vllm.v1.worker.mamba_utils ¶

def cleanup_mamba_state_idx(
    scheduler_output: SchedulerOutput,
    mamba_state_idx: dict[str, int],
) -> None:
    """Pop stale `mamba_state_idx` entries for finished/preempted/resumed reqs.

    Force-preempted requests (e.g., during reset_prefix_cache / KV cache
    flush) appear in resumed_req_ids without a corresponding entry in
    preempted_req_ids, leaving stale entries that can point to block
    indices beyond the new (smaller) block allocation.
    """
    finished_req_ids = scheduler_output.finished_req_ids
    preempted_req_ids = scheduler_output.preempted_req_ids or set()
    resumed_req_ids = scheduler_output.scheduled_cached_reqs.resumed_req_ids
    for req_id in itertools.chain(finished_req_ids, preempted_req_ids, resumed_req_ids):
        mamba_state_idx.pop(req_id, None)

def postprocess_mamba_align_gpu(
    *,
    bufs: "MambaBuffers",
    num_reqs: int,
    num_accepted_tokens_gpu: torch.Tensor,
    num_accepted_tokens_cpu_tensor: torch.Tensor,
    input_batch: GPUInputBatch,
    kv_cache_config: KVCacheConfig,
    forward_context: dict[str, Any],
    mamba_state_copy_funcs: tuple[MambaStateCopyFunc, ...],
) -> None:
    """GPU-side mamba postprocess for spec decode + hybrid + align mode.

    Lazily binds the fused-kernel context to the persistent block tables and
    forward-context state pointers on the first call, runs the fused kernel,
    and async-copies the per-request accepted-token counts back to the input
    batch's CPU tensor for the next iteration's preprocess.
    """
    ctx = bufs.postprocess_align
    # Caller is responsible for gating on spec decode + hybrid; this assert is
    # a tripwire if those gates ever drift apart.
    assert ctx is not None
    assert ctx.mamba_state_idx_buf is not None
    assert ctx.num_scheduled_tokens_buf is not None
    assert ctx.num_computed_tokens_buf is not None
    assert ctx.num_draft_tokens_buf is not None

    if not ctx.is_initialized:
        ctx.initialize_from_forward_context(
            kv_cache_config,
            forward_context,
            mamba_state_copy_funcs,
            [
                input_batch.block_table[gid].get_device_tensor(num_reqs)
                for gid in ctx.mamba_group_ids
            ],
        )

    ctx.run_fused_postprocess(
        num_reqs=num_reqs,
        num_accepted_tokens_gpu=num_accepted_tokens_gpu,
        mamba_state_idx_gpu=ctx.mamba_state_idx_buf.gpu,
        num_scheduled_tokens_gpu=ctx.num_scheduled_tokens_buf.gpu,
        num_computed_tokens_gpu=ctx.num_computed_tokens_buf.gpu,
        num_draft_tokens_gpu=ctx.num_draft_tokens_buf.gpu,
    )

    # ``num_accepted_tokens_out`` is pre-initialized from
    # ``num_accepted_tokens_gpu``; the kernel only overwrites entries to 1
    # when src_block_idx == dest_block_idx (copy within the same block), so
    # the original count is preserved for everyone else.
    num_accepted_tokens_cpu_tensor[:num_reqs].copy_(
        ctx.num_accepted_tokens_out[:num_reqs], non_blocking=True
    )

def postprocess_mamba_all(
    scheduler_output: SchedulerOutput,
    kv_cache_config: KVCacheConfig,
    input_batch: GPUInputBatch,
    requests: dict[str, CachedRequestState],
    mamba_state_idx: dict[str, int],
    num_spec_tokens: int,
    num_reqs: int,
):
    """All-mode postprocess (only meaningful with num_spec_tokens > 0):
    record per-request the block index of the last token scheduled this
    step, so the next step can anchor its in-place writes when accepted
    drafts leave the sequence at a non-block-aligned position.
    """
    if num_spec_tokens <= 0:
        return
    _, mamba_spec = get_mamba_groups(kv_cache_config)
    block_size = mamba_spec.block_size
    full_decode_len = 1 + num_spec_tokens
    scheduled = scheduler_output.num_scheduled_tokens
    for req_id in input_batch.req_ids[:num_reqs]:
        num_query = scheduled.get(req_id, 0)
        if num_query == full_decode_len:
            req = requests[req_id]
            seq_len = req.num_computed_tokens + num_query
            mamba_state_idx[req_id] = max(0, (seq_len - 1) // block_size)
        else:
            mamba_state_idx.pop(req_id, None)

