D3: In-kernel SWA sequence length masking

- Add apply_swa_mask flag to FmhaKernel constructor
- After TMEM load of S, use tTMEM_LOADcS coordinates to map register
  fragment positions to (row, col) in QK matrix
- Mask positions >= swa_lens[batch_idx] to -inf before softmax
- Supports multi-KV-tile (kt*128 + k_coord for absolute position)
- swa_lens parameter passed as CuTe tensor, indexed by block_idx_z
- Dummy tensor (max int) when swa_lens=None (no masking)
- New test: test_d3_inkernel_mask.py with proper in-kernel masking
- Replaces pre-masking approach (BF16 min on K) which can't produce -inf
This commit is contained in:
2026-05-26 10:51:23 +00:00
parent d6a56342cc
commit b6b581777a
2 changed files with 284 additions and 3 deletions

View File

@@ -16,7 +16,7 @@ import math
class FmhaKernel:
def __init__(self, head_dim=64, s_k=128, scale_softmax=None, use_smem_p=None, normalize=True, num_query_heads=1, batch_size=1):
def __init__(self, head_dim=64, s_k=128, scale_softmax=None, use_smem_p=None, normalize=True, num_query_heads=1, batch_size=1, apply_swa_mask=False):
self.head_dim = head_dim
self.s_k = s_k
self.n_kv_tiles = s_k // 128
@@ -31,6 +31,7 @@ class FmhaKernel:
self.num_query_heads = num_query_heads
self.batch_size = batch_size
self.normalize = normalize # D5a: False = emit un-normalized O + lse
self.apply_swa_mask = apply_swa_mask # D3: mask logits at positions >= swa_lens
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
@@ -132,10 +133,12 @@ class FmhaKernel:
if const_expr(lse is None):
lse = cute.make_tensor(c.iterator, cute.make_layout((1,), stride=(0,)))
if const_expr(swa_lens is None):
# No SWA masking — pass a dummy tensor
swa_lens = cute.make_tensor(c.iterator, cute.make_layout((1,), stride=(0,)))
# No SWA masking — pass a dummy tensor with large value (no positions masked)
_swa_dummy = torch.tensor([2147483647], dtype=torch.int32, device='cuda')
swa_lens = ct.from_dlpack(_swa_dummy).mark_layout_dynamic(leading_dim=ct.get_leading_dim(_swa_dummy))
# Grid: (M_tiles, 1, batch) where M = n_h * T packed into M dimension
# For single-head (n_h=1): grid=(1,1,1) — backward compatible
# block_idx_z = batch index, used for swa_lens[batch_idx] in D3 masking
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.p_smem_s,self.c_smem_s,self.epi_tile,lse,swa_lens).launch(grid=(1,1,self.batch_size),block=[self.threads_per_cta,1,1],stream=stream)
@cute.kernel
@@ -407,6 +410,25 @@ class FmhaKernel:
cute.copy(tiled_tmem_load, tTMEM_LOADtS, tTMEM_LOADrS)
cute.arch.fence_view_async_tmem_load()
# D3: In-kernel SWA sequence length masking.
# After loading S from TMEM, mask positions >= swa_lens[batch_idx] to -inf.
# Uses tTMEM_LOADcS coordinate tensor to map register indices to (row, col).
# col = position in KV sequence. For kt > 0, actual pos = kt*128 + col.
# This is the PROPER approach: post-QK masking in the softmax,
# not pre-masking K with BF16 min (which can't produce true -inf).
if const_expr(self.apply_swa_mask):
_bidx, _bidy, _bidz = cute.arch.block_idx()
swa_len = swa_lens[_bidz]
kt_offset = Int32(kt * 128) # KV position offset for this tile
# Iterate using same coordinate indexing as SMEM-P path
for j0 in range(32):
for j1 in range(4):
coord = tTMEM_LOADcS[(j0, 0), j1, 0, 0]
k_coord = coord[1] # position within this KV tile
kv_pos = kt_offset + k_coord # absolute KV position
if kv_pos >= swa_len:
tTMEM_LOADrS[(j0, 0), j1, 0, 0] = -Float32.inf
old_row_max = row_max
frg_cnt = 4
frg_tile = cute.size(tTMEM_LOADrS) // frg_cnt

View File

@@ -0,0 +1,259 @@
"""
FMHA D3: In-kernel SWA sequence length masking.
