Upgrade to vllm 0.17.0 corex v4.1 overlay

vllm/model_executor/models/qwen3_vl_moe.py

@@ -24,6 +24,7 @@
 # limitations under the License.
 """Inference-only Qwen3-VL-MoE model compatible with HuggingFace weights."""
 
+from platform import architecture
 import typing
 from collections.abc import Callable, Iterable
 from itertools import islice
@@ -45,6 +46,7 @@ from vllm.model_executor.model_loader.weight_utils import (
 )
 from vllm.multimodal import MULTIMODAL_REGISTRY
 from vllm.sequence import IntermediateTensors
+from vllm.tokenizers.registry import cached_tokenizer_from_config
 
 from .interfaces import MixtureOfExperts
 from .qwen3_moe import (
@@ -415,6 +417,7 @@ class Qwen3VLMoeForConditionalGeneration(
         multimodal_config = vllm_config.model_config.multimodal_config
 
         self.config = config
+        self._tokenizer = cached_tokenizer_from_config(vllm_config.model_config)
         self.multimodal_config = multimodal_config
         self.use_data_parallel = multimodal_config.mm_encoder_tp_mode == "data"
         self.video_pruning_rate = multimodal_config.video_pruning_rate
@@ -451,14 +454,14 @@ class Qwen3VLMoeForConditionalGeneration(
 
         with self._mark_language_model(vllm_config):
             self.language_model = Qwen3MoeLLMForCausalLM(
-                vllm_config=vllm_config.with_hf_config(config.text_config),
+                vllm_config=vllm_config.with_hf_config(config.text_config, architectures=["Qwen3MoeForCausalLM"]),
                 prefix=maybe_prefix(prefix, "language_model"),
             )
 
         if not get_pp_group().is_first_rank and hasattr(
             config.vision_config, "deepstack_visual_indexes"
         ):
-            assert self.language_model.start_layer >= len(
+            assert self.language_model.model.start_layer >= len(
                 config.vision_config.deepstack_visual_indexes
             ), (
                 "start_layer should be greater than or equal to "