Skip to content

vllm.models.qwen3_5.xpu.specialized_decode.model_state

Qwen3.5 XPU batch semantics over the standard hybrid ModelState.

Classes:

Qwen3_5XPUModelState

Bases: MambaHybridModelState

Retain stock KV/SSM handling and plan only the optional MoE stage.

Source code in vllm/models/qwen3_5/xpu/specialized_decode/model_state.py
class Qwen3_5XPUModelState(MambaHybridModelState):
    """Retain stock KV/SSM handling and plan only the optional MoE stage."""

    def __init__(self, vllm_config, model, encoder_cache, device):
        super().__init__(vllm_config, model, encoder_cache, device)
        language_model = getattr(model, "language_model", model)
        self.decode_runtime = language_model._xpu_moe_decode_runtime

    def prepare_inputs(self, input_batch, req_states):
        inputs = super().prepare_inputs(input_batch, req_states)
        self.decode_runtime.prepare_step(input_batch)
        return inputs

    def prepare_dummy_inputs(self, num_reqs, num_tokens):
        inputs = super().prepare_dummy_inputs(num_reqs, num_tokens)
        self.decode_runtime.prepare_capture(num_reqs, num_tokens)
        return inputs