[cuda backend] int5 quantization support#20757

Merged

Gasoonjia merged 20 commits into

mainfrom

Jul 9, 2026

Commits on Jun 28, 2026

[cuda] Compact int4/int6 weight quant metadata (bf16 -> uint8 + per-row super-scale)
Gasoonjia
committed

Commits on Jun 30, 2026

Merge branch 'main' into cuda-int4-int6-metadata-opt
Gasoonjia
authored

Commits on Jul 1, 2026

Commits on Jul 2, 2026

[cuda][int4] per-256 fp16 scale-step with super-block-cooperative decode
Gasoonjia
committed
[cuda][int6] per-256 fp16 scale step + super-block code-load-hoist decode
Gasoonjia
committed

Commits on Jul 6, 2026

Commits on Jul 7, 2026

Merge branch 'main' into cuda-int4-int6-metadata-opt
Gasoonjia
authored

Commits on Jul 9, 2026