@triton.jit
def postprocess_mamba_fused_kernel(
    # Decision inputs (per-request)
    num_accepted_tokens_ptr,
    mamba_state_idx_ptr,
    num_scheduled_tokens_ptr,
    num_computed_tokens_ptr,
    num_draft_tokens_ptr,
    # Per-group block table base addresses: int64[num_groups]. Each entry is
    # the data_ptr of that group's persistent [max_reqs, max_blocks] int32
    # block table.
    block_table_ptrs_ptr,
    block_table_stride_req: tl.int64,  # stride between requests (in elements)
    # Mamba state metadata (per-layer, per-state-type)
    # These are 1D arrays indexed by (layer_idx * num_state_types + state_type_idx)
    state_base_addrs_ptr,  # base address of each state tensor
    state_block_strides_ptr,  # bytes per block for each state
    state_elem_sizes_ptr,  # element size for each state
    state_inner_sizes_ptr,  # number of elements in inner dimensions
    state_conv_widths_ptr,  # conv width for conv states (0 for temporal)
    state_group_indices_ptr,  # maps state_idx to group index in block table
    # Output: num_accepted_tokens update (for src==dst case)
    num_accepted_tokens_out_ptr,
    # Runtime parameter (varies per batch - NOT constexpr to avoid recompilation)
    num_reqs,
    # Compile-time constants (fixed after model initialization)
    # block_size: determined by model config, constant for all invocations
    block_size: tl.constexpr,
    # COPY_BLOCK_SIZE: fixed tuning parameter for memory copy loop
    COPY_BLOCK_SIZE: tl.constexpr,
):
    """
    Fused GPU kernel for postprocess_mamba that computes decisions AND performs
    mamba state copies without any CPU-GPU synchronization.

    Grid: (num_reqs, num_layers * num_state_types)
    - program_id(0) = request index
    - program_id(1) = state_idx (flattened index into layer/state_type metadata)

    Note: num_layers and num_state_types are not passed as kernel parameters
    because the kernel indexes directly into pre-flattened metadata arrays
    using program_id(1). The grid dimensions encode the total state count.
    """
    req_idx = tl.program_id(0)
    state_idx = tl.program_id(1)

    # Bounds check
    if req_idx >= num_reqs:
        return

    # Compute decision logic (mirrors postprocess_mamba Python reference)
    num_accepted = tl.load(num_accepted_tokens_ptr + req_idx)
    src_block_idx = tl.load(mamba_state_idx_ptr + req_idx)
    num_scheduled = tl.load(num_scheduled_tokens_ptr + req_idx)
    num_computed = tl.load(num_computed_tokens_ptr + req_idx)
    num_draft = tl.load(num_draft_tokens_ptr + req_idx)

    num_tokens_running_state = num_computed + num_scheduled - num_draft
    new_num_computed = num_tokens_running_state + num_accepted - 1
    aligned_new_computed = (new_num_computed // block_size) * block_size

    needs_copy = aligned_new_computed >= num_tokens_running_state

    if not needs_copy:
        return

    # Compute copy parameters
    accept_token_bias = aligned_new_computed - num_tokens_running_state
    dest_block_idx = aligned_new_computed // block_size - 1

