem_kv_scales = PatternVisitor([&](const uint32_t& i) {
+    auto smem_kv_scales = utils::PatternVisitor([&](const uint32_t& i) {
         return reinterpret_cast<float*>(smem_buffer +
             SMEM_Q_SIZE_PER_STAGE * kNumQStages + SMEM_KV_SIZE_PER_STAGE * kNumKVStages +
             SMEM_WEIGHT_SIZE_PER_STAGE * kNumQStages + SMEM_KV_SCALE_SIZE_PER_STAGE * i);
@@ -94,13 +89,13 @@ void sm90_fp8_mqa_logits(const uint32_t seq_len, const uint32_t seq_len_kv,
 
     // TMA barriers
     auto barrier_ptr = reinterpret_cast<Barrier*>(smem_kv_scales[kNumKVStages]);
-    auto full_q_barriers   = PatternVisitor([&](const uint32_t& i) { return barrier_ptr + i; });
-    auto empty_q_barriers  = PatternVisitor([&](const uint32_t& i) { return barrier_ptr + (kNumQStages + i); });
-    auto full_kv_barriers  = PatternVisitor([&](const uint32_t& i) { return barrier_ptr + (kNumQStages * 2 + i); });
-    auto empty_kv_barriers = PatternVisitor([&](const uint32_t& i) { return barrier_ptr + (kNumQStages * 2 + kNumKVStages + i); });
+    auto full_q_barriers   = utils::PatternVisitor([&](const uint32_t& i) { return barrier_ptr + i; });
+    auto empty_q_barriers  = utils::PatternVisitor([&](const uint32_t& i) { return barrier_ptr + (kNumQStages + i); });
+    auto full_kv_barriers  = utils::PatternVisitor([&](const uint32_t& i) { return barrier_ptr + (kNumQStages * 2 + i); });
+    auto empty_kv_barriers = utils::PatternVisitor([&](const uint32_t& i) { return barrier_ptr + (kNumQStages * 2 + kNumKVStages + i); });
 
     // Initialize barriers
-    const bool& is_tma_load_warp = kNumMathThreads <= threadIdx.x and threadIdx.x < kNumMathThreads + 32;
+    const bool is_tma_load_warp = kNumMathThreads <= threadIdx.x and threadIdx.x < kNumMathThreads + 32;
     if (is_tma_load_warp and cute::elect_one_sync()) {
         #pragma unroll
         for (uint32_t i = 0; i < kNumQStages; ++ i) {
@@ -123,38 +118,43 @@ void sm90_fp8_mqa_logits(const uint32_t seq_len, const uint32_t seq_len_kv,
     constexpr uint32_t kNumMathRegisters = 112;
 
     // Block scheduler
-    uint32_t block_q_idx = blockIdx.x, q_iter_idx = 0;
-    const auto& get_next_block_q_idx = [&]() -> cute::tuple<uint32_t, uint32_t> {
-        return {block_q_idx + gridDim.x, q_iter_idx + 1};
+    const auto sm_idx = blockIdx.x;
+    uint32_t block_q_idx = sm_idx, q_iter_idx = 0;
+    const auto get_next_block_q_idx = [&]() -> cute::tuple<uint32_t, uint32_t> {
+        return {block_q_idx + kNumSMs, q_iter_idx + 1};
     };
     uint32_t seq_k_start[BLOCK_Q], seq_k_end[BLOCK_Q];
-    const auto& load_schedule = [&](const uint32_t& q_iter_offset = 0) -> cute::tuple<uint32_t, uint32_t, uint32_t, uint32_t> {
+    const auto load_schedule = [&](const uint32_t& q_iter_offset = 0) -> cute::tuple<uint32_t, uint32_t, uint32_t, uint32_t> {
         uint32_t start = cute::numeric_limits<uint32_t>::max();
         uint32_t end = cute::numeric_limits<uint32_t>::min();
 
         #pragma unroll
         for (uint32_t i = 0; i < BLOCK_Q; ++ i) {
-            const auto& q_idx = min(block_q_idx * BLOCK_Q + i, seq_len - 1);
-            seq_k_start[i] = __ldg(cu_seq_len_k_start + q_idx);
-            seq_k_end[i] = __ldg(cu_seq_len_k_end + q_idx);
+            const auto q_idx = min(block_q_idx * BLOCK_Q + i, seq_len - 1);
+            seq_k_start[i] = cu_seq_len_k_start[q_idx];
+            seq_k_end[i] = cu_seq_len_k_end[q_idx];
             start = min(start, min(seq_k_start[i], seq_len_kv));
             end = max(end, min(seq_k_end[i], seq_len_kv));
         }
+        // TMA alignment requirements for SF KV
         start = start / 4 * 4;
         return {(q_iter_idx + q_iter_offset) % kNumQStages,       // Q pipeline stage
                 ((q_iter_idx + q_iter_offset) / kNumQStages) & 1, // Q pipeline phase
-                start, ceil_div(end - start, BLOCK_KV)};          // Task info
+                start, math::ceil_div(end - start, BLOCK_KV)};          // Task info
     };
 
     // KV pipeline
     uint32_t num_total_kv_blocks = 0;
-    const auto& get_kv_pipeline = [&](const uint32_t& kv_block_idx) -> cute::tuple<uint32_t, uint32_t> {
+    const auto get_kv_pipeline = [&](const uint32_t& kv_block_idx) -> cute::tuple<uint32_t, uint32_t> {
         return {
             (num_total_kv_blocks + kv_block_idx) % kNumKVStages,         // KV pipeline stage
             ((num_total_kv_blocks + kv_block_idx) / kNumKVStages) & 1    // KV pipeline phase
         };
     };
 
+    // Wait for primary kernel completion
+    cudaGridDependencySynchronize();
+
     if (threadIdx.x >= kNumMathThreads) {
         // TMA warp-group for loading data
         cutlass::arch::warpgroup_reg_dealloc<kNumTMARegisters>();
@@ -165,8 +165,8 @@ void sm90_fp8_mqa_logits(const uint32_t seq_len, const uint32_t seq_len_kv,
 
         // Prefetch
         const auto& issue_tma_q = [&](const uint32_t& stage_idx, const auto& block_idx) {
-            tma_copy<kHeadDim, BLOCK_Q * kNumHeads, kHeadDim>(&tensor_map_q, full_q_barriers[stage_idx], smem_q[stage_idx], 0, block_idx * BLOCK_Q * kNumHeads);
-            tma_copy<kNumHeads, BLOCK_Q, 0>(&tensor_map_weights, full_q_barriers[stage_idx], smem_weights[stage_idx], 0, block_idx * BLOCK_Q);
+            tma::copy<kHeadDim, BLOCK_Q * kNumHeads, kHeadDim>(&tensor_map_q, full_q_barriers[stage_idx], smem_q[stage_idx], 0, block_idx * BLOCK_Q * kNumHeads);
+            tma::copy<kNumHeads, BLOCK_Q, 0>(&tensor_map_weights, full_q_barriers[stage_idx], smem_weights[stage_idx], 0, block_idx * BLOCK_Q);
             full_q_barriers[stage_idx]->arrive_and_expect_tx(SMEM_Q_SIZE_PER_STAGE + SMEM_WEIGHT_SIZE_PER_STAGE);
         };
         if (cute::elect_one_sync() and block_q_idx < num_q_blocks)
@@ -192,9 +192,9 @@ void sm90_fp8_mqa_logits(const uint32_t seq_len, const uint32_t seq_len_kv,
                     empty_kv_barriers[kv_stage_idx]->wait(kv_phase ^ 1);
 
                     // Issue TMA KV
-                    tma_copy<kHeadDim, BLOCK_KV, kHeadDim>(&tensor_map_kv, full_kv_barriers[kv_stage_idx],
+                    tma::copy<kHeadDim, BLOCK_KV, kHeadDim>(&tensor_map_kv, full_kv_barriers[kv_stage_idx],
                              smem_kv[kv_stage_idx], 0, kv_start + kv_block_idx * BLOCK_KV);
-                    tma_copy<BLOCK_KV, 1, 0>(&tensor_map_kv_scales, full_kv_barriers[kv_stage_idx],
+                    tma::copy<BLOCK_KV, 1, 0>(&tensor_map_kv_scales, full_kv_barriers[kv_stage_idx],
                              smem_kv_scales[kv_stage_idx], kv_start + kv_block_idx * BLOCK_KV, 0);
                     full_kv_barriers[kv_stage_idx]->arrive_and_expect_tx(SMEM_KV_SIZE_PER_STAGE + SMEM_KV_SCALE_SIZE_PER_STAGE);
                 }
@@ -212,7 +212,7 @@ void sm90_fp8_mqa_logits(const uint32_t seq_len, const uint32_t seq_len_kv,
         const auto& thread_idx = threadIdx.x % kNumMathThreads;
         const auto& warp_idx = __shfl_sync(0xffffffff, thread_idx / 32, 0);
         const auto& warpgroup_idx = warp_idx / 4;
-        const auto& lane_idx = get_lane_idx();
+        const auto& lane_idx = ptx::get_lane_idx();
         float accum[WGMMA::kNumAccum], weights[BLOCK_Q][kNumHeads / 4];
 
         const auto& warp_offset = warp_idx * 16;
@@ -230,7 +230,7 @@ void sm90_fp8_mqa_logits(const uint32_t seq_len, const uint32_t seq_len_kv,
             for (uint32_t i = 0; i < BLOCK_Q; ++ i) {
                 #pragma unroll
                 for (uint32_t j = 0; j < kNumHeads / 4; ++ j)
-                    weights[i][j] = ld_shared(smem_weights[q_stage_idx] + i * kNumHeads + (j / 2) * 8 + (j & 1) + (lane_idx % 4) * 2);
+                    weights[i][j] = ptx::ld_shared(smem_weights[q_stage_idx] + i * kNumHeads + (j / 2) * 8 + (j & 1) + (lane_idx % 4) * 2);
             }
 
             // Compute over KV blocks
@@ -242,29 +242,31 @@ void sm90_fp8_mqa_logits(const uint32_t seq_len, const uint32_t seq_len_kv,
                 full_kv_barriers[kv_stage_idx]->wait(kv_phase);
 
                 // Read per-KV scales
-                float scale_kv_0 = ld_shared(smem_kv_scales[kv_stage_idx] + warp_offset + v_0_offset);
-                float scale_kv_1 = ld_shared(smem_kv_scales[kv_stage_idx] + warp_offset + v_1_offset);
+                float scale_kv_0 = ptx::ld_shared(smem_kv_scales[kv_stage_idx] + warp_offset + v_0_offset);
+                float scale_kv_1 = ptx::ld_shared(smem_kv_scales[kv_stage_idx] + warp_offset + v_1_offset);
 
                 // Issue WGMMA
                 DG_STATIC_ASSERT(BLOCK_KV == kNumMathThreads / 2, "Invalid block size");
                 DG_STATIC_ASSERT(kHeadDim % WGMMA::K == 0, "Invalid head dim");
                 #pragma unroll
                 for (uint32_t i = 0; i < WGMMA::kNumAccum; ++ i)
-                    warpgroup_fence_operand(accum[i]);
-                warpgroup_arrive();
+                    ptx::warpgroup_fence_operand(accum[i]);
+                ptx::warpgroup_arrive();
                 #pragma unroll
                 for (uint32_t k = 0; k < kHeadDim / WGMMA::K; ++ k) {
-                    auto desc_a = make_smem_desc(smem_kv[kv_stage_idx] + (warpgroup_idx * WGMMA::M) * kHeadDim + k * WGMMA::K,
-                                                 to_swizzle_cute_type<kHeadDim>(), 0, kHeadDim * 8);
-                    auto desc_b = make_smem_desc(smem_q[q_stage_idx] + k * WGMMA::K,
-                                                 to_swizzle_cute_type<kHeadDim>(), 0, kHeadDim * 8);
+                    auto desc_a = mma::sm90::make_smem_desc(
+                        smem_kv[kv_stage_idx] + (warpgroup_idx * WGMMA::M) * kHeadDim + k * WGMMA::K,
+                        mma::sm90::to_swizzle_cute_type<kHeadDim>(), 0, kHeadDim * 8);
+                    auto desc_b = mma::sm90::make_smem_desc(
+                        smem_q[q_stage_idx] + k * WGMMA::K,
+                        mma::sm90::to_swizzle_cute_type<kHeadDim>(), 0, kHeadDim * 8);
                     WGMMA::wgmma(desc_a, desc_b, accum, k);
                 }
-                warpgroup_commit_batch();
+                ptx::warpgroup_commit_batch();
                 #pragma unroll
                 for (uint32_t i = 0; i < WGMMA::kNumAccum; ++ i)
-                    warpgroup_fence_operand(accum[i]);
-                warpgroup_wait<0>();
+                    ptx::warpgroup_fence_operand(accum[i]);
+                ptx::warpgroup_wait<0>();
 
                 // Release KV empty
                 empty_kv_barriers[kv_stage_idx]->arrive();
@@ -278,7 +280,7 @@ void sm90_fp8_mqa_logits(const uint32_t seq_len, const uint32_t seq_len_kv,
                 #pragma unroll
                 for (uint32_t i = 0; i < BLOCK_Q; ++ i) {
                     auto shifted_accum = accum + i * kNumAccumPerReduce;
-                    const auto& transform = [&](const uint32_t& j) {
+                    const auto transform = [&](const uint32_t& j) {
                         return fmaxf(shifted_accum[j], 0) * weights[i][(j / 4) * 2 + (j & 1)];
                     };
 
@@ -302,16 +304,15 @@ void sm90_fp8_mqa_logits(const uint32_t seq_len, const uint32_t seq_len_kv,
                     }
 
                     // Store into the global memory
-                    // NOTES: we have redundant writes here, consider more carefully
-                    const uint32_t& q_idx = block_q_idx * BLOCK_Q + i;
+                    const auto q_offset = (block_q_idx * BLOCK_Q + i) * static_cast<uint64_t>(stride_logits);
                     if constexpr (kIsCompressedLogits) {
                         if (seq_k_start[i] <= kv_offset + v_0_offset and kv_offset + v_0_offset < seq_k_end[i])
-                            logits[q_idx * stride_logits + kv_offset + v_0_offset - seq_k_start[i]] = v_0;
+                            logits[q_offset + kv_offset + v_0_offset - seq_k_start[i]] = static_cast<logits_dtype_t>(v_0);
                         if (seq_k_start[i] <= kv_offset + v_1_offset and kv_offset + v_1_offset < seq_k_end[i])
-                            logits[q_idx * stride_logits + kv_offset + v_1_offset - seq_k_start[i]] = v_1;
+                            logits[q_offset + kv_offset + v_1_offset - seq_k_start[i]] = static_cast<logits_dtype_t>(v_1);
                     } else {
-                        logits[q_idx * stride_logits + kv_offset + v_0_offset] = v_0;
-                        logits[q_idx * stride_logits + kv_offset + v_1_offset] = v_1;
+                        logits[q_offset + kv_offset + v_0_offset] = static_cast<logits_dtype_t>(v_0);
+                        logits[q_offset + kv_offset + v_1_offset] = static_cast<logits_dtype_t>(v_1);
                     }
                 }
             }

deep_gemm/include/deep_gemm/impls/sm90_fp8_paged_mqa_logits.cuh

@@ -6,133 +6,43 @@
 #include <cute/arch/cluster_sm90.hpp>
 #include <cute/arch/copy_sm90_desc.hpp>
 
+#include <deep_gemm/common/cute_tie.cuh>
+#include <deep_gemm/common/math.cuh>
 #include <deep_gemm/common/utils.cuh>
-#include <deep_gemm/common/sm90_utils.cuh>
-#include <deep_gemm/impls/sm90_fp8_mqa_logits.cuh>
+#include <deep_gemm/common/tma_copy.cuh>
+#include <deep_gemm/common/types.cuh>
+#include <deep_gemm/mma/sm90.cuh>
+#include <deep_gemm/ptx/ld_st.cuh>
+#include <deep_gemm/ptx/utils.cuh>
+#include <deep_gemm/ptx/wgmma.cuh>
+#include <deep_gemm/scheduler/paged_mqa_logits.cuh>
 
 namespace deep_gemm {
 
-template <uint32_t kAlignedBatchSize, uint32_t SPLIT_KV, uint32_t kNumSMs>
-__global__ __launch_bounds__(32, 1)
-void smxx_paged_mqa_logits_metadata(const uint32_t batch_size, const uint32_t next_n, const bool is_context_lens_2d,
-                                    const uint32_t* context_lens, uint32_t* schedule_metadata) {
-    DG_STATIC_ASSERT(kAlignedBatchSize % 32 == 0, "Invalid aligned batch size");
-    const uint32_t lane_idx = get_lane_idx();
-
-    uint32_t num_segs[kAlignedBatchSize / 32];
-    #pragma unroll
-    for (uint32_t k = 0; k < kAlignedBatchSize / 32; ++ k) {
-        const uint32_t q_idx = k * 32 + lane_idx;
-        const uint32_t lens_idx = (is_context_lens_2d ? q_idx * next_n + next_n - 1 : q_idx);
-        const uint32_t& context_len = (q_idx < batch_size ? __ldg(context_lens + lens_idx) : 0);
-        num_segs[k] = ceil_div(context_len, SPLIT_KV);
-    }
-
-    __shared__ uint32_t prefix_sum[kAlignedBatchSize];
-    uint32_t sum = 0;
-    #pragma unroll
-    for (uint32_t k = 0; k < kAlignedBatchSize / 32; ++ k) {
-        uint32_t x = num_segs[k];
-        #pragma unroll
-        for (uint32_t offset = 1; offset < 32; offset <<= 1) {
-            const uint32_t& y = __shfl_up_sync(0xffffffff, x, offset);
-            x += (lane_idx >= offset ? y : 0);
-        }
-        x += sum;
-        prefix_sum[k * 32 + lane_idx] = x;
-        sum = __shfl_sync(0xffffffff, x, 31);
-    }
-
-    const uint32_t& q = sum / kNumSMs, r = sum % kNumSMs;
-    for (uint32_t sm_idx = lane_idx; sm_idx <= kNumSMs; sm_idx += 32) {
-        uint32_t seg_starts = sm_idx * q + min(sm_idx, r);
-        uint32_t q_idx = 0;
-        while (q_idx < batch_size and prefix_sum[q_idx] <= seg_starts)
-            ++ q_idx;
-        const uint32_t& kv_split_idx = (q_idx == 0 ? seg_starts : seg_starts - prefix_sum[q_idx - 1]);
-        __syncwarp();
-
-        schedule_metadata[sm_idx * 2] = q_idx;
-        schedule_metadata[sm_idx * 2 + 1] = kv_split_idx;
-    }
-}
-
-template <uint32_t kNextN, bool kIsContextLens2D,
-          uint32_t BLOCK_KV, uint32_t kNumBlocksPerSplit>
-struct PagedMQALogitsScheduler {
-    uint32_t batch_size;
-    const uint32_t* context_lens;
-
-    uint32_t current_q_idx, current_kv_idx;
-    uint32_t end_q_idx, end_kv_idx;
-    uint32_t current_num_kv;
-
-    __device__ __forceinline__ uint32_t get_num_kv(const uint32_t& q_idx) {
-        const auto& lens_idx = (kIsContextLens2D ? q_idx * kNextN + kNextN - 1 : q_idx);
-        return q_idx < batch_size ? ceil_div(__ldg(context_lens + lens_idx), BLOCK_KV) : 0;
-    }
-
-    __device__ __forceinline__ explicit PagedMQALogitsScheduler(const uint32_t& batch_size, const uint32_t& sm_idx,
-                                                                const uint32_t* context_lens, const uint32_t* schedule_meta) {
-        this->batch_size = batch_size;
-        this->context_lens = context_lens;
-
-        const auto& current_pack = __ldg(reinterpret_cast<const uint2*>(schedule_meta) + sm_idx);
-        const auto& end_pack = __ldg(reinterpret_cast<const uint2*>(schedule_meta) + sm_idx + 1);
-        current_q_idx = current_pack.x, current_kv_idx = current_pack.y * kNumBlocksPerSplit;
-        end_q_idx = end_pack.x, end_kv_idx = end_pack.y * kNumBlocksPerSplit;
-
-        current_num_kv = get_num_kv(current_q_idx);
-    }
-
-    __device__ __forceinline__ bool fetch_next_task(uint32_t &q_idx, uint32_t &kv_idx, uint32_t &num_kv) {
-        q_idx = current_q_idx;
-        kv_idx = current_kv_idx;
-        num_kv = current_num_kv;
-
-        if (q_idx == end_q_idx and kv_idx == end_kv_idx)
-            return false;
-
-        current_kv_idx += kNumBlocksPerSplit;
-        if (current_kv_idx >= current_num_kv) {
-            ++ current_q_idx;
-            current_kv_idx = 0;
-            current_num_kv = get_num_kv(current_q_idx);
-        }
-
-        return true;
-    }
-
-    __device__ __forceinline__ bool exist_q_idx(const uint32_t& q_idx) const {
-        return q_idx < end_q_idx or q_idx == end_q_idx and 0 < end_kv_idx;
-    }
-};
-
-using namespace deep_gemm::sm90;
-
 template <uint32_t kNextN, uint32_t kNumHeads,
           uint32_t kHeadDim, uint32_t BLOCK_KV,
           bool kIsContextLens2D,
           uint32_t kNumQStages, uint32_t kNumKVStages,
           uint32_t SPLIT_KV,
-          uint32_t kNumTMAThreads, uint32_t kNumMathThreads>
-__global__ __launch_bounds__(kNumTMAThreads + kNumMathThreads, 1)
+          uint32_t kNumTMAThreads, uint32_t kNumMathThreads,
+          typename logits_dtype_t>
+CUTLASS_GLOBAL __launch_bounds__(kNumTMAThreads + kNumMathThreads, 1)
 void sm90_fp8_paged_mqa_logits(const uint32_t batch_size,
-                               const uint64_t logits_stride, const uint64_t block_table_stride,
-                               const uint32_t* context_lens, float* logits,
+                               const uint32_t logits_stride, const uint32_t block_table_stride,
+                               const uint32_t* context_lens, logits_dtype_t* logits,
                                const uint32_t* block_table, const uint32_t* schedule_meta,
                                const __grid_constant__ cute::TmaDescriptor tensor_map_q,
                                const __grid_constant__ cute::TmaDescriptor tensor_map_kv,
                                const __grid_constant__ cute::TmaDescriptor tensor_map_kv_scales,
                                const __grid_constant__ cute::TmaDescriptor tensor_map_weights) {
     // Types
-    using WGMMA = typename FP8MMASelector<kNextN * kNumHeads>::type;
+    using WGMMA = typename mma::sm90::FP8MMASelector<kNextN * kNumHeads>::type;
     using Barrier = cutlass::arch::ClusterTransactionBarrier;
 
