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

fix: weight reshape bug + igs double-count in BF16 reference

Browse Source
This commit is contained in:
biondizzle
2026-05-15 15:46:16 +00:00
parent c421a668f3
commit 2fd55a94c6
1 changed files with 13 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files

20
src/nvfp4_megamoe_kernel/nvfp4_mega_moe.py
View File

@@ -401,33 +401,39 @@ def nvfp4_mega_moe_full(
x_u8 = x_fp4[s0].view(torch.uint8)
lo = (x_u8 & 0x0F).long()
hi = ((x_u8 >> 4) & 0x0F).long()
x_nib = torch.stack([lo, hi], dim=-1).reshape(-1)
x_nib = torch.stack([lo, hi], dim=-1).reshape(-1) # (K,) — 1D so simple flatten works
x_signs = (x_nib >> 3).float() * -2 + 1
x_mags = _E2M1_MAGNITUDES.to(device=x_u8.device)[(x_nib & 0x07)]
x_deq = x_signs * x_mags
sf_exp = x_sf[s0].to(torch.float32).repeat_interleave(16, dim=-1)
x_deq = x_signs * x_mags # (K,) = (7168,)
sf_exp = x_sf[s0].to(torch.float32).repeat_interleave(16, dim=-1) # (K,)
igs = float(l1_global_scale) if not isinstance(l1_global_scale, float) else l1_global_scale
x_bf16 = (x_deq * sf_exp * igs).to(torch.bfloat16)
# Dequantize L1 weight for expert e0
w_u8 = l1_w[e0].view(torch.uint8)
wlo = (w_u8 & 0x0F).long()
whi = ((w_u8 >> 4) & 0x0F).long()
w_nib = torch.stack([wlo, whi], dim=-1).reshape(w_u8.shape[0], -1)
w_nib = torch.stack([wlo, whi], dim=-1).reshape(w_u8.shape[0] * 2, w_u8.shape[1]) # (K, N)
w_signs = (w_nib >> 3).float() * -2 + 1
w_mags = _E2M1_MAGNITUDES.to(device=w_u8.device)[(w_nib & 0x07)]
w_deq = w_signs * w_mags
w_sf_exp = l1_sf[e0].to(torch.float32).repeat_interleave(16, dim=0)
w_deq = w_signs * w_mags # (K, N) = (7168, 6144)
# Weight SF: (sf_k, N) = (448, 6144). Each SF covers 16 FP4 values (8 bytes).
# repeat_interleave(16) on dim 0 gives (7168, 6144) — but wait,
# sf_k=448, and K=7168, so 448*16=7168. This is per-FP4-value already.
w_sf_exp = l1_sf[e0].to(torch.float32).repeat_interleave(16, dim=0) # (K, N)
gs = l1_global_sf[e0]
if gs.dim() == 0:
w_bf16 = (w_deq * w_sf_exp * gs.item()).to(torch.bfloat16)
else:
w_bf16 = (w_deq * w_sf_exp).to(torch.bfloat16)
# GEMM computes: (w * w_sf) @ (x * x_sf) * alpha * gs
# We compute: (x * x_sf * igs) @ (w * w_sf * gs_per_half)
# which equals igs * gs * (x_sf * w_sf) @ (x * w) = same as GEMM
ref_out = torch.nn.functional.linear(x_bf16.unsqueeze(0), w_bf16.T).squeeze(0)
if gs.dim() > 0:
gn = ref_out.shape[0] // 2
ref_out[:gn] = ref_out[:gn] * gs[0].item()
ref_out[gn:] = ref_out[gn:] * gs[1].item()
ref_out = ref_out * igs
# DON'T multiply by igs again — already in x_bf16
nvfp4_mega_moe_full._ref_l1 = (s0, e0, ref_out)
print(f"[BF16-REF-L1] expert={e0} amax={ref_out.abs().max():.4e} mean={ref_out.mean():.4e}")
except Exception as ex: