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

[AMD] Remove vllm's scaled_fp8_quant and moe_sum when SGLANG_USE_AITER=1 (#7484)

Browse Source
This commit is contained in:
Hubert Lu
2025-07-21 17:33:19 -07:00
committed by GitHub
parent 69adc4f81c
commit e50109f2ed
8 changed files with 156 additions and 69 deletions
Download Patch File Download Diff File Expand all files Collapse all files

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

@@ -12,7 +12,7 @@ import torch
from sglang.srt.layers.quantization.fp8_kernel import scaled_fp8_quant
from sglang.srt.layers.quantization.scalar_type import ScalarType, scalar_types
from sglang.srt.utils import cpu_has_amx_support, is_cpu, is_cuda, is_npu
from sglang.srt.utils import cpu_has_amx_support, is_cpu, is_cuda, is_hip, is_npu
if TYPE_CHECKING:
from sglang.srt.layers.quantization.base_config import QuantizationConfig
@@ -21,8 +21,9 @@ _is_cuda = is_cuda()
_is_npu = is_npu()
_is_cpu_amx_available = cpu_has_amx_support()
_is_cpu = is_cpu()
_is_hip = is_hip()
if not (_is_cuda or _is_npu or (_is_cpu and _is_cpu_amx_available)):
if not (_is_cuda or _is_npu or (_is_cpu and _is_cpu_amx_available) or _is_hip):
from vllm._custom_ops import scaled_fp8_quant