     // NOTES: use `__shfl_sync` to encourage NVCC to use unified registers
-    const auto& warp_idx = __shfl_sync(0xffffffff, threadIdx.x / 32, 0);
-    const auto& warpgroup_idx = warp_idx / 4;
-    const auto& lane_idx = get_lane_idx();
+    const auto warp_idx = __shfl_sync(0xffffffff, threadIdx.x / 32, 0);
+    const auto warpgroup_idx = warp_idx / 4;
+    const auto lane_idx = ptx::get_lane_idx();
 
     // Prefetch TMA descriptors
     static constexpr uint32_t kNumMathWarpGroups = kNumMathThreads / 128;
@@ -150,15 +60,15 @@ void sm90_fp8_paged_mqa_logits(const uint32_t batch_size,
     static constexpr uint32_t kSwizzleAlignment = kHeadDim * 8;
     static constexpr uint32_t SMEM_Q_SIZE_PER_STAGE = kNextN * kNumHeads * kHeadDim * sizeof(__nv_fp8_e4m3);
     static constexpr uint32_t SMEM_WEIGHT_SIZE_PER_STAGE = kNextN * kNumHeads * sizeof(float);
-    static constexpr uint32_t ALIGNED_SMEM_WEIGHT_SIZE_PER_STAGE = constexpr_align(SMEM_WEIGHT_SIZE_PER_STAGE, kSwizzleAlignment);
+    static constexpr uint32_t ALIGNED_SMEM_WEIGHT_SIZE_PER_STAGE = math::constexpr_align(SMEM_WEIGHT_SIZE_PER_STAGE, kSwizzleAlignment);
     static constexpr uint32_t SMEM_Q_PIPE_SIZE = kNumQStages * (SMEM_Q_SIZE_PER_STAGE + ALIGNED_SMEM_WEIGHT_SIZE_PER_STAGE) +
-                                                 constexpr_align(kNumQStages * 8 * 2, kSwizzleAlignment);
+                                                 math::constexpr_align(kNumQStages * 8 * 2, kSwizzleAlignment);
 
     static constexpr uint32_t SMEM_KV_SIZE_PER_STAGE = BLOCK_KV * kHeadDim * sizeof(__nv_fp8_e4m3);
     static constexpr uint32_t SMEM_KV_SCALE_SIZE_PER_STAGE = BLOCK_KV * sizeof(float);
-    static constexpr uint32_t ALIGNED_SMEM_KV_SCALE_SIZE_PER_STAGE = constexpr_align(SMEM_KV_SCALE_SIZE_PER_STAGE, kSwizzleAlignment);
+    static constexpr uint32_t ALIGNED_SMEM_KV_SCALE_SIZE_PER_STAGE = math::constexpr_align(SMEM_KV_SCALE_SIZE_PER_STAGE, kSwizzleAlignment);
     static constexpr uint32_t SMEM_KV_PIPE_SIZE = kNumKVStages * (SMEM_KV_SIZE_PER_STAGE + ALIGNED_SMEM_KV_SCALE_SIZE_PER_STAGE) +
-                                                  constexpr_align(kNumKVStages * 8 * 2, kSwizzleAlignment);
+                                                  math::constexpr_align(kNumKVStages * 8 * 2, kSwizzleAlignment);
 
     // Align to swizzling alignment bytes
     extern __shared__ __align__(kSwizzleAlignment) uint8_t smem_buffer[];
@@ -166,31 +76,31 @@ void sm90_fp8_paged_mqa_logits(const uint32_t batch_size,
     DG_STATIC_ASSERT(SMEM_KV_SIZE_PER_STAGE % kSwizzleAlignment == 0, "Unaligned TMA swizzling");
 
     // Q data and barriers on shared memory
-    auto smem_q = PatternVisitor([&](const uint32_t& i) {
+    auto smem_q = utils::PatternVisitor([&](const uint32_t& i) {
         return reinterpret_cast<__nv_fp8_e4m3*>(smem_buffer + SMEM_Q_SIZE_PER_STAGE * i);
     });
-    auto smem_weights = PatternVisitor([&](const uint32_t& i) {
+    auto smem_weights = utils::PatternVisitor([&](const uint32_t& i) {
         return reinterpret_cast<float*>(smem_buffer + SMEM_Q_SIZE_PER_STAGE * kNumQStages + ALIGNED_SMEM_WEIGHT_SIZE_PER_STAGE * i);
     });
     auto q_barrier_ptr = reinterpret_cast<Barrier*>(smem_weights[kNumQStages]);
-    auto full_q_barriers  = PatternVisitor([&](const uint32_t& i) { return q_barrier_ptr + i; });
-    auto empty_q_barriers = PatternVisitor([&](const uint32_t& i) { return q_barrier_ptr + (kNumQStages + i); });
+    auto full_q_barriers  = utils::PatternVisitor([&](const uint32_t& i) { return q_barrier_ptr + i; });
+    auto empty_q_barriers = utils::PatternVisitor([&](const uint32_t& i) { return q_barrier_ptr + (kNumQStages + i); });
 
     // Separate math warpgroups and tma load warps into KV groups
     // Each math warpgroup corresponds to a tma load warp
-    const auto& kv_group_idx = __shfl_sync(0xffffffff, threadIdx.x >= kNumMathThreads ? (threadIdx.x - kNumMathThreads) / 32 : warpgroup_idx, 0);
+    const auto kv_group_idx = __shfl_sync(0xffffffff, threadIdx.x >= kNumMathThreads ? (threadIdx.x - kNumMathThreads) / 32 : warpgroup_idx, 0);
 
     // Per group KV data and barriers on shared memory
-    const auto& smem_offset = SMEM_Q_PIPE_SIZE + SMEM_KV_PIPE_SIZE * kv_group_idx;
-    auto smem_kv = PatternVisitor([&](const uint32_t& i) {
+    const auto smem_offset = SMEM_Q_PIPE_SIZE + SMEM_KV_PIPE_SIZE * kv_group_idx;
+    auto smem_kv = utils::PatternVisitor([&](const uint32_t& i) {
         return reinterpret_cast<__nv_fp8_e4m3*>(smem_buffer + smem_offset + SMEM_KV_SIZE_PER_STAGE * i);
     });
-    auto smem_kv_scales =  PatternVisitor([&](const uint32_t& i) {
+    auto smem_kv_scales = utils::PatternVisitor([&](const uint32_t& i) {
         return reinterpret_cast<float*>(smem_buffer + smem_offset + SMEM_KV_SIZE_PER_STAGE * kNumKVStages + ALIGNED_SMEM_KV_SCALE_SIZE_PER_STAGE * i);
     });
     auto kv_barrier_ptr = reinterpret_cast<Barrier*>(smem_kv_scales[kNumKVStages]);
-    auto full_kv_barriers  = PatternVisitor([&](const uint32_t& i) { return kv_barrier_ptr + i; });
-    auto empty_kv_barriers = PatternVisitor([&](const uint32_t& i) { return kv_barrier_ptr + kNumKVStages + i; });
+    auto full_kv_barriers  = utils::PatternVisitor([&](const uint32_t& i) { return kv_barrier_ptr + i; });
+    auto empty_kv_barriers = utils::PatternVisitor([&](const uint32_t& i) { return kv_barrier_ptr + kNumKVStages + i; });
 
     // Initialize barriers
     if (warp_idx >= kNumMathThreads / 32 and cute::elect_one_sync()) {
@@ -218,15 +128,19 @@ void sm90_fp8_paged_mqa_logits(const uint32_t batch_size,
     constexpr uint32_t kNumTMARegisters = 64;
     constexpr uint32_t kNumMathRegisters = 104;
 
+    // Wait for primary kernel completion
+    cudaGridDependencySynchronize();
+
     // Scheduler
-    auto scheduler = PagedMQALogitsScheduler<kNextN, kIsContextLens2D, BLOCK_KV, kNumMathWarpGroups>(batch_size, blockIdx.x, context_lens, schedule_meta);
+    auto scheduler = sched::PagedMQALogitsScheduler<kNextN, kIsContextLens2D, BLOCK_KV, kNumMathWarpGroups, 1>(
+        blockIdx.x, context_lens, schedule_meta);
     DG_STATIC_ASSERT(SPLIT_KV % BLOCK_KV == 0, "Unaligned SPLIT_KV");
 
     // Q and KV pipeline
-    const auto& get_q_pipeline = [=](const uint32_t& q_iter_idx) -> cute::tuple<uint32_t, uint32_t> {
+    const auto get_q_pipeline = [=](const uint32_t& q_iter_idx) -> cute::tuple<uint32_t, uint32_t> {
         return {q_iter_idx % kNumQStages, (q_iter_idx / kNumQStages) & 1}; // Q pipeline stage and phase
     };
-    const auto& get_kv_pipeline = [=](const uint32_t& kv_iter_idx) -> cute::tuple<uint32_t, uint32_t> {
+    const auto get_kv_pipeline = [=](const uint32_t& kv_iter_idx) -> cute::tuple<uint32_t, uint32_t> {
         return {kv_iter_idx % kNumKVStages, (kv_iter_idx / kNumKVStages) & 1}; // KV pipeline stage and phase
     };
     uint32_t q_iter_idx = 0, kv_iter_idx = 0;
@@ -237,10 +151,10 @@ void sm90_fp8_paged_mqa_logits(const uint32_t batch_size,
         if (kv_group_idx >= kNumMathWarpGroups)
             return;
 
-        const auto& issue_tma_q = [&](const uint32_t& stage_idx, const uint32_t& q_idx) {
+        const auto issue_tma_q = [&](const uint32_t& stage_idx, const uint32_t& q_idx) {
             if (kv_group_idx == 0 and cute::elect_one_sync()) {
-                tma_copy<kHeadDim, kNextN * kNumHeads, kHeadDim>(&tensor_map_q, full_q_barriers[stage_idx], smem_q[stage_idx], 0, q_idx * kNextN * kNumHeads);
-                tma_copy<kNextN * kNumHeads, 1, 0>(&tensor_map_weights, full_q_barriers[stage_idx], smem_weights[stage_idx], 0, q_idx);
+                tma::copy<kHeadDim, kNextN * kNumHeads, kHeadDim>(&tensor_map_q, full_q_barriers[stage_idx], smem_q[stage_idx], 0, q_idx * kNextN * kNumHeads);
+                tma::copy<kNextN * kNumHeads, 1, 0>(&tensor_map_weights, full_q_barriers[stage_idx], smem_weights[stage_idx], 0, q_idx * kNextN);
                 full_q_barriers[stage_idx]->arrive_and_expect_tx(SMEM_Q_SIZE_PER_STAGE + SMEM_WEIGHT_SIZE_PER_STAGE);
             }
         };
@@ -259,7 +173,7 @@ void sm90_fp8_paged_mqa_logits(const uint32_t batch_size,
 
         while (fetched_next_task) {
             // Prefetch next Q when current Q changes
-            bool prefetch_q = (q_idx != next_q_idx and scheduler.exist_q_idx(next_q_idx + 1));
+            bool prefetch_q = (q_idx != next_q_idx and scheduler.exist_q_atom_idx(next_q_idx + 1));
             q_idx = next_q_idx;
             kv_idx = next_kv_idx;
             num_kv = next_num_kv;
@@ -276,9 +190,9 @@ void sm90_fp8_paged_mqa_logits(const uint32_t batch_size,
             if (kv_idx == 0 or kv_block_idx_ptr == 32) {
                 kv_block_idx_ptr = 0;
                 kv_block_idx_storage = (kv_idx + kv_group_idx + lane_idx * kNumMathWarpGroups < num_kv ?
-                    __ldg(block_table + q_idx * block_table_stride + (kv_idx + kv_group_idx + lane_idx * kNumMathWarpGroups)) : 0);
+                    block_table[q_idx * static_cast<uint64_t>(block_table_stride) + (kv_idx + kv_group_idx + lane_idx * kNumMathWarpGroups)] : 0);
             }
-            const auto& kv_block_idx = __shfl_sync(0xffffffff, kv_block_idx_storage, kv_block_idx_ptr ++);
+            const auto kv_block_idx = __shfl_sync(0xffffffff, kv_block_idx_storage, kv_block_idx_ptr ++);
 
             // Wait KV consumer release
             CUTE_TIE_DECL(get_kv_pipeline(kv_iter_idx ++), kv_stage_idx, kv_phase);
@@ -286,10 +200,10 @@ void sm90_fp8_paged_mqa_logits(const uint32_t batch_size,
 
             // Issue TMA KV
             if (cute::elect_one_sync()) {
-                tma_copy<kHeadDim, BLOCK_KV, 0, __nv_fp8_e4m3, true>(&tensor_map_kv, full_kv_barriers[kv_stage_idx],
-                                                                     smem_kv[kv_stage_idx], 0, 0, 1, kv_block_idx);
-                tma_copy<BLOCK_KV, 1, 0>(&tensor_map_kv_scales, full_kv_barriers[kv_stage_idx],
-                                         smem_kv_scales[kv_stage_idx], 0, kv_block_idx);
+                tma::copy<kHeadDim, BLOCK_KV, 0, __nv_fp8_e4m3, true>(&tensor_map_kv, full_kv_barriers[kv_stage_idx],
+                                                                      smem_kv[kv_stage_idx], 0, 0, 1, kv_block_idx);
+                tma::copy<BLOCK_KV, 1, 0>(&tensor_map_kv_scales, full_kv_barriers[kv_stage_idx],
+                                          smem_kv_scales[kv_stage_idx], 0, kv_block_idx);
                 full_kv_barriers[kv_stage_idx]->arrive_and_expect_tx(SMEM_KV_SIZE_PER_STAGE + SMEM_KV_SCALE_SIZE_PER_STAGE);
             }
 
@@ -301,9 +215,9 @@ void sm90_fp8_paged_mqa_logits(const uint32_t batch_size,
         cutlass::arch::warpgroup_reg_alloc<kNumMathRegisters>();
 
         float accum[WGMMA::kNumAccum], weights[kNextN][kNumHeads / 4];
-        const auto& sub_warp_offset = (warp_idx % 4) * 16;
-        const auto& v_0_offset = lane_idx / 4 + 0;
-        const auto& v_1_offset = lane_idx / 4 + 8;
+        const auto sub_warp_offset = (warp_idx % 4) * 16;
+        const auto v_0_offset = lane_idx / 4 + 0;
+        const auto v_1_offset = lane_idx / 4 + 8;
 
         // Initialize `q_idx` outside `[0, batch_size)` to indicate it was none
         uint32_t q_idx = batch_size, kv_idx;
@@ -326,7 +240,7 @@ void sm90_fp8_paged_mqa_logits(const uint32_t batch_size,
                 for (uint32_t i = 0; i < kNextN; ++ i) {
                     #pragma unroll
                     for (uint32_t j = 0; j < kNumHeads / 4; ++ j)
-                        weights[i][j] = ld_shared(smem_weights[q_stage_idx] + i * kNumHeads + (j / 2) * 8 + (j & 1) + (lane_idx % 4) * 2);
+                        weights[i][j] = ptx::ld_shared(smem_weights[q_stage_idx] + i * kNumHeads + (j / 2) * 8 + (j & 1) + (lane_idx % 4) * 2);
                 }
             }
 
@@ -335,7 +249,7 @@ void sm90_fp8_paged_mqa_logits(const uint32_t batch_size,
             kv_idx = next_kv_idx;
 
             // Calculate KV offset in advance
-            auto kv_offset = q_idx * kNextN * logits_stride + ((kv_idx + kv_group_idx) * BLOCK_KV + sub_warp_offset);
+            auto kv_offset = q_idx * kNextN * static_cast<uint64_t>(logits_stride) + ((kv_idx + kv_group_idx) * BLOCK_KV + sub_warp_offset);
 
             // Compute `[kNextN * kNumHeads, kHeadDim] @ [BLOCK_KV, kHeadDim] -> [kNextN, BLOCK_KV]`
             // Wait TMA KV arrival
@@ -347,25 +261,29 @@ void sm90_fp8_paged_mqa_logits(const uint32_t batch_size,
             DG_STATIC_ASSERT(kHeadDim % WGMMA::K == 0, "Invalid head dim");
             #pragma unroll
             for (uint32_t i = 0; i < WGMMA::kNumAccum; ++ i)
-                warpgroup_fence_operand(accum[i]);
-            warpgroup_arrive();
+                ptx::warpgroup_fence_operand(accum[i]);
+            ptx::warpgroup_arrive();
             #pragma unroll
             for (uint32_t k = 0; k < kHeadDim / WGMMA::K; ++ k) {
-                auto desc_a = make_smem_desc(smem_kv[kv_stage_idx] + k * WGMMA::K, to_swizzle_cute_type<kHeadDim>(), 0, kHeadDim * 8);
-                auto desc_b = make_smem_desc(smem_q[q_stage_idx] + k * WGMMA::K, to_swizzle_cute_type<kHeadDim>(), 0, kHeadDim * 8);
+                auto desc_a = mma::sm90::make_smem_desc(
+                    smem_kv[kv_stage_idx] + k * WGMMA::K,
+                    mma::sm90::to_swizzle_cute_type<kHeadDim>(), 0, kHeadDim * 8);
+                auto desc_b = mma::sm90::make_smem_desc(
+                    smem_q[q_stage_idx] + k * WGMMA::K,
+                    mma::sm90::to_swizzle_cute_type<kHeadDim>(), 0, kHeadDim * 8);
                 WGMMA::wgmma(desc_a, desc_b, accum, k);
             }
-            warpgroup_commit_batch();
+            ptx::warpgroup_commit_batch();
             #pragma unroll
             for (uint32_t i = 0; i < WGMMA::kNumAccum; ++ i)
-                warpgroup_fence_operand(accum[i]);
+                ptx::warpgroup_fence_operand(accum[i]);
 
             // Read per-KV scales
-            float scale_kv_0 = ld_shared(smem_kv_scales[kv_stage_idx] + sub_warp_offset + v_0_offset);
-            float scale_kv_1 = ld_shared(smem_kv_scales[kv_stage_idx] + sub_warp_offset + v_1_offset);
+            float scale_kv_0 = ptx::ld_shared(smem_kv_scales[kv_stage_idx] + sub_warp_offset + v_0_offset);
+            float scale_kv_1 = ptx::ld_shared(smem_kv_scales[kv_stage_idx] + sub_warp_offset + v_1_offset);
 
             // Wait WGMMA
-            warpgroup_wait<0>();
+            ptx::warpgroup_wait<0>();
 
             // Release KV empty
             empty_kv_barriers[kv_stage_idx]->arrive();
@@ -378,7 +296,7 @@ void sm90_fp8_paged_mqa_logits(const uint32_t batch_size,
             #pragma unroll
             for (uint32_t i = 0; i < kNextN; ++ i) {
                 auto shifted_accum = accum + i * kNumAccumPerReduce;
-                const auto& transform = [&](const uint32_t& j) {
+                const auto transform = [&](const uint32_t& j) {
                     return fmaxf(shifted_accum[j], 0) * weights[i][(j / 4) * 2 + (j & 1)];
                 };
 
@@ -396,15 +314,15 @@ void sm90_fp8_paged_mqa_logits(const uint32_t batch_size,
                 // Inter-thread reduction
                 #pragma unroll
                 for (uint32_t j = 0; j < 2; ++ j) {
-                    const auto& offset = static_cast<int>(1u << j);
+                    const auto offset = static_cast<int>(1u << j);
                     v_0 += __shfl_xor_sync(0xffffffffu, v_0, offset);
                     v_1 += __shfl_xor_sync(0xffffffffu, v_1, offset);
                 }
 
                 // Store into the global memory
                 // NOTES: we have redundant writes here, consider more carefully
-                logits[kv_offset + i * logits_stride + v_0_offset] = v_0;
-                logits[kv_offset + i * logits_stride + v_1_offset] = v_1;
+                logits[kv_offset + i * static_cast<uint64_t>(logits_stride) + v_0_offset] = static_cast<logits_dtype_t>(v_0);
+                logits[kv_offset + i * static_cast<uint64_t>(logits_stride) + v_1_offset] = static_cast<logits_dtype_t>(v_1);
             }
         }
     }

deep_gemm/include/deep_gemm/impls/sm90_tf32_hc_prenorm_gemm.cuh

@@ -5,20 +5,23 @@
 #include <cutlass/arch/barrier.h>
 #include <cutlass/arch/reg_reconfig.h>
 
-#include <deep_gemm/common/reduction.cuh>
+#include <deep_gemm/common/math.cuh>
 #include <deep_gemm/common/utils.cuh>
-#include <deep_gemm/common/sm90_utils.cuh>
+#include <deep_gemm/common/tma_copy.cuh>
+#include <deep_gemm/common/types.cuh>
+#include <deep_gemm/mma/sm90.cuh>
+#include <deep_gemm/ptx/ld_st.cuh>
+#include <deep_gemm/ptx/utils.cuh>
+#include <deep_gemm/ptx/wgmma.cuh>
 
 namespace deep_gemm {
 
-using namespace deep_gemm::sm90;
-
 template <uint32_t kSwizzleMode, uint32_t kSwizzleBase = 16>
-__device__ __forceinline__
+CUTLASS_DEVICE
 uint32_t get_swizzled_bank_group_idx(const uint32_t& offset, const uint32_t& lane_idx) {
     constexpr uint32_t kGroupsInSwizzleRange = kSwizzleMode / kSwizzleBase;
 
-    const auto& bank_group_idx = offset + lane_idx * kGroupsInSwizzleRange;
+    const auto bank_group_idx = offset + lane_idx * kGroupsInSwizzleRange;
 
     constexpr uint32_t kNumBankGroups = 128 / kSwizzleBase;
     constexpr bool kHasShortcut = kGroupsInSwizzleRange == kNumBankGroups;
@@ -35,7 +38,7 @@ template <uint32_t SHAPE_N, uint32_t SHAPE_K,
           uint32_t kSwizzleCDMode,
           uint32_t kNumStages,
           uint32_t kNumMathThreads, uint32_t kNumTMAThreads>
-__global__ void __launch_bounds__(kNumMathThreads + kNumTMAThreads, 1)
+CUTLASS_GLOBAL void __launch_bounds__(kNumMathThreads + kNumTMAThreads, 1)
 sm90_tf32_hc_prenorm_gemm_impl(const uint32_t shape_m,
                                const __grid_constant__ cute::TmaDescriptor tensor_map_a,
                                const __grid_constant__ cute::TmaDescriptor tensor_map_b,
@@ -56,7 +59,7 @@ sm90_tf32_hc_prenorm_gemm_impl(const uint32_t shape_m,
 
