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

vLLM issue #34694: BF16 NVFP4 Marlin produces garbled output on GPUs without native FP4 support

forum

Fields

id
sources/gotcha-vllm-issue-34694-nvfp4-bf16-marlin
URL
https://github.com/vllm-project/vllm/issues/34694
Publisher
vLLM GitHub issues
Kind
forum
Published
2026-02-17
Accessed
2026-10-03
notes
Root-caused to exponent widening in dequant_fp8_scales<nv_bfloat162, kFE4M3fn> in csrc/quantization/marlin/dequant.h; fix is PR #34577.

Notes

Root-caused to exponent widening in dequant_fp8_scales<nv_bfloat162, kFE4M3fn> in csrc/quantization/marlin/dequant.h; fix is PR #34577.

Cross-references

This record references no other records.