[AMD] Add an opt-in AITER CK blockscale GEMM for gfx942 block-FP8 linears (#41071)

Signed-off-by: siliangchen-amd <SiLiang.Chen@amd.com
This commit is contained in:
siliangchen-amd
2026-10-10 01:31:20 -07:00
committed by GitHub
parent e35f60b9bc
commit ff19159e16
3 changed files with 11 additions and 0 deletions
+2
View File
@@ -978,6 +978,8 @@ class Envs:
SGLANG_AITER_MLA_VERIFY_BACKEND = EnvStr("asm")
# Let aiter plan the KV splits for the asm persistent MLA decode.
SGLANG_AITER_MLA_AUTO_KV_SPLITS = EnvBool(False)
# Run gfx942 block-FP8 linears with the AITER CK blockscale GEMM instead of Triton.
SGLANG_AITER_GFX942_BLOCKSCALE_USE_CK = EnvBool(False)
# DSV4 Aiter flags
SGLANG_OPT_USE_AITER_SILU_MUL = EnvBool(False)
@@ -45,6 +45,7 @@ from sglang.srt.utils import (
is_cuda,
is_flashinfer_available,
is_gfx95_supported,
is_gfx942_supported,
is_gfx1250_supported,
is_hip,
is_musa,
@@ -74,6 +75,7 @@ _use_aiter = (
get_bool_env_var("SGLANG_USE_AITER") and _is_hip and not _is_gfx1250_supported
)
_use_aiter_gfx95 = _use_aiter and _is_gfx95_supported
_use_aiter_gfx942 = _use_aiter and is_gfx942_supported()
# Conservative, not a tuned crossover: ptpc beat block-fp8 up to M=512 on
# MiniMax-M3 TP4 gfx950 shapes, but lost past M=64 on narrow ones (K=768).
MXFP8_DENSE_PTPC_DECODE_MAX_M = 128
@@ -1476,6 +1478,8 @@ def aiter_w8a8_block_fp8_linear(
use_triton = use_aiter_triton_gemm_w8a8_tuned_gfx950(n, k) or (
_ck_safe_m is not None and input_2d.shape[0] > _ck_safe_m
)
elif _use_aiter_gfx942:
use_triton = not envs.SGLANG_AITER_GFX942_BLOCKSCALE_USE_CK.get()
else:
use_triton = True