[Bugfix][V1] Read ModelState max_model_len from model config (#58149)

Signed-off-by: Chenglun Hu <chenglunhu@gmail.com>
Signed-off-by: mgoin <mgoin64@gmail.com>
Co-authored-by: mgoin <mgoin64@gmail.com>

tests/v1/worker/test_mamba_hybrid_model_state.py

@@ -21,7 +21,7 @@ from vllm.v1.worker.gpu.model_states.recoverssm import RecoverSSMState
 def test_prepare_attn_forwards_positions(monkeypatch: pytest.MonkeyPatch) -> None:
     state = object.__new__(MambaHybridModelState)
     state.vllm_config = SimpleNamespace(num_speculative_tokens=0)
-    state.max_model_len = 8192
+    state.model_config = SimpleNamespace(max_model_len=8192)
     state._align_mode = False
     state.recoverssm = None
 

vllm/v1/worker/gpu/model_states/interface.py

@@ -59,7 +59,6 @@ class ModelState(ABC):
         self.model = model
         self.device = device
 
-        self.max_model_len = self.model_config.max_model_len
         self.max_num_reqs = self.scheduler_config.max_num_seqs
         self.max_num_tokens = self.scheduler_config.max_num_batched_tokens
         self.inputs_embeds_size = self.model_config.get_inputs_embeds_size()
@@ -99,6 +98,11 @@ class ModelState(ABC):
                 ),
             )
 
+    @property
+    def max_model_len(self) -> int:
+        # Auto-fit can reduce the limit after model state initialization.
+        return self.model_config.max_model_len
+
     def get_supported_generation_tasks(self) -> tuple[GenerationTask, ...]:
         from vllm.model_executor.models.interfaces import (
             supports_realtime,