forward_context.py

# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project

import time
from collections import defaultdict
from contextlib import contextmanager
from dataclasses import dataclass, field
from typing import Any

import torch

import vllm.envs as envs
import vllm.ir
from vllm.config import CUDAGraphMode, ParallelConfig, VllmConfig
from vllm.logger import init_logger
from vllm.platforms import current_platform
from vllm.v1.attention.backend import AttentionMetadata
from vllm.v1.worker.dp_utils import coordinate_batch_across_dp
from vllm.v1.worker.ubatch_utils import UBatchSlices

logger = init_logger(__name__)

track_batchsize: bool = envs.VLLM_LOG_BATCHSIZE_INTERVAL >= 0
last_logging_time: float = 0
forward_start_time: float = 0
batchsize_logging_interval: float = envs.VLLM_LOG_BATCHSIZE_INTERVAL
batchsize_forward_time: defaultdict = defaultdict(list)


@dataclass(frozen=True)
class BatchDescriptor:
    """
    Batch descriptor for cudagraph dispatching. We should keep the num of
    items as minimal as possible to properly and uniquely describe the padded
    batch for cudagraph.
    """

    num_tokens: int
    num_reqs: int | None = None
    """
    Number of requests in the batch. Can be None for PIECEWISE cudagraphs where
    the cudagraphs can handle any number of requests.
    """
    uniform: bool = False
    """
    True if all the requests in the batch have the same number of tokens.
    """
    has_lora: bool = False
    """
    Whether this batch has active LoRA adapters.
    """
    num_active_loras: int = 0
    """
    Number of distinct active LoRA adapters in this batch.
    When cudagraph_specialize_lora_count is enabled, separate CUDA graphs
    are captured for each num_active_loras value. This allows kernels
    (like fused_moe_lora) whose grid size depends on num_active_loras
    to be properly captured.
    """


def _compute_sp_num_tokens(
    num_tokens_across_dp_cpu: torch.Tensor, sequence_parallel_size: int
) -> list[int]:
    sp_tokens = (
        num_tokens_across_dp_cpu + sequence_parallel_size - 1
    ) // sequence_parallel_size

    sp_tokens = sp_tokens.repeat_interleave(sequence_parallel_size)
    return sp_tokens.tolist()


def _compute_chunked_local_num_tokens(
    num_tokens_across_dp_cpu: torch.Tensor,
    sequence_parallel_size: int,
    max_num_tokens: int,
    chunk_idx: int,
) -> list[int]:
    sp_tokens = _compute_sp_num_tokens(num_tokens_across_dp_cpu, sequence_parallel_size)
    sp_size = len(sp_tokens)

    local_size = [-1] * sp_size
    for i in range(sp_size):
        # Take into account sharding if MoE activation is sequence parallel.
        local_size[i] = min(max_num_tokens, sp_tokens[i] - (max_num_tokens * chunk_idx))
        if local_size[i] <= 0:
            local_size[i] = 1  # ensure lockstep even if done
    return local_size


@dataclass
class DPMetadata:
    max_tokens_across_dp_cpu: torch.Tensor
    num_tokens_across_dp_cpu: torch.Tensor

    # NOTE: local_sizes should only be set by the chunked_sizes context manager
    local_sizes: list[int] | None = None

    @staticmethod
    def make(
        parallel_config: ParallelConfig,
        num_tokens: int,
        num_tokens_across_dp_cpu: torch.Tensor,
    ) -> "DPMetadata":
        assert num_tokens_across_dp_cpu is not None
        assert parallel_config.data_parallel_size > 1
        assert parallel_config.is_moe_model is not False
        dp_rank = parallel_config.data_parallel_rank
        batchsize = num_tokens

        # If num_tokens_across_dp is None, it will be computed by all_reduce
        # Otherwise, num_tokens_across_dp[dp_rank] should be equal to batchsize
        assert num_tokens_across_dp_cpu[dp_rank] == batchsize, (
            f"{num_tokens_across_dp_cpu[dp_rank]} {batchsize}"
        )
        max_tokens_across_dp_cpu = torch.max(num_tokens_across_dp_cpu)
        return DPMetadata(max_tokens_across_dp_cpu, num_tokens_across_dp_cpu)

    @contextmanager
    def chunked_sizes(
        self, sequence_parallel_size: int, max_chunk_size_per_rank: int, chunk_idx: int
    ):
        """
        Context manager to compute and temporarily set the per-rank local token
        sizes for a specific chunk during chunked forward execution.

