biondizzle/nvfp4-megamoe-kernel
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Reset to working_softmax_maybe.py + TMA fix only

Browse Source 
Previous O rescale attempt broke n=128 (0.464773).
Revert to known-good softmax code, only apply TMA fix:
tBgK[(None,None,0,0)] → tBgK[(None,0,None,0)]

Expected: n=128 cos 0.999998 (same as working), n=256 cos 0.71 (TMA fix loads 2 tiles but no O rescale)
This commit is contained in:
biondizzle
2026-05-23 00:27:41 +00:00
parent 524b5b1840
commit fae9f6fbb5
1 changed files with 61 additions and 68 deletions
Download Patch File Download Diff File Expand all files Collapse all files

129
tests/unit/test_fmha_v3_stage_c.py
View File

@@ -180,13 +180,7 @@ class FmhaV3StageCMulti:
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))
# After tma_partition, tBgK has 4 modes: (V_grouped, ?, KV_tiles, ?).
# Mode 4 is the GMEM tile iteration axis (size = n_kv_tiles).
# 8-None no-op pre-slice opens the full TMA coord space.
# tAgQ is fine with 4-mode slice (Q has only 1 tile).
tAgQ = tAgQ[(None,0,None,0)]
tBgK = tBgK[(None,0,None,0)]
tVgV = tVgV[(None,0,None,0)]
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)
@@ -210,13 +204,9 @@ class FmhaV3StageCMulti:
pipeline.pipeline_init_wait(cluster_shape_mn=cl_vmnk)
# ===== TMA LOAD warp =====
# GMEM tile coordinate: use the cutlass.range induction variable kt
# directly. CuTeDSL's `cutlass.range` doesn't auto-detect a Python `+=`
# rebinding as a loop-carried iter_args update — the JIT traces the
# body once and captures whatever value `kv_coord` had at trace time,
# so an outer `kv_coord = Int32(0)` plus a `kv_coord += 1` inside the
# loop bakes 0 into every iteration's TMA descriptor at runtime.
# The induction variable IS the loop-carried state, properly tracked.
# NOTE: using kt from cutlass.range works for n=128 (single tile).
# Multi-tile (n>128) loads from tile 0 only — the JIT constant-folds kt.
# TODO: fix multi-tile TMA indexing (kv_coord pattern from diag test).
if warp_idx == self.tma_warp_id:
qp.reset(); qh = qp.acquire_and_advance()
cute.copy(tma_q, tAgQ[(None, Int32(0))], tAsQ[(None, qh.index)], tma_bar_ptr=qh.barrier)
@@ -224,8 +214,8 @@ class FmhaV3StageCMulti:
kvp.reset(); pk = kvp.try_acquire()
for kt in cutlass.range(0, n_kv_tiles, 1, unroll=1):
kvh = kvp.acquire_and_advance(pk)
cute.copy(tma_k, tBgK[None, kt], tBsK[(None, kvh.index)], tma_bar_ptr=kvh.barrier)
cute.copy(tma_v, tVgV[None, kt], tVsV[(None, kvh.index)], tma_bar_ptr=kvh.barrier)
cute.copy(tma_k, tBgK[(None, kt)], tBsK[(None, kvh.index)], tma_bar_ptr=kvh.barrier)
cute.copy(tma_v, tVgV[(None, kt)], tVsV[(None, kvh.index)], tma_bar_ptr=kvh.barrier)
pk = cutlass.Boolean(1)
kvp.tail()
@@ -238,7 +228,7 @@ class FmhaV3StageCMulti:
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(self.n_kv_tiles):
for kt in range(n_kv_tiles):
kvh = kvc.wait_and_advance(pk); pk = cutlass.Boolean(1)
sh = s_prod.acquire_and_advance()
qk_mma.set(tcgen05.Field.ACCUMULATE, False)
@@ -296,31 +286,7 @@ class FmhaV3StageCMulti:
row_sum = Float32(0.0)
scale_log2 = Float32(self.scale_softmax_log2)
# === O rescale / correction_rescale setup (needed inside softmax loop) ===
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)
tmem_load_o_atom = cute.make_copy_atom(
tcgen05.copy.Ld32x32bOp(tcgen05.copy.Repetition(corr_tile_size)),
self.acc_dtype,
)
tmem_store_o_atom = cute.make_copy_atom(
tcgen05.copy.St32x32bOp(tcgen05.copy.Repetition(corr_tile_size)),
self.acc_dtype,
)
tiled_tmem_load_o = tcgen05.make_tmem_copy(tmem_load_o_atom, tOtO_i)
tiled_tmem_store_o = tcgen05.make_tmem_copy(tmem_store_o_atom, tOtO_i)
thr_tmem_load_o = tiled_tmem_load_o.get_slice(sfw_idx)
thr_tmem_store_o = tiled_tmem_store_o.get_slice(sfw_idx)
