vLLM issue #56770: compressed-tensors MXFP4 W4A16 is misclassified as W4A4 and dispatched to a W4A4 kernel on SM100+

Fields

id

sources/qgotcha-vllm-56770-mxfp4-w4a16-as-w4a4

URL

https://github.com/vllm-project/vllm/issues/56770

Publisher

vLLM GitHub issues

Kind

forum

Accessed

2026-10-03

notes

AIME-24/25, GSM8K, MATH-500 comparison of BF16 vs MXFP4 W4A16 on the W4A4 kernel vs MarlinMxfp4Kernel. The quality gap on the W4A4 dispatch (avg 70.58%) vs Marlin (78.16%) vs BF16 (82.36%) is the silent-corruption evidence. Fix proposed separately as vLLM PR #56803.

Notes

AIME-24/25, GSM8K, MATH-500 comparison of BF16 vs MXFP4 W4A16 on the W4A4 kernel vs MarlinMxfp4Kernel. The quality gap on the W4A4 dispatch (avg 70.58%) vs Marlin (78.16%) vs BF16 (82.36%) is the silent-corruption evidence. Fix proposed separately as vLLM PR #56803.

Cross-references

This record references no other records.