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

[XPU][CPU] Enable the native path of DeepSeek (#4086)

Browse Source 
Co-authored-by: Zhang, Liangang <liangang.zhang@intel.com>
This commit is contained in:
Meng, Hengyu
2025-03-13 13:26:29 +08:00
committed by GitHub
parent c76040e31b
commit 71046fcd71
16 changed files with 501 additions and 223 deletions
Download Patch File Download Diff File Expand all files Collapse all files

20
python/sglang/srt/model_loader/loader.py
View File

@@ -108,15 +108,25 @@ def _get_quantization_config(
quant_config = get_quant_config(model_config, load_config)
major, minor = get_device_capability()
if major is not None and minor is not None:
assert 0 <= minor < 10
capability = major * 10 + minor
if capability < quant_config.get_min_capability():
if not hasattr(quant_config, "get_availability"):
# Update VLLM to support get_available
if major is not None and minor is not None:
assert 0 <= minor < 10
capability = major * 10 + minor
if capability < quant_config.get_min_capability():
raise ValueError(
f"The quantization method {model_config.quantization} "
"is not supported for the current GPU. "
f"Minimum capability: {quant_config.get_min_capability()}. "
f"Current capability: {capability}."
)
else:
if not quant_config.get_availability():
raise ValueError(
f"The quantization method {model_config.quantization} "
"is not supported for the current GPU. "
f"Minimum capability: {quant_config.get_min_capability()}. "
f"Current capability: {capability}."
f"Current capability: {major, minor}."
)
supported_dtypes = quant_config.get_supported_act_dtypes()
if model_config.dtype not in supported_dtypes: