[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>
hcl committed
Sep 23, 2026 at 03:45 UTC
71a239284507e4f9477f8256e04fd28d1aa9a2db
2 files changed
+6
-2
tests/v1/worker/test_mamba_hybrid_model_state.py
+1
-1
@@ -21,7 +21,7 @@ from vllm.v1.worker.gpu.model_states.recoverssm import RecoverSSMState
21
def test_prepare_attn_forwards_positions(monkeypatch: pytest.MonkeyPatch) -> None:
22
state = object.__new__(MambaHybridModelState)
23
state.vllm_config = SimpleNamespace(num_speculative_tokens=0)
24
- state.max_model_len = 8192
24
+ state.model_config = SimpleNamespace(max_model_len=8192)
25
state._align_mode = False
26
state.recoverssm = None
27
vllm/v1/worker/gpu/model_states/interface.py
+5
-1
@@ -59,7 +59,6 @@ class ModelState(ABC):
59
self.model = model
60
self.device = device
61
62
- self.max_model_len = self.model_config.max_model_len
62
self.max_num_reqs = self.scheduler_config.max_num_seqs
63
self.max_num_tokens = self.scheduler_config.max_num_batched_tokens
64
self.inputs_embeds_size = self.model_config.get_inputs_embeds_size()
@@ -99,6 +98,11 @@ class ModelState(ABC):
98
),
99
)
100
101
+ @property
102
+ def max_model_len(self) -> int:
103
+ # Auto-fit can reduce the limit after model state initialization.
104
+ return self.model_config.max_model_len
105
+
106
def get_supported_generation_tasks(self) -> tuple[GenerationTask, ...]:
107
from vllm.model_executor.models.interfaces import (
108
supports_realtime,