biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Bugfix] Fix vllm serve failure with Nemotron Nano V3 FP8 (#31960)

Browse Source 
Signed-off-by: Daniel Serebrenik <daserebrenik@nvidia.com>
This commit is contained in:
danisereb
2026-01-08 18:08:37 +02:00
committed by GitHub
parent eaba8ece77
commit b8112c1d85
1 changed files with 4 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
vllm/model_executor/layers/quantization/utils/fp8_utils.py
View File

@@ -1619,11 +1619,12 @@ def process_fp8_weight_tensor_strategy_moe(
"""Process moe weights for tensor-wise quantization strategy.""" """Process moe weights for tensor-wise quantization strategy."""
max_scales = weight_scales.max(dim=1).values max_scales = weight_scales.max(dim=1).values
# For w1 case (i.e. not w13): just collapse the last dim since # For w1 case (i.e. not w13): there is already just one scale per expert.
# there is already just one scale per expert in this case.
if not is_act_and_mul: if not is_act_and_mul:
assert weight_scales.shape[1] == 1 assert weight_scales.shape[1] == 1
return weight, weight_scales.max() # One scale per expert
assert max_scales.shape == (num_experts,)
return weight, max_scales
# For w13 case (common): require single scale for w13 per expert, but # For w13 case (common): require single scale for w13 per expert, but
# on disk there is a scale for w1 and w3. Use the max to requantize. # on disk there is a scale for w1 and w3. Use the max to requantize.