[JIT] Trait-driven per_token_group_quant: unify the quant kernel family (flat + masked) - #30924

Merged

BBuf merged 6 commits into

sgl-project:mainfrom

DarkSharpness:per-token-group-quant-v3

Jul 22, 2026

Commits on Jul 21, 2026

[JIT] Add trait-driven per_token_group_quant_v3 (flat + masked)

claude
authored and
DarkSharpness
committed
c31d32c
[JIT] Fix per_token_group_quant_v3 Hopper regression with fixed 32B/lane tiling

claude
authored and
DarkSharpness
committed
f6a2608
[JIT] Route all per-token-group quant call sites through v3 (RFC #29630)

claude
authored and
DarkSharpness
committed
a08ae1d
[MoE] Unify masked silu-quant Triton wrappers; clarify ue8m0 layout note

claude
authored and
DarkSharpness
committed
0deab89
[JIT] Drop the v3 suffix: the kernel is per_token_group_quant

claude
authored and
DarkSharpness
committed
3ccaf26
[JIT] Keep fp32-pow-2 row-major UE8M0 on the v2 baseline in the srt entry

DarkSharpness
and
claude
committed
1f14b6e