chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[ROCM] Optimized deepseek-r1 fp8 model with + triton_gemm_a8w8 + batch_gemm_a8w8 + fused set_mla_kv_buffer kernel (#13617)

Browse Source 
Co-authored-by: root <root@smci355-ccs-aus-m12-17.cs-aus.dcgpu>
Co-authored-by: jacky.cheng <yichiche@amd.com>
This commit is contained in:
yctseng0211
2025-11-20 16:29:56 +08:00
committed by GitHub
parent 19729f723e
commit c8ede0e93c
2 changed files with 61 additions and 15 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
python/sglang/srt/layers/quantization/fp8_utils.py
View File

@@ -45,7 +45,10 @@ _use_aiter = get_bool_env_var("SGLANG_USE_AITER") and _is_hip
if _use_aiter:
import aiter
from aiter import gemm_a8w8_blockscale, gemm_a8w8_bpreshuffle, get_hip_quant
# from aiter import gemm_a8w8_blockscale, gemm_a8w8_bpreshuffle, get_hip_quant
from aiter import gemm_a8w8_bpreshuffle, get_hip_quant
from aiter.ops.triton.gemm_a8w8_blockscale import gemm_a8w8_blockscale
aiter_per1x128_quant = get_hip_quant(aiter.QuantType.per_1x128)