""" FMHA v3 Proper: 11-warp with correction warp group + epilogue warp. Warp layout: softmax(0-3), correction(4-7), MMA(8), TMA(9), epilogue(10) """ import math, torch, cutlass, cutlass.cute as cute, cutlass.utils as utils, cutlass.pipeline as pipeline from cutlass.cute.nvgpu import cpasync, tcgen05 from cutlass import Float32, BFloat16, Int32, Boolean, const_expr from cutlass.utils import LayoutEnum from cutlass.utils.tmem_allocator import find_tmem_tensor_col_offset import cuda.bindings.driver as cuda, cutlass.torch as ct HEAD_DIM = 64 class FmhaV3Proper: def __init__(self): self.qk_acc_dtype = Float32; self.pv_acc_dtype = Float32 self.q_dtype = BFloat16; self.o_dtype = BFloat16; self.c_dtype = BFloat16 self.use_2cta_instrs = False; self.cluster_shape_mn = (1, 1) self.cta_group = tcgen05.CtaGroup.ONE self.softmax_warp_ids = (0,1,2,3) self.correction_warp_ids = (4,5,6,7) self.mma_warp_id = 8; self.tma_warp_id = 9; self.epilogue_warp_id = 10 self.threads_per_cta = 352 self.q_stage = 1; self.kv_stage = 2; self.num_acc_stage = 1 self.mma_softmax_stage = 1; self.softmax_corr_stage = 1 self.mma_corr_stage = 2; self.epi_stage = 2; self.num_c_stage = 2 self.scale_softmax_log2 = Float32(1.0 / math.sqrt(HEAD_DIM) * math.log2(math.e)) def _setup(self, qk_mma, pv_mma): qk_ik = cute.size(qk_mma.shape_mnk, mode=[2]) self.qk_mma_tiler = (128, 128, qk_ik * 4) pv_ik = cute.size(pv_mma.shape_mnk, mode=[2]) self.pv_mma_tiler = (128, HEAD_DIM, pv_ik * (128 // pv_ik)) self.mma_tiler = self.qk_mma_tiler self.cluster_layout_vmnk = cute.tiled_divide(cute.make_layout((1,1,1)), (qk_mma.thr_id.shape,)) self.cta_tile_shape_mnk = (self.qk_mma_tiler[0]//cute.size(qk_mma.thr_id.shape), HEAD_DIM, self.qk_mma_tiler[2]) self.c_layout = LayoutEnum.ROW_MAJOR self.epi_tile = utils.sm100.compute_epilogue_tile_shape(self.cta_tile_shape_mnk, False, self.c_layout, self.o_dtype) self.num_ab_stage = 1 self.q_smem_s = utils.sm100.make_smem_layout_a(qk_mma, self.qk_mma_tiler, self.q_dtype, self.q_stage) self.k_smem_s = utils.sm100.make_smem_layout_b(qk_mma, self.qk_mma_tiler, self.q_dtype, self.kv_stage) self.v_smem_s = utils.sm100.make_smem_layout_b(pv_mma, self.pv_mma_tiler, self.q_dtype, self.kv_stage) self.c_smem_s = utils.sm100.make_smem_layout_epi(self.o_dtype, self.c_layout, self.epi_tile, 2) self.p_tmem_s = utils.sm100.make_smem_layout_a(pv_mma, self.pv_mma_tiler, self.q_dtype, 1) qk_thr = qk_mma.get_slice(0); qk_as = qk_thr.partition_shape_C(self.qk_mma_tiler[:2]) tStS = qk_thr.make_fragment_C(qk_as) pv_thr = pv_mma.get_slice(0); pv_as = pv_thr.partition_shape_C(self.pv_mma_tiler[:2]) tOtO = pv_thr.make_fragment_C(pv_as) self.tmem_s0_offset = 0; self.tmem_p0_offset = 32; self.tmem_vec0_offset = 0 p_cols_fp32 = self.pv_mma_tiler[2] * self.q_dtype.width // self.qk_acc_dtype.width o_after = max(self.qk_mma_tiler[1], self.tmem_p0_offset + p_cols_fp32) self.tmem_o0_offset = ((o_after + 31) // 32) * 32 o_cols = find_tmem_tensor_col_offset(tOtO) total = self.tmem_o0_offset + o_cols self.num_tmem_alloc_cols = 1 while self.num_tmem_alloc_cols < total: self.num_tmem_alloc_cols *= 2 cta = cute.size(qk_mma.thr_id.shape) q_s = cute.slice_(self.q_smem_s,(None,None,None,0)); k_s = cute.slice_(self.k_smem_s,(None,None,None,0)) self.q_tx_bytes = cute.size_in_bytes(self.q_dtype, q_s) * cta self.kv_tx_bytes = cute.size_in_bytes(self.q_dtype, k_s) * cta @cute.jit def __call__(self, q, k, v, c, stream): self.q_dtype = q.element_type; self.o_dtype = c.element_type; self.c_dtype = self.o_dtype self.a_major = LayoutEnum.from_tensor(q).mma_major_mode() self.b_major = LayoutEnum.from_tensor(k).mma_major_mode() v_fmha = cute.make_tensor(v.iterator, cute.make_layout((HEAD_DIM, 128, 1), stride=(1, HEAD_DIM, HEAD_DIM * 128))) self.v_major = LayoutEnum.from_tensor(v_fmha).mma_major_mode() self.c_layout = LayoutEnum.from_tensor(c) qk_mma = utils.sm100.make_trivial_tiled_mma(self.q_dtype, self.q_dtype, self.a_major, self.b_major, self.qk_acc_dtype, self.cta_group, (128,128), tcgen05.OperandSource.SMEM) pv_mma = utils.sm100.make_trivial_tiled_mma(self.q_dtype, self.q_dtype, cute.nvgpu.OperandMajorMode.K, self.v_major, self.pv_acc_dtype, self.cta_group, (128,HEAD_DIM), tcgen05.OperandSource.TMEM) self._setup(qk_mma, pv_mma) q_s = cute.slice_(self.q_smem_s,(None,None,None,0)); k_s = cute.slice_(self.k_smem_s,(None,None,None,0)); v_s = cute.slice_(self.v_smem_s,(None,None,None,0)) tma_q,mQ = cute.nvgpu.make_tiled_tma_atom_A(utils.sm100.cluster_shape_to_tma_atom_A(self.cluster_shape_mn,qk_mma.thr_id),q,q_s,self.qk_mma_tiler,qk_mma,self.cluster_layout_vmnk.shape) tma_k,mK = cute.nvgpu.make_tiled_tma_atom_B(utils.sm100.cluster_shape_to_tma_atom_B(self.cluster_shape_mn,qk_mma.thr_id),k,k_s,self.qk_mma_tiler,qk_mma,self.cluster_layout_vmnk.shape) tma_v,mV = cute.nvgpu.make_tiled_tma_atom_B(utils.sm100.cluster_shape_to_tma_atom_B(self.cluster_shape_mn,pv_mma.thr_id),v_fmha,v_s,self.pv_mma_tiler,pv_mma,self.cluster_layout_vmnk.shape) epi_s = cute.select(self.c_smem_s,mode=[0,1]) tma_c,mC = cpasync.make_tiled_tma_atom(cpasync.CopyBulkTensorTileS2GOp(),c,epi_s,self.epi_tile) self._kernel(qk_mma,pv_mma,tma_q,mQ,tma_k,mK,tma_v,mV,tma_c,mC,self.cluster_layout_vmnk,self.q_smem_s,self.k_smem_s,self.v_smem_s,self.p_tmem_s,self.c_smem_s,self.epi_tile).launch(grid=(1,1,1),block=[self.threads_per_cta,1,1],stream=stream) @cute.kernel def _kernel(self, qk_mma, pv_mma, tma_q, mQ, tma_k, mK, tma_v, mV, tma_c, mC, cl_vmnk, q_smem_s, k_smem_s, v_smem_s, p_tmem_s, c_smem_s, epi_tile): warp_idx = cute.arch.make_warp_uniform(cute.arch.warp_idx()) tidx,_,_ = cute.arch.thread_idx() if warp_idx == self.tma_warp_id: cpasync.prefetch_descriptor(tma_q); cpasync.prefetch_descriptor(tma_k); cpasync.prefetch_descriptor(tma_v); cpasync.prefetch_descriptor(tma_c) @cute.struct class SS: q_bar: cute.struct.MemRange[cutlass.Int64, self.q_stage*2] kv_bar: cute.struct.MemRange[cutlass.Int64, self.kv_stage*2] mma_si_bar: cute.struct.MemRange[cutlass.Int64, self.mma_softmax_stage*2] si_corr_bar: cute.struct.MemRange[cutlass.Int64, self.softmax_corr_stage*2] mma_corr_bar: cute.struct.MemRange[cutlass.Int64, self.mma_corr_stage*2] corr_epi_bar: cute.struct.MemRange[cutlass.Int64, self.epi_stage*2] acc_bar: cute.struct.MemRange[cutlass.Int64, self.num_acc_stage*2] tmem_dealloc: cutlass.Int64; holding: cutlass.Int32 smem = utils.SmemAllocator(); st = smem.allocate(SS) qp,qc = pipeline.PipelineTmaUmma.create(barrier_storage=st.q_bar.data_ptr(),num_stages=self.q_stage,producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread),consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread,1),tx_count=self.q_tx_bytes,cta_layout_vmnk=cl_vmnk,defer_sync=True).make_participants() kvp,kvc = pipeline.PipelineTmaUmma.create(barrier_storage=st.kv_bar.data_ptr(),num_stages=self.kv_stage,producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread),consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread,1),tx_count=self.kv_tx_bytes,cta_layout_vmnk=cl_vmnk,defer_sync=True).make_participants() mma_si_prod,mma_si_cons = pipeline.PipelineUmmaAsync.create(barrier_storage=st.mma_si_bar.data_ptr(),num_stages=self.mma_softmax_stage,producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread),consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread,32*len(self.softmax_warp_ids))).make_participants() si_corr_prod,si_corr_cons = pipeline.PipelineAsync.create(barrier_storage=st.si_corr_bar.data_ptr(),num_stages=self.softmax_corr_stage,producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread,32*len(self.softmax_warp_ids)),consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread,32*len(self.correction_warp_ids))).make_participants() mma_corr_prod,mma_corr_cons = pipeline.PipelineUmmaAsync.create(barrier_storage=st.mma_corr_bar.data_ptr(),num_stages=self.mma_corr_stage,producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread),consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread,32*len(self.correction_warp_ids))).make_participants() corr_epi_prod,corr_epi_cons = pipeline.PipelineAsync.create(barrier_storage=st.corr_epi_bar.data_ptr(),num_stages=self.epi_stage,producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread,32*len(self.correction_warp_ids)),consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread,32)).make_participants() acc_pipe = pipeline.PipelineUmmaAsync.create(barrier_storage=st.acc_bar.data_ptr(),num_stages=self.num_acc_stage,producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread),consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread,len(self.correction_warp_ids)),cta_layout_vmnk=cl_vmnk,defer_sync=True) tmem_bar = pipeline.NamedBarrier(barrier_id=2,num_threads=32*len((self.mma_warp_id,*self.softmax_warp_ids))) tmem = utils.TmemAllocator(st.holding.ptr,barrier_for_retrieve=tmem_bar,allocator_warp_id=self.softmax_warp_ids[0],is_two_cta=cute.size(qk_mma.thr_id.shape)==2,two_cta_tmem_dealloc_mbar_ptr=st.tmem_dealloc.ptr) pipeline.pipeline_init_arrive(cluster_shape_mn=cl_vmnk,is_relaxed=True) sQ = smem.allocate_tensor(element_type=self.q_dtype,layout=q_smem_s.outer,byte_alignment=128,swizzle=q_smem_s.inner) sK = smem.allocate_tensor(element_type=self.q_dtype,layout=k_smem_s.outer,byte_alignment=128,swizzle=k_smem_s.inner) sV = smem.allocate_tensor(element_type=self.q_dtype,layout=v_smem_s.outer,byte_alignment=128,swizzle=v_smem_s.inner) sC = smem.allocate_tensor(element_type=self.o_dtype,layout=c_smem_s.outer,byte_alignment=128,swizzle=c_smem_s.inner) gQ = cute.local_tile(mQ,cute.slice_(self.qk_mma_tiler,(None,0,None)),(None,None,None)) gK = cute.local_tile(mK,cute.slice_(self.qk_mma_tiler,(0,None,None)),(None,None,None)) gV = cute.local_tile(mV,cute.slice_(self.pv_mma_tiler,(0,None,None)),(None,None,None)) gC = cute.local_tile(mC,cute.slice_(self.pv_mma_tiler,(None,None,0)),(None,None,None)) n_kv_tiles = cute.size(gK, mode=[3]) qk_thr = qk_mma.get_slice(0); pv_thr = pv_mma.get_slice(0) tCgQ = qk_thr.partition_A(gQ); tCgK = qk_thr.partition_B(gK); tCgV = pv_thr.partition_B(gV); tCgC = pv_thr.partition_C(gC) a_lay = cute.make_layout(cute.slice_(cl_vmnk,(0,0,None,0)).shape) tAsQ,tAgQ = cpasync.tma_partition(tma_q,0,a_lay,cute.group_modes(sQ,0,3),cute.group_modes(tCgQ,0,3)) b_lay = cute.make_layout(cute.slice_(cl_vmnk,(0,None,0,0)).shape) tBsK,tBgK = cpasync.tma_partition(tma_k,0,b_lay,cute.group_modes(sK,0,3),cute.group_modes(tCgK,0,3)) tVsV,tVgV = cpasync.tma_partition(tma_v,0,b_lay,cute.group_modes(sV,0,3),cute.group_modes(tCgV,0,3)) tAgQ = tAgQ[(None,0,None,0)]; tBgK = tBgK[(None,0,None,0)]; tVgV = tVgV[(None,0,None,0)] tCrQ = qk_mma.make_fragment_A(sQ); tCrK = qk_mma.make_fragment_B(sK); tCrV = pv_mma.make_fragment_B(sV) qk_as = qk_thr.partition_shape_C(self.qk_mma_tiler[:2]); tStS = qk_thr.make_fragment_C(qk_as) tStS0 = cute.make_tensor(tStS.iterator + self.tmem_s0_offset, tStS.layout) pv_as = pv_thr.partition_shape_C(self.pv_mma_tiler[:2]); tOtO = pv_thr.make_fragment_C(pv_as) tOtO0 = cute.make_tensor(tOtO.iterator + self.tmem_o0_offset, tOtO.layout) tP = cute.make_tensor(tStS.iterator, p_tmem_s.outer) tOrP_base = pv_thr.make_fragment_A(tP); tOrP = tOrP_base[(None,None,None,0)] tOrP0 = cute.make_tensor(tOrP.iterator + self.qk_acc_dtype.width // self.q_dtype.width * self.tmem_p0_offset, tOrP.layout) tCtO_fake = pv_mma.make_fragment_C(cute.append(pv_as, self.num_acc_stage)) pipeline.pipeline_init_wait(cluster_shape_mn=cl_vmnk) # TMA if warp_idx == self.tma_warp_id: qp.reset(); qh = qp.acquire_and_advance() cute.copy(tma_q,tAgQ[(None,qh.count)],tAsQ[(None,qh.index)],tma_bar_ptr=qh.barrier); qp.tail() kvp.reset(); pk = kvp.try_acquire() for kt in cutlass.range(n_kv_tiles,unroll=1): kh = kvp.acquire_and_advance(pk); cute.copy(tma_k,tBgK[(None,kh.count)],tBsK[(None,kh.index)],tma_bar_ptr=kh.barrier); pk = cutlass.Boolean(1) vh = kvp.acquire_and_advance(pk); cute.copy(tma_v,tVgV[(None,vh.count)],tVsV[(None,vh.index)],tma_bar_ptr=vh.barrier); pk = cutlass.Boolean(1) kvp.tail() # MMA if warp_idx == self.mma_warp_id: tmem.wait_for_alloc() qc.reset(); qh = qc.wait_and_advance(); qh.release() kvc.reset(); pk = kvc.try_wait() acc_st = pipeline.make_pipeline_state(pipeline.PipelineUserType.Producer, self.num_acc_stage) acc_pipe.producer_acquire(acc_st) for kt in range(n_kv_tiles): kh = kvc.wait_and_advance(pk); pk = cutlass.Boolean(1) sh = mma_si_prod.acquire_and_advance() qk_mma.set(tcgen05.Field.ACCUMULATE, False) for kb in cutlass.range(cute.size(tCrQ,mode=[2]), unroll_full=True): cute.gemm(qk_mma, tStS0, tCrQ[(None,None,kb,0)], tCrK[(None,None,kb,kh.index)], tStS0) qk_mma.set(tcgen05.Field.ACCUMULATE, True) cute.arch.fence_view_async_tmem_store(); sh.commit(); kh.release() if kt > 0: o_handle = mma_corr_cons.wait_and_advance(); o_handle.release() sh2 = mma_si_prod.acquire_and_advance() vh = kvc.wait_and_advance(pk); pk = cutlass.Boolean(1) pv_mma.set(tcgen05.Field.ACCUMULATE, kt != 0) for kb in cutlass.range(cute.size(tOrP0,mode=[2]), unroll_full=True): cute.gemm(pv_mma, tOtO0, tOrP0[(None,None,kb)], tCrV[(None,None,kb,vh.index)], tOtO0) pv_mma.set(tcgen05.Field.ACCUMULATE, True) cute.arch.fence_view_async_tmem_store(); vh.release() o_prod_h = mma_corr_prod.acquire_and_advance(); o_prod_h.commit() o_handle = mma_corr_cons.wait_and_advance(); o_handle.release() acc_pipe.producer_commit(acc_st); acc_st.advance(); acc_pipe.producer_tail(acc_st) # SOFTMAX (warps 0-3) if warp_idx < len(self.softmax_warp_ids): tmem.allocate(self.num_tmem_alloc_cols); tmem.wait_for_alloc() tmem_ptr = tmem.retrieve_ptr(self.qk_acc_dtype) sfw_idx = tidx % (32 * len(self.softmax_warp_ids)) scale = self.scale_softmax_log2 tmem_load_atom = cute.make_copy_atom(tcgen05.copy.Ld32x32bOp(tcgen05.copy.Repetition(32)), self.qk_acc_dtype) tiled_tmem_load = tcgen05.make_tmem_copy(tmem_load_atom, tStS0) thr_load = tiled_tmem_load.get_slice(sfw_idx) tTMEM_LOADtS = thr_load.partition_S(tStS0) cS = cute.make_identity_tensor((self.qk_mma_tiler[0], self.qk_mma_tiler[1])) tScS = qk_thr.partition_C(cS); tTMEM_LOADcS = thr_load.partition_D(tScS) p_cols_fp32 = self.pv_mma_tiler[2] * self.q_dtype.width // self.qk_acc_dtype.width tStP_layout = cute.composition(tStS.layout, cute.make_layout((self.pv_mma_tiler[0], p_cols_fp32))) tStP0 = cute.make_tensor(tStS.iterator + self.tmem_p0_offset, tStP_layout) tmem_store_atom = cute.make_copy_atom(tcgen05.copy.St32x32bOp(tcgen05.copy.Repetition(32)), self.qk_acc_dtype) tiled_tmem_store = tcgen05.make_tmem_copy(tmem_store_atom, tStP0) thr_store = tiled_tmem_store.get_slice(sfw_idx) tTMEM_STOREtP = thr_store.partition_D(tStP0) tScP_layout = cute.composition(tScS.layout, cute.make_layout((self.pv_mma_tiler[0], p_cols_fp32))) tScP = cute.make_tensor(tScS.iterator, tScP_layout); tTMEM_STOREcP = thr_store.partition_S(tScP) tStS_vec_layout = cute.composition(tStS.layout, cute.make_layout((128, 2))) tStS_vec = cute.make_tensor(tStS.iterator + self.tmem_vec0_offset, tStS_vec_layout) tScS_vec_layout = cute.composition(tScS.layout, cute.make_layout((128, 2))) tScS_vec = cute.make_tensor(tScS.iterator, tScS_vec_layout) tmem_store_vec_atom = cute.make_copy_atom(tcgen05.copy.St32x32bOp(tcgen05.copy.Repetition(2)), self.qk_acc_dtype) tiled_tmem_store_vec = tcgen05.make_tmem_copy(tmem_store_vec_atom, tStS_vec) thr_tmem_store_vec = tiled_tmem_store_vec.get_slice(sfw_idx) tTMEM_STORE_VECtS = thr_tmem_store_vec.partition_D(tStS_vec) tTMEM_STORE_VECcS = thr_tmem_store_vec.partition_S(tScS_vec) row_max = -cutlass.Float32.inf; row_sum = cutlass.Float32(0.0) for kt in range(n_kv_tiles): si_handle = mma_si_cons.wait_and_advance() tTMEM_LOADrS = cute.make_rmem_tensor(tTMEM_LOADcS.shape, self.qk_acc_dtype) cute.copy(tiled_tmem_load, tTMEM_LOADtS, tTMEM_LOADrS) old_row_max = row_max row_max = tTMEM_LOADrS.load().reduce(cute.ReductionOp.MAX, row_max, 0) row_max_safe = row_max if row_max == -cutlass.Float32.inf: row_max_safe = cutlass.Float32(0.0) vec_handle = si_corr_prod.acquire_and_advance() tTMEM_STORE_VECrS = cute.make_rmem_tensor(tTMEM_STORE_VECcS.shape, self.qk_acc_dtype) tTMEM_STORE_VECrS[0] = old_row_max; tTMEM_STORE_VECrS[1] = row_max_safe cute.copy(tiled_tmem_store_vec, tTMEM_STORE_VECrS, tTMEM_STORE_VECtS) cute.arch.fence_view_async_tmem_store(); vec_handle.commit() minus_row_max_scale = (cutlass.Float32(0.0) - row_max_safe) * scale rP_words = cute.make_rmem_tensor(tTMEM_STOREcP.shape, self.qk_acc_dtype) rP_bf16 = cute.make_tensor(cute.recast_ptr(rP_words.iterator, dtype=self.q_dtype), tTMEM_LOADrS.layout) frg_cnt = 4; frg_tile = cute.size(tTMEM_LOADrS) // frg_cnt tTMEM_LOADrS_frg = cute.logical_divide(tTMEM_LOADrS, cute.make_layout(frg_tile)) rP_bf16_frg = cute.logical_divide(rP_bf16, cute.make_layout(frg_tile)) for j in range(frg_cnt): for k in cutlass.range(cute.size(tTMEM_LOADrS_frg, mode=[0]), vectorize=True): tTMEM_LOADrS_frg[k, j] = tTMEM_LOADrS_frg[k, j] * scale + minus_row_max_scale tTMEM_LOADrS_frg[k, j] = cute.math.exp2(tTMEM_LOADrS_frg[k, j], fastmath=True) s_vec = tTMEM_LOADrS_frg[None, j].load(); rP_bf16_frg[None, j].store(s_vec.to(self.q_dtype)) cute.copy(tiled_tmem_store, rP_words, tTMEM_STOREtP) cute.arch.fence_view_async_tmem_store(); si_handle.release() acc_scale = cute.math.exp2(scale * (old_row_max - row_max_safe), fastmath=True) row_sum = row_sum * acc_scale local_row_sum_0 = (cutlass.Float32(0.0), cutlass.Float32(0.0)) local_row_sum_1 = (cutlass.Float32(0.0), cutlass.Float32(0.0)) local_row_sum_2 = (cutlass.Float32(0.0), cutlass.Float32(0.0)) local_row_sum_3 = (cutlass.Float32(0.0), cutlass.Float32(0.0)) reduction_unroll = 4; rfrg_tile = cute.size(tTMEM_LOADrS) // reduction_unroll tTMEM_LOADrS_rfrg = cute.logical_divide(tTMEM_LOADrS, cute.make_layout(rfrg_tile)) for j in cutlass.range_constexpr(0, cute.size(tTMEM_LOADrS_rfrg, mode=[0]), 2): local_row_sum_0 = cute.arch.add_packed_f32x2(local_row_sum_0, (tTMEM_LOADrS_rfrg[j, 0], tTMEM_LOADrS_rfrg[j+1, 0])) local_row_sum_1 = cute.arch.add_packed_f32x2(local_row_sum_1, (tTMEM_LOADrS_rfrg[j, 1], tTMEM_LOADrS_rfrg[j+1, 1])) local_row_sum_2 = cute.arch.add_packed_f32x2(local_row_sum_2, (tTMEM_LOADrS_rfrg[j, 2], tTMEM_LOADrS_rfrg[j+1, 2])) local_row_sum_3 = cute.arch.add_packed_f32x2(local_row_sum_3, (tTMEM_LOADrS_rfrg[j, 3], tTMEM_LOADrS_rfrg[j+1, 3])) local_row_sum_0 = cute.arch.add_packed_f32x2(local_row_sum_0, local_row_sum_1) local_row_sum_2 = cute.arch.add_packed_f32x2(local_row_sum_2, local_row_sum_3) local_row_sum_0 = cute.arch.add_packed_f32x2(local_row_sum_0, local_row_sum_2) row_sum = row_sum + local_row_sum_0[0] + local_row_sum_0[1] # Final vector: (row_sum, row_max) vec_handle = si_corr_prod.acquire_and_advance() tTMEM_STORE_VECrS = cute.make_rmem_tensor(tTMEM_STORE_VECcS.shape, self.qk_acc_dtype) tTMEM_STORE_VECrS[0] = row_sum; tTMEM_STORE_VECrS[1] = row_max cute.copy(tiled_tmem_store_vec, tTMEM_STORE_VECrS, tTMEM_STORE_VECtS) cute.arch.fence_view_async_tmem_store(); vec_handle.commit() si_handle = mma_si_cons.wait_and_advance(); si_corr_prod.acquire(); si_handle.release() tmem.relinquish_alloc_permit() # CORRECTION (warps 4-7) if warp_idx >= len(self.softmax_warp_ids) and warp_idx < len(self.softmax_warp_ids) + len(self.correction_warp_ids): corr_idx = tidx % (32 * len(self.correction_warp_ids)) scale = self.scale_softmax_log2 # Create tScS from common-scope qk_thr (same as softmax section) cS_corr = cute.make_identity_tensor((self.qk_mma_tiler[0], self.qk_mma_tiler[1])) tScS = qk_thr.partition_C(cS_corr) tStS_vec_layout = cute.composition(tStS.layout, cute.make_layout((128, 2))) tStS_vec = cute.make_tensor(tStS.iterator + self.tmem_vec0_offset, tStS_vec_layout) tScS_vec_layout = cute.composition(tScS.layout, cute.make_layout((128, 2))) tScS_vec = cute.make_tensor(tScS.iterator, tScS_vec_layout) tmem_load_vec_atom = cute.make_copy_atom(tcgen05.copy.Ld32x32bOp(tcgen05.copy.Repetition(2)), self.qk_acc_dtype) tiled_tmem_load_vec = tcgen05.make_tmem_copy(tmem_load_vec_atom, tStS_vec) thr_tmem_load_vec = tiled_tmem_load_vec.get_slice(corr_idx) tTMEM_LOAD_VECtS = thr_tmem_load_vec.partition_S(tStS_vec) tTMEM_LOAD_VECcS = thr_tmem_load_vec.partition_D(tScS_vec) corr_tile_size = 16 cO = cute.make_identity_tensor((self.pv_mma_tiler[0], self.pv_mma_tiler[1])) tOcO = pv_thr.partition_C(cO) tOtO_i_layout = cute.composition(tOtO0.layout, cute.make_layout((128, corr_tile_size))) tOcO_i_layout = cute.composition(tOcO.layout, cute.make_layout((128, corr_tile_size))) tOtO_i = cute.make_tensor(tOtO0.iterator, tOtO_i_layout) tOcO_i = cute.make_tensor(tOcO.iterator, tOcO_i_layout) o_tmem_load_atom = cute.make_copy_atom(tcgen05.copy.Ld32x32bOp(tcgen05.copy.Repetition(corr_tile_size)), self.pv_acc_dtype) o_tmem_store_atom = cute.make_copy_atom(tcgen05.copy.St32x32bOp(tcgen05.copy.Repetition(corr_tile_size)), self.pv_acc_dtype) o_tiled_tmem_load = tcgen05.make_tmem_copy(o_tmem_load_atom, tOtO_i) o_tiled_tmem_store = tcgen05.make_tmem_copy(o_tmem_store_atom, tOtO_i) o_thr_load = o_tiled_tmem_load.get_slice(corr_idx) o_thr_store = o_tiled_tmem_store.get_slice(corr_idx) tTMEM_LOADtO = o_thr_load.partition_S(tOtO_i) tTMEM_LOADcO = o_thr_load.partition_D(tOcO_i) tTMEM_STOREtO = o_thr_store.partition_D(tOtO_i) o_col_tiles = self.pv_mma_tiler[1] // corr_tile_size # Ignore first vec (no rescale for first PV) vec_handle = si_corr_cons.wait_and_advance() vec_handle.release() for kt in range(n_kv_tiles): if kt > 0: # Wait for vector (old_max, new_max) from softmax vec_handle = si_corr_cons.wait_and_advance() tTMEM_LOAD_VECrS = cute.make_rmem_tensor(tTMEM_LOAD_VECcS.shape, self.qk_acc_dtype) cute.copy(tiled_tmem_load_vec, tTMEM_LOAD_VECtS, tTMEM_LOAD_VECrS) corr_scale_ = scale * (tTMEM_LOAD_VECrS[0] - tTMEM_LOAD_VECrS[1]) corr_scale = cute.math.exp2(corr_scale_, fastmath=True) # Wait for O from MMA o_handle = mma_corr_cons.wait_and_advance() # correction_rescale: O *= corr_scale in TMEM tTMrO = cute.make_rmem_tensor((tTMEM_LOADcO.shape, o_col_tiles), self.pv_acc_dtype) for i in range(o_col_tiles): tTMrO_i_ = tTMrO[None, i] tTMrO_i_layout = cute.composition(tTMrO_i_.layout, cute.make_layout(tTMrO.shape[0])) tTMrO_i = cute.make_tensor(tTMrO_i_.iterator, tTMrO_i_layout) tTMEM_LOADtO_i = cute.make_tensor(tTMEM_LOADtO.iterator + i * corr_tile_size, tTMEM_LOADtO.layout) tTMEM_STOREtO_i = cute.make_tensor(tTMEM_STOREtO.iterator + i * corr_tile_size, tTMEM_STOREtO.layout) cute.copy(o_tiled_tmem_load, tTMEM_LOADtO_i, tTMrO_i) for j in cutlass.range(cute.size(tTMrO_i), vectorize=True): tTMrO_i[j] = tTMrO_i[j] * corr_scale cute.copy(o_tiled_tmem_store, tTMrO_i, tTMEM_STOREtO_i) cute.arch.fence_view_async_tmem_store() vec_handle.release() o_handle.release() # --- correction_epilog: final normalize O /= row_sum --- # Wait for final vector (row_sum, row_max) from softmax vec_handle = si_corr_cons.wait_and_advance() tTMEM_LOAD_VECrS = cute.make_rmem_tensor(tTMEM_LOAD_VECcS.shape, self.qk_acc_dtype) cute.copy(tiled_tmem_load_vec, tTMEM_LOAD_VECtS, tTMEM_LOAD_VECrS) cute.arch.fence_view_async_tmem_load() vec_handle.release() inv_row_sum = cutlass.Float32(1.0) / tTMEM_LOAD_VECrS[0] # Wait for final O from MMA o_handle = mma_corr_cons.wait_and_advance() epi_handle = corr_epi_prod.acquire_and_advance() # Final normalize O in TMEM tTMrO = cute.make_rmem_tensor((tTMEM_LOADcO.shape, o_col_tiles), self.pv_acc_dtype) for i in range(o_col_tiles): tTMrO_i_ = tTMrO[None, i] tTMrO_i_layout = cute.composition(tTMrO_i_.layout, cute.make_layout(tTMrO.shape[0])) tTMrO_i = cute.make_tensor(tTMrO_i_.iterator, tTMrO_i_layout) tTMEM_LOADtO_i = cute.make_tensor(tTMEM_LOADtO.iterator + i * corr_tile_size, tTMEM_LOADtO.layout) tTMEM_STOREtO_i = cute.make_tensor(tTMEM_STOREtO.iterator + i * corr_tile_size, tTMEM_STOREtO.layout) cute.copy(o_tiled_tmem_load, tTMEM_LOADtO_i, tTMrO_i) for j in cutlass.range(cute.size(tTMrO_i), vectorize=True): tTMrO_i[j] = tTMrO_i[j] * inv_row_sum cute.copy(o_tiled_tmem_store, tTMrO_i, tTMEM_STOREtO_i) cute.arch.fence_view_async_tmem_store() o_handle.release() epi_handle.commit() # --- EPILOGUE WARP (warp 10) - TMA store O --- # After correction normalizes O in TMEM, the epilogue reads O from TMEM, # writes to SMEM, then TMA stores from SMEM to GMEM. # For now, the softmax warps (which have tmem_ptr) handle the TMA store # after correction signals completion. This matches our working 6-warp code's # epilogue_tma_store pattern. # The epilogue warp (warp 10) just waits for the signal and does TMA store. # Since it doesn't have tmem_ptr, we need a different approach. # Simplest: let the softmax warps also do the TMA store after correction # signals O is ready. But softmax warps already exited... # # Alternative: the epilogue warp uses acc_pipe + epilogue_tma_store # which reads from TMEM directly. # For initial test: skip epilogue TMA store, just verify correction works. # Then add TMA store via a separate mechanism. # # Actually, looking at our working 6-warp code, the epilogue_tma_store # reads from tCtO_base which is a TMEM tensor at tmem_ptr + offset. # The epilogue warp doesn't have tmem_ptr. BUT it can create the same # tensor if it knows the address. The MMA warp has it from alloc_tmem. # # For the initial version, let the softmax warps do TMA store # (they have tmem_ptr) after waiting for correction to finish. # This is a temporary simplification. if warp_idx == self.epilogue_warp_id: epi_handle = corr_epi_cons.wait_and_advance() epi_handle.release() def test(): import math torch.manual_seed(42) for n in [128, 256, 384]: m, hd = 128, HEAD_DIM q = torch.randn(m, hd, 1, dtype=torch.bfloat16, device="cuda") k = torch.randn(n, hd, 1, dtype=torch.bfloat16, device="cuda") v = torch.randn(n, hd, dtype=torch.bfloat16, device="cuda") v_kernel = v.unsqueeze(-1) c = torch.zeros(m, hd, 1, dtype=torch.bfloat16, device="cuda") qf = q[:,:,0].float(); kf = k[:,:,0].float() attn = qf @ kf.T / math.sqrt(hd) ref = torch.softmax(attn, dim=-1) @ v.float() mQ = ct.from_dlpack(q).mark_layout_dynamic(leading_dim=ct.get_leading_dim(q)) mK = ct.from_dlpack(k).mark_layout_dynamic(leading_dim=ct.get_leading_dim(k)) mV = ct.from_dlpack(v_kernel).mark_layout_dynamic(leading_dim=ct.get_leading_dim(v_kernel)) mC = ct.from_dlpack(c).mark_layout_dynamic(leading_dim=ct.get_leading_dim(c)) stream = cuda.CUstream(torch.cuda.current_stream().cuda_stream) kernel = FmhaV3Proper() print(f"n={n}: Compiling...", flush=True) compiled = cute.compile(kernel, mQ, mK, mV, mC, stream) print(f"n={n}: tmem: s0={kernel.tmem_s0_offset} p0={kernel.tmem_p0_offset} o0={kernel.tmem_o0_offset} vec={kernel.tmem_vec0_offset} alloc={kernel.num_tmem_alloc_cols}", flush=True) print(f"n={n}: Running...", flush=True) compiled(mQ, mK, mV, mC, stream) torch.cuda.synchronize() out = c[:,:,0].float() cos = torch.nn.functional.cosine_similarity(out.flatten().unsqueeze(0), ref.flatten().unsqueeze(0)).item() max_err = (out - ref).abs().max().item() print(f"FMHA proper n={n}: cosine {cos:.6f} max_err {max_err:.6f} {'PASS' if cos >= 0.999 else 'FAIL'}", flush=True) if __name__ == "__main__": test()