     // Utils
     const auto warp_idx = cutlass::canonical_warp_idx_sync();
-    const auto lane_idx = get_lane_idx();
+    const auto lane_idx = ptx::get_lane_idx();
 
     // Align to 1024 bytes for swizzle-128B
     extern __shared__ __align__(1024) uint8_t smem_buffer[];
@@ -76,17 +79,17 @@ sm90_tf32_hc_prenorm_gemm_impl(const uint32_t shape_m,
     // Data on shared memory (layout as ordered below)
     // Fill D/A/B pointers
     auto smem_cd = reinterpret_cast<float*>(smem_buffer);
-    auto smem_a = PatternVisitor([&](const uint32_t& i) {
+    auto smem_a = utils::PatternVisitor([&](const uint32_t& i) {
         return reinterpret_cast<nv_bfloat16*>(smem_buffer + (SMEM_CD_SIZE + i * SMEM_A_SIZE_PER_STAGE));
     });
-    auto smem_b = PatternVisitor([&](const uint32_t& i) {
+    auto smem_b = utils::PatternVisitor([&](const uint32_t& i) {
         return reinterpret_cast<float*>(smem_buffer + (SMEM_CD_SIZE + kNumStages * SMEM_A_SIZE_PER_STAGE + i * SMEM_B_SIZE_PER_STAGE));
     });
 
     // Fill barriers
     auto barrier_start_ptr = reinterpret_cast<Barrier*>(smem_buffer + SMEM_CD_SIZE + kNumStages * (SMEM_A_SIZE_PER_STAGE + SMEM_B_SIZE_PER_STAGE));
-    auto full_barriers           = PatternVisitor([=](const uint32_t& i) { return barrier_start_ptr + (i); });
-    auto empty_barriers          = PatternVisitor([=](const uint32_t& i) { return barrier_start_ptr + (kNumStages + i); });
+    auto full_barriers           = utils::PatternVisitor([=](const uint32_t& i) { return barrier_start_ptr + (i); });
+    auto empty_barriers          = utils::PatternVisitor([=](const uint32_t& i) { return barrier_start_ptr + (kNumStages + i); });
 
     // Initialize barriers
     if (warp_idx == 1 and cute::elect_one_sync()) {
@@ -101,7 +104,7 @@ sm90_tf32_hc_prenorm_gemm_impl(const uint32_t shape_m,
     }
     __syncthreads();
 
-    constexpr uint32_t kNumKBlocks = constexpr_ceil_div(SHAPE_K, BLOCK_K);
+    constexpr uint32_t kNumKBlocks = math::constexpr_ceil_div(SHAPE_K, BLOCK_K);
     constexpr uint32_t kNumKBlocksPerSplit = kNumKBlocks / kNumSplits;
     constexpr uint32_t kRemainKBlocks = kNumKBlocks % kNumSplits;
     const uint32_t block_idx = __shfl_sync(0xffffffff, blockIdx.x, 0);
@@ -113,12 +116,15 @@ sm90_tf32_hc_prenorm_gemm_impl(const uint32_t shape_m,
     constexpr uint32_t kNumTMARegisters = 40;
     constexpr uint32_t kNumMathRegisters = 256;
 
+    // Wait for primary kernel completion
+    cudaGridDependencySynchronize();
+
     // TMA load warp
     if (warp_idx == kNumMathThreads / 32 and cute::elect_one_sync()) {
         cutlass::arch::warpgroup_reg_dealloc<kNumTMARegisters>();
         for (uint32_t s = 0; s < num_total_stages; ++ s) {
             // Wait consumer release
-            const auto& stage_idx = s % kNumStages;
+            const auto stage_idx = s % kNumStages;
             empty_barriers[stage_idx]->wait(((s / kNumStages) & 1) ^ 1);
 
             // Compute offsets
@@ -126,8 +132,8 @@ sm90_tf32_hc_prenorm_gemm_impl(const uint32_t shape_m,
             uint32_t k_idx = k_offset + s * BLOCK_K;
 
             // Issue TMAs
-            tma_copy<BLOCK_K, BLOCK_M, kSwizzleAMode>(&tensor_map_a, full_barriers[stage_idx], smem_a[stage_idx], k_idx, m_idx);
-            tma_copy<BLOCK_K, BLOCK_N, kSwizzleBMode>(&tensor_map_b, full_barriers[stage_idx], smem_b[stage_idx], k_idx, 0);
+            tma::copy<BLOCK_K, BLOCK_M, kSwizzleAMode>(&tensor_map_a, full_barriers[stage_idx], smem_a[stage_idx], k_idx, m_idx);
+            tma::copy<BLOCK_K, BLOCK_N, kSwizzleBMode>(&tensor_map_b, full_barriers[stage_idx], smem_b[stage_idx], k_idx, 0);
 
             // Arrive at full barriers
             constexpr uint32_t kNumArrivalBytes = SMEM_A_SIZE_PER_STAGE + SMEM_B_SIZE_PER_STAGE;
@@ -135,7 +141,7 @@ sm90_tf32_hc_prenorm_gemm_impl(const uint32_t shape_m,
         }
 
         for (uint32_t s = num_total_stages; s < num_total_stages + kNumStages; ++ s) {
-            const auto& stage_idx = s % kNumStages;
+            const auto stage_idx = s % kNumStages;
             empty_barriers[stage_idx]->wait(((s / kNumStages) & 1) ^ 1);
         }
     } else if (warp_idx < kNumMathThreads / 32) {
@@ -148,7 +154,7 @@ sm90_tf32_hc_prenorm_gemm_impl(const uint32_t shape_m,
         constexpr uint32_t WGMMA_N = BLOCK_N;
         constexpr uint32_t WGMMA_K = 8;
 
-        using WGMMA = typename TF32MMASelector<WGMMA_N, true>::type;
+        using WGMMA = typename mma::sm90::TF32MMASelector<WGMMA_N, true>::type;
         float accum[WGMMA::kNumAccum] = {0};
 
         constexpr uint32_t kNumBankGroupBytes = 16;
@@ -196,14 +202,14 @@ sm90_tf32_hc_prenorm_gemm_impl(const uint32_t shape_m,
                 sqr_sum_acc_1 += a_float2_0.y * a_float2_0.y + a_float2_1.y * a_float2_1.y;
             }
 
-            warpgroup_wait<0>();
+            ptx::warpgroup_wait<0>();
             if (s > 0)
                 empty_barriers[(s - 1) % kNumStages]->arrive();
 
             #pragma unroll
             for (uint32_t i = 0; i < WGMMA::kNumAccum; ++ i)
-                warpgroup_fence_operand(accum[i]);
-            warpgroup_arrive();
+                ptx::warpgroup_fence_operand(accum[i]);
+            ptx::warpgroup_arrive();
 
             constexpr int kNumElemsInSwizzleRange = 128 / sizeof(float);
             constexpr uint32_t kNumWgmmaInSwizzleRange = kNumElemsInSwizzleRange / WGMMA::K;
@@ -213,18 +219,19 @@ sm90_tf32_hc_prenorm_gemm_impl(const uint32_t shape_m,
             for (int i = 0; i < BLOCK_K / kNumElemsInSwizzleRange; i++) {
                 #pragma unroll
                 for (int k = 0; k < kNumElemsInSwizzleRange / WGMMA::K; k++) {
-                    auto b_desc = make_smem_desc(smem_b[stage_idx] + i * BLOCK_N * kNumElemsInSwizzleRange + k * WGMMA::K, 1);
+                    auto b_desc = mma::sm90::make_smem_desc(
+                        smem_b[stage_idx] + i * BLOCK_N * kNumElemsInSwizzleRange + k * WGMMA::K, 1);
                     WGMMA::wgmma(a + (i * kNumWgmmaInSwizzleRange + k) * kNumRegPerWgmma, b_desc, accum, 1);
                 }
             }
-            warpgroup_commit_batch();
+            ptx::warpgroup_commit_batch();
             #pragma unroll
             for (uint32_t i = 0; i < WGMMA::kNumAccum; ++ i)
-                warpgroup_fence_operand(accum[i]);
+                ptx::warpgroup_fence_operand(accum[i]);
         }
 
-        const auto& reduced_sum_0 = warp_reduce_sum<4>(sqr_sum_acc_0);
-        const auto& reduced_sum_1 = warp_reduce_sum<4>(sqr_sum_acc_1);
+        const auto& reduced_sum_0 = math::warp_reduce_sum<4>(sqr_sum_acc_0);
+        const auto& reduced_sum_1 = math::warp_reduce_sum<4>(sqr_sum_acc_1);
 
         const auto& m_idx = m_block_idx * BLOCK_M + (warp_idx * BLOCK_M_PER_WARP + lane_idx / 4);
         if (lane_idx % 4 == 0) {
@@ -233,7 +240,7 @@ sm90_tf32_hc_prenorm_gemm_impl(const uint32_t shape_m,
             if (m_idx + 8 < shape_m)
                 sqr_sum[m_offset + m_idx + 8] = reduced_sum_1;
         }
-        warpgroup_wait<0>();
+        ptx::warpgroup_wait<0>();
         empty_barriers[(num_total_stages-1) % kNumStages]->arrive();
 
         // Write accum to shared memory
@@ -260,8 +267,8 @@ sm90_tf32_hc_prenorm_gemm_impl(const uint32_t shape_m,
 
             // 0/1 write to the same row, 2/3 write to another row
             auto values = reinterpret_cast<uint32_t*>(accum + i * 2);
-            st_shared(smem_ptr, values[0], values[1]);
-            st_shared(smem_ptr + 8 * kSwizzleCDMode, values[2], values[3]);
+            ptx::st_shared(smem_ptr, values[0], values[1]);
+            ptx::st_shared(smem_ptr + 8 * kSwizzleCDMode, values[2], values[3]);
         }
         cute::tma_store_fence();
         cutlass::arch::NamedBarrier::sync(128, 1);

deep_gemm/include/deep_gemm/impls/smxx_clean_logits.cuh

@@ -3,21 +3,24 @@
 #include <cutlass/arch/barrier.h>
 #include <cute/arch/cluster_sm90.hpp>
 
-#include <deep_gemm/common/utils.cuh>
+#include <deep_gemm/common/cute_tie.cuh>
+#include <deep_gemm/common/math.cuh>
 
 namespace deep_gemm {
 
-template <uint32_t kNextN, uint32_t BLOCK_KV, uint32_t kNumWarps>
-__global__ __launch_bounds__(kNumWarps * 32, 1)
+template <uint32_t kNextN, uint32_t BLOCK_KV, uint32_t kNumWarps, typename logits_dtype_t>
+CUTLASS_GLOBAL __launch_bounds__(kNumWarps * 32, 1)
 void smxx_clean_logits(const uint32_t seq_len, const uint32_t seq_len_kv, const uint64_t stride_logits,
-                       const uint32_t* cu_seq_len_k_start, const uint32_t* cu_seq_len_k_end, float* logits) {
-    const uint32_t& num_sms = gridDim.x;
-    const uint32_t& sm_idx = blockIdx.x;
-    const uint32_t& warp_idx = __shfl_sync(0xffffffff, threadIdx.x / 32, 0);
-    constexpr float neg_inf = -cute::numeric_limits<float>::infinity();
+                       const uint32_t* cu_seq_len_k_start, const uint32_t* cu_seq_len_k_end, logits_dtype_t* logits) {
+    const uint32_t num_sms = gridDim.x;
+    const uint32_t sm_idx = blockIdx.x;
+    const uint32_t warp_idx = __shfl_sync(0xffffffff, threadIdx.x / 32, 0);
+
+    constexpr uint32_t kAlignment = 16 / sizeof(logits_dtype_t);
+    const logits_dtype_t neg_inf = -cute::numeric_limits<logits_dtype_t>::infinity();
 
     // Allocate filled `-inf` shared memory
-    extern __shared__ __align__(1024) float smem_buffer[];
+    extern __shared__ __align__(1024) logits_dtype_t smem_buffer[];
     #pragma unroll
     for (uint32_t i = threadIdx.x; i < BLOCK_KV; i += kNumWarps * 32)
         smem_buffer[i] = neg_inf;
@@ -25,38 +28,42 @@ void smxx_clean_logits(const uint32_t seq_len, const uint32_t seq_len_kv, const
     __syncthreads();
 
     // Assign sequence to each warp
-    const auto& assign_task = [&](const uint32_t& num, const uint32_t& idx,
-                                  const uint32_t& start, const uint32_t& total) -> cute::tuple<uint32_t, uint32_t> {
-        const auto& per = total / num, rem = total % num;
-        return {start + idx * per + min(idx, rem), per + (idx < rem)};
+    const auto assign_task = [&](const uint32_t& num, const uint32_t& idx,
+                                 const uint32_t& start, const uint32_t& total) -> cute::tuple<uint32_t, uint32_t> {
+        const auto per = total / num, rem = total % num;
+        return {start + idx * per + cute::min(idx, rem), per + (idx < rem)};
     };
     CUTE_TIE_DECL(assign_task(num_sms, sm_idx, 0, seq_len), sm_seq_start, sm_seq_len);
     CUTE_TIE_DECL(assign_task(kNumWarps, warp_idx, sm_seq_start, sm_seq_len), warp_seq_start, warp_seq_len);
 
+    // Wait for primary kernel completion
+    cudaGridDependencySynchronize();
+
     if (cute::elect_one_sync()) {
         for (uint32_t i = warp_seq_start; i < warp_seq_start + warp_seq_len; ++ i) {
-            const auto& ks = cu_seq_len_k_start == nullptr ? 0 : __ldg(cu_seq_len_k_start + i / kNextN);
-            const auto& ke = __ldg(cu_seq_len_k_end + i / kNextN) - kNextN + i % kNextN + 1;
-            const auto& aligned_ks = ks / 4 * 4, aligned_ke = (ke + 3) / 4 * 4;
+            const auto ks = cu_seq_len_k_start == nullptr ? 0 : cu_seq_len_k_start[i / kNextN];
+            const auto ke = cu_seq_len_k_end[i / kNextN] - kNextN + i % kNextN + 1;
+            const auto aligned_ks = ks / kAlignment * kAlignment, aligned_ke = (ke + kAlignment - 1) / kAlignment * kAlignment;
 
             for (uint32_t left = 0; left < seq_len_kv; left += BLOCK_KV) {
-                const auto& right = min(left + BLOCK_KV, static_cast<uint32_t>(stride_logits));
+                const auto right = cute::min(left + BLOCK_KV, static_cast<uint32_t>(stride_logits));
                 if (right <= ks or ke <= left) {
-                    cute::SM90_BULK_COPY_S2G::copy(smem_buffer, logits + i * stride_logits + left, (right - left) * sizeof(float));
+                    cute::SM90_BULK_COPY_S2G::copy(smem_buffer, logits + i * stride_logits + left, (right - left) * sizeof(logits_dtype_t));
                 } else {
                     if (left < aligned_ks)
-                        cute::SM90_BULK_COPY_S2G::copy(smem_buffer, logits + i * stride_logits + left, (aligned_ks - left) * sizeof(float));
+                        cute::SM90_BULK_COPY_S2G::copy(smem_buffer, logits + i * stride_logits + left, (aligned_ks - left) * sizeof(logits_dtype_t));
                     if (aligned_ke < right)
-                        cute::SM90_BULK_COPY_S2G::copy(smem_buffer, logits + i * stride_logits + aligned_ke, (right - aligned_ke) * sizeof(float));
+                        cute::SM90_BULK_COPY_S2G::copy(smem_buffer, logits + i * stride_logits + aligned_ke, (right - aligned_ke) * sizeof(logits_dtype_t));
                 }
             }
         }
     }
+    __syncwarp();
 
     for (uint32_t i = warp_seq_start; i < warp_seq_start + warp_seq_len; ++ i) {
-        const auto& ks = cu_seq_len_k_start == nullptr ? 0 : __ldg(cu_seq_len_k_start + i / kNextN);
-        const auto& ke = __ldg(cu_seq_len_k_end + i / kNextN) - kNextN + i % kNextN + 1;
-        const auto& aligned_ks = ks / 4 * 4, aligned_ke = (ke + 3) / 4 * 4;
+        const auto ks = cu_seq_len_k_start == nullptr ? 0 : cu_seq_len_k_start[i / kNextN];
+        const auto ke = cu_seq_len_k_end[i / kNextN] - kNextN + i % kNextN + 1;
+        const auto aligned_ks = ks / kAlignment * kAlignment, aligned_ke = (ke + kAlignment - 1) / kAlignment * kAlignment;
         for (uint32_t j = aligned_ks; j < ks; ++ j)
             logits[i * stride_logits + j] = neg_inf;
         for (uint32_t j = ke; j < aligned_ke; ++ j)

deep_gemm/include/deep_gemm/impls/smxx_layout.cuh

@@ -1,13 +1,16 @@
 #pragma once
 
+#include <deep_gemm/common/math.cuh>
 #include <deep_gemm/common/utils.cuh>
+#include <deep_gemm/ptx/ld_st.cuh>
+#include <deep_gemm/ptx/utils.cuh>
 
 namespace deep_gemm {
 
 template <uint32_t kNumThreads, uint32_t BLOCK_MN, uint32_t SF_K,
           uint32_t PADDED_SF_K = SF_K + (1 - (SF_K % 2))>
-__global__ void transpose_fp32(const float* sf, float* out, const uint32_t mn) {
-    typedef typename Vectorized<sizeof(float) * SF_K>::vec_t in_vec_t;
+CUTLASS_GLOBAL void transpose_fp32(const float* sf, float* out, const uint32_t mn) {
+    typedef typename utils::Vectorized<sizeof(float) * SF_K>::vec_t in_vec_t;
     constexpr static uint32_t kNumElemsPerVec = sizeof(in_vec_t) / sizeof(float);
     constexpr static uint32_t SF_VEC_K = SF_K / kNumElemsPerVec;
 
@@ -15,16 +18,19 @@ __global__ void transpose_fp32(const float* sf, float* out, const uint32_t mn) {
     extern __shared__ float smem_buffer[];
     constexpr auto kNumTMAAlignedElems = static_cast<uint32_t>(16 / sizeof(float));
     const auto in_block_mn = min(BLOCK_MN, mn - blockIdx.x * BLOCK_MN);
-    const auto tma_aligned_mn = align<uint32_t>(mn, kNumTMAAlignedElems);
+    const auto tma_aligned_mn = math::align<uint32_t>(mn, kNumTMAAlignedElems);
 
     // Shift into the block
     sf = sf + static_cast<uint64_t>(blockIdx.y) * mn * SF_K;
     out = out + static_cast<uint64_t>(blockIdx.y) * tma_aligned_mn * SF_K;
     const auto& local_sf = reinterpret_cast<const in_vec_t*>(sf + static_cast<uint64_t>(blockIdx.x) * (BLOCK_MN * SF_K));
 
+    // Wait for primary kernel completion
+    cudaGridDependencySynchronize();
+
     // Load
     for (uint32_t i = threadIdx.x; i < in_block_mn * SF_VEC_K; i += kNumThreads) {
-        auto in_vec = __ldg(local_sf + i);
+        auto in_vec = local_sf[i];
         const auto& in_values = reinterpret_cast<float*>(&in_vec);
 
         const auto& row = i / SF_VEC_K, col = (i % SF_VEC_K) * kNumElemsPerVec;
@@ -39,26 +45,29 @@ __global__ void transpose_fp32(const float* sf, float* out, const uint32_t mn) {
     for (uint32_t i = threadIdx.x; i < in_block_mn * SF_K; i += kNumThreads) {
         const auto& sf_k_idx = i / in_block_mn, mn_idx = i % in_block_mn;
         const auto& global_mn_idx = blockIdx.x * BLOCK_MN + mn_idx;
-        out[sf_k_idx * tma_aligned_mn + global_mn_idx] = ld_shared(smem_buffer + mn_idx * PADDED_SF_K + sf_k_idx);
+        out[sf_k_idx * tma_aligned_mn + global_mn_idx] = ptx::ld_shared(smem_buffer + mn_idx * PADDED_SF_K + sf_k_idx);
     }
 }
 
 // NOTES: the two kernels below always pack the K dimension
 
 template <uint32_t kNumThreads, uint32_t BLOCK_MN, uint32_t SF_K>
-__global__ void transpose_and_pack_fp32_into_ue8m0(float* sf, uint32_t* out, const uint32_t mn) {
+CUTLASS_GLOBAL void transpose_and_pack_fp32_into_ue8m0(float* sf, uint32_t* out, const uint32_t mn) {
     extern __shared__ uint32_t smem_buffer[];
 
     // Shapes and strides
-    constexpr auto kNumPackedSFK = constexpr_ceil_div(SF_K, 4u);
+    constexpr auto kNumPackedSFK = math::constexpr_ceil_div(SF_K, 4u);
     constexpr auto kNumTMAAlignedElems = static_cast<uint32_t>(16 / sizeof(int));
     const auto in_block_mn = min(BLOCK_MN, mn - blockIdx.x * BLOCK_MN);
-    const auto tma_aligned_mn = align<uint64_t>(mn, kNumTMAAlignedElems);
+    const auto tma_aligned_mn = math::align<uint64_t>(mn, kNumTMAAlignedElems);
 
     // Shift into the group
     sf = sf + static_cast<uint64_t>(blockIdx.y) * mn * SF_K;
     out = out + static_cast<uint64_t>(blockIdx.y) * tma_aligned_mn * kNumPackedSFK;
 
+    // Wait for primary kernel completion
+    cudaGridDependencySynchronize();
+
     // Load FP32 SFs
     DG_STATIC_ASSERT(BLOCK_MN % 4 == 0, "Invalid block size");
     const auto local_sf = reinterpret_cast<uint32_t*>(sf + static_cast<uint64_t>(blockIdx.x) * (BLOCK_MN * SF_K));
@@ -66,13 +75,13 @@ __global__ void transpose_and_pack_fp32_into_ue8m0(float* sf, uint32_t* out, con
     const auto num_uint4 = num_values / 4;
     #pragma unroll
     for (uint32_t i = threadIdx.x; i < num_uint4; i += kNumThreads) {
-        const auto& [x, y, z, w] = __ldg(reinterpret_cast<uint4*>(local_sf) + i);
-        st_shared(reinterpret_cast<uint4*>(smem_buffer) + i, x, y, z, w);
+        const auto& [x, y, z, w] = reinterpret_cast<const uint4*>(local_sf)[i];
+        ptx::st_shared(reinterpret_cast<uint4*>(smem_buffer) + i, x, y, z, w);
     }
 
