EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

[Dockerfile] Multi-arch support for ROCm (#7902)

Browse Source 
Co-authored-by: Lin, Soga <soga.lin@amd.com>
Co-authored-by: HaiShaw <hixiao@gmail.com>
This commit is contained in:
mqhc2020
2025-07-14 14:13:09 +08:00
committed by GitHub
parent cb736df854
commit a562c8a35c
3 changed files with 105 additions and 46 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
python/pyproject.toml
View File

@@ -79,7 +79,6 @@ blackwell = [
srt_hip = [
"sglang[runtime_common]",
"torch",
"vllm==0.6.7.dev2",
]
# xpu is not enabled in public vllm and torch whl,

4
python/sglang/srt/layers/quantization/fp8_utils.py
View File

@@ -42,7 +42,7 @@ _use_aiter = get_bool_env_var("SGLANG_USE_AITER") and _is_hip
if _use_aiter:
import aiter
from aiter import gemm_a8w8_blockscale_CK, get_hip_quant
from aiter import gemm_a8w8_blockscale, get_hip_quant
aiter_per1x128_quant = get_hip_quant(aiter.QuantType.per_1x128)
@@ -274,7 +274,7 @@ def aiter_w8a8_block_fp8_linear(
output_shape = [*input.shape[:-1], weight.shape[0]]
q_input, x_scale = aiter_per1x128_quant(input_2d, quant_dtype=aiter.dtypes.fp8)
output = gemm_a8w8_blockscale_CK(
output = gemm_a8w8_blockscale(
q_input, weight, x_scale, weight_scale, dtype=input.dtype
)