xc-llm-ascend/vllm_ascend/ascend_forward_context.py

import math
from contextlib import contextmanager
from enum import Enum
from typing import Any

import torch
from vllm.config import CUDAGraphMode, VllmConfig
from vllm.distributed import get_dp_group, get_ep_group, get_tensor_model_parallel_world_size
from vllm.forward_context import BatchDescriptor, get_forward_context, set_forward_context

import vllm_ascend.envs as envs_ascend
from vllm_ascend.ascend_config import get_ascend_config
from vllm_ascend.utils import (
    AscendDeviceType,
    enable_sp,
    flashcomm2_enable,
    get_ascend_device_type,
    has_layer_idx,
    is_drafter_moe_model,
    is_moe_model,
    speculative_enable_dispatch_gmm_combine_decode,
)


class MoECommType(Enum):
    ALLGATHER = 0
    MC2 = 1
    ALLTOALL = 2
    FUSED_MC2 = 3


@contextmanager
def set_ascend_forward_context(
    attn_metadata: Any,
    vllm_config: VllmConfig,
    virtual_engine: int = 0,
    num_tokens: int = 0,
    num_tokens_across_dp: torch.Tensor | None = None,
    in_profile_run: bool = False,
    num_actual_tokens: int | None = None,
    aclgraph_runtime_mode: CUDAGraphMode = CUDAGraphMode.NONE,
    batch_descriptor: BatchDescriptor | None = None,
    model_instance: torch.nn.Module = None,
    is_draft_model=False,
):
    """A context manager that stores the current forward context,
    can be attention metadata, etc.
    We add some additional param into forward_context.
    """
    with set_forward_context(
        attn_metadata,
        vllm_config,
        virtual_engine=virtual_engine,
        num_tokens=num_tokens,
        num_tokens_across_dp=num_tokens_across_dp,
        cudagraph_runtime_mode=aclgraph_runtime_mode,
        batch_descriptor=batch_descriptor,
    ):
        forward_context = get_forward_context()

        from vllm_ascend.ops.fused_moe.moe_comm_method import get_moe_comm_method

        moe_comm_type = select_moe_comm_method(num_tokens, vllm_config, is_draft_model)
        forward_context.moe_comm_type = moe_comm_type
        forward_context.moe_comm_method = get_moe_comm_method(moe_comm_type)

        tp_world_size = get_tensor_model_parallel_world_size()

        forward_context.in_profile_run = in_profile_run

        # NOTE: This cannot be set using set_forward_context
        # due to multiple warmups before actual capturing
        forward_context.capturing = False

        # TODO: remove it when torch_npu.npu_mm_reduce_scatter_base supports tp_size >= 16.
        mmrs_fusion = tp_world_size <= 8

        # set for sequence parallelism, 1000 is the batch size concurrency threshold
        # for enabling the flashcomm_v1 or sequence_parallelism feature.
        # Currently, it is an empirical value. In normal scenarios, if the concurrency
        # exceeds this threshold, the performance benefits can be maximized.
        # Conversely, if the concurrency is below the threshold,
        # the performance may degrade due to the switching of communication methods.

        # main model and drafter model may have different architecture
        is_context_moe_model = is_drafter_moe_model(vllm_config) if is_draft_model else is_moe_model(vllm_config)
        if is_context_moe_model:
            sp_enabled = enable_sp(vllm_config) and num_tokens is not None
            mmrs_fusion = False
        elif is_draft_model:
            # TODO: for dense drafter, `sp` is redundant and is not compatible with `dp` and `graph`.
            # Disable it to avoid more problems.
            sp_enabled = False
        else:
            sp_enabled = enable_sp(vllm_config) and num_tokens is not None and num_tokens > 1000

        forward_context.mmrs_fusion = mmrs_fusion
        forward_context.num_tokens = num_tokens
        forward_context.sp_enabled = sp_enabled
        # TODO(Levi-JQ): another PR to normalize the enabling logic for sp/fc2
        forward_context.flashcomm_v2_enabled = flashcomm2_enable() and tp_world_size > 1 and num_tokens is not None

        if forward_context.sp_enabled or forward_context.flashcomm_v2_enabled:
            pad_size = (tp_world_size - (num_tokens % tp_world_size)) % tp_world_size
            forward_context.pad_size = pad_size

        # set this for rope forward_oot using
        forward_context.is_first_layer = True

        # set layer_idx to enable optimization features that depend on this information.
        # This is only applicable to models that contain these necessary attributes.
        forward_context.layer_idx = None
        if has_layer_idx(model_instance):
            forward_context.layer_idx = model_instance.model.start_layer

        # TODO(rjg-lyh): refactor mlp weight prefetch method
        # set for mlp weight prefetch
        prefetch_mlp_enabled = (
            envs_ascend.VLLM_ASCEND_ENABLE_PREFETCH_MLP
            and forward_context.layer_idx is not None
            and num_tokens is not None
            and num_tokens < 500
        )
        if prefetch_mlp_enabled:
            forward_context.prefetch_mlp_gate_up_proj = False
            forward_context.prefetch_mlp_down_proj = False
        forward_context.prefetch_mlp_enabled = prefetch_mlp_enabled
        forward_context.model_instance = model_instance
        forward_context.is_draft_model = is_draft_model

        if num_tokens is None and attn_metadata is not None:
            num_tokens = attn_metadata.num_actual_tokens

        dp_world_size = get_dp_group().world_size
        if dp_world_size > 1 and forward_context.dp_metadata is not None:
            max_tokens_across_dp = forward_context.dp_metadata.max_tokens_across_dp_cpu.item()
            if forward_context.sp_enabled or forward_context.flashcomm_v2_enabled:
                padded_length = (max_tokens_across_dp + tp_world_size - 1) // tp_world_size * tp_world_size
                pad_size = padded_length - num_tokens
                forward_context.padded_length = padded_length
                forward_context.pad_size = pad_size
        else:
            max_tokens_across_dp = num_tokens

        forward_context.max_tokens_across_dp = max_tokens_across_dp

        if num_tokens is not None:
            if num_actual_tokens is None:
                num_actual_tokens = num_tokens
            # NOTE: token num which need to pad to when mc2
            forward_context.padded_num_tokens = math.ceil(max_tokens_across_dp / tp_world_size) * tp_world_size
            reserved_mc2_mask = get_mc2_mask()
            if reserved_mc2_mask is not None:
                mc2_mask = reserved_mc2_mask[: forward_context.padded_num_tokens]
                mc2_mask[:num_actual_tokens] = True
                mc2_mask[num_actual_tokens:] = False
                forward_context.mc2_mask = mc2_mask

        try:
            yield
        finally:
            pass


_mc2_tokens_capacity: int | None = None
_reserved_mc2_mask: torch.Tensor | None = None
_sin: torch.Tensor | None = None
_cos: torch.Tensor | None = None


def set_mc2_tokens_capacity(vllm_config, max_num_reqs, uniform_decode_query_len):
    global _mc2_tokens_capacity
    if _mc2_tokens_capacity is not None:
        return
    if vllm_config.compilation_config.cudagraph_capture_sizes:
        max_num_tokens = vllm_config.compilation_config.max_cudagraph_capture_size
    else:
        # NOTE: To save memory, we cap the max number of tokens to 512.
        max_num_tokens = min(max_num_reqs * uniform_decode_query_len, 512)
    tp_size = vllm_config.parallel_config.tensor_parallel_size
    # Use integer arithmetic for ceiling division.
    num_tokens_per_tp_rank = (max_num_tokens + tp_size - 1) // tp_size
    _mc2_tokens_capacity = num_tokens_per_tp_rank * tp_size


def get_mc2_tokens_capacity():
    return _mc2_tokens_capacity


def set_mc2_mask(vllm_config, device):
    global _reserved_mc2_mask
    if _reserved_mc2_mask is not None:
        return
    if is_moe_model(vllm_config):
        _reserved_mc2_mask = torch.zeros(get_mc2_tokens_capacity(), dtype=torch.bool, device=device)
    else:
        _reserved_mc2_mask = None


def get_mc2_mask():
    return _reserved_mc2_mask


def select_moe_comm_method(num_tokens: int, vllm_config: VllmConfig, is_draft_model=False) -> MoECommType | None:
    """Select the MoE communication method according to parallel settings,
    device generation, token count, and quantization.