     // Fill unaligned values as well
     if (const auto unaligned_idx = num_uint4 * 4 + threadIdx.x; unaligned_idx < num_values)
-        st_shared(smem_buffer + unaligned_idx, __ldg(local_sf + unaligned_idx));
+        ptx::st_shared(smem_buffer + unaligned_idx, local_sf[unaligned_idx]);
     __syncthreads();
 
     // Pack into UE8M0 and store
@@ -85,7 +94,7 @@ __global__ void transpose_and_pack_fp32_into_ue8m0(float* sf, uint32_t* out, con
         #pragma unroll
         for (uint32_t j = 0; j < 4; ++ j) {
             const auto sf_k_idx = sf_k_pack_idx * 4 + j;
-            values[j] = sf_k_idx < SF_K ? ld_shared(smem_buffer + mn_idx * SF_K + sf_k_idx) : 0;
+            values[j] = sf_k_idx < SF_K ? ptx::ld_shared(smem_buffer + mn_idx * SF_K + sf_k_idx) : 0;
         }
 
         // Pack and store
@@ -101,8 +110,9 @@ __global__ void transpose_and_pack_fp32_into_ue8m0(float* sf, uint32_t* out, con
 
 template <uint32_t kNumGroups, uint32_t kNumThreads,
           uint32_t BLOCK_MN, uint32_t BLOCK_PACKED_SF_K, bool kTransposed = true>
-__global__ void pack_fp32_into_ue8m0(float* sf, uint32_t* out, uint32_t* ks,
-                                     const uint32_t mn, uint32_t sf_k, const uint32_t packed_sf_k) {
+CUTLASS_GLOBAL void pack_fp32_into_ue8m0(float* sf, uint32_t* out, uint32_t* ks,
+                                         const uint32_t mn, uint32_t sf_k, const uint32_t packed_sf_k,
+                                         const uint32_t gran_k) {
     // Always packing the K dimension
     // NOTES: should also assert `mn % 4 == 0` at launch
     DG_STATIC_ASSERT(kTransposed, "Currently only support transposed SFs (MN-major)");
@@ -120,11 +130,14 @@ __global__ void pack_fp32_into_ue8m0(float* sf, uint32_t* out, uint32_t* ks,
 
     // Each warp is responsible for a packed row
     const auto warp_idx = threadIdx.x / 32;
-    const auto lane_idx = get_lane_idx();
+    const auto lane_idx = ptx::get_lane_idx();
     const auto packed_sf_k_idx = static_cast<uint64_t>(blockIdx.y) * BLOCK_PACKED_SF_K + warp_idx;
     if (warp_idx >= in_block_packed_sf_k)
         return;
 
+    // Wait for primary kernel completion
+    cudaGridDependencySynchronize();
+
     // Make an offset on the input
     uint32_t input_offset = 0;
     if constexpr (kNumGroups > 1) {
@@ -134,18 +147,18 @@ __global__ void pack_fp32_into_ue8m0(float* sf, uint32_t* out, uint32_t* ks,
         #pragma unroll
         for (uint32_t i = 0; i < 4; ++ i) {
             const auto group_idx = lane_idx * 4 + i;
-            group_ks[i] = group_idx < kNumGroups ? __ldg(ks + group_idx) : 0;
+            group_ks[i] = group_idx < kNumGroups ? ks[group_idx] : 0;
         }
         __syncwarp();
 
         // Make the offset
         sf_k = 0;
-        auto sum_packed_sf_k = 0;
+        uint32_t sum_packed_sf_k = 0;
         #pragma unroll
         for (uint32_t i = 0; i < kNumGroups; ++ i) {
-            const auto sf_k_in_group = __shfl_sync(0xffffffff, group_ks[i % 4] / 128, i / 4);
+            const auto sf_k_in_group = __shfl_sync(0xffffffff, group_ks[i % 4] / gran_k, i / 4);
             sf_k += sf_k_in_group;
-            sum_packed_sf_k += ceil_div(sf_k_in_group, 4u);
+            sum_packed_sf_k += math::ceil_div(sf_k_in_group, 4u);
             if (packed_sf_k_idx < sum_packed_sf_k)
                 break;
             if (const auto remainder = sf_k_in_group % 4; remainder > 0)
@@ -153,14 +166,14 @@ __global__ void pack_fp32_into_ue8m0(float* sf, uint32_t* out, uint32_t* ks,
         }
     }
 
-    for (uint32_t mn_idx = get_lane_idx(); mn_idx < in_block_mn_uint4; mn_idx += 32) {
+    for (uint32_t mn_idx = ptx::get_lane_idx(); mn_idx < in_block_mn_uint4; mn_idx += 32) {
         // Load
         uint4 values[4];
         #pragma unroll
         for (uint32_t j = 0; j < 4; ++ j) {
             values[j] = make_uint4(0, 0, 0, 0);
             if (const auto sf_k_idx = packed_sf_k_idx * 4 + j - input_offset; sf_k_idx < sf_k)
-                values[j] = __ldg(reinterpret_cast<uint4*>(sf + sf_k_idx * mn) + mn_idx);
+                values[j] = reinterpret_cast<const uint4*>(sf + sf_k_idx * mn)[mn_idx];
         }
 
         // Pack and store

deep_gemm/include/deep_gemm/layout/mega_moe.cuh

@@ -0,0 +1,255 @@
+#pragma once
+
+#include <deep_gemm/common/math.cuh>
+#include <deep_gemm/common/exception.cuh>
+
+namespace deep_gemm::layout {
+
+// Pool capacity for shared expert token pool: worst-case total tokens + per-expert BLOCK_M alignment padding
+template <typename T>
+CUTLASS_HOST_DEVICE constexpr T get_num_max_pool_tokens(T num_ranks, T num_max_tokens_per_rank, T num_topk,
+                                                        T num_experts_per_rank, T block_m) {
+    const auto num_max_recv_tokens = num_ranks * num_max_tokens_per_rank;
+    const auto num_max_experts_per_token = math::constexpr_min(num_topk, num_experts_per_rank);
+    return math::constexpr_align(
+        num_max_recv_tokens * num_max_experts_per_token + num_experts_per_rank * (block_m - 1),
+        block_m);
+}
+
+// SF pool capacity: all experts share a contiguous SF region, sized by pool blocks × SF_BLOCK_M
+template <typename T>
+CUTLASS_HOST_DEVICE constexpr T get_num_padded_sf_pool_tokens(T num_max_pool_tokens, T block_m) {
+    return (num_max_pool_tokens / block_m) * math::constexpr_align(block_m, static_cast<T>(128));
+}
+
+// Per-token source metadata for combine write-back
+struct TokenSrcMetadata {
+    uint32_t rank_idx;
+    uint32_t token_idx;
+    uint32_t topk_idx;
+};
+
+struct Workspace {
+    void* base;
+    uint32_t num_ranks, num_experts;
+    uint32_t num_experts_per_rank;
+    uint32_t num_max_tokens_per_rank;
+    uint32_t num_max_recv_tokens_per_expert;
+
+    // Pool capacity: all local experts share a contiguous token pool
+    uint32_t num_max_pool_tokens;
+    uint32_t num_max_pool_blocks;
+
+    // For both grid barrier and NVLink barrier
+    static constexpr uint64_t kNumBarrierSignalBytes = 32;
+
+    CUTLASS_HOST_DEVICE
+    Workspace(void* base,
+              const uint32_t& num_ranks,
+              const uint32_t& num_experts,
+              const uint32_t& num_max_tokens_per_rank,
+              const uint32_t& num_topk,
+              const uint32_t& block_m):
+        base(base),
+        num_ranks(num_ranks), num_experts(num_experts),
+        num_max_tokens_per_rank(num_max_tokens_per_rank) {
+        num_experts_per_rank = num_experts / num_ranks;
+        num_max_recv_tokens_per_expert = num_ranks * num_max_tokens_per_rank;
+        num_max_pool_tokens = get_num_max_pool_tokens(
+            num_ranks, num_max_tokens_per_rank, num_topk, num_experts_per_rank, block_m);
+        num_max_pool_blocks = num_max_pool_tokens / block_m;
+        DG_UNIFIED_ASSERT(num_max_tokens_per_rank % block_m == 0);
+    }
+
+    CUTLASS_HOST_DEVICE
+    uint64_t get_num_bytes() const {
+        uint64_t num_bytes = 0;
+
+        // Barrier
+        num_bytes += kNumBarrierSignalBytes;
+
+        // Expert send/recv count
+        num_bytes += num_experts * sizeof(uint64_t) * 2;
+
+        // Expert recv count sum
+        num_bytes += num_experts_per_rank * sizeof(uint64_t);
+
+        // L1 arrival count (padded to even entry count for `uint64_t` alignment of L2 mask)
+        num_bytes += math::align(num_max_pool_blocks, 2u) * sizeof(uint32_t);
+
+        // L2 block arrival mask
+        num_bytes += num_max_pool_blocks * sizeof(uint64_t);
+
+        // Dispatch pulling source token-topk
+        num_bytes += num_experts_per_rank * num_ranks * num_max_recv_tokens_per_expert * sizeof(int);
+
+        // Combine push source indices
+        num_bytes += num_max_pool_tokens * sizeof(TokenSrcMetadata);
+
+        // Align to TMA descriptor requirements
+        num_bytes = math::align<uint64_t>(num_bytes, 16);
+        return num_bytes;
+    }
+
+    CUTLASS_HOST_DEVICE
+    void* get_end_ptr() const {
+        return math::advance_ptr(base, get_num_bytes());
+    }
+
+    // Grid sync counters: `kNumBarrierSignalBytes` layout
+    // [ 0..15]: 4 x `uint32_t` grid sync counters
+    // [16..20]: `uint32_t` NVLink barrier counter
+    // [20..27]: 2 x `int` NVLink barrier signals (phase 0 and 1)
+    static constexpr uint32_t kNumMaxGridSyncCounters = 4;
+
+    template <uint32_t kIndex = 0>
+    CUTLASS_DEVICE
+    uint32_t* get_grid_sync_count_ptr() const {
+        DG_STATIC_ASSERT(kIndex < kNumMaxGridSyncCounters, "Grid sync index out of bounds");
+        return static_cast<uint32_t*>(base) + kIndex;
+    }
+
+    CUTLASS_DEVICE
+    uint32_t* get_nvl_barrier_counter_ptr() const {
+        return static_cast<uint32_t*>(base) + kNumMaxGridSyncCounters;
+    }
+
+    CUTLASS_DEVICE
+    int* get_nvl_barrier_signal_ptr(const uint32_t& phase) const {
+        // NOTES: the signal is signed, as we may minus
+        return math::advance_ptr<int>(base, (kNumMaxGridSyncCounters + 1) * sizeof(uint32_t) + phase * sizeof(int));
+    }
+
+    CUTLASS_DEVICE
+    uint64_t* get_expert_send_count_ptr(const uint32_t& expert_idx = 0) const {
+        return math::advance_ptr<uint64_t>(base, kNumBarrierSignalBytes) + expert_idx;
+    }
+
+    CUTLASS_DEVICE
+    uint64_t* get_expert_recv_count_ptr(
+        const uint32_t& rank_idx = 0, const uint32_t& expert_idx = 0) const {
+        return get_expert_send_count_ptr(num_experts) + rank_idx * num_experts_per_rank + expert_idx;
+    }
+
+    CUTLASS_DEVICE
+    uint64_t* get_expert_recv_count_sum_ptr(const uint32_t& expert_idx = 0) const {
+        return get_expert_send_count_ptr(num_experts * 2) + expert_idx;
+    }
+
+    CUTLASS_DEVICE
+    uint32_t* get_l1_arrival_count_ptr(const uint32_t& pool_block_idx = 0) const {
+        const auto base = get_expert_recv_count_sum_ptr(num_experts_per_rank);
+        return reinterpret_cast<uint32_t*>(base) + pool_block_idx;
+    }
+
+    CUTLASS_DEVICE
+    uint64_t* get_l2_arrival_mask_ptr(const uint32_t& pool_block_idx = 0) const {
+        // Pad L1 entry count to even so that the `l2_arrival_mask` is 8-byte aligned
+        const auto base = get_l1_arrival_count_ptr(math::align(num_max_pool_blocks, 2u));
+        return reinterpret_cast<uint64_t*>(base) + pool_block_idx;
+    }
+
+    // For dispatch pulling
+    CUTLASS_DEVICE
+    uint32_t* get_src_token_topk_idx_ptr(
+        const uint32_t& expert_idx = 0, const uint32_t& rank_idx = 0, const uint32_t& token_idx = 0) const {
+        const auto base = get_l2_arrival_mask_ptr(num_max_pool_blocks);
+        return reinterpret_cast<uint32_t*>(base) +
+            expert_idx * (num_ranks * num_max_recv_tokens_per_expert) +
+            rank_idx * num_max_recv_tokens_per_expert + token_idx;
+    }
+
+    // For combine usages
+    CUTLASS_DEVICE
+    TokenSrcMetadata* get_token_src_metadata_ptr(const uint32_t& pool_token_idx = 0) const {
+        const auto base = reinterpret_cast<TokenSrcMetadata*>(get_src_token_topk_idx_ptr(num_experts_per_rank));
+        return base + pool_token_idx;
+    }
+};
+
+struct Data {
+    uint32_t num_bytes;
+    bool require_tma_alignment;
+    void* base;
+
+    CUTLASS_HOST_DEVICE
+    constexpr explicit Data(
+        const uint32_t& num_bytes,
+        const bool& require_tma_alignment = true,
+        void* base = nullptr) :
+        num_bytes(num_bytes), require_tma_alignment(require_tma_alignment), base(base) {
+        DG_UNIFIED_ASSERT(num_bytes % 16 == 0 or not require_tma_alignment);
+    }
+
+    template <typename dtype_t = uint32_t>
+    CUTLASS_HOST_DEVICE constexpr dtype_t get_num_bytes() const {
+        return static_cast<dtype_t>(num_bytes);
+    }
+
+    template <typename dtype_t = void>
+    CUTLASS_HOST_DEVICE dtype_t* get_base_ptr() const {
+        return static_cast<dtype_t*>(base);
+    }
+
+    CUTLASS_HOST_DEVICE void set_base_ptr(void* ptr) {
+        base = ptr;
+    }
+};
+
+struct Buffer {
+    Data data_layout;
+    uint32_t num_ranks;
+    uint32_t num_max_tokens_per_rank;
+
+    void* base;
+
+    CUTLASS_HOST_DEVICE
+    Buffer(const Data& data_layout,
+           const uint32_t& num_ranks,
+           const uint32_t& max_num_tokens_per_rank,
+           void* base = nullptr) :
+        data_layout(data_layout),
+        num_ranks(num_ranks), num_max_tokens_per_rank(max_num_tokens_per_rank),
+        base(base) {}
+
+    CUTLASS_HOST_DEVICE
+    uint64_t get_num_bytes_per_rank() const {
+        return num_max_tokens_per_rank * data_layout.get_num_bytes<uint64_t>();
+    }
+
+    CUTLASS_HOST_DEVICE
+    uint64_t get_num_bytes() const {
+        return get_num_bytes_per_rank() * num_ranks;
+    }
+
+    template <typename dtype_t = void>
+    CUTLASS_HOST_DEVICE dtype_t* get_base_ptr() const {
+        return static_cast<dtype_t*>(base);
+    }
+
+    CUTLASS_HOST_DEVICE
+    void* get_end_ptr() const {
+        return math::advance_ptr(base, get_num_bytes());
+    }
+
+    CUTLASS_HOST_DEVICE
+    Buffer get_rank_buffer(const uint32_t& rank_idx) const {
+        return {
+            data_layout,
+            1, num_max_tokens_per_rank,
+            math::advance_ptr(base, get_num_bytes_per_rank() * rank_idx)
+        };
+    }
+
+    CUTLASS_HOST_DEVICE
+    Data get_data_buffer(const uint32_t& token_idx, const bool& global = false) const {
+        DG_DEVICE_ASSERT(num_ranks == 1 or global);
+        return Data(
+            data_layout.num_bytes,
+            data_layout.require_tma_alignment,
+            math::advance_ptr(base, data_layout.get_num_bytes<uint64_t>() * token_idx)
+        );
+    }
+};
+
+} // namespace deep_gemm::layout

deep_gemm/include/deep_gemm/layout/sym_buffer.cuh

@@ -0,0 +1,41 @@
+#pragma once
+
+#include <deep_gemm/common/exception.cuh>
+
+namespace deep_gemm::layout {
+
+constexpr static uint32_t kNumMaxRanks = 72;
+
+template <uint32_t kNumRanks = kNumMaxRanks>
+struct SymBuffer {
+    int64_t base;
+    int64_t offsets[kNumMaxRanks];
+    uint32_t rank_idx;
+
+    DG_STATIC_ASSERT(kNumRanks <= kNumMaxRanks, "Too many ranks");
+
+    SymBuffer() = default;
+
+    template <typename Container>
+    explicit SymBuffer(const Container& c, const uint32_t& rank_idx): rank_idx(rank_idx) {
+        const auto size = static_cast<uint32_t>(c.size());
+        base = c[rank_idx];
+        for (uint32_t i = 0; i < kNumMaxRanks; ++ i)
+            offsets[i] = i < size ? (c[i] - base) : 0;
+    }
+
+#if defined(__CUDA_ARCH__) or defined(__CLION_IDE__)
+    template <typename ptr_t = void*>
+    CUTLASS_DEVICE ptr_t get_base_ptr() const {
+        return reinterpret_cast<ptr_t>(base);
+    }
+
+    template <typename ptr_t>
+    CUTLASS_DEVICE ptr_t map(const ptr_t& ptr, const uint32_t& dst_rank_idx) const {
+        int64_t mapped_ptr = offsets[dst_rank_idx] + reinterpret_cast<int64_t>(ptr);
+        return *reinterpret_cast<ptr_t*>(&mapped_ptr);
+    }
+#endif
+};
+
+} // namespace deep_gemm::layout

