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

D1: Simplified debug test

Browse Source
This commit is contained in:
biondizzle
2026-05-24 03:29:14 +00:00
parent 1f1b16ad07
commit dd7356afc6
1 changed files with 15 additions and 27 deletions
Download Patch File Download Diff File Expand all files Collapse all files

42
tests/unit/test_d1_hd128_debug.py
View File

@@ -1,4 +1,4 @@
"""D1: Debug hd=128 — check if the QK output is correct."""
"""D1: Debug hd=128 — check if the pipeline works with TMEM-P."""
import torch, math
import cutlass.cute as cute
import cutlass.torch as ct
@@ -6,16 +6,13 @@ import cuda.bindings.driver as cuda
from dsv4.kernels.attention.fmha import FmhaKernel
def test_hd128_debug():
hd = 128
n_kv = 128
def test_hd(hd, n_kv=128, use_smem_p=False):
m = 128
torch.manual_seed(42)
q = torch.randn(m, hd, 1, dtype=torch.bfloat16, device='cuda')
k = torch.randn(n_kv, hd, 1, dtype=torch.bfloat16, device='cuda')
v = torch.randn(n_kv, hd, dtype=torch.bfloat16, device='cuda')
# Reference: just the QK @ V attention (un-normalized)
qf = q[:, :, 0].float()
kf = k[:, :, 0].float()
scale = 1.0 / math.sqrt(hd)
@@ -26,17 +23,11 @@ def test_hd128_debug():
ref_unnorm = attn_exp @ v.float()
ref_lse = (torch.log(attn_sum.squeeze(-1)) + attn_max.squeeze(-1))[0].item()
# Run kernel with TMEM-P (force)
lse_tensor = torch.zeros(m, 1, 1, dtype=torch.float32, device='cuda')
kernel = FmhaKernel(head_dim=hd, s_k=n_kv, use_smem_p=False)
kernel = FmhaKernel(head_dim=hd, s_k=n_kv, use_smem_p=use_smem_p)
pv_n_tile = kernel.pv_n_tile
print(f'pv_n_tile={pv_n_tile}, n_pv_tiles={kernel.n_pv_tiles}')
print(f'tmem_o0_offset={kernel.tmem_o0_offset}, tmem_p0_offset={kernel.tmem_p0_offset}')
print(f'tOrP0_offset={kernel.tOrP0_offset}')
print(f'num_tmem_alloc_cols={kernel.num_tmem_alloc_cols}')
print(f'scale_softmax={kernel.scale_softmax}, scale_softmax_log2={kernel.scale_softmax_log2}')
stream = cuda.CUstream(torch.cuda.current_stream().cuda_stream)
v_tile = v[:, 0:pv_n_tile].contiguous().unsqueeze(-1)
c_tile = torch.zeros(m, pv_n_tile, 1, dtype=torch.bfloat16, device='cuda')
mQ = ct.from_dlpack(q).mark_layout_dynamic(leading_dim=ct.get_leading_dim(q))
@@ -45,6 +36,8 @@ def test_hd128_debug():
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))
mode = "SMEM-P" if use_smem_p else "TMEM-P"
print(f'hd={hd} {mode}: Compiling...', flush=True)
compiled = cute.compile(kernel, mQ, mK, mV, mC, stream, mLSE)
compiled(mQ, mK, mV, mC, stream, mLSE)
torch.cuda.synchronize()
@@ -52,22 +45,17 @@ def test_hd128_debug():
out = c_tile[:, :, 0].float()
kernel_lse = lse_tensor[0, 0, 0].item()
cos = torch.nn.functional.cosine_similarity(out.flatten().unsqueeze(0), ref_unnorm.flatten().unsqueeze(0)).item()
lse_err = abs(kernel_lse - ref_lse)
print(f'\nResults:')
print(f' cos_unnorm={cos:.6f}')
print(f' kernel_lse={kernel_lse:.6f} ref_lse={ref_lse:.6f} err={abs(kernel_lse - ref_lse):.6f}')
print(f' out[0,:4]={out[0,:4].tolist()}')
print(f' ref[0,:4]={ref_unnorm[0,:4].tolist()}')
# Check: is the output roughly the right magnitude?
print(f' out.abs().max()={out.abs().max().item():.4f}')
print(f' ref.abs().max()={ref_unnorm.abs().max().item():.4f}')
# Check row-by-row: is O[0] proportional to ref[0]?
print(f'hd={hd} {mode}: cos={cos:.6f} lse_err={lse_err:.6f} {"PASS" if cos >= 0.99 else "FAIL"}')
if cos < 0.99:
row_cos = torch.nn.functional.cosine_similarity(out[0].unsqueeze(0), ref_unnorm[0].unsqueeze(0)).item()
print(f' row0_cos={row_cos:.6f}')
print(f' out[0,:4]={out[0,:4].tolist()}')
print(f' ref[0,:4]={ref_unnorm[0,:4].tolist()}')
return cos
if __name__ == '__main__':
test_hd128_debug()
print("=== D1 Debug: TMEM-P at various hd ===\n")
test_hd(64, use_smem_p=False)
test_hd(128, use_smem_p=False)
test_hd(256, use_smem_p=False)