[Bugfix] Fix mamba cache dtype for Qwen3.5 (#34200)

Signed-off-by: Roger Wang <hey@rogerw.io>
2026-02-10 13:12:31 -08:00
parent 4293c00b84
commit 9615575afc
1 changed files with 2 additions and 1 deletions
--- a/vllm/model_executor/models/qwen3_5.py
+++ b/vllm/model_executor/models/qwen3_5.py
@@ -870,8 +870,9 @@ class Qwen3_5ForConditionalGeneration(Qwen3VLForConditionalGeneration, IsHybrid)
        cls,
        vllm_config: "VllmConfig",
    ) -> tuple[torch.dtype, torch.dtype]:
+        mamba_ssm_dtype = vllm_config.model_config.hf_text_config.mamba_ssm_dtype
        return MambaStateDtypeCalculator.gated_delta_net_state_dtype(
-            vllm_config.model_config.dtype, vllm_config.cache_config.mamba_cache_dtype
+            vllm_config.model_config.dtype, mamba_ssm_dtype
        )

    @classmethod