vllm.v1.worker.gpu.warmup ¶

def _reserved_block_count(
    num_tokens: int,
    kvcache_spec: KVCacheSpec,
    *,
    num_lookahead_tokens: int,
    max_model_len: int,
    max_encoder_len: int,
) -> int:
    """Number of blocks the scheduler would hold for a request of `num_tokens`.

    Warmup hand-builds its `SchedulerOutput`s, so it must reserve what
    `KVCacheManager.allocate_slots` reserves: the token range plus
    `num_lookahead_tokens`, where the speculator writes the KV of its drafts.
    """
    if isinstance(kvcache_spec, UniformTypeKVCacheSpecs):
        kvcache_spec = kvcache_spec.first_spec
    if isinstance(kvcache_spec, CircularBufferSpec):
        # Circular caches keep one physical ring block for the request lifetime.
        return 1
    if isinstance(kvcache_spec, CrossAttentionSpec):
        # Cross-attention blocks cover the encoder sequence only.
        return cdiv(max_encoder_len, kvcache_spec.block_size)
    num_speculative_blocks = 0
    if isinstance(kvcache_spec, MambaSpec):
        # MambaManager appends speculative running-state blocks in every cache
        # mode; align mode sizes from the uncapped, lookahead-free token range.
        num_speculative_blocks = kvcache_spec.num_speculative_blocks
        if kvcache_spec.mamba_cache_mode == "align":
            return cdiv(num_tokens, kvcache_spec.block_size) + num_speculative_blocks
    num_tokens = min(num_tokens + num_lookahead_tokens, max_model_len)
    return cdiv(num_tokens, kvcache_spec.block_size) + num_speculative_blocks

def _warmup_block_counter(
    model_runner: GPUModelRunner,
) -> Callable[[int, KVCacheSpec], int]:
    """Bind `_reserved_block_count` to `model_runner`'s reservation policy."""
    num_lookahead_tokens = model_runner.vllm_config.num_lookahead_tokens
    max_model_len = model_runner.max_model_len
    max_encoder_len = getattr(model_runner.model_state, "max_encoder_len", 0)

    def block_count(num_tokens: int, kvcache_spec: KVCacheSpec) -> int:
        return _reserved_block_count(
            num_tokens,
            kvcache_spec,
            num_lookahead_tokens=num_lookahead_tokens,
            max_model_len=max_model_len,
            max_encoder_len=max_encoder_len,
        )

    return block_count

def run_mixed_prefill_decode_warmup(
    model_runner: GPUModelRunner,
    worker_execute_model: Callable[[SchedulerOutput], Any],
    worker_sample_tokens: Callable[[GrammarOutput | None], Any],
    num_tokens: int,
    *,
    mixed_step_context: AbstractContextManager[object] | None = None,
    req_id_prefix: str = "_v2_mixed_warmup",
) -> bool:
    """Run a V2 mixed prefill+decode step through normal scheduler inputs."""
    if model_runner.is_pooling_model or model_runner.max_num_reqs < 2 or num_tokens < 3:
        return False

    decode_req_id = f"{req_id_prefix}_decode_"
    prefill_req_id = f"{req_id_prefix}_prefill_"
    decode_prompt_len = 2
    decode_scheduled_tokens = 1
    prefill_len = num_tokens - decode_scheduled_tokens
    decode_token_ids = list(range(decode_prompt_len))
    prefill_token_ids = list(range(prefill_len))

    kv_cache_groups = model_runner.kv_cache_config.kv_cache_groups
    num_kv_cache_groups = len(kv_cache_groups)
    block_count = _warmup_block_counter(model_runner)
    kv_cache_specs = [g.kv_cache_spec for g in kv_cache_groups]
    decode_prefill_block_counts = [
        block_count(decode_prompt_len, s) for s in kv_cache_specs
    ]
    decode_block_counts = [
        block_count(decode_prompt_len + decode_scheduled_tokens, s)
        for s in kv_cache_specs
    ]
    decode_block_deltas = [
        decode - prefill
        for decode, prefill in zip(decode_block_counts, decode_prefill_block_counts)
    ]
    prefill_block_counts = [block_count(prefill_len, s) for s in kv_cache_specs]
    required_blocks = sum(decode_block_counts) + sum(prefill_block_counts)
    if model_runner.kv_cache_config.num_blocks <= required_blocks:
        logger.warning(
            "Skipping V2 mixed prefill+decode warmup because only %d KV blocks "
            "are available for %d required warmup blocks.",
            model_runner.kv_cache_config.num_blocks,
            required_blocks,
        )
        return False