deep_gemm/include/deep_gemm/mma/sm100.cuh

@@ -0,0 +1,151 @@
+#pragma once
+
+#include <cute/atom/mma_traits_sm100.hpp>
+#include <cute/arch/mma_sm100_umma.hpp>
+
+#include <deep_gemm/common/exception.cuh>
+#include <deep_gemm/common/math.cuh>
+#include <deep_gemm/common/tma_copy.cuh>
+
+namespace deep_gemm::mma::sm100 {
+
+/// Shared memory descriptor
+CUTLASS_DEVICE
+cute::UMMA::SmemDescriptor make_smem_desc(cute::UMMA::LayoutType layout, void* smem_ptr,
+                                          const uint32_t& stride_byte_offset, const uint32_t& leading_byte_offset) {
+    cute::UMMA::SmemDescriptor desc;
+
+    // Set the version for SM100
+    desc.version_ = 1;
+
+    // Legacy mode
+    desc.lbo_mode_ = 0;
+
+    // Layout
+    desc.layout_type_ = static_cast<uint8_t>(layout);
+
+    // Start address
+    const auto uint_ptr = cute::cast_smem_ptr_to_uint(smem_ptr);
+    desc.start_address_ = static_cast<uint16_t>(uint_ptr >> 4);
+
+    // Base offset
+    desc.base_offset_ = 0;
+
+    // SBO and LBO
+    desc.stride_byte_offset_ = stride_byte_offset >> 4;
+    desc.leading_byte_offset_ = leading_byte_offset >> 4;
+
+    return desc;
+}
+
+CUTLASS_DEVICE
+cute::UMMA::SmemDescriptor make_sf_desc(void* smem_ptr) {
+    // NOTES: the UTCCP layout is K-major by default
+    // Atom size: 8 x 128 bits
+    // {SBO, LBO} means the byte stride between atoms on {MN, K}
+    // Since the UTCCP we used is 128b-wide (only 1 atom on K), so LBO can be zero
+    return make_smem_desc(cute::UMMA::LayoutType::SWIZZLE_NONE, smem_ptr, 8 * 16, 0);
+}
+
+CUTLASS_DEVICE
+void replace_smem_desc_addr(cute::UMMA::SmemDescriptor& desc, const void* smem_ptr) {
+    const auto uint_ptr = cute::cast_smem_ptr_to_uint(smem_ptr);
+    desc.start_address_ = static_cast<uint16_t>(uint_ptr >> 4);
+}
+
+CUTLASS_DEVICE
+static uint32_t get_atom_base(const cute::UMMA::LayoutType& layout_type) {
+    return layout_type == cute::UMMA::LayoutType::SWIZZLE_128B_BASE32B ? 32 : 16;
+}
+
+/// UMMA descriptors
+// ReSharper disable once CppNotAllPathsReturnValue
+template <cute::UMMA::Major kMajorMode, uint32_t kSwizzleMode, bool kUseBase32, typename dtype_t>
+constexpr static cute::UMMA::LayoutType to_umma_layout_type() {
+    DG_STATIC_ASSERT(kSwizzleMode == 0 or kSwizzleMode == 16 or
+                     kSwizzleMode == 32 or kSwizzleMode == 64 or
+                     kSwizzleMode == 128, "Invalid swizzling mode");
+    // A special case
+    if constexpr ((cute::is_same_v<dtype_t, float> and kMajorMode == cute::UMMA::Major::MN) or kUseBase32) {
+        DG_STATIC_ASSERT(kUseBase32, "Invalid swizzling base");
+        return cute::UMMA::LayoutType::SWIZZLE_128B_BASE32B;
+    }
+
+    // Normal cases
+    if constexpr (kSwizzleMode == 0)   return cute::UMMA::LayoutType::SWIZZLE_NONE;
+    if constexpr (kSwizzleMode == 16)  return cute::UMMA::LayoutType::SWIZZLE_NONE;
+    if constexpr (kSwizzleMode == 32)  return cute::UMMA::LayoutType::SWIZZLE_32B;
+    if constexpr (kSwizzleMode == 64)  return cute::UMMA::LayoutType::SWIZZLE_64B;
+    if constexpr (kSwizzleMode == 128) return cute::UMMA::LayoutType::SWIZZLE_128B;
+}
+
+template <cute::UMMA::Major kMajorMode, uint32_t BLOCK_MN, uint32_t kSwizzleMode, typename dtype_t>
+CUTLASS_DEVICE
+constexpr uint32_t get_umma_desc_stride_k() {
+    return kMajorMode == cute::UMMA::Major::K ? 1 : tma::get_inner_block_atom_size<BLOCK_MN, kSwizzleMode, dtype_t>();
+}
+
+template <cute::UMMA::Major kMajorMode, uint32_t BLOCK_MN, uint32_t kSwizzleMode, typename dtype_t>
+CUTLASS_DEVICE
+uint32_t advance_umma_desc_lo(const uint32_t& base, const uint32_t& offset, const uint32_t& k_idx) {
+    return base + (((offset + k_idx * get_umma_desc_stride_k<kMajorMode, BLOCK_MN, kSwizzleMode, dtype_t>()) * static_cast<uint32_t>(sizeof(dtype_t))) >> 4u);
+}
+
+template <cute::UMMA::Major kMajorMode, uint32_t BLOCK_MN, uint32_t BLOCK_K, uint32_t kSwizzleMode, bool kUseBase32 = false, typename dtype_t>
+CUTLASS_DEVICE
+cute::UMMA::SmemDescriptor make_umma_desc(dtype_t* base_smem_ptr, uint32_t mn_idx, uint32_t k_idx) {
+    const uint32_t stride_k = get_umma_desc_stride_k<kMajorMode, BLOCK_MN, kSwizzleMode, dtype_t>();
+    const auto layout_type = to_umma_layout_type<kMajorMode, kSwizzleMode, kUseBase32, dtype_t>();
+    const auto num_non_contiguous = 128 / get_atom_base(layout_type);
+    if constexpr (kMajorMode == cute::UMMA::Major::K) {
+        // NOTES: for K-major layout, the swizzle must be the same as `BLOCK_K * sizeof(dtype_t)`
+        // also, atom index must be 0, so that each block has exactly one swizzle atom on the K axis
+        DG_STATIC_ASSERT(kSwizzleMode == BLOCK_K * sizeof(dtype_t), "Unexpected value");
+
+        // Atom size: 8 x `kSwizzleMode` (in bytes, on K)
+        // {SBO, LBO} means the byte stride between atoms on {MN, K}
+        // NOTES: on K, there is only 1 atom as asserted previously, so LBO can be 0
+        const uint32_t stride_byte_offset = num_non_contiguous * BLOCK_K * sizeof(dtype_t);
+        const uint32_t leading_byte_offset = 0;
+        return make_smem_desc(layout_type,
+                              base_smem_ptr + mn_idx * BLOCK_K + k_idx * stride_k,
+                              stride_byte_offset, leading_byte_offset);
+    } else {
+        constexpr uint32_t BLOCK_MN_ATOM = tma::get_inner_block_atom_size<BLOCK_MN, kSwizzleMode, dtype_t>();
+
+        // Must have no in-atom MN-idx
+        // NOTES: no worries for the runtime assert, the `mn_idx` are constants at compilation time
+        DG_DEVICE_ASSERT(mn_idx % BLOCK_MN_ATOM == 0);
+        DG_STATIC_ASSERT(kSwizzleMode > 0, "Invalid swizzling");
+
+        // Atom size: `kSwizzleMode` (in bytes, on MN) x 8
+        // NOTES: `kSwizzleMode == 16` mean non-swizzling but interleaving
+        // {SBO, LBO} means the byte stride between atoms on {K, MN} for swizzling
+        // {SBO, LBO} means the byte stride between atoms on {MN, K} for non-swizzling
+        uint32_t stride_byte_offset = num_non_contiguous * BLOCK_MN_ATOM * sizeof(dtype_t);
+        uint32_t leading_byte_offset = BLOCK_K * BLOCK_MN_ATOM * sizeof(dtype_t);
+        if constexpr (kSwizzleMode == 16)
+            math::swap(stride_byte_offset, leading_byte_offset);
+        return make_smem_desc(layout_type,
+                              base_smem_ptr + mn_idx * BLOCK_K + k_idx * stride_k,
+                              stride_byte_offset, leading_byte_offset);
+    }
+}
+
+CUTLASS_DEVICE uint64_t make_runtime_instr_desc_with_sf_id(
+    cute::UMMA::InstrDescriptorBlockScaled desc, const uint32_t& sfa_id, const uint32_t& sfb_id) {
+    desc.a_sf_id_ = sfa_id, desc.b_sf_id_ = sfb_id;
+    return static_cast<uint64_t>(static_cast<uint32_t>(desc)) << 32;
+}
+
+CUTLASS_DEVICE void update_instr_desc_with_umma_n(
+    cute::UMMA::InstrDescriptorBlockScaled& desc, const uint32_t& umma_n) {
+    desc.n_dim_ = umma_n >> 3;
+}
+
+CUTLASS_DEVICE void update_instr_desc_with_umma_n(
+    cute::UMMA::InstrDescriptor& desc, const uint32_t& umma_n) {
+    desc.n_dim_ = umma_n >> 3;
+}
+
+} // namespace deep_gemm::mma::sm100

deep_gemm/include/deep_gemm/mma/sm90.cuh

@@ -0,0 +1,293 @@
+#pragma once
+
+#include <cute/arch/cluster_sm90.hpp>
+#include <cute/arch/mma_sm90_desc.hpp>
+#include <cute/arch/mma_sm90_gmma.hpp>
+#include <cute/arch/mma_sm90_gmma_ext.hpp>
+#include <cute/arch/mma_sm100_desc.hpp>
+
+#include <deep_gemm/common/exception.cuh>
+
+namespace deep_gemm::mma::sm90 {
+
+/// MMA
+template <int N_, typename MMA>
+struct FP8MMA {
+    template <size_t ...Idx>
+    CUTLASS_DEVICE static void call_fma_impl(uint64_t const& desc_a, uint64_t const& desc_b, float* d, bool scale_d, cute::index_sequence<Idx...>) {
+        using namespace cute::SM90::GMMA;
+        MMA::fma(desc_a, desc_b, d[Idx]..., (scale_d ? ScaleOut::One : ScaleOut::Zero));
+    }
+
+    CUTLASS_DEVICE static void wgmma(uint64_t const& desc_a, uint64_t const& desc_b, float* d, bool scale_d) {
+        call_fma_impl(desc_a, desc_b, d, scale_d, cute::make_index_sequence<N_ / 2>{});
+    }
+
+    static constexpr int M = 64;
+    static constexpr int N = N_;
+    static constexpr int K = 32;
+    static constexpr int kNumAccum = M * N / 128;
+};
+
+template <int N>
+struct FP8MMASelector {
+    static constexpr auto select_mma() {
+        using namespace cute::SM90::GMMA;
+        if constexpr (N == 8) return MMA_64x8x32_F32E4M3E4M3_SS_TN();
+        if constexpr (N == 16) return MMA_64x16x32_F32E4M3E4M3_SS_TN();
+        if constexpr (N == 24) return MMA_64x24x32_F32E4M3E4M3_SS_TN();
+        if constexpr (N == 32) return MMA_64x32x32_F32E4M3E4M3_SS_TN();
+        if constexpr (N == 40) return MMA_64x40x32_F32E4M3E4M3_SS_TN();
+        if constexpr (N == 48) return MMA_64x48x32_F32E4M3E4M3_SS_TN();
+        if constexpr (N == 56) return MMA_64x56x32_F32E4M3E4M3_SS_TN();
+        if constexpr (N == 64) return MMA_64x64x32_F32E4M3E4M3_SS_TN();
+        if constexpr (N == 72) return MMA_64x72x32_F32E4M3E4M3_SS_TN();
+        if constexpr (N == 80) return MMA_64x80x32_F32E4M3E4M3_SS_TN();
+        if constexpr (N == 88) return MMA_64x88x32_F32E4M3E4M3_SS_TN();
+        if constexpr (N == 96) return MMA_64x96x32_F32E4M3E4M3_SS_TN();
+        if constexpr (N == 104) return MMA_64x104x32_F32E4M3E4M3_SS_TN();
+        if constexpr (N == 112) return MMA_64x112x32_F32E4M3E4M3_SS_TN();
+        if constexpr (N == 120) return MMA_64x120x32_F32E4M3E4M3_SS_TN();
+        if constexpr (N == 128) return MMA_64x128x32_F32E4M3E4M3_SS_TN();
+        if constexpr (N == 136) return MMA_64x136x32_F32E4M3E4M3_SS_TN();
+        if constexpr (N == 144) return MMA_64x144x32_F32E4M3E4M3_SS_TN();
+        if constexpr (N == 152) return MMA_64x152x32_F32E4M3E4M3_SS_TN();
+        if constexpr (N == 160) return MMA_64x160x32_F32E4M3E4M3_SS_TN();
+        if constexpr (N == 168) return MMA_64x168x32_F32E4M3E4M3_SS_TN();
+        if constexpr (N == 176) return MMA_64x176x32_F32E4M3E4M3_SS_TN();
+        if constexpr (N == 184) return MMA_64x184x32_F32E4M3E4M3_SS_TN();
+        if constexpr (N == 192) return MMA_64x192x32_F32E4M3E4M3_SS_TN();
+        if constexpr (N == 200) return MMA_64x200x32_F32E4M3E4M3_SS_TN();
+        if constexpr (N == 208) return MMA_64x208x32_F32E4M3E4M3_SS_TN();
+        if constexpr (N == 216) return MMA_64x216x32_F32E4M3E4M3_SS_TN();
+        if constexpr (N == 224) return MMA_64x224x32_F32E4M3E4M3_SS_TN();
+        if constexpr (N == 232) return MMA_64x232x32_F32E4M3E4M3_SS_TN();
+        if constexpr (N == 240) return MMA_64x240x32_F32E4M3E4M3_SS_TN();
+        if constexpr (N == 248) return MMA_64x248x32_F32E4M3E4M3_SS_TN();
+        if constexpr (N == 256) return MMA_64x256x32_F32E4M3E4M3_SS_TN();
+    }
+
+    static constexpr auto select_type() {
+        return FP8MMA<N, decltype(select_mma())>();
+    }
+
+    using type = decltype(select_type());
+};
+
+template <int N_, typename MMA>
+struct BF16MMA {
+    template <size_t ...Idx>
+    CUTLASS_DEVICE static void call_fma_impl(uint64_t const& desc_a, uint64_t const& desc_b, float* d, bool scale_d, cute::index_sequence<Idx...>) {
+        using namespace cute::SM90::GMMA;
+        MMA::fma(desc_a, desc_b, d[Idx]..., (scale_d ? ScaleOut::One : ScaleOut::Zero));
+    }
+
+    CUTLASS_DEVICE static void wgmma(uint64_t const& desc_a, uint64_t const& desc_b, float* d, bool scale_d) {
+        call_fma_impl(desc_a, desc_b, d, scale_d, cute::make_index_sequence<N_/2>{});
+    }
+
+    static constexpr int M = 64;
+    static constexpr int N = N_;
+    static constexpr int K = 16;
+    static constexpr int kNumAccum = M * N / 128;
+};
+
+template <cute::UMMA::Major kMajor>
+constexpr cute::SM90::GMMA::Major to_sm90_major() {
+    DG_STATIC_ASSERT(kMajor == cute::UMMA::Major::K or kMajor == cute::UMMA::Major::MN, "Invalid major-ness");
+    return kMajor == cute::UMMA::Major::K ? cute::SM90::GMMA::Major::K : cute::SM90::GMMA::Major::MN;
+}
+
+template <int N,
+          cute::UMMA::Major kMajorA = cute::UMMA::Major::K,
+          cute::UMMA::Major kMajorB = cute::UMMA::Major::K>
+struct BF16MMASelector {
+    static constexpr auto select_mma() {
+        using namespace cute::SM90::GMMA;
+        constexpr auto kGMMAMajorA = to_sm90_major<kMajorA>();
+        constexpr auto kGMMAMajorB = to_sm90_major<kMajorB>();
+        if constexpr (N == 8) return MMA_64x8x16_F32BF16BF16_SS<kGMMAMajorA, kGMMAMajorB>();
+        if constexpr (N == 16) return MMA_64x16x16_F32BF16BF16_SS<kGMMAMajorA, kGMMAMajorB>();
+        if constexpr (N == 24) return MMA_64x24x16_F32BF16BF16_SS<kGMMAMajorA, kGMMAMajorB>();
+        if constexpr (N == 32) return MMA_64x32x16_F32BF16BF16_SS<kGMMAMajorA, kGMMAMajorB>();
+        if constexpr (N == 40) return MMA_64x40x16_F32BF16BF16_SS<kGMMAMajorA, kGMMAMajorB>();
+        if constexpr (N == 48) return MMA_64x48x16_F32BF16BF16_SS<kGMMAMajorA, kGMMAMajorB>();
+        if constexpr (N == 56) return MMA_64x56x16_F32BF16BF16_SS<kGMMAMajorA, kGMMAMajorB>();
+        if constexpr (N == 64) return MMA_64x64x16_F32BF16BF16_SS<kGMMAMajorA, kGMMAMajorB>();
+        if constexpr (N == 72) return MMA_64x72x16_F32BF16BF16_SS<kGMMAMajorA, kGMMAMajorB>();
+        if constexpr (N == 80) return MMA_64x80x16_F32BF16BF16_SS<kGMMAMajorA, kGMMAMajorB>();
+        if constexpr (N == 88) return MMA_64x88x16_F32BF16BF16_SS<kGMMAMajorA, kGMMAMajorB>();
+        if constexpr (N == 96) return MMA_64x96x16_F32BF16BF16_SS<kGMMAMajorA, kGMMAMajorB>();
+        if constexpr (N == 104) return MMA_64x104x16_F32BF16BF16_SS<kGMMAMajorA, kGMMAMajorB>();
+        if constexpr (N == 112) return MMA_64x112x16_F32BF16BF16_SS<kGMMAMajorA, kGMMAMajorB>();
+        if constexpr (N == 120) return MMA_64x120x16_F32BF16BF16_SS<kGMMAMajorA, kGMMAMajorB>();
+        if constexpr (N == 128) return MMA_64x128x16_F32BF16BF16_SS<kGMMAMajorA, kGMMAMajorB>();
+        if constexpr (N == 136) return MMA_64x136x16_F32BF16BF16_SS<kGMMAMajorA, kGMMAMajorB>();
+        if constexpr (N == 144) return MMA_64x144x16_F32BF16BF16_SS<kGMMAMajorA, kGMMAMajorB>();
+        if constexpr (N == 152) return MMA_64x152x16_F32BF16BF16_SS<kGMMAMajorA, kGMMAMajorB>();
+        if constexpr (N == 160) return MMA_64x160x16_F32BF16BF16_SS<kGMMAMajorA, kGMMAMajorB>();
+        if constexpr (N == 168) return MMA_64x168x16_F32BF16BF16_SS<kGMMAMajorA, kGMMAMajorB>();
+        if constexpr (N == 176) return MMA_64x176x16_F32BF16BF16_SS<kGMMAMajorA, kGMMAMajorB>();
+        if constexpr (N == 184) return MMA_64x184x16_F32BF16BF16_SS<kGMMAMajorA, kGMMAMajorB>();
+        if constexpr (N == 192) return MMA_64x192x16_F32BF16BF16_SS<kGMMAMajorA, kGMMAMajorB>();
+        if constexpr (N == 200) return MMA_64x200x16_F32BF16BF16_SS<kGMMAMajorA, kGMMAMajorB>();
+        if constexpr (N == 208) return MMA_64x208x16_F32BF16BF16_SS<kGMMAMajorA, kGMMAMajorB>();
+        if constexpr (N == 216) return MMA_64x216x16_F32BF16BF16_SS<kGMMAMajorA, kGMMAMajorB>();
+        if constexpr (N == 224) return MMA_64x224x16_F32BF16BF16_SS<kGMMAMajorA, kGMMAMajorB>();
+        if constexpr (N == 232) return MMA_64x232x16_F32BF16BF16_SS<kGMMAMajorA, kGMMAMajorB>();
+        if constexpr (N == 240) return MMA_64x240x16_F32BF16BF16_SS<kGMMAMajorA, kGMMAMajorB>();
+        if constexpr (N == 248) return MMA_64x248x16_F32BF16BF16_SS<kGMMAMajorA, kGMMAMajorB>();
+        if constexpr (N == 256) return MMA_64x256x16_F32BF16BF16_SS<kGMMAMajorA, kGMMAMajorB>();
+    }
+
+    static constexpr auto select_type() {
+        return BF16MMA<N, decltype(select_mma())>();
+    }
+
+    using type = decltype(select_type());
+};
+
+template <int N_, typename MMA>
+struct TF32MMARS {
+    template <size_t ...Idx>
+    CUTLASS_DEVICE static void call_fma_impl(uint32_t* a, uint64_t const& desc_b, float* d, bool scale_d, cute::index_sequence<Idx...>) {
+        using namespace cute::SM90::GMMA;
+        MMA::fma(a[0], a[1], a[2], a[3], desc_b, d[Idx]..., (scale_d ? ScaleOut::One : ScaleOut::Zero));
+    }
+
+    CUTLASS_DEVICE static void wgmma(float* a, uint64_t const& desc_b, float* d, bool scale_d) {
+        call_fma_impl(reinterpret_cast<uint32_t*>(a), desc_b, d, scale_d, cute::make_index_sequence<N_/2>{});
+    }
+
+    static constexpr int M = 64;
+    static constexpr int N = N_;
+    static constexpr int K = 8;
+    static constexpr int kNumAccum = M * N / 128;
+};
+
+template <int N, bool kUseRS = true>
+struct TF32MMASelector {
+    static constexpr auto select_mma() {
+        using namespace cute::SM90::GMMA;
+        if constexpr (kUseRS) {
+            if constexpr (N == 8) return MMA_64x8x8_F32TF32TF32_RS_TN();
+            if constexpr (N == 16) return MMA_64x16x8_F32TF32TF32_RS_TN();
+            if constexpr (N == 32) return MMA_64x32x8_F32TF32TF32_RS_TN();
+            if constexpr (N == 64) return MMA_64x64x8_F32TF32TF32_RS_TN();
+            if constexpr (N == 128) return MMA_64x128x8_F32TF32TF32_RS_TN();
+            if constexpr (N == 256) return MMA_64x256x8_F32TF32TF32_RS_TN();
+            DG_STATIC_ASSERT(N == 8 or N == 16 or N == 32 or N == 64 or N == 128 or N == 256, "Invalid N");
+        }
+    }
+
+    static constexpr auto select_type() {
+        if constexpr (kUseRS) {
+            return TF32MMARS<N, decltype(select_mma())>();
+        } else {
+            DG_STATIC_ASSERT(kUseRS, "SS mode is not supported for TF32MMASelector for now");
+        }
+    }
+
+    using type = decltype(select_type());
+};
+
+/// Shared memory descriptor
+template <class PointerType>
+CUTLASS_DEVICE cute::GmmaDescriptor
+make_smem_desc(PointerType smem_ptr, const int& layout_type,
+               const uint32_t& leading_byte_offset = 0,
+               const uint32_t& stride_byte_offset = 1024) {
+    // NOTES: the default LBO and SBO are for K-major types
+    cute::GmmaDescriptor desc;
+    const auto uint_ptr = static_cast<uint32_t>(__cvta_generic_to_shared(smem_ptr));
+    desc.bitfield.start_address_ = uint_ptr >> 4;
+    desc.bitfield.layout_type_ = layout_type;
+    desc.bitfield.leading_byte_offset_ = leading_byte_offset >> 4;
+    desc.bitfield.stride_byte_offset_ = stride_byte_offset >> 4;
+    desc.bitfield.base_offset_ = 0;
+    return desc;
+}
+
+template <uint32_t BLOCK_INNER, uint32_t kSwizzleMode, typename dtype_t>
+constexpr uint32_t get_inner_block_atom_size() {
+    return kSwizzleMode == 0 ? BLOCK_INNER : kSwizzleMode / sizeof(dtype_t);
+}
+
+template <cute::UMMA::Major kMajorMode, uint32_t BLOCK_MN, uint32_t kSwizzleMode, typename dtype_t>
+CUTLASS_DEVICE
+constexpr uint32_t get_gmma_desc_stride_k() {
+    return kMajorMode == cute::UMMA::Major::K ? 1 : get_inner_block_atom_size<BLOCK_MN, kSwizzleMode, dtype_t>();
+}
+
+// ReSharper disable once CppNotAllPathsReturnValue
+template <cute::UMMA::Major kMajorMode, uint32_t kSwizzleMode, typename dtype_t>
+constexpr static cute::SM90::GMMA::LayoutType to_gmma_layout_type() {
+    DG_STATIC_ASSERT(kSwizzleMode == 0 or kSwizzleMode == 16 or
+                     kSwizzleMode == 32 or kSwizzleMode == 64 or
+                     kSwizzleMode == 128, "Invalid swizzling mode");
+
+    // Normal cases
+    if constexpr (kSwizzleMode == 0)   return cute::SM90::GMMA::LayoutType::INTERLEAVE;
+    if constexpr (kSwizzleMode == 16)  return cute::SM90::GMMA::LayoutType::INTERLEAVE;
+    if constexpr (kSwizzleMode == 32)  return cute::SM90::GMMA::LayoutType::B32;
+    if constexpr (kSwizzleMode == 64)  return cute::SM90::GMMA::LayoutType::B64;
+    if constexpr (kSwizzleMode == 128) return cute::SM90::GMMA::LayoutType::B128;
+}
+
+template <cute::UMMA::Major kMajorMode, uint32_t BLOCK_MN, uint32_t BLOCK_K, uint32_t kSwizzleMode, typename dtype_t>
+CUTLASS_DEVICE
+uint32_t advance_gmma_desc_lo(const uint32_t& base, const uint32_t& mn_idx, const uint32_t& k_idx, const uint32_t& offset = 0) {
+    return base + (((offset + mn_idx * BLOCK_K + k_idx * get_gmma_desc_stride_k<kMajorMode, BLOCK_MN, kSwizzleMode, dtype_t>()) * static_cast<uint32_t>(sizeof(dtype_t))) >> 4u);
+}
+
+template <cute::UMMA::Major kMajorMode, uint32_t BLOCK_MN, uint32_t BLOCK_K, uint32_t kSwizzleMode, typename dtype_t>
+CUTLASS_DEVICE
+cute::GmmaDescriptor make_gmma_desc(dtype_t* base_smem_ptr, uint32_t mn_idx, uint32_t k_idx) {
+    const uint32_t stride_k = get_gmma_desc_stride_k<kMajorMode, BLOCK_MN, kSwizzleMode, dtype_t>();
+    const auto layout_type = to_gmma_layout_type<kMajorMode, kSwizzleMode, dtype_t>();
+    constexpr uint32_t num_non_contiguous = 128 / 16;
+    if constexpr (kMajorMode == cute::UMMA::Major::K) {
+        // NOTES: for K-major layout, the swizzle must be 128B (also, atom index must be 0), as `BLOCK_K` is always 128
+        DG_STATIC_ASSERT(kSwizzleMode == BLOCK_K * sizeof(dtype_t), "Unexpected value");
+
+        // Atom size: 8 x `kSwizzleMode` (in bytes, on K)
+        // {SBO, LBO} means the byte stride between atoms on {MN, K}
+        // NOTES: on K, there is only 1 atom as asserted previously, so LBO can be 0
+        const uint32_t stride_byte_offset = num_non_contiguous * BLOCK_K * sizeof(dtype_t);
+        const uint32_t leading_byte_offset = 0;
+        return make_smem_desc(base_smem_ptr + mn_idx * BLOCK_K + k_idx * stride_k, static_cast<uint32_t>(layout_type),
+                              leading_byte_offset, stride_byte_offset);
+    } else {
+        constexpr uint32_t BLOCK_MN_ATOM = get_inner_block_atom_size<BLOCK_MN, kSwizzleMode, dtype_t>();
+
+        // Must have no in-atom MN-idx
+        // NOTES: no worries for the runtime assert, the `mn_idx` are constants at compilation time
+        DG_DEVICE_ASSERT(mn_idx % BLOCK_MN_ATOM == 0);
+        DG_STATIC_ASSERT(kSwizzleMode > 0, "Invalid swizzling");
+
+        // Atom size: `kSwizzleMode` (in bytes, on MN) x 8
+        // NOTES: `kSwizzleMode == 16` mean non-swizzling but interleaving
+        // {SBO, LBO} means the byte stride between atoms on {K, MN} for swizzling
+        // {SBO, LBO} means the byte stride between atoms on {MN, K} for non-swizzling
+        uint32_t stride_byte_offset = num_non_contiguous * BLOCK_MN_ATOM * sizeof(dtype_t);
+        uint32_t leading_byte_offset = BLOCK_K * BLOCK_MN_ATOM * sizeof(dtype_t);
+        if constexpr (kSwizzleMode == 16)
+            math::swap(stride_byte_offset, leading_byte_offset);
+        return make_smem_desc(base_smem_ptr + mn_idx * BLOCK_K + k_idx * stride_k, static_cast<uint32_t>(layout_type),
+                              leading_byte_offset, stride_byte_offset);
+    }
+}
+
+// ReSharper disable once CppNotAllPathsReturnValue
+template <uint32_t kHeadDim>
+static constexpr int to_swizzle_cute_type() {
+    DG_STATIC_ASSERT(kHeadDim == 32 or kHeadDim == 64 or kHeadDim == 128, "Invalid swizzling");
+    if constexpr (kHeadDim == 32)
+        return static_cast<int>(cute::SM90::GMMA::LayoutType::B32);
+    if constexpr (kHeadDim == 64)
+        return static_cast<int>(cute::SM90::GMMA::LayoutType::B64);
+    if constexpr (kHeadDim == 128)
+        return static_cast<int>(cute::SM90::GMMA::LayoutType::B128);
+}
+
+} // namespace deep_gemm::mma::sm90

