vllm.v1.worker.utils ¶

@triton.jit
def _zero_kv_blocks_kernel(
    seg_addrs_ptr,
    seg_block_strides_ptr,
    seg_page_sizes_ptr,
    block_ids_ptr,
    BLOCK_SIZE: tl.constexpr,
):
    """Zero KV cache blocks across all segments in a single launch.

    Each segment is a contiguous region of one block's data.  Layer-compact
    layouts have one segment per layer buffer; dimensions physically outside
    the block dim (separate head groups under LHBNC) and virtual block splits
    each get their own segment.

    Segments may have different block strides and page sizes (e.g. packed
    KV views or models with multiple KV cache groups like MLA + DSA
    indexer). Each segment's block stride determines where a logical block
    begins, while its page size determines how many elements are cleared.

    seg_addrs_ptr holds absolute byte addresses (int64) for each segment,
    allowing segments to live in different CUDA allocations.

    Programs are mapped directly onto a 3-D grid as
    (block_index, seg_index, chunk_index).
    """
    block_index = tl.program_id(0)
    seg_index = tl.program_id(1)
    chunk_index = tl.program_id(2)
    block_stride_el = tl.load(seg_block_strides_ptr + seg_index)
    page_size_el = tl.load(seg_page_sizes_ptr + seg_index)
    chunk_offset = chunk_index.to(tl.int64) * BLOCK_SIZE
    if chunk_offset >= page_size_el:
        return
    block_id = tl.load(block_ids_ptr + block_index)
    seg_addr = tl.load(seg_addrs_ptr + seg_index)
    ptr = tl.cast(seg_addr, tl.pointer_type(tl.int32))
    block_offset = block_id.to(tl.int64) * block_stride_el.to(tl.int64)
    cols = chunk_offset + tl.arange(0, BLOCK_SIZE).to(tl.int64)
    tl.store(
        ptr + block_offset + cols,
        tl.zeros([BLOCK_SIZE], dtype=tl.int32),
        mask=cols < page_size_el,
    )

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.
    """
    if not shared_kv_cache_layers:
        return

    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 allocate_kv_cache(
    kv_cache_config: KVCacheConfig,
    device: torch.device,
    layout: KVCacheLayout,
    kernel_block_sizes: list[int] | None = None,
) -> dict[str, torch.Tensor]:
    """Allocate the KV cache and view it as ``[B, H, N, C]`` per layer.

    Every KVCacheTensor places its layers in the same backing allocation: layer ``l`` of
    block ``b`` starts at ``offset + l * layer_stride + b * block_stride``. Cache
    groups overlay each other, so tensors may address the same bytes.
    """
    if not kv_cache_config.kv_cache_tensors:
        return {}

    sizes = {tensor.size for tensor in kv_cache_config.kv_cache_tensors}
    assert len(sizes) == 1, "KV cache tensors must share one backing allocation."
    buf = torch.zeros(sizes.pop(), dtype=torch.int8, device=device)

    kv_caches: dict[str, torch.Tensor] = {}
    for tensor in kv_cache_config.kv_cache_tensors:
        layer_name = tensor.layers[0]
        group_id, group = next(
            (group_id, group)
            for group_id, group in enumerate(kv_cache_config.kv_cache_groups)
            if layer_name in group.layer_names
        )
        spec = group.kv_cache_spec
        if isinstance(spec, UniformTypeKVCacheSpecs):
            spec = spec.kv_cache_specs[layer_name]

        num_blocks = kv_cache_config.num_blocks
        kernel_block_size = None
        if kernel_block_sizes is not None and group_id < len(kernel_block_sizes):
            kernel_block_size = kernel_block_sizes[group_id]

        views = create_kv_cache_views(
            buf,
            spec,
            num_blocks,
            layout,
            tensor,
            kernel_block_size=kernel_block_size,
        )
        kv_caches.update(zip(tensor.layers, views))
    return kv_caches

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.
            current_platform.check_runner_kv_caches_multi_layer()
        for layer_name in layer_names:
            runner_kv_caches.append(kv_caches[layer_name])

    # Bind kv_caches to forward context. Each layer's bind_kv_cache unpacks
    # its raw allocation into the per-layer view(s) it needs (e.g. Mamba
    # splits conv/ssm), so the kv_caches dict can hold a single tensor per
    # layer for the KV connector to register.
    for layer_name, kv_cache in kv_caches.items():
        forward_context[layer_name].bind_kv_cache(kv_cache)

def get_uniform_decode_token_count(
    num_reqs: int, num_tokens: int, max_query_len: int, has_prefill: bool
) -> int | None:
    """Per-request token count of a uniform decode batch, or None."""
    if not has_prefill and is_uniform_query_len(num_reqs, num_tokens, max_query_len):
        return max_query_len
    return None

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. SP is only supported in
    full-graph compilation mode.
    """
    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

    assert (
        vllm_config.compilation_config.use_inductor_graph_partition
        or not vllm_config.compilation_config.splitting_ops
    ), "Sequence parallelism requires full-graph compilation"

    # 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

    return True

def is_uniform_query_len(num_reqs: int, num_tokens: int, max_query_len: int) -> bool:
    """Whether every request in the batch has the same query length.

