From fd6b1e82d8a562ad2dd0ef7fd187e2cce5111a02 Mon Sep 17 00:00:00 2001 From: biondizzle Date: Fri, 22 May 2026 21:55:03 +0000 Subject: [PATCH] TMA shape diag: pure Python, no JIT --- tests/unit/test_tma_shapes.py | 231 +++++++++++++--------------------- 1 file changed, 84 insertions(+), 147 deletions(-) diff --git a/tests/unit/test_tma_shapes.py b/tests/unit/test_tma_shapes.py index 89d49ae8..ec3430fb 100644 --- a/tests/unit/test_tma_shapes.py +++ b/tests/unit/test_tma_shapes.py @@ -1,173 +1,110 @@ """ Diagnostic: print tBgK and tVgV shapes BEFORE pre-slicing. -This runs at JIT trace time, so Python print gives us static shape info. +We do this at Python trace time (before JIT), not inside the kernel. """ -import torch, cutlass, cutlass.cute as cute, cutlass.utils as utils, cutlass.pipeline as pipeline +import torch, cutlass, cutlass.cute as cute, cutlass.utils as utils from cutlass.cute.nvgpu import cpasync, tcgen05 -from cutlass import Float32, BFloat16, Int32, const_expr +from cutlass import Float32, BFloat16, Int32 from cutlass.utils import LayoutEnum -from cutlass.utils.tmem_allocator import find_tmem_tensor_col_offset import cuda.bindings.driver as cuda import cutlass.torch as ct import math HEAD_DIM = 64 - -class TmaShapeDiag: - def __init__(self, s_k=256): - self.s_k = s_k - self.n_kv_tiles = s_k // 128 - self.acc_dtype = Float32; self.qk_acc_dtype = Float32 - self.q_dtype = BFloat16; self.o_dtype = BFloat16; self.c_dtype = BFloat16 - self.use_2cta_instrs = False; self.epilog_sync_bar_id = 1 - self.cluster_shape_mn = (1, 1); self.cta_group = tcgen05.CtaGroup.ONE - self.epilogue_warp_id = (0,1,2,3); self.mma_warp_id = 4; self.tma_warp_id = 5 - self.threads_per_cta = 192; self.num_c_stage = 2 - self.kv_stage = 2; self.q_stage = 1; self.num_c_stage = 2 - self.scale_softmax = 1.0 / math.sqrt(HEAD_DIM) - self.scale_softmax_log2 = self.scale_softmax * 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.num_acc_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 - p_cols_fp32 = self.pv_mma_tiler[2] * self.q_dtype.width // self.qk_acc_dtype.width - p_end = self.tmem_p0_offset + p_cols_fp32 - s_cols = self.qk_mma_tiler[1] - o_after = max(s_cols, p_end) - 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)) - v_s = cute.slice_(self.v_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) + - cute.size_in_bytes(self.q_dtype, v_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, self.s_k, 1), - stride=(1, HEAD_DIM, HEAD_DIM * self.s_k), - ), - ) - 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.qk_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) - # Stop here — just check shapes - self._diag_kernel(qk_mma,pv_mma,tma_q,mQ,tma_k,mK,tma_v,mV,self.cluster_layout_vmnk).launch(grid=(1,1,1),block=[self.threads_per_cta,1,1],stream=stream) - - @cute.kernel - def _diag_kernel(self, qk_mma, pv_mma, tma_q, mQ, tma_k, mK, tma_v, mV, cl_vmnk): - 