Repository navigation

Add a bfloat16 option for the GPTQ and local-Hessian Hessian GEMMs - #3019

Merged

kevalmorabia97 merged 4 commits into

mainfrom

chenjiel/gptq-bf16-hessian

Oct 27, 2026

Merged

Add a bfloat16 option for the GPTQ and local-Hessian Hessian GEMMs#3019
kevalmorabia97 merged 4 commits into
mainfrom
chenjiel/gptq-bf16-hessian

Commits on Oct 27, 2026
- Accumulate GPTQ Hessians from BF16/FP16 inputs with a float32-output GEMM
  
  cjluo-nv
  and
  claude
  committed
  e9c1f4a
- Make the bfloat16 Hessian GEMM opt-in with hessian_input_dtype
  
  cjluo-nv
  and
  claude
  committed
  7a35f3f
- Offer hessian_input_dtype for local-Hessian calibration too
  
  cjluo-nv
  and
  claude
  committed
  bf0cec2
- Test the bfloat16 Hessian fallback on CPU
  
  cjluo-nv
  and
  claude
  committed
  eabd003