[3/N] model runner pass the whole config to model (#9958)

Signed-off-by: youkaichao <youkaichao@gmail.com>
This commit is contained in:
youkaichao
2024-11-02 12:08:49 -07:00
committed by GitHub
parent 74b529ceee
commit cea808f325
9 changed files with 87 additions and 140 deletions
+4 -5
View File
@@ -248,11 +248,10 @@ def llama_2_7b_engine_extra_embeddings():
cleanup_dist_env_and_memory(shutdown_ray=True)
get_model_old = get_model
def get_model_patched(*, model_config, device_config, **kwargs):
kwargs["lora_config"] = LoRAConfig(max_loras=4, max_lora_rank=8)
return get_model_old(model_config=model_config,
device_config=device_config,
**kwargs)
def get_model_patched(**kwargs):
kwargs["vllm_config"].lora_config = LoRAConfig(max_loras=4,
max_lora_rank=8)
return get_model_old(**kwargs)
with patch("vllm.worker.model_runner.get_model", get_model_patched):
engine = vllm.LLM("meta-llama/Llama-2-7b-hf", enable_lora=False)