    next_block_id = 1

    def _alloc_blocks(num_blocks: int) -> list[int]:
        nonlocal next_block_id
        block_ids = list(range(next_block_id, next_block_id + num_blocks))
        next_block_id += num_blocks
        return block_ids

    sampling_params = SamplingParams(max_tokens=2, temperature=0.0)

    decode_prefill_output = SchedulerOutput.make_empty()
    decode_prefill_output.scheduled_new_reqs = [
        NewRequestData(
            req_id=decode_req_id,
            prompt_token_ids=decode_token_ids,
            mm_features=[],
            sampling_params=sampling_params,
            pooling_params=None,
            block_ids=tuple(_alloc_blocks(n) for n in decode_prefill_block_counts),
            num_computed_tokens=0,
            lora_request=None,
            prefill_token_ids=decode_token_ids,
        ),
    ]
    decode_prefill_output.num_scheduled_tokens = {
        decode_req_id: decode_prompt_len,
    }
    decode_prefill_output.total_num_scheduled_tokens = decode_prompt_len
    decode_prefill_output.num_common_prefix_blocks = [0] * num_kv_cache_groups

    decode_new_blocks = tuple(_alloc_blocks(n) for n in decode_block_deltas)
    cached_decode_req = CachedRequestData.make_empty()
    cached_decode_req.req_ids = [decode_req_id]
    cached_decode_req.num_computed_tokens = [decode_prompt_len]
    cached_decode_req.num_output_tokens = [1]
    cached_decode_req.new_block_ids = [
        decode_new_blocks if any(decode_block_deltas) else None
    ]

    mixed_output = SchedulerOutput.make_empty()
    mixed_output.scheduled_cached_reqs = cached_decode_req
    mixed_output.scheduled_new_reqs = [
        NewRequestData(
            req_id=prefill_req_id,
            prompt_token_ids=prefill_token_ids,
            mm_features=[],
            sampling_params=sampling_params,
            pooling_params=None,
            block_ids=tuple(_alloc_blocks(n) for n in prefill_block_counts),
            num_computed_tokens=0,
            lora_request=None,
            prefill_token_ids=prefill_token_ids,
        ),
    ]
    mixed_output.num_scheduled_tokens = {
        decode_req_id: decode_scheduled_tokens,
        prefill_req_id: prefill_len,
    }
    mixed_output.total_num_scheduled_tokens = num_tokens
    mixed_output.num_common_prefix_blocks = [0] * num_kv_cache_groups

    cleanup_output = SchedulerOutput.make_empty()
    cleanup_output.finished_req_ids = {decode_req_id, prefill_req_id}

    context = mixed_step_context or nullcontext()
    model_runner.kv_connector.set_disabled(True)
    try:
        worker_execute_model(decode_prefill_output)
        worker_sample_tokens(None)
        with context:
            worker_execute_model(mixed_output)
            worker_sample_tokens(None)
        worker_execute_model(cleanup_output)
    finally:
        model_runner.kv_connector.set_disabled(False)
    return True

@torch.inference_mode()
def warmup_kernels(
    model_runner: GPUModelRunner,
    worker_execute_model: Callable[[SchedulerOutput], Any],
    worker_sample_tokens: Callable[[GrammarOutput | None], Any],
) -> None:
    """Run scheduler-realistic prefill and decode steps to JIT compile kernels.

    We must call the provided worker's execute_model for pipeline parallel
    coordination.
    """
    # Adaptive costs are calibrated during capture, after this warmup. Exercise
    # fixed draft counts here, then restore the manager for capture and serving.
    adaptive_verification = model_runner.adaptive_verification
    model_runner.adaptive_verification = None
    rejection_sampler = model_runner.rejection_sampler
    adaptive_sampling = (
        rejection_sampler is not None and rejection_sampler.enable_adaptive_verification
    )
    if adaptive_sampling:
        assert rejection_sampler is not None
        rejection_sampler.enable_adaptive_verification = False
    try:
        _warmup_kernels(model_runner, worker_execute_model, worker_sample_tokens)
    finally:
        model_runner.adaptive_verification = adaptive_verification
        if adaptive_sampling:
            assert rejection_sampler is not None
            rejection_sampler.enable_adaptive_verification = True