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

DEBUG: fix VERBOSE reference in shared_expert, always print L2 gsa debug

Browse Source
This commit is contained in:
biondizzle
2026-06-04 00:15:38 +00:00
parent f0f87df906
commit 993bb345d1
1 changed files with 1 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
dsv4/layers/shared_expert.py
View File

@@ -368,8 +368,7 @@ class Nvfp4SharedExpert:
x_fp4, x_sf, gsa_l2_gpu = quantize_nvfp4_gpu_fused(intermediate)
# DEBUG: verify no CUDA errors from quantize kernel
torch.cuda.synchronize() # DEBUG: catch async errors
if VERBOSE >= 1:
print(f" SE L2 gsa: gsa_gpu shape={tuple(gsa_l2_gpu.shape)} dtype={gsa_l2_gpu.dtype} dev={gsa_l2_gpu.device} _l2_gsa_buf shape={tuple(self._l2_gsa_buf.shape)} dev={self._l2_gsa_buf.device}", flush=True)
print(f" SE L2 gsa: gsa_gpu shape={tuple(gsa_l2_gpu.shape)} dtype={gsa_l2_gpu.dtype} dev={gsa_l2_gpu.device} _l2_gsa_buf shape={tuple(self._l2_gsa_buf.shape)} dev={self._l2_gsa_buf.device}", flush=True)
self._l2_gsa_buf.copy_(gsa_l2_gpu[:1].reshape(1)) # GPU → GPU, no sync
else:
x_fp4, x_sf = quantize_activation_nvfp4(