deep_gemm/include/deep_gemm/ptx/ld_st.cuh

@@ -0,0 +1,247 @@
+#pragma once
+
+#include <cuda/std/cstdint>
+#include <cuda_bf16.h>
+
+namespace deep_gemm::ptx {
+
+// Compatibility: 256 bits LD/ST instructions
+#if defined(CUDART_VERSION) and CUDART_VERSION >= 13000
+using longlong4_t = longlong4_32a;
+#define make_longlong4_t make_longlong4_32a
+#else
+struct alignas(32) longlong4_t { long long x, y, z, w; };
+CUTLASS_HOST_DEVICE longlong4_t make_longlong4_t(
+    const long long& x, const long long& y, const long long& z, const long long& w) {
+    return {x, y, z, w};
+}
+#endif
+
+/// LD/ST matrix
+// TODO: remove `struct`
+struct SM90_U32x2_LDSM_N {
+    CUTLASS_DEVICE static void
+    copy(uint32_t& dst_0, uint32_t& dst_1, void* smem_src) {
+        asm volatile("ldmatrix.sync.aligned.x2.m8n8.shared.b16 {%0, %1}, [%2];\n"
+                     : "=r"(dst_0), "=r"(dst_1)
+                     : "l"(__cvta_generic_to_shared(smem_src)));
+    }
+};
+
+struct SM90_U32x4_LDSM_N {
+    CUTLASS_DEVICE static void
+    copy(uint32_t& dst_0, uint32_t& dst_1, uint32_t& dst_2, uint32_t& dst_3, void* smem_src) {
+        asm volatile("ldmatrix.sync.aligned.x4.m8n8.shared.b16 {%0, %1, %2, %3}, [%4];\n"
+                     : "=r"(dst_0), "=r"(dst_1), "=r"(dst_2), "=r"(dst_3)
+                     : "l"(__cvta_generic_to_shared(smem_src)));
+    }
+};
+
+template <typename dtype_t>
+struct SM90_U32x2_STSM_N {
+    CUTLASS_DEVICE static void
+    copy(dtype_t src_0, dtype_t src_1, void* smem_dst) {
+        DG_STATIC_ASSERT(sizeof(dtype_t) == sizeof(uint32_t), "Invalid dtype");
+        const uint32_t src[2] = {*reinterpret_cast<uint32_t*>(&src_0), *reinterpret_cast<uint32_t*>(&src_1)};
+        asm volatile("stmatrix.sync.aligned.x2.m8n8.shared.b16 [%0], {%1, %2};\n"
+                     :: "l"(__cvta_generic_to_shared(smem_dst)), "r"(src[0]), "r"(src[1]));
+    }
+};
+
+template <typename dtype_t>
+struct SM90_U32x4_STSM_T {
+    CUTLASS_DEVICE static void
+    copy(dtype_t src_0, dtype_t src_1, dtype_t src_2, dtype_t src_3, void* smem_dst) {
+        DG_STATIC_ASSERT(sizeof(dtype_t) == sizeof(uint32_t), "Invalid dtype");
+        const uint32_t src[4] = {*reinterpret_cast<uint32_t*>(&src_0), *reinterpret_cast<uint32_t*>(&src_1),
+                                 *reinterpret_cast<uint32_t*>(&src_2), *reinterpret_cast<uint32_t*>(&src_3)};
+        asm volatile("stmatrix.sync.aligned.x4.m8n8.shared.b16.trans [%0], {%1, %2, %3, %4};\n"
+                     :: "l"(__cvta_generic_to_shared(smem_dst)),
+                        "r"(src[0]), "r"(src[1]), "r"(src[2]), "r"(src[3]));
+    }
+};
+
+template <typename dtype_t>
+struct SM100_U8x4_STSM_T {
+    __device__ __forceinline__ static void
+    copy(dtype_t src_0, void* smem_dst) {
+        DG_STATIC_ASSERT(sizeof(dtype_t) == sizeof(uint32_t), "Invalid dtype");
+        const uint32_t src = *reinterpret_cast<uint32_t*>(&src_0);
+        asm volatile("stmatrix.sync.aligned.m16n8.x1.trans.shared.b8 [%0], {%1};\n"
+                     :: "l"(__cvta_generic_to_shared(smem_dst)), "r"(src));
+    }
+};
+
+template <typename dtype_t>
+struct SM100_U8x8_STSM_T {
+    __device__ __forceinline__ static void
+    copy(dtype_t src_0, dtype_t src_1, void* smem_dst) {
+        DG_STATIC_ASSERT(sizeof(dtype_t) == sizeof(uint32_t), "Invalid dtype");
+        const uint32_t src[2] = {*reinterpret_cast<uint32_t*>(&src_0), *reinterpret_cast<uint32_t*>(&src_1)};
+        asm volatile("stmatrix.sync.aligned.m16n8.x2.trans.shared.b8 [%0], {%1, %2};\n"
+                     :: "l"(__cvta_generic_to_shared(smem_dst)), "r"(src[0]), "r"(src[1]));
+    }
+};
+
+/// Shared memory
+CUTLASS_DEVICE uint32_t ld_shared(const uint32_t* ptr) {
+    uint32_t ret;
+    asm volatile("ld.shared.u32 %0, [%1];" : "=r"(ret) : "l"(__cvta_generic_to_shared(ptr)));
+    return ret;
+}
+
+CUTLASS_DEVICE float2 ld_shared(const float2* ptr) {
+    float2 ret;
+    asm volatile("ld.shared.v2.f32 {%0, %1}, [%2];" : "=f"(ret.x), "=f"(ret.y) : "l"(__cvta_generic_to_shared(ptr)));
+    return ret;
+}
+
+CUTLASS_DEVICE float4 ld_shared(const float4* ptr) {
+    float4 ret;
+    asm volatile("ld.shared.v4.f32 {%0, %1, %2, %3}, [%4];" : "=f"(ret.x), "=f"(ret.y), "=f"(ret.z), "=f"(ret.w) : "l"(__cvta_generic_to_shared(ptr)));
+    return ret;
+}
+
+CUTLASS_DEVICE uint4 ld_shared(const uint4* ptr) {
+    uint4 ret;
+    asm volatile("ld.shared.v4.u32 {%0, %1, %2, %3}, [%4];" : "=r"(ret.x), "=r"(ret.y), "=r"(ret.z), "=r"(ret.w) : "l"(__cvta_generic_to_shared(ptr)));
+    return ret;
+}
+
+CUTLASS_DEVICE float ld_shared(const float* ptr) {
+    float ret;
+    asm volatile("ld.shared.f32 %0, [%1];" : "=f"(ret) : "l"(__cvta_generic_to_shared(ptr)));
+    return ret;
+}
+
+CUTLASS_DEVICE void st_shared(const float* ptr, float val) {
+    asm volatile("st.shared.f32 [%0], %1;" :: "l"(__cvta_generic_to_shared(ptr)), "f"(val));
+}
+
+CUTLASS_DEVICE void st_shared(const float2* ptr, float2 val) {
+    asm volatile("st.shared.v2.f32 [%0], {%1, %2};" :: "l"(__cvta_generic_to_shared(ptr)), "f"(val.x), "f"(val.y));
+}
+
+CUTLASS_DEVICE void st_shared(const uint32_t* ptr, uint32_t val) {
+    asm volatile("st.shared.u32 [%0], %1;" :: "l"(__cvta_generic_to_shared(ptr)), "r"(val));
+}
+
+CUTLASS_DEVICE void st_shared(const void* ptr, uint32_t x, uint32_t y) {
+    asm volatile("st.shared.v2.u32 [%0], {%1, %2};" :: "l"(__cvta_generic_to_shared(ptr)), "r"(x), "r"(y));
+}
+
+CUTLASS_DEVICE void st_shared(const void* ptr, uint32_t x, uint32_t y, uint32_t z, uint32_t w) {
+    asm volatile("st.shared.v4.u32 [%0], {%1, %2, %3, %4};" :: "l"(__cvta_generic_to_shared(ptr)), "r"(x), "r"(y), "r"(z), "r"(w));
+}
+
+CUTLASS_DEVICE void st_shared(const __int128_t* ptr, __int128_t val) {
+    asm volatile("st.shared.b128 [%0], %1;" :: "l"(__cvta_generic_to_shared(ptr)), "q"(val));
+}
+
+CUTLASS_DEVICE void st_shared_bulk(void* smem_ptr, const uint32_t& num_bytes) {
+    // `size` must be 64-bit before PTX ISA 9.0
+    asm volatile("st.bulk.weak.shared::cta [%0], %1, 0;" ::
+                 "l"(__cvta_generic_to_shared(smem_ptr)), "l"(static_cast<uint64_t>(num_bytes)));
+}
+
+/// Global memory
+CUTLASS_DEVICE uint64_t ld_volatile(const uint64_t* ptr) {
+    uint64_t ret;
+    asm volatile("ld.volatile.global.b64 %0, [%1];" : "=l"(ret) : "l"(ptr));
+    return ret;
+}
+
+CUTLASS_DEVICE uint32_t ld_acq(const uint32_t* ptr) {
+    uint32_t ret;
+    asm volatile("ld.acquire.gpu.global.b32 %0, [%1];" : "=r"(ret) : "l"(ptr));
+    return ret;
+}
+
+CUTLASS_DEVICE uint64_t ld_acq_sys(const uint64_t* ptr) {
+    uint64_t ret;
+    asm volatile("ld.acquire.sys.global.b64 %0, [%1];" : "=l"(ret) : "l"(ptr));
+    return ret;
+}
+
+CUTLASS_DEVICE void st_relaxed_sys(const uint64_t* ptr, const uint64_t& value) {
+    asm volatile("st.L1::no_allocate.relaxed.sys.u64 [%0], %1;" :: "l"(ptr), "l"(value));
+}
+
+/// Atomics
+CUTLASS_DEVICE uint64_t atomic_add(const uint64_t* ptr, const uint64_t& value) {
+    uint64_t ret;
+    asm volatile("atom.global.add.u64 %0, [%1], %2;" : "=l"(ret) : "l"(ptr), "l"(value));
+    return ret;
+}
+
+CUTLASS_DEVICE uint64_t atomic_add_sys(const uint64_t* ptr, const uint64_t& value) {
+    uint64_t ret;
+    asm volatile("atom.sys.global.add.u64 %0, [%1], %2;" : "=l"(ret) : "l"(ptr), "l"(value));
+    return ret;
+}
+
+CUTLASS_DEVICE uint32_t atomic_add_rel(const uint32_t* ptr, const uint32_t& value) {
+    uint32_t ret;
+    asm volatile("atom.release.gpu.global.add.u32 %0, [%1], %2;" : "=r"(ret) : "l"(ptr), "r"(value));
+    return ret;
+}
+
+__forceinline__ __device__ void red_add(const uint32_t* ptr, const uint32_t& value) {
+    asm volatile("red.gpu.global.add.u32 [%0], %1;" :: "l"(ptr), "r"(value));
+}
+
+CUTLASS_DEVICE void red_or_rel_sys(const uint64_t* ptr, const uint64_t& value) {
+    asm volatile("red.release.sys.global.or.b64 [%0], %1;" :: "l"(ptr), "l"(value));
+}
+
+CUTLASS_DEVICE void red_or_rel_gpu(uint64_t* ptr, const uint64_t& value) {
+    asm volatile("red.release.gpu.global.or.b64 [%0], %1;" :: "l"(ptr), "l"(value));
+}
+
+CUTLASS_DEVICE void red_add_rel(const uint32_t* ptr, const uint32_t& value) {
+    asm volatile("red.release.gpu.global.add.u32 [%0], %1;" :: "l"(ptr), "r"(value));
+}
+
+CUTLASS_DEVICE void red_add_rel_sys(const int* ptr, const int& value) {
+    asm volatile("red.release.sys.global.add.s32 [%0], %1;" :: "l"(ptr), "r"(value));
+}
+
+CUTLASS_DEVICE int ld_acq_sys(const int* ptr) {
+    int ret;
+    asm volatile("ld.acquire.sys.global.s32 %0, [%1];" : "=r"(ret) : "l"(ptr));
+    return ret;
+}
+
+CUTLASS_DEVICE uint32_t ld_acq_sys(const uint32_t* ptr) {
+    uint32_t ret;
+    asm volatile("ld.acquire.sys.global.u32 %0, [%1];" : "=r"(ret) : "l"(ptr));
+    return ret;
+}
+
+CUTLASS_DEVICE uint64_t ld_acq_gpu(const uint64_t* ptr) {
+    uint64_t ret;
+    asm volatile("ld.acquire.gpu.global.u64 %0, [%1];" : "=l"(ret) : "l"(ptr));
+    return ret;
+}
+
+/// Predicated loads
+CUTLASS_DEVICE longlong4_t ld_gez_pred(const longlong4_t* ptr, const int& pred) {
+    longlong4_t ret = make_longlong4_t(0, 0, 0, 0);
+    asm volatile(
+        "{\n\t"
+        "  .reg .pred p;\n\t"
+        "  setp.ge.s32 p, %5, 0;\n\t"
+        "  @p ld.global.L2::256B.v4.s64 {%0, %1, %2, %3}, [%4];\n\t"
+        "}"
+        : "+l"(ret.x), "+l"(ret.y), "+l"(ret.z), "+l"(ret.w)
+        : "l"(ptr), "r"(pred)
+        : "memory");
+    return ret;
+}
+
+/// Prefetch
+CUTLASS_DEVICE void prefetch_l1(void *ptr) {
+    asm volatile("prefetch.global.L1 [%0];" :: "l"(ptr));
+}
+
+} // namespace deep_gemm::ptx

deep_gemm/include/deep_gemm/ptx/tcgen05.cuh

@@ -0,0 +1,168 @@
+#pragma once
+
+namespace deep_gemm::ptx {
+
+/// UMMA versions with relaxed assertions
+struct SM100_MMA_F16BF16_SS {
+    CUTLASS_DEVICE static void
+    fma(uint64_t const& desc_a,
+        uint64_t const& desc_b,
+        uint32_t const& tmem_c,
+        uint32_t const& scale_c,
+        uint64_t const& desc) {
+        asm volatile(
+            "{\n\t"
+            ".reg .pred p;\n\t"
+            "setp.ne.b32 p, %4, 0;\n\t"
+            "tcgen05.mma.cta_group::1.kind::f16 [%0], %1, %2, %3, p; \n\t"
+            "}\n"
+            :: "r"(tmem_c), "l"(desc_a), "l"(desc_b), "r"(static_cast<uint32_t>(desc >> 32)), "r"(scale_c));
+    }
+};
+
+struct SM100_MMA_F16BF16_2x1SM_SS {
+    CUTLASS_DEVICE static void
+    fma(uint64_t const& desc_a,
+        uint64_t const& desc_b,
+        uint32_t const& tmem_c,
+        uint32_t const& scale_c,
+        uint64_t const& desc) {
+        asm volatile(
+            "{\n\t"
+            ".reg .pred p;\n\t"
+            "setp.ne.b32 p, %4, 0;\n\t"
+            "tcgen05.mma.cta_group::2.kind::f16 [%0], %1, %2, %3, p; \n\t"
+            "}\n"
+            :: "r"(tmem_c), "l"(desc_a), "l"(desc_b), "r"(static_cast<uint32_t>(desc >> 32)), "r"(scale_c));
+    }
+};
+
+struct SM100_MMA_MXF8F6F4_SS {
+    CUTLASS_DEVICE static void
+    fma(uint64_t const& desc_a,
+        uint64_t const& desc_b,
+        uint32_t const& tmem_c,
+        uint32_t const& scale_c,
+        uint64_t const& desc,
+        uint32_t const& tmem_sfa,
+        uint32_t const& tmem_sfb) {
+        asm volatile(
+          "{\n\t"
+          ".reg .pred p;\n\t"
+          "setp.ne.b32 p, %4, 0;\n\t"
+          "tcgen05.mma.cta_group::1.kind::mxf8f6f4.block_scale [%0], %1, %2, %3, [%5], [%6], p; \n\t"
+          "}\n"
+          :
+          : "r"(tmem_c), "l"(desc_a), "l"(desc_b), "r"(static_cast<uint32_t>(desc >> 32)), "r"(scale_c),
+            "r"(tmem_sfa), "r"(tmem_sfb));
+    }
+};
+
+struct SM100_MMA_MXF8F6F4_2x1SM_SS {
+    CUTLASS_DEVICE static void
+    fma(uint64_t const& desc_a,
+        uint64_t const& desc_b,
+        uint32_t const& tmem_c,
+        uint32_t const& scale_c,
+        uint64_t const& desc,
+        uint32_t const& tmem_sfa,
+        uint32_t const& tmem_sfb) {
+        asm volatile(
+          "{\n\t"
+          ".reg .pred p;\n\t"
+          "setp.ne.b32 p, %4, 0;\n\t"
+          "tcgen05.mma.cta_group::2.kind::mxf8f6f4.block_scale [%0], %1, %2, %3, [%5], [%6], p; \n\t"
+          "}\n"
+          :
+          : "r"(tmem_c), "l"(desc_a), "l"(desc_b), "r"(static_cast<uint32_t>(desc >> 32)), "r"(scale_c),
+            "r"(tmem_sfa), "r"(tmem_sfb));
+    }
+};
+
+struct SM100_MMA_F8F6F4_SS {
+    CUTLASS_DEVICE static void
+    fma(uint64_t const& desc_a,
+        uint64_t const& desc_b,
+        uint32_t const& tmem_c,
+        uint32_t const& scale_c,
+        uint64_t const& desc) {
+        asm volatile(
+          "{\n\t"
+          ".reg .pred p;\n\t"
+          "setp.ne.b32 p, %4, 0;\n\t"
+          "tcgen05.mma.cta_group::1.kind::f8f6f4 [%0], %1, %2, %3, p; \n\t"
+          "}\n"
+          :
+          : "r"(tmem_c), "l"(desc_a), "l"(desc_b), "r"(static_cast<uint32_t>(desc >> 32)), "r"(scale_c));
+    }
+};
+
+struct SM100_MMA_F8F6F4_2x1SM_SS {
+    CUTLASS_DEVICE static void
+    fma(uint64_t const& desc_a,
+        uint64_t const& desc_b,
+        uint32_t const& tmem_c,
+        uint32_t const& scale_c,
+        uint64_t const& desc) {
+        asm volatile(
+          "{\n\t"
+          ".reg .pred p;\n\t"
+          "setp.ne.b32 p, %4, 0;\n\t"
+          "tcgen05.mma.cta_group::2.kind::f8f6f4 [%0], %1, %2, %3, p; \n\t"
+          "}\n"
+          :
+          : "r"(tmem_c), "l"(desc_a), "l"(desc_b), "r"(static_cast<uint32_t>(desc >> 32)), "r"(scale_c));
+    }
+};
+
+struct SM100_MMA_MXF4_SS {
+    CUTLASS_DEVICE static void
+    fma(uint64_t const& desc_a,
+        uint64_t const& desc_b,
+        uint32_t const& tmem_c,
+        uint32_t const& scale_c,
+        uint64_t const& desc,
+        uint32_t const& tmem_sfa,
+        uint32_t const& tmem_sfb) {
+        asm volatile(
+            "{\n\t"
+            ".reg .pred p;\n\t"
+            "setp.ne.b32 p, %4, 0;\n\t"
+#if (__CUDACC_VER_MAJOR__ > 12) || (__CUDACC_VER_MAJOR__ == 12 && __CUDACC_VER_MINOR__ >= 9)
+            "tcgen05.mma.cta_group::1.kind::mxf4.block_scale.block32 [%0], %1, %2, %3, [%5], [%6], p; \n\t"
+#else
+            "tcgen05.mma.cta_group::1.kind::mxf4.block_scale.scale_vec::2X [%0], %1, %2, %3, [%5], [%6], p; \n\t"
+#endif
+            "}\n"
+            :: "r"(tmem_c), "l"(desc_a), "l"(desc_b), "r"(static_cast<uint32_t>(desc >> 32)), "r"(scale_c),
+               "r"(tmem_sfa), "r"(tmem_sfb));
+    }
+};
+
+struct SM100_MMA_F16BF16_WS_SS {
+    CUTLASS_DEVICE static void
+    fma(uint64_t const& desc_a,
+        uint64_t const& desc_b,
+        uint32_t const& tmem_c,
+        uint32_t const& scale_c,
+        uint64_t const& desc) {
+        asm volatile(
+            "{\n\t"
+            ".reg .pred p;\n\t"
+            "setp.ne.b32 p, %4, 0;\n\t"
+            "tcgen05.mma.ws.cta_group::1.kind::f16 [%0], %1, %2, %3, p; \n\t"
+            "}\n"
+            :: "r"(tmem_c), "l"(desc_a), "l"(desc_b), "r"(static_cast<uint32_t>(desc >> 32)), "r"(scale_c));
+    }
+};
+
+/// Tensor memory operations
+CUTLASS_DEVICE void tcgen05_before_thread_sync() {
+    asm volatile("tcgen05.fence::before_thread_sync;");
+}
+
+CUTLASS_DEVICE void tcgen05_after_thread_sync() {
+    asm volatile("tcgen05.fence::after_thread_sync;");
+}
+
+} // namespace deep_gemm::ptx

