File size: 2,497 Bytes
93e76f3 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 | self._handle_gpu_memory_settings(gpu_mem)
File "/mnt/data/yiran/sglang/python/sglang/srt/server_args.py", line 1011, in _handle_gpu_memory_settings
if not self.use_mla_backend():
^^^^^^^^^^^^^^^^^^^^^^
File "/mnt/data/yiran/sglang/python/sglang/srt/server_args.py", line 5124, in use_mla_backend
from sglang.srt.configs.model_config import AttentionArch
File "/mnt/data/yiran/sglang/python/sglang/srt/configs/model_config.py", line 27, in <module>
from sglang.srt.layers.quantization import QUANTIZATION_METHODS
File "/mnt/data/yiran/sglang/python/sglang/srt/layers/quantization/__init__.py", line 19, in <module>
from sglang.srt.layers.quantization.auto_round import AutoRoundConfig
File "/mnt/data/yiran/sglang/python/sglang/srt/layers/quantization/auto_round.py", line 12, in <module>
from sglang.srt.layers.quantization.utils import get_scalar_types
File "/mnt/data/yiran/sglang/python/sglang/srt/layers/quantization/utils.py", line 13, in <module>
from sglang.srt.layers.quantization.fp8_kernel import scaled_fp8_quant
File "/mnt/data/yiran/sglang/python/sglang/srt/layers/quantization/fp8_kernel.py", line 50, in <module>
from sgl_kernel import sgl_per_token_quant_fp8
File "/mnt/data/miniconda3/envs/slime/lib/python3.12/site-packages/sgl_kernel/__init__.py", line 5, in <module>
common_ops = _load_architecture_specific_ops()
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/mnt/data/miniconda3/envs/slime/lib/python3.12/site-packages/sgl_kernel/load_utils.py", line 188, in _load_architecture_specific_ops
raise ImportError(error_msg)
ImportError:
[sgl_kernel] CRITICAL: Could not load any common_ops library!
Attempted locations:
1. Architecture-specific pattern: /mnt/data/miniconda3/envs/slime/lib/python3.12/site-packages/sgl_kernel/sm100/common_ops.* - found files: ['/mnt/data/miniconda3/envs/slime/lib/python3.12/site-packages/sgl_kernel/sm100/common_ops.abi3.so']
2. Fallback pattern: /mnt/data/miniconda3/envs/slime/lib/python3.12/site-packages/sgl_kernel/common_ops.* - found files: []
3. Standard Python import: common_ops - failed
GPU Info:
- Compute capability: 80
- Expected variant: SM80 (precise math for compatibility)
Please ensure sgl_kernel is properly installed with:
pip install --upgrade sgl_kernel
Error details from previous import attempts:
- ImportError: libnuma.so.1: cannot open shared object file: No such file or directory
- ModuleNotFoundError: No module named 'common_ops' |