Proper approach: the kernel receives swa_lens and masks logits to -inf
inside the softmax, using the tTMEM_LOADcS coordinate tensor to map
register fragment positions to (row, col) in the QK matrix.
This replaces the pre-masking approach (BF16 min on K) which cannot
produce true -inf QK scores.
Run: ~/.openclaw/workspace/fire_b200_test tests/unit/test_d3_inkernel_mask.py
"""
import torch
import math
import cutlass.cute as cute
import cutlass.torch as ct
import cuda.bindings.driver as cuda
from dsv4.kernels.attention.fmha import FmhaKernel
def reference_swa_attention(q, k, v, swa_lens, scale):
"""FP32 reference with proper -inf masking.
Args:
q: (M, hd) BF16
k: (s_k, hd) BF16
v: (s_k, hd) BF16
swa_lens: (M,) int32 — per-row number of valid KV positions
scale: float
Returns:
o: (M, hd) BF16
"""
scores = torch.matmul(q.float(), k.float().T) * scale
for i in range(q.shape[0]):
sl = swa_lens[i].item()
if sl < k.shape[0]:
scores[i, sl:] = float('-inf')
max_s = scores.max(dim=-1, keepdim=True).values
exp_s = (scores - max_s).exp()
sum_s = exp_s.sum(dim=-1, keepdim=True)
p = exp_s / sum_s
o = torch.matmul(p, v.float())
return o.to(torch.bfloat16)
def _run_fmha_masked(q_3d, k_3d, v, m, s_k, hd, swa_lens_tensor, use_smem_p=False):
"""Run FMHA with in-kernel SWA masking and return normalized output.
Args:
q_3d: (M, hd, 1) BF16
k_3d: (s_k, hd, 1) BF16
v: (s_k, hd) BF16
swa_lens_tensor: (1,) int32 — number of valid KV positions
Returns:
o_norm: (M, hd) BF16
"""
scale = 1.0 / math.sqrt(hd)
kernel = FmhaKernel(
head_dim=hd, s_k=s_k, use_smem_p=use_smem_p,
apply_swa_mask=True, normalize=False,
)
pv_n_tile = kernel.pv_n_tile
n_pv_tiles = kernel.n_pv_tiles
stream = cuda.CUstream(torch.cuda.current_stream().cuda_stream)
# swa_lens as CuTe tensor (1D, int32)
mSwaLens = ct.from_dlpack(swa_lens_tensor).mark_layout_dynamic(
leading_dim=ct.get_leading_dim(swa_lens_tensor)
)
o_unnorm = torch.zeros(m, hd, dtype=torch.float32, device='cuda')
lse_all = torch.zeros(m, dtype=torch.float32, device='cuda')
for pv in range(n_pv_tiles):
v_tile = v[:, pv * pv_n_tile:(pv + 1) * pv_n_tile].contiguous().unsqueeze(-1)
c_tile = torch.zeros(m, pv_n_tile, 1, dtype=torch.bfloat16, device='cuda')
lse_tensor = torch.zeros(m, 1, 1, dtype=torch.float32, device='cuda')
mQ = ct.from_dlpack(q_3d).mark_layout_dynamic(leading_dim=ct.get_leading_dim(q_3d))
mK = ct.from_dlpack(k_3d).mark_layout_dynamic(leading_dim=ct.get_leading_dim(k_3d))
mV = ct.from_dlpack(v_tile).mark_layout_dynamic(leading_dim=ct.get_leading_dim(v_tile))
mC = ct.from_dlpack(c_tile).mark_layout_dynamic(leading_dim=ct.get_leading_dim(c_tile))
mLSE = ct.from_dlpack(lse_tensor).mark_layout_dynamic(leading_dim=ct.get_leading_dim(lse_tensor))
if pv == 0:
compiled = cute.compile(kernel, mQ, mK, mV, mC, stream, mLSE, mSwaLens)
compiled(mQ, mK, mV, mC, stream, mLSE, mSwaLens)
o_unnorm[:, pv * pv_n_tile:(pv + 1) * pv_n_tile] = c_tile[:, :, 0].float()
lse_all += lse_tensor[:, 0, 0]
