(
self,
model: str,
tokenizer: str,
tokenizer_mode: str,
trust_remote_code: bool,
download_dir: Optional[str],
load_format: str,
dtype: Union[str, torch.dtype],
seed: int,
attention_backend: str,
revision: Optional[str] = None,
tokenizer_revision: Optional[str] = None,
max_model_len: Optional[int] = None,
quantization: Optional[str] = None,
)
| 52 | """ |
| 53 | |
| 54 | def __init__( |
| 55 | self, |
| 56 | model: str, |
| 57 | tokenizer: str, |
| 58 | tokenizer_mode: str, |
| 59 | trust_remote_code: bool, |
| 60 | download_dir: Optional[str], |
| 61 | load_format: str, |
| 62 | dtype: Union[str, torch.dtype], |
| 63 | seed: int, |
| 64 | attention_backend: str, |
| 65 | revision: Optional[str] = None, |
| 66 | tokenizer_revision: Optional[str] = None, |
| 67 | max_model_len: Optional[int] = None, |
| 68 | quantization: Optional[str] = None, |
| 69 | ) -> None: |
| 70 | self.model = model |
| 71 | self.tokenizer = tokenizer |
| 72 | self.tokenizer_mode = tokenizer_mode |
| 73 | self.trust_remote_code = trust_remote_code |
| 74 | self.download_dir = download_dir |
| 75 | self.load_format = load_format |
| 76 | self.seed = seed |
| 77 | self.revision = revision |
| 78 | self.tokenizer_revision = tokenizer_revision |
| 79 | self.quantization = quantization |
| 80 | self.attention_backend = attention_backend |
| 81 | |
| 82 | if os.environ.get("VLLM_USE_MODELSCOPE", "False").lower() == "true": |
| 83 | # download model from ModelScope hub, |
| 84 | # lazy import so that modelscope is not required for normal use. |
| 85 | from modelscope.hub.snapshot_download import snapshot_download # pylint: disable=C |
| 86 | model_path = snapshot_download(model_id=model, |
| 87 | cache_dir=download_dir, |
| 88 | revision=revision) |
| 89 | self.model = model_path |
| 90 | self.download_dir = model_path |
| 91 | self.tokenizer = model_path |
| 92 | |
| 93 | self.hf_config = get_config(self.model, trust_remote_code, revision) |
| 94 | self.dtype = _get_and_verify_dtype(self.hf_config, dtype) |
| 95 | self.max_model_len = _get_and_verify_max_len(self.hf_config, |
| 96 | max_model_len) |
| 97 | self._verify_load_format() |
| 98 | self._verify_tokenizer_mode() |
| 99 | self._verify_quantization() |
| 100 | |
| 101 | def _verify_load_format(self) -> None: |
| 102 | load_format = self.load_format.lower() |
nothing calls this directly
no test coverage detected