Historical tool output | archived output | JSONL line 6097 | 2026-09-08T01:29:21.212Z

=== speculator.py 100-160 ===
            self.use_fused_multi_step_decode = False
            return

        if not self.advance_draft_positions:
            self.use_fused_multi_step_decode = True
            return

        unsupported_backends = sorted(
            {
                attn_group.backend.get_name()
                for attn_groups in self.attn_groups
                for attn_group in attn_groups
                if not attn_group.supports_draft_decode_metadata_update
            }
        )
        self.use_fused_multi_step_decode = not unsupported_backends
        if unsupported_backends:
            logger.info_once(
                "Fused multi-step draft decode is not supported by attention "
                "backend(s) %s; falling back to rebuilding attention metadata "
                "between draft steps.",
                ", ".join(unsupported_backends),
            )

    def init_cudagraph_manager(self, cudagraph_mode: CUDAGraphMode) -> None:
        # Initialize cudagraph manager for draft prefill (draft position 0).
        self.prefill_cudagraph_manager = SpeculatorCudaGraphManager(
            self.vllm_config,
            self.device,
            cudagraph_mode,
            self.num_speculative_steps + 1,
        )

        # PIECEWISE cudagraphs are not supported for draft decodes.
        if cudagraph_mode.decode_mode() == CUDAGraphMode.FULL:
            cudagraph_mode = CUDAGraphMode.FULL_DECODE_ONLY
        else:
            cudagraph_mode = CUDAGraphMode.NONE

        # Initialize cudagraph manager for draft decodes (draft positions > 0).
        self.decode_cudagraph_manager = SpeculatorCudaGraphManager(
            self.vllm_config,
            self.device,
            cudagraph_mode,
            decode_query_len=1,
        )

    def capture(self) -> None:
        logger.info("Capturing model for speculator...")
        # Reset indices to zeros to prevent stale values from prior
        # dummy runs to cause out-of-bounds indexing during capture.
        self.last_token_indices.zero_()
        self.idx_mapping.zero_()

        # Capture the prefill routine (model forward + compute_logits +
        # sample).
        # For FULL graphs, the entire routine is recorded as one graph.
        # For PIECEWISE, only the model's compiled regions are captured
        # and the rest (compute_logits, gumbel_sample) runs eagerly.
        # Draft prefill reuses the target model's attention metadata at
        # runtime, so capture builds its dummy metadata through the target
=== 哪个文件 ===
/usr/local/lib/python3.12/dist-packages/vllm/v1/worker/gpu/spec_decode/autoregressive/speculator.py
/usr/local/lib/python3.12/dist-packages/vllm/v1/worker/gpu/spec_decode/autoregressive/__pycache__/speculator.cpython-312.pyc