Unverified50% confidenceFactExact time
In the same test, using the Marlin kernel to keep quantized weights while keeping activations in bf16 lowered Qwen FP8's KLD from 0.0039 to 0.0028, a reduction of about 30%.
1
Sources
50%
Confidence
Long-term
Relevance
11/7/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedvLLM's FP8 KV cache adds 0.0004 to 0.0006 of KLD at every tested checkpoint on Qwen3.8 27B, a very small value that nonetheless appears consistently on almost every prompt.72% similarUnverifiedIn tests on Qwen3.8 27B, even with activations kept in bf16, vLLM's NVFP4 checkpoint only matches UD-Q5_K_XL in KLD, while the latter is 8.7 GiB smaller in size.71% similarUnverifiedQ64's 74.32% accuracy is still below the native 27B baseline's 75.68%, and also below AnyJev Qwen3-32B L0's 76.68%.69% similarUnverifiedCompared to FP16 original precision, FP8 quantization typically causes no more than 1–2% performance degradation on mainstream benchmarks.66% similarUnverifiedAfter adopting a conservative fusion rule (outputting Support or Contradict only when the NLI and LLM judgments agree, and otherwise outputting Not Enough Evidence), Macro-F1 dropped to 0.7392 (Qwen3), 0.6730 (Llama 3.1) and 0.5991 (Mistral), all lower than the scores achieved when each was used alone.66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1146821API
curl https://kongchang.com/api/v1/knowledge/claims/1146821MCP
get_claim(id=1146821)