Support download models from www.modelscope.cn (#1588)
This commit is contained in:
+13
-1
@@ -1,4 +1,5 @@
|
||||
from typing import Optional, Union
|
||||
import os
|
||||
|
||||
import torch
|
||||
from transformers import PretrainedConfig
|
||||
@@ -76,7 +77,18 @@ class ModelConfig:
|
||||
self.tokenizer_revision = tokenizer_revision
|
||||
self.quantization = quantization
|
||||
|
||||
self.hf_config = get_config(model, trust_remote_code, revision)
|
||||
if os.environ.get("VLLM_USE_MODELSCOPE", "False").lower() == "true":
|
||||
# download model from ModelScope hub,
|
||||
# lazy import so that modelscope is not required for normal use.
|
||||
from modelscope.hub.snapshot_download import snapshot_download # pylint: disable=C
|
||||
model_path = snapshot_download(model_id=model,
|
||||
cache_dir=download_dir,
|
||||
revision=revision)
|
||||
self.model = model_path
|
||||
self.download_dir = model_path
|
||||
self.tokenizer = model_path
|
||||
|
||||
self.hf_config = get_config(self.model, trust_remote_code, revision)
|
||||
self.dtype = _get_and_verify_dtype(self.hf_config, dtype)
|
||||
self.max_model_len = _get_and_verify_max_len(self.hf_config,
|
||||
max_model_len)
|
||||
|
||||
@@ -648,9 +648,10 @@ if __name__ == "__main__":
|
||||
max_model_len = engine_model_config.max_model_len
|
||||
|
||||
# A separate tokenizer to map token IDs to strings.
|
||||
tokenizer = get_tokenizer(engine_args.tokenizer,
|
||||
tokenizer_mode=engine_args.tokenizer_mode,
|
||||
trust_remote_code=engine_args.trust_remote_code)
|
||||
tokenizer = get_tokenizer(
|
||||
engine_model_config.tokenizer,
|
||||
tokenizer_mode=engine_model_config.tokenizer_mode,
|
||||
trust_remote_code=engine_model_config.trust_remote_code)
|
||||
|
||||
uvicorn.run(app,
|
||||
host=args.host,
|
||||
|
||||
Reference in New Issue
Block a user