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)) - sQ = cute.make_tensor(BFloat16, q_s.outer) - sK = cute.make_tensor(BFloat16, k_s.outer) - sV = cute.make_tensor(BFloat16, v_s.outer) - - 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)) - - 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) - - 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)) - - # Print shapes BEFORE any pre-slice - print(f"=== TMA partition shapes (n_kv_tiles={self.n_kv_tiles}) ===") - print(f"tAgQ: shape={cute.shape(tAgQ)} rank={tAgQ.rank}") - print(f"tBgK: shape={cute.shape(tBgK)} rank={tBgK.rank}") - print(f"tVgV: shape={cute.shape(tVgV)} rank={tVgV.rank}") - print(f"tAsQ: shape={cute.shape(tAsQ)} rank={tAsQ.rank}") - print(f"tBsK: shape={cute.shape(tBsK)} rank={tBsK.rank}") - print(f"tVsV: shape={cute.shape(tVsV)} rank={tVsV.rank}") - - # Print per-mode sizes - for name, t in [("tAgQ", tAgQ), ("tBgK", tBgK), ("tVgV", tVgV)]: - for i in range(t.rank): - sz = cute.size(t, mode=[i]) - print(f" {name} mode {i} size={sz}") - - # Also print after the original pre-slice - tAgQ2 = tAgQ[(None,0,None,0)] - tBgK2 = tBgK[(None,None,0,0)] - tVgV2 = tVgV[(None,0,None,0)] - print(f"\nAfter pre-slice:") - print(f"tAgQ[(None,0,None,0)]: shape={cute.shape(tAgQ2)} rank={tAgQ2.rank}") - print(f"tBgK[(None,None,0,0)]: shape={cute.shape(tBgK2)} rank={tBgK2.rank}") - print(f"tVgV[(None,0,None,0)]: shape={cute.shape(tVgV2)} rank={tVgV2.rank}") - for name, t in [("tAgQ2", tAgQ2), ("tBgK2", tBgK2), ("tVgV2", tVgV2)]: - for i in range(t.rank): - sz = cute.size(t, mode=[i]) - print(f" {name} mode {i} size={sz}") - - -def test(): +def diag(): n = 256 m, hd = 128, HEAD_DIM + s_k = n + n_kv_tiles = s_k // 128 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') + qk_mma_tiler = (128, 128, 4) + pv_mma_tiler = (128, HEAD_DIM, 4) + cluster_shape_mn = (1, 1) + cta_group = tcgen05.CtaGroup.ONE + qk_acc_dtype = Float32 + q_dtype = BFloat16 + + a_major = LayoutEnum.from_tensor(q).mma_major_mode() + b_major = LayoutEnum.from_tensor(k).mma_major_mode() + v_fmha = cute.make_tensor( + v_kernel, + cute.make_layout( + (HEAD_DIM, s_k, 1), + stride=(1, HEAD_DIM, HEAD_DIM * s_k), + ), + ) + v_major = LayoutEnum.from_tensor(v_fmha).mma_major_mode() + + qk_mma = utils.sm100.make_trivial_tiled_mma(q_dtype, q_dtype, a_major, b_major, qk_acc_dtype, cta_group, (128,128), tcgen05.OperandSource.SMEM) + pv_mma = utils.sm100.make_trivial_tiled_mma(q_dtype, q_dtype, cute.nvgpu.OperandMajorMode.K, v_major, qk_acc_dtype, cta_group, (128,HEAD_DIM), tcgen05.OperandSource.TMEM) + + kv_stage = 2; q_stage = 1 + q_smem_s = utils.sm100.make_smem_layout_a(qk_mma, qk_mma_tiler, q_dtype, q_stage) + k_smem_s = utils.sm100.make_smem_layout_b(qk_mma, qk_mma_tiler, q_dtype, kv_stage) + v_smem_s = utils.sm100.make_smem_layout_b(pv_mma, pv_mma_tiler, q_dtype, kv_stage) + + q_s = cute.slice_(q_smem_s,(None,None,None,0)) + k_s = cute.slice_(k_smem_s,(None,None,None,0)) + v_s = cute.slice_(v_smem_s,(None,None,None,0)) + + cluster_layout_vmnk = cute.tiled_divide(cute.make_layout((1,1,1)), (qk_mma.thr_id.shape,)) + 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) - diag = TmaShapeDiag(s_k=n) - print('Compiling...', flush=True) - compiled = cute.compile(diag, mQ, mK, mV, mC, stream) - print('Running...', flush=True) - compiled(mQ, mK, mV, mC, stream) - torch.cuda.synchronize() - print('Done.') + tma_q,tma_mQ = cute.nvgpu.make_tiled_tma_atom_A(utils.sm100.cluster_shape_to_tma_atom_A(cluster_shape_mn,qk_mma.thr_id),mQ,q_s,qk_mma_tiler,qk_mma,cluster_layout_vmnk.shape) + tma_k,tma_mK = cute.nvgpu.make_tiled_tma_atom_B(utils.sm100.cluster_shape_to_tma_atom_B(cluster_shape_mn,qk_mma.thr_id),mK,k_s,qk_mma_tiler,qk_mma,cluster_layout_vmnk.shape) + tma_v,tma_mV = cute.nvgpu.make_tiled_tma_atom_B(utils.sm100.cluster_shape_to_tma_atom_B(cluster_shape_mn,pv_mma.thr_id),mV,v_s,pv_mma_tiler,pv_mma,cluster_layout_vmnk.shape) + + gQ = cute.local_tile(tma_mQ,cute.slice_(qk_mma_tiler,(None,0,None)),(None,None,None)) + gK = cute.local_tile(tma_mK,cute.slice_(qk_mma_tiler,(0,None,None)),(None,None,None)) + gV = cute.local_tile(tma_mV,cute.slice_(pv_mma_tiler,(0,None,None)),(None,None,None)) + + 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) + + a_lay = cute.make_layout(cute.slice_(cluster_layout_vmnk,(0,0,None,0)).shape) + b_lay = cute.make_layout(cute.slice_(cluster_layout_vmnk,(0,None,0,0)).shape) + + # Use the full SMEM layouts (not sliced) for group_modes + tAsQ,tAgQ = cpasync.tma_partition(tma_q,0,a_lay,cute.group_modes(q_s,0,3),cute.group_modes(tCgQ,0,3)) + tBsK,tBgK = cpasync.tma_partition(tma_k,0,b_lay,cute.group_modes(k_s,0,3),cute.group_modes(tCgK,0,3)) + tVsV,tVgV = cpasync.tma_partition(tma_v,0,b_lay,cute.group_modes(v_s,0,3),cute.group_modes(tCgV,0,3)) + + print(f"=== TMA partition shapes (n_kv_tiles={n_kv_tiles}) ===") + print(f"tAgQ: shape={cute.shape(tAgQ)}") + print(f"tBgK: shape={cute.shape(tBgK)}") + print(f"tVgV: shape={cute.shape(tVgV)}") + print(f"tAsQ: shape={cute.shape(tAsQ)}") + print(f"tBsK: shape={cute.shape(tBsK)}") + print(f"tVsV: shape={cute.shape(tVsV)}") + + for name, t in [("tAgQ", tAgQ), ("tBgK", tBgK), ("tVgV", tVgV)]: + for i in range(t.rank): + sz = cute.size(t, mode=[i]) + print(f" {name} mode {i} size={sz}") + + # After pre-slice + tAgQ2 = tAgQ[(None,0,None,0)] + tBgK2 = tBgK[(None,None,0,0)] + tVgV2 = tVgV[(None,0,None,0)] + print(f"\nAfter pre-slice (None,0,None,0) / (None,None,0,0) / (None,0,None,0):") + print(f"tAgQ: shape={cute.shape(tAgQ2)}") + print(f"tBgK: shape={cute.shape(tBgK2)}") + print(f"tVgV: shape={cute.shape(tVgV2)}") + for name, t in [("tAgQ2", tAgQ2), ("tBgK2", tBgK2), ("tVgV2", tVgV2)]: + for i in range(t.rank): + sz = cute.size(t, mode=[i]) + print(f" {name} mode {i} size={sz}") if __name__ == '__main__': - test() + diag()