    1. Non-MoE models return `None`.
    2. Without expert parallel, fall back to all-gather.
    3. On A2 with expert parallel, pick MC2 when tokens fit the MC2 capacity
       and the DP size is large enough; otherwise use all-gather.
    4. On A3 with expert parallel, prefer fused MC2 when using w8a8_dynamic
       quantization with small EP size, no dynamic_eplb, and not in MTP
       mode; otherwise use MC2 within capacity or all-to-all.

    Args:
        num_tokens (int): The number of tokens in the current batch.
        vllm_config (VllmConfig): Runtime configuration for the model.
        is_draft_model (bool): Whether the model runs in MTP mode (disables fused MC2).

    Raises:
        ValueError: If the soc version is unsupported.

    Returns:
        MoECommType | None: The selected MoE communication method.
    """
    if not is_moe_model(vllm_config):
        return None
    mc2_tokens_capacity = get_mc2_tokens_capacity()
    soc_version = get_ascend_device_type()
    quant_type = getattr(
        vllm_config.model_config.hf_text_config,
        "moe_quantize",
        getattr(vllm_config.model_config.hf_text_config, "quantize", None),
    )

    if not vllm_config.parallel_config.enable_expert_parallel or get_ep_group().world_size == 1:
        moe_comm_type = MoECommType.ALLGATHER
    elif soc_version in {AscendDeviceType.A2}:
        if (
            num_tokens <= mc2_tokens_capacity
            and vllm_config.parallel_config.world_size_across_dp / vllm_config.parallel_config.pipeline_parallel_size
            >= 16
        ):
            moe_comm_type = MoECommType.MC2
        else:
            moe_comm_type = MoECommType.ALLGATHER

    elif soc_version in {AscendDeviceType.A3}:
        dynamic_eplb = get_ascend_config().eplb_config.dynamic_eplb
        # TODO: drop the EP-size guard when dispatch_ffn_combine supports larger EP sizes
        # TODO: drop speculative method guard when dispatch_gmm_combine_decode supports w16a16
        fused_mc2_enable = envs_ascend.VLLM_ASCEND_ENABLE_FUSED_MC2 and quant_type == "w8a8_dynamic"
        dispatch_ffn_combine_enable = get_ep_group().world_size <= 32 and (not is_draft_model) and (not dynamic_eplb)
        if num_tokens <= mc2_tokens_capacity:
            fused_decode_enable = fused_mc2_enable
            if envs_ascend.VLLM_ASCEND_ENABLE_FUSED_MC2 == 1:
                fused_decode_enable = fused_mc2_enable and dispatch_ffn_combine_enable
            elif envs_ascend.VLLM_ASCEND_ENABLE_FUSED_MC2 == 2:
                fused_decode_enable = fused_mc2_enable and speculative_enable_dispatch_gmm_combine_decode(vllm_config)
            moe_comm_type = MoECommType.FUSED_MC2 if fused_decode_enable else MoECommType.MC2
        else:
            fused_prefill_enable = fused_mc2_enable
            if envs_ascend.VLLM_ASCEND_ENABLE_FUSED_MC2 == 1:
                fused_prefill_enable = fused_mc2_enable and dispatch_ffn_combine_enable
            elif envs_ascend.VLLM_ASCEND_ENABLE_FUSED_MC2 == 2:
                fused_prefill_enable = False
            moe_comm_type = MoECommType.FUSED_MC2 if fused_prefill_enable else MoECommType.ALLTOALL

    else:
        raise ValueError(f"Unsupported soc_version: {soc_version}")
    return moe_comm_type