    # Load state metadata for this layer/state_type
    state_base_addr = tl.load(state_base_addrs_ptr + state_idx)
    state_block_stride = tl.load(state_block_strides_ptr + state_idx)
    state_elem_size = tl.load(state_elem_sizes_ptr + state_idx)
    state_inner_size = tl.load(state_inner_sizes_ptr + state_idx)
    conv_width = tl.load(state_conv_widths_ptr + state_idx)

    # Load the group index for this state, then index into the correct
    # group's block table. Each mamba group has independently allocated
    # physical blocks.
    group_idx = tl.load(state_group_indices_ptr + state_idx).to(tl.int64)

    # block_table_ptrs_ptr holds one pointer per group (each group owns its own
    # block table). Reinterpret as int32* since block ids are int32.
    group_base_addr = tl.load(block_table_ptrs_ptr + group_idx)
    block_table_typed = group_base_addr.to(tl.pointer_type(tl.int32))
    block_table_base = block_table_typed + req_idx * block_table_stride_req

    # Widen block ids to int64 before they reach `block_id * state_block_stride`
    # below: state_block_stride can exceed 2**31 bytes for large mamba caches,
    # and Triton would otherwise do the multiply in int32 and wrap.
    src_block_id = tl.load(block_table_base + src_block_idx).to(tl.int64)
    dest_block_id = tl.load(block_table_base + dest_block_idx).to(tl.int64)

    # Compute source and destination addresses based on state type
    # conv_width > 0 means this is a conv state (get_conv_copy_spec logic)
    # conv_width == 0 means this is a temporal state (get_temporal_copy_spec logic)
    is_conv_state = conv_width > 0

    if is_conv_state:
        # Conv state: copy
        #   state[block_table[req_idx, src_block_idx],  accept_token_bias:]
        # to
        #   state[block_table[req_idx, dest_block_idx], :conv_width - accept_token_bias]
        src_offset = accept_token_bias.to(tl.int64) * state_inner_size * state_elem_size
        src_addr = state_base_addr + src_block_id * state_block_stride + src_offset
        dst_addr = state_base_addr + dest_block_id * state_block_stride
        # Number of elements to copy:
        # (conv_width - accept_token_bias) * inner_size
        num_elems_to_copy = (conv_width - accept_token_bias).to(
            tl.int64
        ) * state_inner_size
        copy_size = num_elems_to_copy * state_elem_size
    else:
        # Temporal state: copy
        #   state[block_table[req_idx, src_block_idx + accept_token_bias]]
        # to
        #   state[block_table[req_idx, dest_block_idx]]
        actual_src_block_idx = src_block_idx + accept_token_bias
        actual_src_block_id = tl.load(block_table_base + actual_src_block_idx).to(
            tl.int64
        )
        src_addr = state_base_addr + actual_src_block_id * state_block_stride
        dst_addr = state_base_addr + dest_block_id * state_block_stride
        # Use natural block data size (inner_size * elem_size), NOT
        # state_block_stride which is the page stride and can exceed the
        # actual data when the state tensor uses as_strided page padding.
        copy_size = state_inner_size * state_elem_size

    # Mirror postprocess_mamba's trailing
    #     if src_block_idx == dest_block_idx: num_accepted_tokens_cpu[i] = 1
    # This runs whether or not the copy below is skipped (it's per-request, so
    # only state_idx == 0 writes).
    if src_block_idx == dest_block_idx and state_idx == 0:
        tl.store(num_accepted_tokens_out_ptr + req_idx, 1)

    # Mirror collect_mamba_copy_meta's early return: src==dst with no token
    # bias means source and destination ranges coincide, so the copy is a
    # no-op.
    if src_block_idx == dest_block_idx and accept_token_bias == 0:
        return

    offsets = tl.arange(0, COPY_BLOCK_SIZE)
    for i in range(0, copy_size, COPY_BLOCK_SIZE):
        mask = (i + offsets) < copy_size
        curr_src = (src_addr + i + offsets).to(tl.pointer_type(tl.uint8))
        curr_dst = (dst_addr + i + offsets).to(tl.pointer_type(tl.uint8))
        data = tl.load(curr_src, mask=mask)
        tl.store(curr_dst, data, mask=mask)

