A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 16, 07:00 UTC
[core][misc] keep compatibility for old-style classes (#10356)
Signed-off-by: youkaichao <youkaichao@gmail.com>
youkaichao committed
Nov 15, 2024 at 05:55 UTC
3a763ba0c3a92fdde78e855ded94f9ff29e02088
1 file changed
+28
-12
vllm/model_executor/model_loader/loader.py
+28
-12
@@ -94,18 +94,34 @@ def _initialize_model(vllm_config: VllmConfig, prefix: str = "") -> nn.Module:
94
model_config = vllm_config.model_config
95
model_class, _ = get_model_architecture(model_config)
96
signatures = inspect.signature(model_class.__init__)
97
- # collect all kw-only parameters
98
- kw_only_params = [
99
- param.name for param in signatures.parameters.values()
100
- if param.kind == inspect.Parameter.KEYWORD_ONLY
101
- ]
102
- assert "vllm_config" in kw_only_params and "prefix" in kw_only_params, \
103
- ("vLLM model class must accept `vllm_config` and `prefix` as kw-only "
104
- "arguments. Possibly you have an old-style model class registered from "
105
- "out of tree and it is used for new vLLM version. "
106
- "Please check https://docs.vllm.ai/en/latest/design/class_hierarchy.html "
107
- "for the design and update the model class accordingly.")
108
- return model_class(vllm_config=vllm_config, prefix=prefix)
97
+ all_params = [param.name for param in signatures.parameters.values()]
98
+ if "vllm_config" in all_params and "prefix" in all_params:
99
+ # new-style model class
100
+ return model_class(vllm_config=vllm_config, prefix=prefix)
101
+ msg = ("vLLM model class should accept `vllm_config` and `prefix` as "
102
+ "input arguments. Possibly you have an old-style model class"
103
+ " registered from out of tree and it is used for new vLLM version. "
104
+ "Check https://docs.vllm.ai/en/latest/design/class_hierarchy.html "
105
+ "for the design and update the model class accordingly.")
106
+ logger.warning(msg)
107
+ logger.warning(
108
+ "Trying to guess the arguments for old-style model class %s",
109
+ model_class)
110
+ # try to be compatible with old-style model class
111
+ kwargs = {}
112
+ if "prefix" in all_params:
113
+ kwargs["prefix"] = prefix
114
+ if "config" in all_params:
115
+ kwargs["config"] = model_config.hf_config
116
+ if "cache_config" in all_params:
117
+ kwargs["cache_config"] = vllm_config.cache_config
118
+ if "quant_config" in all_params:
119
+ kwargs["quant_config"] = vllm_config.quant_config
120
+ if "lora_config" in all_params:
121
+ kwargs["lora_config"] = vllm_config.lora_config
122
+ if "scheduler_config" in all_params:
123
+ kwargs["scheduler_config"] = vllm_config.scheduler_config
124
+ return model_class(**kwargs)
125
126
127
class BaseModelLoader(ABC):