(
self,
args,
)
| 197 | """ |
| 198 | |
| 199 | def __init__( |
| 200 | self, |
| 201 | args, |
| 202 | ): |
| 203 | self.model = "" |
| 204 | self.is_quantized = False |
| 205 | self.is_moe_quantized = False |
| 206 | self.max_model_len = 0 |
| 207 | self.dtype = "bfloat16" |
| 208 | self.enable_logprob = False |
| 209 | self.max_logprobs = 20 |
| 210 | self.logprobs_mode = "raw_logprobs" |
| 211 | self.enable_keep_sampling_mask = False |
| 212 | self.redundant_experts_num = 0 |
| 213 | self.seed = 0 |
| 214 | self.quantization = None |
| 215 | self.pad_token_id: int = -1 |
| 216 | self.eos_tokens_lens: int = 2 |
| 217 | self.lm_head_fp32: bool = False |
| 218 | self.model_format = "auto" |
| 219 | self.runner = "auto" |
| 220 | self.convert = "auto" |
| 221 | self.pooler_config: Optional["PoolerConfig"] = field(init=False) |
| 222 | self.override_pooler_config: Optional[Union[dict, "PoolerConfig"]] = None |
| 223 | self.revision = None |
| 224 | self.prefix_layer_name = "layers" |
| 225 | self.kv_cache_quant_scale_path = "" |
| 226 | self.enable_entropy = False |
| 227 | self.model_impl: ModelImpl = "auto" |
| 228 | |
| 229 | self.partial_rotary_factor: float = 1.0 |
| 230 | self.num_nextn_predict_layers = 0 |
| 231 | self.mm_max_tokens_per_item = None |
| 232 | for key, value in args.items(): |
| 233 | if hasattr(self, key) and value != "None": |
| 234 | setattr(self, key, value) |
| 235 | |
| 236 | assert self.model != "" |
| 237 | pretrained_config, _ = PretrainedConfig.get_config_dict(self.model) |
| 238 | self.pretrained_config = PretrainedConfig.from_dict(pretrained_config) |
| 239 | |
| 240 | # Some exported configs (e.g. Qwen3-VL) embed the text model's configuration under a `text_config` key. |
| 241 | if "text_config" in pretrained_config and isinstance(pretrained_config["text_config"], dict): |
| 242 | text_fg = pretrained_config.pop("text_config") |
| 243 | for key, value in text_fg.items(): |
| 244 | if not hasattr(self, key): |
| 245 | setattr(self, key, value) |
| 246 | |
| 247 | # set attribute from pretrained_config |
| 248 | for key, value in pretrained_config.items(): |
| 249 | setattr(self, key, value) |
| 250 | # we need set default value when not exist |
| 251 | for key, value in PRETRAINED_INIT_CONFIGURATION.items(): |
| 252 | if not hasattr(self, key): |
| 253 | setattr(self, key, value) |
| 254 | |
| 255 | if not hasattr(self, "head_dim"): |
| 256 | self.head_dim = self.hidden_size // self.num_attention_heads |
nothing calls this directly
no test coverage detected