[Bugfix] Fix the max_seq_len limit of 16384 for DeepSeek models (#20322)

Signed-off-by: Wang Huaqiang <huaqiang.wang@intel.com>

tests/test_config.py

@@ -412,6 +412,8 @@ def test_load_config_pt_load_map_location(pt_load_map_location):
         ("BAAI/bge-reranker-base", None, 512, False),
         ("BAAI/bge-reranker-base", 256, 256, False),
         ("BAAI/bge-reranker-base", 513, 512, True),
+        ("deepseek-ai/DeepSeek-R1-Distill-Qwen-7B", None, 131072, False),
+        ("deepseek-ai/DeepSeek-R1-Distill-Qwen-7B", 131073, 131072, True),
     ])
 def test_get_and_verify_max_len(model_id, max_model_len, expected_max_len,
                                 should_raise):

vllm/config.py

@@ -1442,10 +1442,17 @@ class ModelConfig:
         return getattr(self.hf_config, "matryoshka_dimensions", None)
 
     def get_and_verify_max_len(self, max_model_len: int):
-        tokenizer_config = try_get_tokenizer_config(
-            self.tokenizer,
-            trust_remote_code=self.trust_remote_code,
-            revision=self.tokenizer_revision)
+        # For pooling models, the tokenizer's `model_max_length` is often a
+        # reliable source for the maximum sequence length. However, for
+        # generative models, this can be incorrect and unduly limit the
+        # context window (e.g., DeepSeek-R1). Therefore, we only consider
+        # tokenizer_config for pooling models.
+        tokenizer_config = None
+        if self.runner_type == "pooling":
+            tokenizer_config = try_get_tokenizer_config(
+                self.tokenizer,
+                trust_remote_code=self.trust_remote_code,
+                revision=self.tokenizer_revision)
         max_model_len = _get_and_verify_max_len(
             hf_config=self.hf_text_config,
             tokenizer_config=tokenizer_config,