deep_gemm/include/deep_gemm/ptx/tma.cuh

@@ -0,0 +1,112 @@
+#pragma once
+
+#include <cutlass/arch/barrier.h>
+#include <cute/arch/copy_sm90_desc.hpp>
+
+namespace deep_gemm::ptx {
+
+// Tensor-map instructions
+CUTLASS_DEVICE void tensor_map_release_gpu() {
+    asm volatile ("fence.proxy.tensormap::generic.release.gpu;" ::: "memory");
+}
+
+CUTLASS_DEVICE void tensor_map_acquire_gpu(const cute::TmaDescriptor* gmem_desc_ptr) {
+    auto gmem_int_desc = reinterpret_cast<uint64_t>(gmem_desc_ptr);
+    asm volatile ("fence.proxy.tensormap::generic.acquire.gpu [%0], 128;" :: "l"(gmem_int_desc) : "memory");
+}
+
+CUTLASS_DEVICE void tensor_map_replace_global_addr_in_smem(cute::TmaDescriptor* smem_desc, const void* new_addr) {
+    auto smem_int_desc = static_cast<uint32_t>(__cvta_generic_to_shared(smem_desc));
+    const auto new_int64_addr = reinterpret_cast<uint64_t>(new_addr);
+    asm volatile ("tensormap.replace.tile.global_address.shared::cta.b1024.b64 [%0], %1;" :: "r"(smem_int_desc), "l"(new_int64_addr));
+}
+
+CUTLASS_DEVICE void tensor_map_replace_global_inner_dim_stride_in_smem(cute::TmaDescriptor* smem_desc, const uint32_t& new_dim, const uint64_t& new_stride) {
+    auto smem_int_desc = __cvta_generic_to_shared(smem_desc);
+    asm volatile ("tensormap.replace.tile.global_dim.shared::cta.b1024.b32 [%0], 0, %1;" :: "l"(smem_int_desc), "r"(new_dim));
+#if ((__CUDACC_VER_MAJOR__ > 12) or ((__CUDACC_VER_MAJOR__ == 12) and (__CUDACC_VER_MINOR__ >= 3)))
+    asm volatile("tensormap.replace.tile.global_stride.shared::cta.b1024.b64 [%0], 0, %1;" :: "l"(smem_int_desc), "l"(new_stride));
+#else
+    DG_STATIC_ASSERT(false, "Invalid CUDA version");
+#endif
+}
+
+/// TMA instructions
+CUTLASS_DEVICE void mbarrier_arrive(
+    cutlass::arch::ClusterTransactionBarrier* ptr) {
+    asm volatile("mbarrier.arrive.shared::cta.b64 _, [%0]; \n\t" ::
+                 "r"(static_cast<uint32_t>(__cvta_generic_to_shared(ptr))));
+}
+
+CUTLASS_DEVICE void mbarrier_arrive_and_set_tx(
+    cutlass::arch::ClusterTransactionBarrier* ptr, const uint32_t& num_bytes) {
+    asm volatile("mbarrier.arrive.expect_tx.shared::cta.b64 _, [%1], %0; \n\t" ::
+                 "r"(num_bytes), "r"(static_cast<uint32_t>(__cvta_generic_to_shared(ptr))));
+}
+
+CUTLASS_DEVICE void mbarrier_wait_and_flip_phase(
+    cutlass::arch::ClusterTransactionBarrier* ptr, uint32_t& phase) {
+    asm volatile(
+        "{\n\t"
+        ".reg .pred       P1; \n\t"
+        "LAB_WAIT: \n\t"
+        "mbarrier.try_wait.parity.shared::cta.b64 P1, [%0], %1, %2; \n\t"
+        "@P1 bra DONE; \n\t"
+        "bra     LAB_WAIT; \n\t"
+        "DONE: \n\t"
+        "}" ::
+        "r"(static_cast<uint32_t>(__cvta_generic_to_shared(ptr))),
+        "r"(phase), "r"(0x989680));
+    phase ^= 1;
+}
+
+CUTLASS_DEVICE void tma_load_1d(
+    const void* dst_ptr, const void* src_ptr,
+    cutlass::arch::ClusterTransactionBarrier* mbarrier_ptr,
+    const uint32_t& num_bytes,
+    const cute::TMA::CacheHintSm90& hint = cute::TMA::CacheHintSm90::EVICT_FIRST) {
+    // NOTES: normally, the loaded part will be evicted soon
+    asm volatile(
+        "cp.async.bulk.shared::cluster.global.mbarrier::complete_tx::bytes.L2::cache_hint [%0], [%1], %2, [%3], %4;\n" ::
+        "r"(static_cast<uint32_t>(__cvta_generic_to_shared(dst_ptr))),
+        "l"(src_ptr),
+        "r"(num_bytes),
+        "r"(static_cast<uint32_t>(__cvta_generic_to_shared(mbarrier_ptr))),
+        "l"(hint)
+        : "memory");
+}
+
+CUTLASS_DEVICE void tma_store_1d(
+    const void* dst_ptr, const void* src_ptr, const uint32_t& num_bytes,
+    const cute::TMA::CacheHintSm90& hint = cute::TMA::CacheHintSm90::EVICT_NORMAL) {
+    // NOTES: normally, the stored part will be used soon
+    asm volatile("cp.async.bulk.global.shared::cta.bulk_group.L2::cache_hint [%0], [%1], %2, %3;\n" ::
+                 "l"(dst_ptr),
+                 "r"(static_cast<uint32_t>(__cvta_generic_to_shared(src_ptr))),
+                 "r"(num_bytes),
+                 "l"(hint)
+                 : "memory");
+}
+
+template <int kNumRemainingWaits = 0>
+__forceinline__ __device__ void tma_store_wait() {
+    // NOTES: this function does not have `.read`
+    asm volatile("cp.async.bulk.wait_group %0;" ::"n"(kNumRemainingWaits) : "memory");
+}
+
+CUTLASS_DEVICE
+void tma_gather4(const void* desc_ptr, cutlass::arch::ClusterTransactionBarrier& mbarrier,
+                 void* smem_ptr, const uint32_t& col_idx, const int4& row_idxs, const uint64_t& cache_hint) {
+    const auto smem_addr = cute::cast_smem_ptr_to_uint(smem_ptr);
+    const auto mbarrier_addr = cute::cast_smem_ptr_to_uint(&mbarrier);
+    asm volatile(
+        "cp.async.bulk.tensor.2d.shared::cta.global.tile::gather4.mbarrier::complete_tx::bytes.cta_group::1.L2::cache_hint [%0], [%1, {%2, %3, %4, %5, %6}], [%7], %8;\n"
+        :
+        : "r"(smem_addr), "l"(desc_ptr), "r"(col_idx),
+          "r"(row_idxs.x), "r"(row_idxs.y), "r"(row_idxs.z), "r"(row_idxs.w),
+          "r"(mbarrier_addr), "l"(cache_hint)
+        : "memory"
+    );
+}
+
+} // namespace deep_gemm::ptx

deep_gemm/include/deep_gemm/ptx/utils.cuh

@@ -0,0 +1,53 @@
+#pragma once
+
+#include <cuda/std/cstdint>
+#include <cuda_bf16.h>
+
+#include <deep_gemm/common/exception.cuh>
+
+namespace deep_gemm::ptx {
+
+CUTLASS_DEVICE uint32_t get_sm_idx() {
+    uint32_t sm_idx;
+    asm ("mov.u32 %0, %%smid;" : "=r"(sm_idx));
+    return sm_idx;
+}
+
+CUTLASS_DEVICE uint32_t get_lane_idx() {
+    uint32_t lane_id;
+    asm ("mov.u32 %0, %%laneid;" : "=r"(lane_id));
+    return lane_id;
+}
+
+CUTLASS_DEVICE void sync_aligned(const uint32_t& num_threads, const uint32_t& barrier_idx) {
+    asm volatile("bar.sync %0, %1;" : : "r"(barrier_idx), "r"(num_threads));
+}
+
+CUTLASS_DEVICE void sync_unaligned(const uint32_t& num_threads, const uint32_t& barrier_idx) {
+    asm volatile("barrier.sync %0, %1;" : : "r"(barrier_idx), "r"(num_threads));
+}
+
+template <typename dtype_t>
+CUTLASS_DEVICE dtype_t exchange(dtype_t ptr, const uint32_t& src_lane_idx) {
+    DG_STATIC_ASSERT(sizeof(dtype_t) % sizeof(uint32_t) == 0, "");
+    const auto send_int_values = reinterpret_cast<uint32_t*>(&ptr);
+    dtype_t recv_dtype;
+    auto recv_int_values = reinterpret_cast<uint32_t*>(&recv_dtype);
+    #pragma unroll
+    for (uint32_t i = 0; i < sizeof(dtype_t) / sizeof(uint32_t); ++ i)
+        recv_int_values[i] = __shfl_sync(0xffffffff, send_int_values[i], static_cast<int>(src_lane_idx));
+    return recv_dtype;
+}
+
+CUTLASS_DEVICE void accumulate(float2& a, nv_bfloat162 b) {
+#if defined(__CUDA_ARCH__) and (__CUDA_ARCH__ >= 1000)
+    // Use `add.rn.f32.bf16` instruction to perform fused (cast + add) operation on SM100
+    asm("add.rn.f32.bf16 %0, %1, %0;\n" : "+f"(a.x) : "h"(*reinterpret_cast<uint16_t*>(&b.x)));
+    asm("add.rn.f32.bf16 %0, %1, %0;\n" : "+f"(a.y) : "h"(*reinterpret_cast<uint16_t*>(&b.y)));
+#else
+    const auto [x, y] = __bfloat1622float2(b);
+    a.x += x, a.y += y;
+#endif
+}
+
+} // namespace deep_gemm::ptx

deep_gemm/include/deep_gemm/ptx/wgmma.cuh

@@ -0,0 +1,25 @@
+#pragma once
+
+#include <deep_gemm/common/exception.cuh>
+
+namespace deep_gemm::ptx {
+
+CUTLASS_DEVICE void warpgroup_arrive() {
+    asm volatile("wgmma.fence.sync.aligned;\n" ::: "memory");
+}
+
+CUTLASS_DEVICE void warpgroup_commit_batch() {
+    asm volatile("wgmma.commit_group.sync.aligned;\n" ::: "memory");
+}
+
+CUTLASS_DEVICE void warpgroup_fence_operand(float& reg) {
+    asm volatile("" : "+f"(reg) :: "memory");
+}
+
+template <int N>
+CUTLASS_DEVICE void warpgroup_wait() {
+    DG_STATIC_ASSERT(N >= 0 and N <= 7, "WGMMA wait: N must be in range [0, 7]");
+    asm volatile("wgmma.wait_group.sync.aligned %0;\n" :: "n"(N) : "memory");
+}
+
+} // namespace deep_gemm::ptx

deep_gemm/include/deep_gemm/scheduler/gemm.cuh

@@ -0,0 +1,300 @@
+#pragma once
+
+#include <deep_gemm/common/math.cuh>
+#include <deep_gemm/common/types.cuh>
+
+namespace deep_gemm::sched {
+
+enum class IndexType {
+    MN,
+    K,
+    SF_K,
+};
+
+template <GemmType kGemmType, uint32_t BLOCK_M, uint32_t BLOCK_N, uint32_t kNumSMs, bool kIsMulticastOnA>
+static constexpr uint32_t get_num_1d_blocks_per_group() {
+    // Select the best from candidates
+    uint32_t num_best_blocks = 0, min_usage = cute::numeric_limits<uint32_t>::max();
+    for (const auto candidate: {8u, 16u}) {
+        const auto usage = kIsMulticastOnA ?
+            candidate * BLOCK_N + math::constexpr_ceil_div(kNumSMs, candidate) * BLOCK_M: // Grouping on N
+            candidate * BLOCK_M + math::constexpr_ceil_div(kNumSMs, candidate) * BLOCK_N; // Grouping on M
+        if (usage < min_usage)
+            min_usage = usage, num_best_blocks = candidate;
+    }
+    return num_best_blocks;
+}
+
+#pragma clang diagnostic push
+#pragma ide diagnostic ignored "cppcoreguidelines-pro-type-member-init"
+template <GemmType kGemmType,
+          uint32_t BLOCK_M, uint32_t BLOCK_N,
+          uint32_t kNumGroups,
+          uint32_t kNumMulticast, bool kIsMulticastOnA,
+          uint32_t kNumSMs,
+          uint32_t SF_K_ALIGNMENT = 512u,  // for k-grouped GEMM only: 128 on SM90 (float SF), gran_k * 4 on SM100 (packed UE8M0 SF)
+          uint32_t kNum1DBlocksPerGroup = get_num_1d_blocks_per_group<kGemmType, BLOCK_M, BLOCK_N, kNumSMs, kIsMulticastOnA>()>
+struct Scheduler {
+    int current_iter = -1;
+
+    // Block configs
+    uint32_t num_blocks;
+    uint32_t num_m_blocks;
+    uint32_t num_n_blocks;
+
+    // For SM90 multicast checks
+    uint32_t num_blocks_in_group;
+    bool is_peer_cta_alive = true;
+
+    // For grouped GEMM
+    int* grouped_layout;
+    uint32_t current_group_idx = 0;
+    // Only used for masked layout
+    uint32_t current_m_cumsum = 0;
+    // Only used for contiguous psum layout
+    uint32_t last_psum_m = 0, current_psum_m, current_m_block_cumsum = 0;
+    // Only used for k-grouped layout
+    uint32_t current_shape_k, current_num_valid_groups = 0, current_k_cumsum = 0, current_sf_k_cumsum = 0;
+    uint32_t next_group_idx, next_shape_k;
+
+    // Only used for k-grouped gemm
+    CUTLASS_DEVICE void get_next_k_group(uint32_t &group_idx, uint32_t &shape_k) const {
+        for (; group_idx < kNumGroups; ++ group_idx) {
+            shape_k = grouped_layout[group_idx];
+            if (shape_k > 0)
+                break;
+        }
+    }
+
+    // ReSharper disable once CppPossiblyUninitializedMember
+    CUTLASS_DEVICE explicit Scheduler(const uint32_t& shape_m, const uint32_t& shape_n,
+                                       const uint32_t& shape_k, int* grouped_layout = nullptr) {
+        num_m_blocks = math::ceil_div(shape_m, BLOCK_M);
+        num_n_blocks = math::ceil_div(shape_n, BLOCK_N);
+        current_shape_k = shape_k;
+        if constexpr (kGemmType == GemmType::Normal or kGemmType == GemmType::Batched) {
+            num_blocks = num_m_blocks * num_n_blocks;
+        } else if constexpr (kGemmType == GemmType::MGroupedContiguous) {
+            num_blocks = num_m_blocks * num_n_blocks;
+            this->grouped_layout = grouped_layout;
+        } else if constexpr (kGemmType == GemmType::MGroupedMasked) {
+            this->grouped_layout = grouped_layout;
+        } else if constexpr (kGemmType == GemmType::MGroupedContiguousWithPsumLayout) {
+            this->grouped_layout = grouped_layout;
+            current_psum_m = grouped_layout[0];
+            num_m_blocks = math::ceil_div(current_psum_m, BLOCK_M);
+        } else if constexpr (kGemmType == GemmType::KGroupedContiguous) {
+            num_blocks = num_m_blocks * num_n_blocks;
+            this->grouped_layout = grouped_layout;
+            get_next_k_group(current_group_idx, current_shape_k);
+            next_group_idx = current_group_idx + 1;
+            get_next_k_group(next_group_idx, next_shape_k);
+        }
+    }
+
+    CUTLASS_DEVICE void get_swizzled_block_idx(const uint32_t& block_idx, uint32_t& m_block_idx, uint32_t& n_block_idx) {
+        DG_STATIC_ASSERT(kNum1DBlocksPerGroup % kNumMulticast == 0, "Invalid group size");
+
+        // Swizzle for better L2 usages
+        const auto primary_num_blocks = kIsMulticastOnA ? num_n_blocks : num_m_blocks;
+        const auto secondary_num_blocks = kIsMulticastOnA ? num_m_blocks : num_n_blocks;
+        const auto num_blocks_per_group = secondary_num_blocks * kNum1DBlocksPerGroup;
+        const auto group_idx = block_idx / num_blocks_per_group;
+        auto first_block_idx = group_idx * kNum1DBlocksPerGroup;
+        auto in_group_idx = block_idx % num_blocks_per_group;
+        num_blocks_in_group = min(kNum1DBlocksPerGroup, primary_num_blocks - first_block_idx);
+
+        // Fix unaligned TMA multicast
+        // NOTES: for SM90 only, as SM90 can dynamically disable TMA multicast
+        // while SM100 uses 2-CTA, which can not be dynamically disabled
+#if __CUDA_ARCH__ < 1000
+        if (kNumMulticast > 1 and num_blocks_in_group % 2 != 0) {
+            if (in_group_idx < (num_blocks_in_group ^ 1) * secondary_num_blocks) {
+                num_blocks_in_group = num_blocks_in_group ^ 1;
+            } else {
+                in_group_idx = in_group_idx - (num_blocks_in_group ^ 1) * secondary_num_blocks;
+                first_block_idx += num_blocks_in_group ^ 1;
+                num_blocks_in_group = 1;
+            }
+        }
+#endif
+
+        // Convert to final M/N block indices
+        // `kIsMulticastOnA == true` leads to groups on N
+        if constexpr (kIsMulticastOnA) {
+            m_block_idx = in_group_idx / num_blocks_in_group;
+            n_block_idx = first_block_idx + in_group_idx % num_blocks_in_group;
+        } else {
+            m_block_idx = first_block_idx + in_group_idx % num_blocks_in_group;
+            n_block_idx = in_group_idx / num_blocks_in_group;
+        }
+    }
+
+    template <bool kWithGroupOffset, IndexType kIndexType = IndexType::MN>
+    CUTLASS_DEVICE uint32_t get_global_idx(const uint32_t shape_dim, const uint32_t block_size,
+                                             const uint32_t& block_idx, const uint32_t& m_block_idx = 0) {
+        if constexpr (kGemmType == GemmType::Normal) {
+            return block_idx * block_size;
+        } else if constexpr (kGemmType == GemmType::MGroupedContiguous) {
+            const auto offset = kWithGroupOffset ? cute::max(0, grouped_layout[m_block_idx * BLOCK_M]) : 0;
+            return offset * shape_dim + block_idx * block_size;
+        } else if constexpr (kGemmType == GemmType::MGroupedMasked or kGemmType == GemmType::MGroupedContiguousWithPsumLayout) {
+            const auto offset = kWithGroupOffset ? current_group_idx : 0;
+            return offset * shape_dim + block_idx * block_size;
+        } else if constexpr (kGemmType == GemmType::KGroupedContiguous) {
+            auto offset = 0;
+            if constexpr (kWithGroupOffset) {
+                if constexpr (kIndexType == IndexType::MN)
+                    offset = current_group_idx * shape_dim;
+                else if constexpr (kIndexType == IndexType::K)
+                    offset = current_k_cumsum;
+                else if constexpr (kIndexType == IndexType::SF_K)
+                    offset = current_sf_k_cumsum;
+            }
+            return offset + block_idx * block_size;
+        } else if constexpr (kGemmType == GemmType::Batched) {
+            // Ignore kWithGroupOffset, and apply offset for IndexType::SF_K
+            const auto offset = kIndexType == IndexType::SF_K ? current_group_idx : 0;
+            return offset * shape_dim + block_idx * block_size;
+        }
+    }
+
+    // For swap A/B and psum layout only
+    CUTLASS_DEVICE uint32_t get_aligned_effective_m_in_block(const uint32_t& m_block_idx) const {
+        constexpr uint32_t UMMA_STEP_N = 16;
+        DG_STATIC_ASSERT(BLOCK_M % UMMA_STEP_N == 0, "Invalid alignment");
+        if constexpr (kGemmType == GemmType::MGroupedContiguousWithPsumLayout)
+            return math::align(m_block_idx == last_psum_m / BLOCK_M + num_m_blocks - 1 ? current_psum_m - m_block_idx * BLOCK_M : BLOCK_M, UMMA_STEP_N);
+        return BLOCK_M;
+    }
+
+    CUTLASS_DEVICE bool get_next_block(uint32_t& m_block_idx, uint32_t& n_block_idx) {
+        const auto next_block_idx = (++ current_iter) * kNumSMs + blockIdx.x;
+
+        if constexpr (kGemmType == GemmType::MGroupedMasked) {
+            while (true) {
+                // End of the task
+                if (current_group_idx == kNumGroups)
+                    return false;
+
+                // Within current group
+                num_m_blocks = math::ceil_div(static_cast<uint32_t>(grouped_layout[current_group_idx]), BLOCK_M);
+                const auto current_m_block_cumsum = current_m_cumsum + num_m_blocks;
+                if (next_block_idx < current_m_block_cumsum * num_n_blocks)
+                    break;
+
+                // Move to check the next group
+                current_group_idx ++, current_m_cumsum = current_m_block_cumsum;
+            }
+
+            get_swizzled_block_idx(next_block_idx - current_m_cumsum * num_n_blocks, m_block_idx, n_block_idx);
+        } else if constexpr (kGemmType == GemmType::MGroupedContiguousWithPsumLayout) { 
+            while (true) {
+                // Within current group
+                if (next_block_idx < (current_m_block_cumsum + num_m_blocks) * num_n_blocks)
+                    break;
+
+                // Move to check the next group
+                if (++ current_group_idx == kNumGroups)
+                    return false;
+
+                // NOTES: `num_m_blocks` varies with the increase of the group index
+                last_psum_m = math::align(current_psum_m, BLOCK_M);
+                current_psum_m = grouped_layout[current_group_idx];
+                current_m_block_cumsum += num_m_blocks;
+                num_m_blocks = math::ceil_div(current_psum_m - last_psum_m, BLOCK_M);
+            }
+
+            get_swizzled_block_idx(next_block_idx - current_m_block_cumsum * num_n_blocks, m_block_idx, n_block_idx);
+
+            // NOTES: `last_psum_m` is aligned with block M
+            m_block_idx += last_psum_m / BLOCK_M;
+        } else if constexpr (kGemmType == GemmType::KGroupedContiguous) {
+            while (true) {
+                // End of the task
+                if (current_group_idx == kNumGroups)
+                    return false;
+
+                // Within current group
+                if (next_block_idx < (current_num_valid_groups + 1) * num_blocks)
+                    break;
+
+                // Move to check the next group
+                current_k_cumsum += current_shape_k;
+                current_sf_k_cumsum += math::ceil_div(current_shape_k, SF_K_ALIGNMENT);
+                current_num_valid_groups ++;
+
+                current_group_idx = next_group_idx ++;
+                current_shape_k = next_shape_k;
+                get_next_k_group(next_group_idx, next_shape_k);
+            }
+
+            get_swizzled_block_idx(next_block_idx - current_num_valid_groups * num_blocks, m_block_idx, n_block_idx);
+        } else if constexpr (kGemmType == GemmType::Batched) {
+            if (next_block_idx >= num_blocks * kNumGroups)
+                return false;
+
+            current_group_idx = next_block_idx / num_blocks;
+            const auto block_idx = next_block_idx - current_group_idx * num_blocks;
+            if constexpr (kIsMulticastOnA) {
+                m_block_idx = block_idx / num_n_blocks;
+                n_block_idx = block_idx % num_n_blocks;
+            } else {
+                m_block_idx = block_idx % num_m_blocks;
+                n_block_idx = block_idx / num_m_blocks;
+            }
+        } else {
+            if (next_block_idx >= num_blocks)
+                return false;
+
+            // For SM90 only
+            // NOTES: we don't have to set `is_peer_cta_alive` for masked grouped GEMM, as it must be aligned
+            is_peer_cta_alive = num_n_blocks % kNumMulticast == 0 or                  // Always aligned on N (constant bypass)
+                                num_m_blocks % kNumMulticast == 0 or                  // Always aligned on M (constant bypass)
+                                (next_block_idx ^ 1) < num_blocks;                    // Peer CTA in bound
+            get_swizzled_block_idx(next_block_idx, m_block_idx, n_block_idx);
+        }
+        return true;
+    }
+
+    // For SM90 only
+    CUTLASS_DEVICE bool is_tma_multicast_valid(const uint32_t& m_block_idx) const {
+        if (num_blocks_in_group == 1)
+            return false;
+        if constexpr (kGemmType == GemmType::Normal or kGemmType == GemmType::MGroupedMasked or
+                      kGemmType == GemmType::KGroupedContiguous or kGemmType == GemmType::Batched or
+                      kGemmType == GemmType::MGroupedContiguousWithPsumLayout) {
+            return true;
+        } else {
+            DG_STATIC_ASSERT(kGemmType == GemmType::MGroupedContiguous, "Invalid Gemm type");
+            if constexpr (kIsMulticastOnA) {
+                return true;
+            } else {
+                const auto group_idx = grouped_layout[m_block_idx * BLOCK_M];
+                const auto peer_group_idx = grouped_layout[(m_block_idx ^ 1) * BLOCK_M];
+                return group_idx == peer_group_idx;
+            }
+        }
+    }
+
+    // For SM90 only
+    // ReSharper disable once CppNotAllPathsReturnValue
+    CUTLASS_DEVICE bool is_computation_valid(const uint32_t& m_block_idx, const uint32_t& m_offset) const {
+        if constexpr (kGemmType == GemmType::Normal or kGemmType == GemmType::Batched) {
+            return true;
+        } else if constexpr (kGemmType == GemmType::MGroupedContiguous) {
+            return grouped_layout[m_offset + m_block_idx * BLOCK_M] >= 0;
+        } else if constexpr (kGemmType == GemmType::MGroupedMasked) {
+            return m_offset + m_block_idx * BLOCK_M < grouped_layout[current_group_idx];
+        } else if constexpr (kGemmType == GemmType::MGroupedContiguousWithPsumLayout) {
+            return m_offset + m_block_idx * BLOCK_M < current_psum_m;
+        } else {
+            // Unreachable 
+            DG_TRAP_ONLY_DEVICE_ASSERT(false);
+        }
+    }
+};
+
+#pragma clang diagnostic pop
+
+} // namespace deep_gemm::sched

