vllm.v1.worker.mamba_utils ¶

def postprocess_mamba(
    scheduler_output: SchedulerOutput,
    kv_cache_config: KVCacheConfig,
    input_batch: GPUInputBatch,
    requests: dict[str, CachedRequestState],
    mamba_state_idx: dict[str, int],
    forward_context: dict[str, Any],
    mamba_state_copy_funcs: tuple[MambaStateCopyFunc, ...],
):
    """
    If a blocks is converted from partial block to full block in this step, copy the
    state from the block for running state to the new full block.
    """
    num_scheduled_tokens_dict = scheduler_output.num_scheduled_tokens
    scheduled_spec_decode_tokens_dict = scheduler_output.scheduled_spec_decode_tokens
    num_accepted_tokens_cpu = input_batch.num_accepted_tokens_cpu
    # NOTE: can be optimized as this function always returns the same result
    mamba_group_ids, mamba_spec = get_mamba_groups(kv_cache_config)
    src_state_list: list[int] = []
    dest_state_list: list[int] = []
    num_elements_list: list[int] = []
    for i, req_id in enumerate(input_batch.req_ids):
        req_state = requests[req_id]
        num_computed_tokens = req_state.num_computed_tokens
        num_draft_tokens = len(scheduled_spec_decode_tokens_dict.get(req_id, []))
        num_scheduled_tokens = num_scheduled_tokens_dict[req_id]
        num_accepted_tokens = num_accepted_tokens_cpu[i]
        num_tokens_running_state = (
            num_computed_tokens + num_scheduled_tokens - num_draft_tokens
        )
        new_num_computed_tokens = num_tokens_running_state + num_accepted_tokens - 1
        aligned_new_computed_tokens = (
            new_num_computed_tokens // mamba_spec.block_size * mamba_spec.block_size
        )
        # TODO: how to ensure all blocks that cache_blocks called are cached here?
        if aligned_new_computed_tokens >= num_tokens_running_state:
            accept_token_bias = aligned_new_computed_tokens - num_tokens_running_state
            src_block_idx = mamba_state_idx[req_id]
            dest_block_idx = aligned_new_computed_tokens // mamba_spec.block_size - 1
            collect_mamba_copy_meta(
                src_state_list,
                dest_state_list,
                num_elements_list,
                kv_cache_config,
                mamba_state_copy_funcs,
                mamba_group_ids,
                src_block_idx,
                dest_block_idx,
                accept_token_bias,
                req_state,
                forward_context,
            )
            if src_block_idx == dest_block_idx:
                num_accepted_tokens_cpu[i] = 1
    do_mamba_copy_block(src_state_list, dest_state_list, num_elements_list)

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 the mamba state of previous step to the last
    (1 + num_speculative_blocks) block.
    """
    mamba_group_ids, mamba_spec = get_mamba_groups(kv_cache_config)
    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
    finished_req_ids = scheduler_output.finished_req_ids
    preempted_req_ids = scheduler_output.preempted_req_ids or set()
    for req_id in itertools.chain(finished_req_ids, preempted_req_ids):
        mamba_state_idx.pop(req_id, None)

    src_state_list: list[int] = []
    dest_state_list: list[int] = []
    num_elements_list: list[int] = []
    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_blocks = len(req_state.block_ids[mamba_group_ids[0]])

        # 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(
                src_state_list,
                dest_state_list,
                num_elements_list,
                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(src_state_list, dest_state_list, num_elements_list)