inference/research
verifiedhigh · 0.90updated 2026-10-03

vLLM issue #54974: modelopt NVFP4 MoE mismatched w1/w3 global scales are detected, warned about, and then used anyway

forum

Fields

id
sources/qgotcha-vllm-54974-nvfp4-w13-global-scale
Publisher
vLLM GitHub issues
Kind
forum
Accessed
2026-10-03
notes
Measured from checkpoints on disk with no GPU: GLM-5.3-Flash-ABLITERATED-NVFP4 has 8,563 of 12,384 gate/up global-scale pairs differing (69.1%), ratio median 1.077 / p99 1.69 / max 10.0. Two modelopt 0.46 checkpoints tie exactly. Fix proposed separately as vLLM PR #55073.

Notes

Measured from checkpoints on disk with no GPU: GLM-5.3-Flash-ABLITERATED-NVFP4 has 8,563 of 12,384 gate/up global-scale pairs differing (69.1%), ratio median 1.077 / p99 1.69 / max 10.0. Two modelopt 0.46 checkpoints tie exactly. Fix proposed separately as vLLM PR #55073.

Cross-references

This record references no other records.