NVFP4 support - #1

Open

sirish-gambhira wants to merge 14 commits into

mainfrom

users/sgambhira/nvfp4

Commits on Aug 5, 2026

nvfp4: wip
Sirish Gambhira
committed
9702f30
fix(loader): apply mxfp8 block scales when materializing fp8 checkpoints

Sirish Gambhira
and
claude
committed
3e01e1c
feat(gptq): accumulate the Hessian on quantized activations for w4a4

Sirish Gambhira
and
claude
committed
96cc706
feat(gptq): load NVFP4's static activation scale from an amax sidecar

Sirish Gambhira
and
claude
committed
16422dd

Commits on Aug 6, 2026

fix(nvfp4): add the two-level weight_scale_2 and match torchao's numerics

Sirish Gambhira
and
claude
committed
6466766
feat(scripts): minimax-m3 NVFP4 w4a4 pipeline -- amax sidecar, quantize driver, run.sh

Sirish Gambhira
and
claude
committed
3c476d2
fix(nvfp4): config hardening, NaN-loss RTN fallback, and test-placement repair

Sirish Gambhira
and
claude
committed
7bedb42
chore(scripts): p1b coverage run hardening from the calibration sessions

Sirish Gambhira
and
claude
committed
41b8b17
fix(dequant): route uint8 E8M0 scales in dequantize_fp8 through block dequant

Sirish Gambhira
and
claude
committed
8323b9a
fix(gptq): resolve activation amax keys against NamedModule.full_name

Sirish Gambhira
and
claude
committed
805a04e
feat(scripts): source-format splice phase + smoke-test fixes for the w4a4 pipeline

Sirish Gambhira
and
claude
committed
6a9dda8
fix(nvfp4): two-level block scales in the GPTQ loop via a primed global scale

Sirish Gambhira
and
claude
committed
39ad168
fix(looper): dynamic per-module fallback overrides must win over loop() baseline

Sirish Gambhira
and
claude
committed
e8c5760
fix(nvfp4): remove the single-level scale fallback -- unprimed find_params raises

Sirish Gambhira
and
claude
committed
6e03ec0