Fix various config related issues for Transformers v5 (#37681)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>

tests/entrypoints/openai/chat_completion/test_completion_with_function_calling.py

@@ -231,13 +231,14 @@ def k2_server():
         "--gpu-memory-utilization",
         "0.4",
     ] + ROCM_EXTRA_ARGS
-    # hack to test kimi_k2 tool use tool_id format.
-    # avoid error in is_deepseek_mla check by setting kv_lora_rank=null
+    # Test kimi_k2 tool use tool_id format by overriding model_type.
+    # is_deepseek_mla safely returns False via getattr when kv_lora_rank
+    # is absent from the underlying config.
     with RemoteOpenAIServer(
         MODEL_NAME,
         args,
         env_dict=ROCM_ENV_OVERRIDES,
-        override_hf_configs={"model_type": "kimi_k2", "kv_lora_rank": None},
+        override_hf_configs={"model_type": "kimi_k2"},
     ) as remote_server:
         yield remote_server
 

vllm/transformers_utils/configs/colpali.py

@@ -27,7 +27,6 @@ class ColPaliConfig(PaliGemmaConfig):
         embedding_dim: int | None = None,
         embed_dim: int | None = None,
         dim: int | None = None,
-        projection_dim: int | None = None,
         colbert_dim: int | None = None,
         pooling: str | None = None,
         vlm_config: dict | None = None,
@@ -37,7 +36,6 @@ class ColPaliConfig(PaliGemmaConfig):
         self.embedding_dim = embedding_dim
         self.embed_dim = embed_dim
         self.dim = dim
-        self.projection_dim = projection_dim
         self.colbert_dim = colbert_dim
         self.pooling = pooling
 

vllm/transformers_utils/configs/deepseek_vl2.py

@@ -90,8 +90,6 @@ class MlpProjectorConfig(PretrainedConfig):
 class DeepseekVLV2Config(PretrainedConfig):
     model_type = "deepseek_vl_v2"
     architectures: list[str] | None = None
-    vision_config: VisionEncoderConfig
-    projector_config: MlpProjectorConfig
 
     tile_tag: str = "2D"
     global_view_pos: str = "head"

vllm/transformers_utils/configs/mistral.py

@@ -257,7 +257,6 @@ def _remap_mistral_audio_args(config: dict) -> dict:
             encoder_attention_heads=encoder_args["n_heads"],
             encoder_head_dim=encoder_args["head_dim"],
             vocab_size=encoder_args["vocab_size"],
-            max_source_positions=encoder_args["max_source_positions"],
             is_encoder_decoder=False,  # Override WhisperConfig default
             is_causal=encoder_args.get("causal", False),
             sliding_window=encoder_args.get("sliding_window", None),
@@ -270,6 +269,10 @@ def _remap_mistral_audio_args(config: dict) -> dict:
             max_position_embeddings=block_pool_size * config["max_position_embeddings"],
         ),
     }
+    # Sometimes max_source_positions is explicitly set to None in params.json but this
+    # is not a valid value for WhisperConfig (or downstream code that uses it).
+    if (max_source_positions := encoder_args.get("max_source_positions")) is not None:
+        config["audio_config"].max_source_positions = max_source_positions
     if quant_config:
         config["quantization_config"] = quant_config
     return config

vllm/transformers_utils/configs/parakeet.py

@@ -6,11 +6,21 @@ from transformers import ParakeetEncoderConfig, PretrainedConfig
 
 
 class ParakeetConfig(ParakeetEncoderConfig):
-    llm_hidden_size: int
-    projection_hidden_size: int
-    projection_bias: bool
-    projection_eps: float = 1e-5
-    sampling_rate: int
+    def __init__(
+        self,
+        llm_hidden_size: int,
+        projection_hidden_size: int,
+        projection_bias: bool,
+        sampling_rate: int,
+        projection_eps: float = 1e-5,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.llm_hidden_size = llm_hidden_size
+        self.projection_hidden_size = projection_hidden_size
+        self.projection_bias = projection_bias
+        self.sampling_rate = sampling_rate
+        self.projection_eps = projection_eps
 
     @staticmethod
     def from_hf_config(

vllm/transformers_utils/configs/qwen3_asr.py

@@ -408,7 +408,6 @@ class Qwen3ASRConfig(PretrainedConfig):
         support_languages=None,
         **kwargs,
     ):
-        super().__init__(**kwargs)
         if thinker_config is None:
             thinker_config = {}
             logger.info(
@@ -417,6 +416,7 @@ class Qwen3ASRConfig(PretrainedConfig):
 
         self.thinker_config = Qwen3ASRThinkerConfig(**thinker_config)
         self.support_languages = support_languages
+        super().__init__(**kwargs)
 
     def get_text_config(self, decoder=False) -> "PretrainedConfig":
         """

vllm/transformers_utils/configs/speculators/base.py

@@ -1,6 +1,7 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 import os
+from dataclasses import fields, is_dataclass
 from typing import Any
 
 from transformers import PretrainedConfig
@@ -15,11 +16,21 @@ class SpeculatorsConfig(PretrainedConfig):
     model_type = "speculators"
 
     def __init__(self, **kwargs):
-        """In Transformers v5, `PretrainedConfig` is decorated with `dataclass` and
-        `huggingface_hub.dataclasses.strict(accept_kwargs=True)`.
-        Inheriting classes do not inherit the `accept_kwargs=True` behaviour so we must
-        explicitly pass any kwargs to `PretrainedConfig.__init__`."""
-        super().__init__(**kwargs)
+        # Transformers v4 - super().__init__ which sets all kwargs as attributes
+        if not is_dataclass(PretrainedConfig):
+            return super().__init__(**kwargs)
+        # Transformers v5 - super().__init__ performs some validation before
+        # setting all kwargs as attributes, so we set them first to be safe
+        pre_trained_config_fields = {f.name for f in fields(PretrainedConfig)}
+        super_kwargs = dict()
+        for key, value in kwargs.items():
+            if key == "model_type":
+                continue  # model_type is set as a class variable, so skip it here
+            elif key in pre_trained_config_fields:
+                super_kwargs[key] = value
+            else:
+                setattr(self, key, value)
+        super().__init__(**super_kwargs)
 
     @classmethod
     def from_pretrained(

vllm/transformers_utils/model_arch_config_convertor.py

@@ -228,7 +228,7 @@ class ModelArchConfigConvertorBase:
             "pangu_ultra_moe_mtp",
             "bailing_hybrid",
         ):
-            return self.hf_text_config.kv_lora_rank is not None
+            return getattr(self.hf_text_config, "kv_lora_rank", None) is not None
         elif self.hf_text_config.model_type == "eagle":
             # if the model is an EAGLE module, check for the
             # underlying architecture
@@ -241,7 +241,7 @@ class ModelArchConfigConvertorBase:
                     "deepseek_v32",
                     "deepseek_mtp",
                 )
-                and self.hf_text_config.kv_lora_rank is not None
+                and getattr(self.hf_text_config, "kv_lora_rank", None) is not None
             )
         return False