    Shape test only; use ``get_uniform_decode_token_count`` to classify a
    scheduled batch, since a prompt chunk can have a decode batch's shape.
    """
    return num_reqs > 0 and num_tokens == max_query_len * num_reqs

def prepare_kernel_block_sizes(
    kv_cache_config: KVCacheConfig, attn_groups: list[list[AttentionGroup]]
) -> list[int]:
    """
    Generate kernel_block_sizes that matches each block_size.

    For attention backends that support virtual block splitting,
    use the supported block sizes from the backend.
    For other backends (like Mamba), use the same block size (no splitting).

    Args:
        kv_cache_config: The KV cache configuration.
        attn_groups: Attention groups indexed by KV cache group id.

    Returns:
        List of kernel block sizes for each cache group.
    """
    kernel_block_sizes = []
    for kv_cache_gid, kv_cache_group in enumerate(kv_cache_config.kv_cache_groups):
        kv_cache_spec = kv_cache_group.kv_cache_spec
        if isinstance(kv_cache_spec, UniformTypeKVCacheSpecs):
            # All layers in the UniformTypeKVCacheSpecs have the same type,
            # pick an arbitrary one to dispatch.
            kv_cache_spec = next(iter(kv_cache_spec.kv_cache_specs.values()))
        if isinstance(kv_cache_spec, EncoderOnlyAttentionSpec):
            continue
        if isinstance(kv_cache_spec, AttentionSpec):
            # This is an attention backend that supports virtual block splitting.
            kv_manager_block_size = kv_cache_group.kv_cache_spec.block_size
            group_backends = [g.backend for g in attn_groups[kv_cache_gid]]
            selected_kernel_size = select_common_block_size(
                kv_manager_block_size, group_backends
            )
            kernel_block_sizes.append(selected_kernel_size)
        elif isinstance(kv_cache_spec, MambaSpec):
            # This is likely Mamba or other non-attention cache, no splitting.
            kernel_block_sizes.append(kv_cache_spec.block_size)
        else:
            raise NotImplementedError(
                f"unknown kv cache spec {kv_cache_group.kv_cache_spec}"
            )
    return kernel_block_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."
    )

def select_common_block_size(
    kv_manager_block_size: int,
    backends: list[type[AttentionBackend]],
) -> int:
    """
    Select a block size that is supported by all backends and is a factor of
    kv_manager_block_size.

    If kv_manager_block_size is supported by all backends, return it directly.
    Otherwise, return the max supported size.

    Args:
        kv_manager_block_size: Block size of KV cache.
        backends: List of attention backend classes.

    Returns:
        The selected block size.

    Raises:
        ValueError: If no valid block size found.
    """

    def block_size_is_supported(
        backends: list[type[AttentionBackend]], block_size: int
    ) -> bool:
        """Check if the block size is supported by all backends."""
        for backend in backends:
            is_supported = False
            for supported_size in backend.get_supported_kernel_block_sizes():
                if isinstance(supported_size, int):
                    if block_size == supported_size:
                        is_supported = True
                elif isinstance(supported_size, MultipleOf):
                    if block_size % supported_size.base == 0:
                        is_supported = True
                else:
                    raise ValueError(f"Unknown supported size: {supported_size}")
            if not is_supported:
                return False
        return True

    # Case 1: if the block_size of kv cache manager is supported by all backends,
    # return it directly.
    if block_size_is_supported(backends, kv_manager_block_size):
        return kv_manager_block_size

    # Case 2: otherwise, the block_size must be an `int`-format supported size of
    # at least one backend. Iterate over all `int`-format supported sizes in
    # descending order and return the first one that is supported by all backends.
    # Simple proof:
    # If the supported size b is in MultipleOf(x_i) format for all attention
    # backends i, and b a factor of kv_manager_block_size, then
    # kv_manager_block_size also satisfies MultipleOf(x_i) for all i. We will
    # return kv_manager_block_size in case 1.
    all_int_supported_sizes = set(
        supported_size
        for backend in backends
        for supported_size in backend.get_supported_kernel_block_sizes()
        if isinstance(supported_size, int)
    )

    for supported_size in sorted(all_int_supported_sizes, reverse=True):
        if kv_manager_block_size % supported_size != 0:
            continue
        if block_size_is_supported(backends, supported_size):
            return supported_size
    raise ValueError(f"No common block size for {kv_manager_block_size}. ")