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

Fix marlin model loading compat with autogptq (#290)

Browse Source 
Co-authored-by: LRL <lrl@lbx.dev>
This commit is contained in:
Liurl
2024-03-13 13:15:43 +08:00
committed by GitHub
parent 92e2d74fd0
commit ed31579971
1 changed files with 9 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
python/sglang/srt/managers/router/model_runner.py
View File

@@ -300,9 +300,15 @@ class ModelRunner:
self.model_config.hf_config, "quantization_config", None self.model_config.hf_config, "quantization_config", None
) )
if hf_quant_config is not None: if hf_quant_config is not None:
quant_config_class = QUANTIONCONFIG_MAPPING.get( hf_quant_method = hf_quant_config["quant_method"]
hf_quant_config["quant_method"]
) # compat: autogptq uses is_marlin_format within quant config
if (hf_quant_method == "gptq"
and "is_marlin_format" in hf_quant_config
and hf_quant_config["is_marlin_format"]):
hf_quant_method = "marlin"
quant_config_class = QUANTIONCONFIG_MAPPING.get(hf_quant_method)
if quant_config_class is None: if quant_config_class is None:
raise ValueError( raise ValueError(
f"Unsupported quantization method: {hf_quant_config['quant_method']}" f"Unsupported quantization method: {hf_quant_config['quant_method']}"