vllm.v1.worker.utils ¶

def add_kv_sharing_layers_to_kv_cache_groups(
    shared_kv_cache_layers: dict[str, str],
    kv_cache_groups: list[KVCacheGroupSpec],
    runner_only_attn_layers: set[str] | None = None,
) -> None:
    """
    Sets up KV cache sharing by reusing the allocated KV caches in `kv_caches`
    for layers that do not allocate its own KV cache, based on the mapping in
    `shared_kv_cache_layers`. Adds these layers to the corresponding KV cache
    group, which is needed to ensure that attention metadata is assigned later.

    Args:
        shared_kv_cache_layers: Layer pairings for cross-layer KV sharing.
            If an Attention layer `layer_name` is in the keys of this dict, it
            means this layer will perform attention using the keys and values
            from the KV cache of `shared_kv_cache_layers[layer_name]`.
        kv_cache_groups: The KV cache groups of the model.
    """
    layer_to_kv_cache_group: dict[str, KVCacheGroupSpec] = {}
    for kv_cache_group in kv_cache_groups:
        for layer_name in kv_cache_group.layer_names:
            layer_to_kv_cache_group[layer_name] = kv_cache_group

    for layer_name, target_layer_name in shared_kv_cache_layers.items():
        tgt_kv_cache_group = layer_to_kv_cache_group[target_layer_name]
        tgt_kv_cache_group.layer_names.append(layer_name)

        if runner_only_attn_layers is not None:
            runner_only_attn_layers.add(layer_name)

def bind_kv_cache(
    kv_caches: dict[str, torch.Tensor],
    forward_context: dict[str, Attention],
    runner_kv_caches: list[torch.Tensor],
    num_attn_module: int = 1,
) -> None:
    """
    Bind the allocated KV cache to both ModelRunner and forward context so
    that the KV cache can be used in the forward pass.

    This function:
      1) Fills the ModelRunner's kv cache list (`runner_kv_caches`) with
         kv_caches.
      2) Associates each attention layer in the `forward_context` with its
         corresponding KV cache in kv_caches.

    Args:
        kv_caches: The allocated kv_caches with layer names as keys.
        forward_context: The global forward context containing all Attention
            layers with layer names as keys.
        runner_kv_caches: The kv_cache declared by ModelRunner.
    """
    # Bind kv_caches to ModelRunner
    assert len(runner_kv_caches) == 0

    # Convert kv_caches dict to a list of tensors in the order of layer_index.
    index2name = defaultdict(list)
    for layer_name in kv_caches:
        index2name[extract_layer_index(layer_name, num_attn_module)].append(layer_name)

    for layer_index in sorted(index2name.keys()):
        layer_names = index2name[layer_index]
        if len(layer_names) > 1:
            # One typical case is encoder-decoder model, e.g., bart.
            # The cross attention and self attention in the same decoder layer
            # has different layer_name but the same layer_index.

            # TODO - analyze where runner_kv_caches is used and the right
            # way to ensure it properly reflects multiple attention layers
            # in the same decoder block.
            if (
                current_platform.is_cuda_alike()
                or current_platform.is_xpu()
                or current_platform.is_cpu()
            ):
                # We know that the GPU / CPU runner is not impacted by this
                # case. Some test code depends on runner_kv_caches, but
                # not in a way that's impacted by ignoring this.
                pass
            else:
                raise NotImplementedError
        for layer_name in layer_names:
            runner_kv_caches.append(kv_caches[layer_name])

    # Bind kv_caches to forward context
    for layer_name, kv_cache in kv_caches.items():
        # NOTE: Use list because of v0 PP virtual engine.
        forward_context[layer_name].kv_cache = [kv_cache]

def is_residual_scattered_for_sp(
    vllm_config: VllmConfig, num_input_tokens: int
) -> bool:
    """Check if the residual tensor is scattered for sequence parallelism.

    The residual tensor is scattered across tensor parallel ranks when sequence
    parallelism and tensor parallelism is enabled.

    This follows the same logic as SequenceParallelismPass.is_applicable_for_range():
    - In full-graph compilation mode (no splitting ops or using inductor graph
      partition), SP is always applied
    - Otherwise, SP is only applied for specific shapes in compile_sizes
    """
    if not vllm_config.compilation_config.pass_config.enable_sp:
        return False

    tp = vllm_config.parallel_config.tensor_parallel_size

    if tp == 1:
        return False

    # When sequence parallelism is enabled, we always pad num_input_tokens
    # to be a multiple of tensor_parallel_size (tp) earlier.
    assert num_input_tokens % tp == 0

    if (
        not vllm_config.compilation_config.splitting_ops
        or vllm_config.compilation_config.use_inductor_graph_partition
    ):
        return True
    compile_sizes = vllm_config.compilation_config.compile_sizes
    if compile_sizes is None:
        return False
    return num_input_tokens in compile_sizes

def request_memory(init_snapshot: MemorySnapshot, cache_config: CacheConfig) -> int:
    """
    Calculate the amount of memory required by vLLM, then validate
    that the current amount of free memory is sufficient for that.
    """
    requested_memory = math.ceil(
        init_snapshot.total_memory * cache_config.gpu_memory_utilization
    )

    if init_snapshot.free_memory < requested_memory:
        raise ValueError(
            f"Free memory on device {init_snapshot.device_} "
            f"({format_gib(init_snapshot.free_memory)}/"
            f"{format_gib(init_snapshot.total_memory)} GiB) on startup "
            f"is less than desired GPU memory utilization "
            f"({cache_config.gpu_memory_utilization}, "
            f"{format_gib(requested_memory)} GiB). Decrease GPU memory "
            f"utilization or reduce GPU memory used by other processes."
        )

    return requested_memory

def sanity_check_mm_encoder_outputs(
    mm_embeddings: MultiModalEmbeddings,
    expected_num_items: int,
) -> None:
    """
    Perform sanity checks for the result of
    [`vllm.model_executor.models.SupportsMultiModal.embed_multimodal`][].
    """
    assert isinstance(mm_embeddings, (list, tuple, torch.Tensor)), (
        "Expected multimodal embeddings to be a list/tuple of 2D tensors, "
        f"or a single 3D tensor, but got {type(mm_embeddings)} "
        "instead. This is most likely due to incorrect implementation "
        "of the model's `embed_multimodal` method."
    )

    assert len(mm_embeddings) == expected_num_items, (
        "Expected number of multimodal embeddings to match number of "
        f"input items: {expected_num_items}, but got {len(mm_embeddings)=} "
        "instead. This is most likely due to incorrect implementation "
        "of the model's `embed_multimodal` method."
    )

    assert all(e.ndim == 2 for e in mm_embeddings), (
        "Expected multimodal embeddings to be a sequence of 2D tensors, "
        f"but got tensors with shapes {[e.shape for e in mm_embeddings]} "
        "instead. This is most likely due to incorrect implementation "
        "of the model's `embed_multimodal` method."
    )