def preprocess_mamba(
    scheduler_output: SchedulerOutput,
    kv_cache_config: KVCacheConfig,
    cache_config: CacheConfig,
    mamba_state_idx: dict[str, int],
    input_batch: GPUInputBatch,
    requests: dict[str, CachedRequestState],
    forward_context: dict[str, Any],
    mamba_state_copy_funcs: tuple[MambaStateCopyFunc, ...],
    copy_bufs: MambaCopyBuffers,
):
    """
    Copy the mamba state of previous step to the last
    (1 + num_speculative_blocks) block.
    """
    mamba_group_ids = copy_bufs.mamba_group_ids
    mamba_spec = copy_bufs.mamba_spec
    num_speculative_blocks = mamba_spec.num_speculative_blocks
    # TODO(Chen): we need to optimize this function a lot
    assert cache_config.enable_prefix_caching
    block_size = mamba_spec.block_size
    cleanup_mamba_state_idx(scheduler_output, mamba_state_idx)

    copy_bufs.offset = 0
    for i, req_id in enumerate(input_batch.req_ids):
        req_state = requests[req_id]
        prev_state_idx = mamba_state_idx.get(req_id)
        if prev_state_idx is None:
            # new / resumed request, no previous state
            # if num_computed_tokens is 0, prev_state_idx will be -1
            prev_state_idx = (req_state.num_computed_tokens - 1) // block_size

        num_scheduled_tokens = scheduler_output.num_scheduled_tokens[req_id]
        num_blocks: int = (
            cdiv(req_state.num_computed_tokens + num_scheduled_tokens, block_size)
            + num_speculative_blocks
        )

        # We always save the current running state at the last
        # (1 + num_speculative_blocks) block.
        # A corner case worth mention here: assume we have block_size = 4 and
        # num_speculative_tokens = 2. The request is [A, B, C] and contains 2 draft
        # tokens [draft 1, draft 2]. Then we will have:
        # Block 0: [A, B, C, draft 1]
        # Block 1: [draft 2, TOFILL, TOFILL, TOFILL]
        # Block 2: speculative block
        # Block 3: speculative block
        # And use block 1 to save the running state.
        curr_state_idx = num_blocks - 1 - num_speculative_blocks
        mamba_state_idx[req_id] = curr_state_idx
        if prev_state_idx != -1 and prev_state_idx != curr_state_idx:
            collect_mamba_copy_meta(
                copy_bufs,
                kv_cache_config,
                mamba_state_copy_funcs,
                mamba_group_ids,
                prev_state_idx,
                curr_state_idx,
                input_batch.num_accepted_tokens_cpu[i] - 1,
                req_state,
                forward_context,
            )
            input_batch.num_accepted_tokens_cpu[i] = 1
    do_mamba_copy_block(copy_bufs)

def stage_mamba_state_idx_to_gpu(
    mamba_state_idx: dict[str, int],
    req_ids: list[str],
    num_reqs: int,
    gpu_buf: CpuGpuBuffer,
) -> None:
    """Materialize ``mamba_state_idx`` into ``gpu_buf`` and copy to GPU.

    Walks ``req_ids[:num_reqs]`` in batch order, writing each request's block
    index into the buffer's pinned numpy view, then issues a non-blocking H→D
    copy. The fused kernel indexes the resulting GPU tensor by ``req_idx``.

    Invariant: ``preprocess_mamba`` must have run first for the same batch so
    that every ``req_ids[i]`` has an entry in ``mamba_state_idx``.
    """
    np_view = gpu_buf.np
    for i in range(num_reqs):
        req_id = req_ids[i]
        state_idx = mamba_state_idx.get(req_id)
        assert state_idx is not None, (
            f"mamba_state_idx missing entry for {req_id!r}; "
            "preprocess_mamba must run before stage_mamba_state_idx_to_gpu"
        )
        np_view[i] = state_idx
    gpu_buf.copy_to_gpu(num_reqs)