# External normalization using LSE
# O_norm = O_unnorm / exp(LSE) ... but per-row LSE only row 0 is written.
# Use reference attn_sum for normalization (same as head-packed tests).
q_flat = q_3d[:, :, 0]
k_flat = k_3d[:, :, 0]
scores = torch.matmul(q_flat.float(), k_flat.float().T) * scale
swa_len_val = swa_lens_tensor[0].item()
if swa_len_val < s_k:
scores[:, swa_len_val:] = float('-inf')
max_s = scores.max(dim=-1, keepdim=True).values
attn_sum = (scores - max_s).exp().sum(dim=-1, keepdim=True)
o_norm = (o_unnorm / attn_sum).to(torch.bfloat16)
return o_norm
def test_d3_no_mask():
"""Full window (swa_lens=128): no masking, regression test."""
print("\n=== Test 1: No masking (swa_lens=128, hd=64) ===")
torch.manual_seed(42)
m, s_k, hd = 128, 128, 64
scale = 1.0 / math.sqrt(hd)
q = torch.randn(m, hd, 1, dtype=torch.bfloat16, device='cuda')
k = torch.randn(s_k, hd, 1, dtype=torch.bfloat16, device='cuda')
v = torch.randn(s_k, hd, dtype=torch.bfloat16, device='cuda')
swa_lens = torch.tensor([s_k], dtype=torch.int32, device='cuda')
o = _run_fmha_masked(q, k, v, m, s_k, hd, swa_lens)
ref = reference_swa_attention(q[:, :, 0], k[:, :, 0], v, swa_lens.cpu().expand(m), scale)
cos = torch.nn.functional.cosine_similarity(
o.flatten().float().unsqueeze(0), ref.flatten().float().unsqueeze(0)
).item()
print(f" cos = {cos:.6f}")
assert cos >= 0.995, f"Regression: cosine too low: {cos}"
print(" ✅ PASS")
def test_d3_swa64():
"""SWA with swa_lens=64: mask positions 64-127 to -inf."""
print("\n=== Test 2: swa_lens=64 (hd=64) ===")
torch.manual_seed(42)
m, s_k, hd = 128, 128, 64
scale = 1.0 / math.sqrt(hd)
q = torch.randn(m, hd, 1, dtype=torch.bfloat16, device='cuda')
k = torch.randn(s_k, hd, 1, dtype=torch.bfloat16, device='cuda')
v = torch.randn(s_k, hd, dtype=torch.bfloat16, device='cuda')
swa_lens = torch.tensor([64], dtype=torch.int32, device='cuda')
o = _run_fmha_masked(q, k, v, m, s_k, hd, swa_lens)
ref = reference_swa_attention(q[:, :, 0], k[:, :, 0], v, swa_lens.cpu().expand(m), scale)
cos = torch.nn.functional.cosine_similarity(
o.flatten().float().unsqueeze(0), ref.flatten().float().unsqueeze(0)
).item()
print(f" cos = {cos:.6f}")
assert cos >= 0.99, f"cosine too low: {cos}"
print(" ✅ PASS")
def test_d3_swa32():
"""SWA with swa_lens=32: only 32 valid positions."""
print("\n=== Test 3: swa_lens=32 (hd=64) ===")
torch.manual_seed(42)
m, s_k, hd = 128, 128, 64
scale = 1.0 / math.sqrt(hd)
q = torch.randn(m, hd, 1, dtype=torch.bfloat16, device='cuda')
k = torch.randn(s_k, hd, 1, dtype=torch.bfloat16, device='cuda')
v = torch.randn(s_k, hd, dtype=torch.bfloat16, device='cuda')
swa_lens = torch.tensor([32], dtype=torch.int32, device='cuda')
o = _run_fmha_masked(q, k, v, m, s_k, hd, swa_lens)
ref = reference_swa_attention(q[:, :, 0], k[:, :, 0], v, swa_lens.cpu().expand(m), scale)
cos = torch.nn.functional.cosine_similarity(
o.flatten().float().unsqueeze(0), ref.flatten().float().unsqueeze(0)
).item()
print(f" cos = {cos:.6f}")
assert cos >= 0.99, f"cosine too low: {cos}"
print(" ✅ PASS")
def test_d3_swa1():
"""Edge case: swa_lens=1, only one valid KV position."""