tTMEM_LOADtO = thr_tmem_load_o.partition_S(tOtO_i)
tTMEM_LOADcO = thr_tmem_load_o.partition_D(tOcO_i)
tTMEM_STOREtO = thr_tmem_store_o.partition_D(tOtO_i)
# Per-tile softmax loop with online O rescale.
# Per-tile softmax loop.
# Online softmax row_max/row_sum tracking is maintained, but the
# in-place TMEM O rescale (which would multiply existing O by
# exp2(old_max - new_max) before PV[kt]) is DISABLED — this is the
@@ -330,7 +296,7 @@ class FmhaV3StageCMulti:
# For now: kernel is correct when row_max growth across tiles is
# mild (typical for short n with random data); for very long n
# the missing rescale shows as accuracy drift.
for kt in range(self.n_kv_tiles):
for kt in range(n_kv_tiles):
si_handle = s_cons.wait_and_advance()
# Load S[kt]
@@ -378,41 +344,68 @@ class FmhaV3StageCMulti:
cute.copy(tiled_tmem_store, rP_words, tTMEM_STOREtP)
cute.arch.fence_view_async_tmem_store()
# === Per-tile O rescale: O *= acc_scale for kt > 0 ===
# Must run BEFORE softmax_done_bar.arrive() so MMA's PV[kt]
# reads the rescaled O. Uses same paired atoms as final normalize.
if kt > 0:
for i in range(HEAD_DIM // corr_tile_size):
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,
)
tTMrO = cute.make_rmem_tensor(tTMEM_LOADcO.shape, self.acc_dtype)
cute.copy(tiled_tmem_load_o, tTMEM_LOADtO_i, tTMrO)
cute.arch.fence_view_async_tmem_load()
for k in cutlass.range(cute.size(tTMrO), vectorize=True):
tTMrO[k] = tTMrO[k] * acc_scale
cute.copy(tiled_tmem_store_o, tTMrO, tTMEM_STOREtO_i)
cute.arch.fence_view_async_tmem_store()
si_handle.release()
softmax_done_bar.arrive()
# Wait for MMA's PV[N-1] to commit before reading O for normalize.
# === Reference-style scaled epilogue (no TMEM round-trip) ===
#
# Pattern (mirrors CUTLASS Blackwell FMHA reference's
# correction_epilog): for each column sub-tile,
# 1. TMEM -> registers via PAIRED tmem_load atom
# 2. scale in registers (1/row_sum)
# 3. FP32 -> BF16 conversion in registers
# 4. registers -> SMEM via PAIRED smem_store atom
# Then TMA SMEM -> GMEM as a separate step.
#
# Critical: the load and store atoms MUST be a matched pair.
# Independently constructed Ld32x32bOp + St32x32bOp atoms (the
# previous code) don't preserve the register tile shape, so even a
# no-op load+store corrupts data. Using utils.blackwell_helpers
# (sm100_utils) gives a paired set keyed to the same epi_subtile.
# Wait for MMA's PV[N-1] to commit before reading O.
final_o_bar.arrive_and_wait()
# === Final O normalization: O *= 1/row_sum ===
# Uses the working 2D register tensor pattern from working_softmax_maybe.py.
inv_row_sum = Float32(1.0) / row_sum
# === O normalization via TMEM load → scale → TMEM store ===
# Matches CUTLASS reference's correction_rescale pattern exactly.
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)
tmem_load_atom = cute.make_copy_atom(
tcgen05.copy.Ld32x32bOp(tcgen05.copy.Repetition(corr_tile_size)),
self.acc_dtype,
)
tmem_store_atom = cute.make_copy_atom(
tcgen05.copy.St32x32bOp(tcgen05.copy.Repetition(corr_tile_size)),
self.acc_dtype,
)
tiled_tmem_load_o = tcgen05.make_tmem_copy(tmem_load_atom, tOtO_i)
tiled_tmem_store_o = tcgen05.make_tmem_copy(tmem_store_atom, tOtO_i)
thr_tmem_load_o = tiled_tmem_load_o.get_slice(sfw_idx)
thr_tmem_store_o = tiled_tmem_store_o.get_slice(sfw_idx)
tTMEM_LOADtO = thr_tmem_load_o.partition_S(tOtO_i)
tTMEM_LOADcO = thr_tmem_load_o.partition_D(tOcO_i)
tTMEM_STOREtO = thr_tmem_store_o.partition_D(tOtO_i)
# 2D register tensor: (frg_shape, n_corr_tiles)
tTMrO = cute.make_rmem_tensor(
(tTMEM_LOADcO.shape, 128 // corr_tile_size), self.acc_dtype
)
inv_row_sum = Float32(1.0) / row_sum
for i in range(HEAD_DIM // corr_tile_size):
tTMrO_i_ = tTMrO[None, i]
tTMrO_i_layout = cute.composition(