deep_gemm/include/deep_gemm/scheduler/mega_moe.cuh

@@ -0,0 +1,221 @@
+#pragma once
+
+#include <deep_gemm/common/cute_tie.cuh>
+#include <deep_gemm/common/math.cuh>
+#include <deep_gemm/common/types.cuh>
+#include <deep_gemm/layout/mega_moe.cuh>
+#include <deep_gemm/ptx/ld_st.cuh>
+#include <deep_gemm/ptx/utils.cuh>
+
+namespace deep_gemm::sched {
+
+// Computation phase for the current block
+enum class BlockPhase {
+    None = 0,
+    Linear1 = 1,
+    Linear2 = 2
+};
+
+template <uint32_t BLOCK_M, uint32_t BLOCK_N, uint32_t BLOCK_K,
+          uint32_t L1_SHAPE_N, uint32_t L1_SHAPE_K,
+          uint32_t L2_SHAPE_N, uint32_t L2_SHAPE_K,
+          uint32_t kNumExpertsPerRank,
+          uint32_t kNumExpertsPerWave,
+          uint32_t kNumSMs, uint32_t kNumRanks,
+          uint32_t kNumExpertsPerLane = math::constexpr_ceil_div(kNumExpertsPerRank, 32u),
+          uint32_t kNumL1BlockNs = L1_SHAPE_N / BLOCK_N,
+          uint32_t kNumL2BlockNs = L2_SHAPE_N / BLOCK_N,
+          uint32_t kNumL1BlockKs = L1_SHAPE_K / BLOCK_K,
+          uint32_t kNumL2BlockKs = L2_SHAPE_K / BLOCK_K>
+struct MegaMoEScheduler {
+    DG_STATIC_ASSERT(L1_SHAPE_N % BLOCK_N == 0, "Invalid shape");
+    DG_STATIC_ASSERT(L2_SHAPE_N % BLOCK_N == 0, "Invalid shape");
+    DG_STATIC_ASSERT(L1_SHAPE_K % BLOCK_K == 0, "Invalid shape");
+    DG_STATIC_ASSERT(L2_SHAPE_K % BLOCK_K == 0, "Invalid shape");
+    DG_STATIC_ASSERT(kNumExpertsPerRank % kNumExpertsPerWave == 0, "Invalid wave config");
+
+    // NOTES: N block counts must be even so that 2 adjacent CTAs in a cluster
+    // always land on the same m_block_idx with n_block_idx differing by 1
+    DG_STATIC_ASSERT(kNumSMs % 2 == 0, "Number of SMs must be even for 2-CTA cluster");
+    DG_STATIC_ASSERT(kNumL1BlockNs % 2 == 0, "L1 N block count must be even for 2-CTA cluster");
+    DG_STATIC_ASSERT(kNumL2BlockNs % 2 == 0, "L2 N block count must be even for 2-CTA cluster");
+
+    // Arrival counts
+    const layout::Workspace& workspace;
+
+    // Scheduler state
+    BlockPhase next_phase = BlockPhase::Linear1;
+
+    // Current expert and block indices
+    uint32_t current_local_expert_idx = 0;
+    uint32_t current_num_tokens = 0;
+    uint32_t current_pool_block_offset = 0;
+    uint32_t block_idx = 0;
+    uint32_t m_block_idx = 0;
+    uint32_t n_block_idx = 0;
+
+    // Pre-cached per-expert token counts (filled during `for_each_block` init)
+    // Layout: `stored_num_tokens_per_expert[i]` holds expert (i * 32 + lane_idx)'s count
+    uint32_t stored_num_tokens_per_expert[kNumExpertsPerLane] = {};
+
+    CUTLASS_DEVICE explicit MegaMoEScheduler(const layout::Workspace& workspace): workspace(workspace) {
+        block_idx = blockIdx.x;
+    }
+
+    CUTLASS_DEVICE uint32_t get_wave_expert_end_idx() const {
+        return math::align(current_local_expert_idx + 1, kNumExpertsPerWave);
+    }
+
+    CUTLASS_DEVICE uint32_t get_num_tokens(const uint32_t& expert_idx) const {
+        uint32_t valid_value;
+        #pragma unroll
+        for (uint32_t i = 0; i < kNumExpertsPerLane; ++ i) {
+            valid_value = (expert_idx == i * 32 + ptx::get_lane_idx()) ?
+                stored_num_tokens_per_expert[i] : valid_value;
+        }
+        return ptx::exchange(valid_value, expert_idx % 32);
+    }
+
+    // Get pool block offset for a given expert index from a per-lane token count array
+    CUTLASS_DEVICE uint32_t get_pool_block_offset(const uint32_t& expert_idx) {
+        uint32_t num_blocks = 0;
+        #pragma unroll
+        for (uint32_t i = 0; i < kNumExpertsPerLane; ++ i) {
+            if (i * 32 + ptx::get_lane_idx() < expert_idx)
+                num_blocks += math::ceil_div(stored_num_tokens_per_expert[i], BLOCK_M);
+        }
+        return __reduce_add_sync(0xffffffff, num_blocks);
+    }
+
+    CUTLASS_DEVICE void advance_expert_idx() {
+        current_pool_block_offset += get_current_num_m_blocks();
+        current_local_expert_idx += 1;
+        current_num_tokens = get_num_tokens(current_local_expert_idx);
+    }
+
+    CUTLASS_DEVICE void set_expert_idx(const uint32_t& expert_idx) {
+        current_local_expert_idx = expert_idx;
+        current_num_tokens = get_num_tokens(expert_idx);
+        current_pool_block_offset = get_pool_block_offset(expert_idx);
+    }
+
+    CUTLASS_DEVICE uint32_t get_current_pool_block_offset() const {
+        return current_pool_block_offset;
+    }
+
+    CUTLASS_DEVICE uint32_t get_current_num_m_blocks() const {
+        return math::ceil_div(current_num_tokens, BLOCK_M);
+    }
+
+    template <bool kDoUMMAAligned = false>
+    CUTLASS_DEVICE uint32_t get_valid_m() const {
+        const auto m = cute::min(current_num_tokens - m_block_idx * BLOCK_M, BLOCK_M);
+        return kDoUMMAAligned ? math::align(m, 16u) : m;
+    }
+
+    CUTLASS_DEVICE bool fetch_next_l1_block() {
+        const auto wave_end_expert_idx = get_wave_expert_end_idx();
+        while (current_local_expert_idx < wave_end_expert_idx) {
+            const auto num_m_blocks = get_current_num_m_blocks();
+            m_block_idx = block_idx / kNumL1BlockNs;
+            if (m_block_idx < num_m_blocks)
+                return true;
+
+            // Current expert is fully assigned, move to the next
+            block_idx -= num_m_blocks * kNumL1BlockNs;
+            advance_expert_idx();
+        }
+        return false;
+    }
+
+    CUTLASS_DEVICE bool fetch_next_l2_block() {
+        const auto wave_end_expert_idx = get_wave_expert_end_idx();
+        while (current_local_expert_idx < wave_end_expert_idx) {
+            const auto num_m_blocks = get_current_num_m_blocks();
+            if (block_idx < num_m_blocks * kNumL2BlockNs) {
+                m_block_idx = block_idx / kNumL2BlockNs;
+                return true;
+            }
+
+            // Current expert is fully assigned, move to the next
+            block_idx -= num_m_blocks * kNumL2BlockNs;
+            advance_expert_idx();
+        }
+        return false;
+    }
+
+    // Core state machine: assigns the next block
+    CUTLASS_DEVICE cute::tuple<BlockPhase, uint32_t, uint32_t, uint32_t> get_next_block() {
+        while (true) {
+            if (current_local_expert_idx >= kNumExpertsPerRank)
+                break;
+
+            if (next_phase == BlockPhase::Linear1) {
+                if (fetch_next_l1_block()) {
+                    // Found a new L1 block
+                    n_block_idx = block_idx - m_block_idx * kNumL1BlockNs;
+                    // Jump to next block
+                    block_idx += kNumSMs;
+                    return {BlockPhase::Linear1, current_local_expert_idx, m_block_idx, n_block_idx};
+                } else {
+                    // L1 for the current wave is complete, transition to L2
+                    next_phase = BlockPhase::Linear2;
+                    set_expert_idx(math::align<uint32_t, false>(current_local_expert_idx - 1, kNumExpertsPerWave));
+                }
+            } else {
+                if (fetch_next_l2_block()) {
+                    // Found a new L2 block
+                    n_block_idx = block_idx - m_block_idx * kNumL2BlockNs;
+                    // Jump to next block
+                    block_idx += kNumSMs;
+                    return {BlockPhase::Linear2, current_local_expert_idx, m_block_idx, n_block_idx};
+                } else {
+                    // Move to L1 of the next wave
+                    next_phase = BlockPhase::Linear1;
+                }
+            }
+        }
+
+        // All waves and experts are fully processed
+        return {BlockPhase::None, 0, 0, 0};
+    }
+
+    CUTLASS_DEVICE void fetch_expert_recv_count() {
+        // NOTES: each lane caches experts at indices (i * 32 + lane_idx)
+        #pragma unroll
+        for (uint32_t i = 0; i < kNumExpertsPerLane; ++ i) {
+            const auto expert_idx = i * 32 + ptx::get_lane_idx();
+            uint64_t value = 0;
+            if (expert_idx < kNumExpertsPerRank) {
+                do {
+                    value = ptx::ld_volatile(workspace.get_expert_recv_count_sum_ptr(expert_idx));
+                } while (static_cast<uint32_t>(value >> 32) != kNumSMs * kNumRanks);
+            }
+            stored_num_tokens_per_expert[i] = static_cast<uint32_t>(value);
+        }
+        __syncwarp();
+    }
+
+    template <typename Func>
+    CUTLASS_DEVICE void for_each_block(Func&& func) {
+        // Wait for all expert counters to be finalized
+        fetch_expert_recv_count();
+
+        // Initialize current expert with 0
+        set_expert_idx(0);
+
+        // Iterate over all blocks
+        // TODO: add swizzle within expert waves for better L2 cache utilization
+        while (true) {
+            CUTE_TIE_DECL(get_next_block(), block_phase, current_local_expert_idx, m_block_idx, n_block_idx);
+            if (block_phase == BlockPhase::None)
+                break;
+
+            func(block_phase, current_local_expert_idx,
+                 block_phase == BlockPhase::Linear2 ? kNumL2BlockKs : kNumL1BlockKs,
+                 m_block_idx, n_block_idx);
+        }
+    }
+};
+
+} // namespace deep_gemm::sched

deep_gemm/include/deep_gemm/scheduler/paged_mqa_logits.cuh

@@ -0,0 +1,114 @@
+#pragma once
+
+#include <deep_gemm/common/math.cuh>
+#include <deep_gemm/common/types.cuh>
+#include <deep_gemm/ptx/utils.cuh>
+
+namespace deep_gemm::sched {
+
+template <uint32_t kAlignedBatchSize, uint32_t SPLIT_KV, uint32_t kNumSMs>
+CUTLASS_GLOBAL __launch_bounds__(32, 1)
+void smxx_paged_mqa_logits_metadata(const uint32_t batch_size, const uint32_t next_n, const bool is_context_lens_2d,
+                                    const uint32_t* context_lens, uint32_t* schedule_metadata) {
+    DG_STATIC_ASSERT(kAlignedBatchSize % 32 == 0, "Invalid aligned batch size");
+    const uint32_t lane_idx = ptx::get_lane_idx();
+
+    // Wait for primary kernel completion
+    cudaGridDependencySynchronize();
+
+    uint32_t num_segs[kAlignedBatchSize / 32];
+    #pragma unroll
+    for (uint32_t k = 0; k < kAlignedBatchSize / 32; ++ k) {
+        const uint32_t q_idx = k * 32 + lane_idx;
+        const uint32_t lens_idx = (is_context_lens_2d ? q_idx * next_n + next_n - 1 : q_idx);
+        const uint32_t context_len = (q_idx < batch_size ? context_lens[lens_idx] : 0);
+        num_segs[k] = math::ceil_div(context_len, SPLIT_KV);
+    }
+
+    __shared__ uint32_t prefix_sum[kAlignedBatchSize];
+    uint32_t sum = 0;
+    #pragma unroll
+    for (uint32_t k = 0; k < kAlignedBatchSize / 32; ++ k) {
+        uint32_t x = num_segs[k];
+        #pragma unroll
+        for (uint32_t offset = 1; offset < 32; offset <<= 1) {
+            const uint32_t y = __shfl_up_sync(0xffffffff, x, offset);
+            x += (lane_idx >= offset ? y : 0);
+        }
+        x += sum;
+        prefix_sum[k * 32 + lane_idx] = x;
+        sum = __shfl_sync(0xffffffff, x, 31);
+    }
+
+    const uint32_t num_next_n_atoms = next_n / ((next_n % 2 == 0) ? 2 : 1);
+    const uint32_t total = sum * num_next_n_atoms;
+    const uint32_t q = total / kNumSMs, r = total % kNumSMs;
+    for (uint32_t sm_idx = lane_idx; sm_idx <= kNumSMs; sm_idx += 32) {
+        uint32_t seg_starts = sm_idx * q + min(sm_idx, r);
+        uint32_t q_idx = 0;
+        while (q_idx < batch_size and prefix_sum[q_idx] * num_next_n_atoms <= seg_starts)
+            ++ q_idx;
+        const uint32_t offset_in_q = (q_idx == 0 ? seg_starts : seg_starts - prefix_sum[q_idx - 1] * num_next_n_atoms);
+        const uint32_t num_segs_q = (q_idx == 0 ? prefix_sum[0] : prefix_sum[q_idx] - prefix_sum[q_idx - 1]);
+        const uint32_t atom_idx = num_segs_q > 0 ? offset_in_q / num_segs_q : 0;
+        const uint32_t kv_split_idx = num_segs_q > 0 ? offset_in_q % num_segs_q : 0;
+        const uint32_t q_atom_idx = q_idx * num_next_n_atoms + atom_idx;
+        __syncwarp();
+
+        schedule_metadata[sm_idx * 2] = q_atom_idx;
+        schedule_metadata[sm_idx * 2 + 1] = kv_split_idx;
+    }
+}
+
+template <uint32_t kNextN, bool kIsContextLens2D,
+          uint32_t BLOCK_KV, uint32_t kNumBlocksPerSplit,
+          uint32_t kNumNextNAtoms>
+struct PagedMQALogitsScheduler {
+    const uint32_t* context_lens;
+
+    uint32_t current_q_atom_idx, current_kv_idx;
+    uint32_t end_q_atom_idx, end_kv_idx;
+    uint32_t current_num_kv;
+
+    CUTLASS_DEVICE uint32_t get_num_kv(const uint32_t& q_atom_idx) const {
+        const uint32_t q_idx = q_atom_idx / kNumNextNAtoms;
+        const auto lens_idx = (kIsContextLens2D ? q_idx * kNextN + kNextN - 1 : q_idx);
+        return math::ceil_div(context_lens[lens_idx], BLOCK_KV);
+    }
+
+    CUTLASS_DEVICE explicit PagedMQALogitsScheduler(const uint32_t& sm_idx, const uint32_t* context_lens, const uint32_t* schedule_meta) {
+        this->context_lens = context_lens;
+
+        const auto current_pack = reinterpret_cast<const uint2*>(schedule_meta)[sm_idx];
+        const auto end_pack = reinterpret_cast<const uint2*>(schedule_meta)[sm_idx + 1];
+        current_q_atom_idx = current_pack.x, current_kv_idx = current_pack.y * kNumBlocksPerSplit;
+        end_q_atom_idx = end_pack.x, end_kv_idx = end_pack.y * kNumBlocksPerSplit;
+
+        current_num_kv = get_num_kv(current_q_atom_idx);
+    }
+
+    CUTLASS_DEVICE bool fetch_next_task(uint32_t &q_atom_idx, uint32_t &kv_idx, uint32_t &num_kv) {
+        q_atom_idx = current_q_atom_idx;
+        kv_idx = current_kv_idx;
+        num_kv = current_num_kv;
+
+        if (current_q_atom_idx == end_q_atom_idx and current_kv_idx == end_kv_idx)
+            return false;
+
+        current_kv_idx += kNumBlocksPerSplit;
+        if (current_kv_idx >= current_num_kv) {
+            ++ current_q_atom_idx;
+            current_kv_idx = 0;
+            if (current_q_atom_idx % kNumNextNAtoms == 0 and exist_q_atom_idx(current_q_atom_idx)) {
+                current_num_kv = get_num_kv(current_q_atom_idx);
+            }
+        }
+        return true;
+    }
+
+    CUTLASS_DEVICE bool exist_q_atom_idx(const uint32_t& q_atom_idx) const {
+        return q_atom_idx < end_q_atom_idx or (q_atom_idx == end_q_atom_idx and 0 < end_kv_idx);
+    }
+};
+
+} // namespace deep_gemm::sched