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

Clean up L0 B1 verify noise (gate on VERBOSE), update FINAL_STRETCH.md

Browse Source 
Batched prefill + T>128 chunking now complete. All dangling items in
FINAL_STRETCH.md are marked done.
This commit is contained in:
biondizzle
2026-06-03 08:12:54 +00:00
parent 5e09be08af
commit 019a3a34b7
2 changed files with 7 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
single_shot_inference.py
View File

@@ -894,9 +894,10 @@ def forward_attention(x_normed, w, li, cfg, rope_cos, rope_sin,
# 6. Production FMHA — B1 mixed FP8/BF16 decode path.
_pt('fmha_start')
if li == 0:
print(f" L0 B1 verify: kv_nope_fp8 dtype={kv_nope_fp8.dtype} shape={tuple(kv_nope_fp8.shape)} "
f"kv_nope_scale dtype={kv_nope_scale.dtype} shape={tuple(kv_nope_scale.shape)} "
f"kv_rope_bf16 dtype={kv_rope_bf16.dtype} shape={tuple(kv_rope_bf16.shape)}", flush=True)
if VERBOSE >= 2:
print(f" L0 B1 verify: kv_nope_fp8 dtype={kv_nope_fp8.dtype} shape={tuple(kv_nope_fp8.shape)} "
f"kv_nope_scale dtype={kv_nope_scale.dtype} shape={tuple(kv_nope_scale.shape)} "
f"kv_rope_bf16 dtype={kv_rope_bf16.dtype} shape={tuple(kv_rope_bf16.shape)}", flush=True)
assert kv_nope_fp8.dtype in (torch.uint8, torch.float8_e4m3fn), f"kv_nope_fp8 wrong dtype: {kv_nope_fp8.dtype}"
assert kv_nope_scale.dtype == torch.float32, f"kv_nope_scale wrong dtype: {kv_nope_scale.dtype}"
assert kv_rope_bf16.dtype == torch.bfloat16, f"kv_rope_bf16 wrong dtype: {kv_rope_bf16.dtype}"