print("\n=== Test 4: swa_lens=1 (hd=64) ===")
torch.manual_seed(42)
m, s_k, hd = 128, 128, 64
scale = 1.0 / math.sqrt(hd)
q = torch.randn(m, hd, 1, dtype=torch.bfloat16, device='cuda')
k = torch.randn(s_k, hd, 1, dtype=torch.bfloat16, device='cuda')
v = torch.randn(s_k, hd, dtype=torch.bfloat16, device='cuda')
swa_lens = torch.tensor([1], dtype=torch.int32, device='cuda')
o = _run_fmha_masked(q, k, v, m, s_k, hd, swa_lens)
ref = reference_swa_attention(q[:, :, 0], k[:, :, 0], v, swa_lens.cpu().expand(m), scale)
cos = torch.nn.functional.cosine_similarity(
o.flatten().float().unsqueeze(0), ref.flatten().float().unsqueeze(0)
).item()
print(f" cos = {cos:.6f}")
assert cos >= 0.99, f"cosine too low: {cos}"
print(" ✅ PASS")
def test_d3_hd128():
"""SWA masking at hd=128 (SMEM-P path)."""
print("\n=== Test 5: swa_lens=64 (hd=128) ===")
torch.manual_seed(42)
m, s_k, hd = 128, 128, 128
scale = 1.0 / math.sqrt(hd)
q = torch.randn(m, hd, 1, dtype=torch.bfloat16, device='cuda')
k = torch.randn(s_k, hd, 1, dtype=torch.bfloat16, device='cuda')
v = torch.randn(s_k, hd, dtype=torch.bfloat16, device='cuda')
swa_lens = torch.tensor([64], dtype=torch.int32, device='cuda')
o = _run_fmha_masked(q, k, v, m, s_k, hd, swa_lens, use_smem_p=True)
ref = reference_swa_attention(q[:, :, 0], k[:, :, 0], v, swa_lens.cpu().expand(m), scale)
cos = torch.nn.functional.cosine_similarity(
o.flatten().float().unsqueeze(0), ref.flatten().float().unsqueeze(0)
).item()
print(f" cos = {cos:.6f}")
assert cos >= 0.99, f"cosine too low: {cos}"
print(" ✅ PASS")
def test_d3_swa128_hd128():
"""No masking at hd=128: regression test (should match existing D1 results)."""
print("\n=== Test 6: No masking (swa_lens=128, hd=128) ===")
torch.manual_seed(42)
m, s_k, hd = 128, 128, 128
scale = 1.0 / math.sqrt(hd)
q = torch.randn(m, hd, 1, dtype=torch.bfloat16, device='cuda')
k = torch.randn(s_k, hd, 1, dtype=torch.bfloat16, device='cuda')
v = torch.randn(s_k, hd, dtype=torch.bfloat16, device='cuda')
swa_lens = torch.tensor([s_k], dtype=torch.int32, device='cuda')
o = _run_fmha_masked(q, k, v, m, s_k, hd, swa_lens, use_smem_p=True)
ref = reference_swa_attention(q[:, :, 0], k[:, :, 0], v, swa_lens.cpu().expand(m), scale)
cos = torch.nn.functional.cosine_similarity(
o.flatten().float().unsqueeze(0), ref.flatten().float().unsqueeze(0)
).item()
print(f" cos = {cos:.6f}")
assert cos >= 0.995, f"Regression: cosine too low: {cos}"
print(" ✅ PASS")
def test():
print("=== D3: In-Kernel SWA Sequence Length Mask ===")
test_d3_no_mask()
test_d3_swa64()
test_d3_swa32()
test_d3_swa1()
test_d3_hd128()
test_d3_swa128_hd128()
print("\n=== ALL TESTS PASSED ===")
if __name__ == '__main__':
test()