Skip to content

vllm.v1.worker.gpu.spec_decode.autoregressive.cudagraph_utils

DecodeSpeculatorCudaGraphManager

Bases: CudaGraphManager

CudaGraphManager for draft decode, building its own attention metadata.

Source code in vllm/v1/worker/gpu/spec_decode/autoregressive/cudagraph_utils.py 
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
class DecodeSpeculatorCudaGraphManager(CudaGraphManager):
    """CudaGraphManager for draft decode, building its own attention metadata."""

    def capture(
        self,
        forward_fn: Callable,
        model_state: ModelState,
        input_buffers: InputBuffers,
        block_tables: BlockTables,
        attn_groups: list[list[AttentionGroup]],
        kv_cache_config: KVCacheConfig,
        progress_bar_desc: str = "Capturing CUDA graphs",
    ) -> None:
        def create_forward_fn(
            desc: BatchExecutionDescriptor,
            warmup: bool,
        ) -> tuple[Callable[[CUDAGraphMode], None], AttentionState]:
            num_tokens = desc.num_tokens
            num_reqs = desc.num_reqs or min(num_tokens, self.max_num_reqs)
            num_tokens_across_dp = (
                torch.full((self.dp_size,), num_tokens, dtype=torch.int32, device="cpu")
                if self.dp_size > 1
                else None
            )
            attn_state = prepare_inputs_to_capture(
                num_reqs,
                num_tokens,
                model_state,
                input_buffers,
                block_tables,
                attn_groups,
                kv_cache_config,
                skip_attn=(desc.cg_mode == CUDAGraphMode.PIECEWISE),
            )
            attn_metadata, slot_mappings = attn_state

            fwd = lambda cg_mode: forward_fn(
                num_reqs,
                num_tokens,
                attn_metadata,
                slot_mappings,
                num_tokens_across_dp,
                cg_mode,
            )
            return fwd, attn_state

        super().capture(create_forward_fn, progress_bar_desc)

PrefillSpeculatorCudaGraphManager

Bases: CudaGraphManager

CudaGraphManager for draft prefill, using pre-built attention states from the target model's capture.

Source code in vllm/v1/worker/gpu/spec_decode/autoregressive/cudagraph_utils.py 
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
class PrefillSpeculatorCudaGraphManager(CudaGraphManager):
    """CudaGraphManager for draft prefill, using pre-built attention states
    from the target model's capture."""

    def capture(
        self,
        forward_fn: Callable,
        attn_states: dict[BatchExecutionDescriptor, AttentionStatePair],
        progress_bar_desc: str = "Capturing CUDA graphs",
    ) -> None:
        def create_forward_fn(
            desc: BatchExecutionDescriptor,
            warmup: bool,
        ) -> tuple[Callable[[CUDAGraphMode], None], AttentionState]:
            num_tokens = desc.num_tokens
            num_reqs = desc.num_reqs or min(num_tokens, self.max_num_reqs)
            num_tokens_across_dp = (
                torch.full((self.dp_size,), num_tokens, dtype=torch.int32, device="cpu")
                if self.dp_size > 1
                else None
            )
            attn_state_pair = attn_states[desc]
            attn_state = attn_state_pair.warmup if warmup else attn_state_pair.captured
            attn_metadata, slot_mappings = attn_state
            fwd = lambda cg_mode: forward_fn(
                num_reqs,
                num_tokens,
                attn_metadata,
                slot_mappings,
                num_tokens_across_dp,
                cg_mode,
            )
            return fwd, attn_state

        super().capture(create_forward_fn, progress_bar_desc)