        This is necessary to ensure each DP (data parallel) rank processes its
        designated portion of tokens in lockstep with others, even when the
        token counts are uneven or some ranks have completed their input early.

        For chunked execution, we break up the total tokens on each rank into
        multiple chunks (of at most `max_chunk_size_per_rank`), and for a given
        `chunk_idx`, this context manager sets `self.local_sizes` to the number
        of tokens to process in that chunk on each rank.

        `self.local_sizes` is only valid inside the context.

        Args:
            sequence_parallel_size: When Attn is TP and MoE layers are EP,
                                    we use SP between the layers to avoid
                                    redundant ops. We need this value to
                                    compute the chunked sizes.
            max_chunk_size_per_rank: The max number of tokens each rank is
                                     allowed to process in this chunk.
            chunk_idx: The index of the chunk to compute sizes for.
        """
        self.local_sizes = _compute_chunked_local_num_tokens(
            self.num_tokens_across_dp_cpu,
            sequence_parallel_size,
            max_chunk_size_per_rank,
            chunk_idx,
        )
        try:
            yield self.local_sizes
        finally:
            self.local_sizes = None

    @contextmanager
    def sp_local_sizes(self, sequence_parallel_size: int):
        """
        Context manager for setting self.local_sizes. Same as self.chunked_sizes
        but without any chunking.
        """
        self.local_sizes = _compute_sp_num_tokens(
            self.num_tokens_across_dp_cpu, sequence_parallel_size
        )
        try:
            yield self.local_sizes
        finally:
            self.local_sizes = None

    def get_chunk_sizes_across_dp_rank(self) -> list[int] | None:
        assert self.local_sizes is not None
        return self.local_sizes

    # Get the cumulative tokens across sequence parallel ranks.
    # In this case the input to the MoEs will be distributed w.r.t both
    # DP and TP rank.
    # When sp_size==1, this is just the cumulative num tokens across DP.
    def cu_tokens_across_sp(self, sp_size: int) -> torch.Tensor:
        num_tokens_across_sp_cpu = (
            self.num_tokens_across_dp_cpu - 1 + sp_size
        ) // sp_size
        num_tokens_across_sp_cpu = num_tokens_across_sp_cpu.repeat_interleave(sp_size)
        return torch.cumsum(num_tokens_across_sp_cpu, dim=0)


@dataclass
class ForwardContext:
    # copy from vllm_config.compilation_config.static_forward_context
    no_compile_layers: dict[str, Any]
    attn_metadata: dict[str, AttentionMetadata] | list[dict[str, AttentionMetadata]]
    slot_mapping: dict[str, torch.Tensor] | list[dict[str, torch.Tensor]]
    """
    Type Dict[str, AttentionMetadata] for v1, map from layer_name of each
    attention layer to its attention metadata
    Type List[Dict[str, AttentionMetadata]] for DBO. List of size two, one
    for each microbatch.
    Set dynamically for each forward pass
    """
    # set dynamically for each forward pass
    dp_metadata: DPMetadata | None = None
    # determine the cudagraph style at runtime to be FULL, PIECEWISE, or NONE.
    # by default NONE, no cudagraph is used.
    cudagraph_runtime_mode: CUDAGraphMode = CUDAGraphMode.NONE
    batch_descriptor: BatchDescriptor | None = None

    ubatch_slices: UBatchSlices | None = None

    # If True, bypass the compiled model call, e.g. by using .forward() directly
    skip_compiled: bool = False

    # For torch.compile cold start times, we need to avoid hard-coding
    # any strings into the graph. Right now, the vllm.moe_forward
    # and vllm.moe_forward_shared custom operators hard-code strings into
    # the graph.
    #
    # The workaround is to store a list of the strings that each of those
    # custom ops needs in the ForwardContext (all_moe_layers)
    # as well as a counter (moe_layer_index).
    # The ForwardContext object is alive for the duration of the forward pass.
    # When the custom op needs a layer string, get the next string
    # from all_moe_layers and increment the counter.
    #
    # This assumes that the custom operators will always be executed in
    # order and that torch.compile will not try to reorder these
    # operations with respect to each other.
    #
    # TODO(https://github.com/vllm-project/vllm/issues/31985):
    # There are longer-term solutions, like unwrapping the moe custom operator,
    # that aren't ready yet.
    # We could also treat the string as a "symbolic input" to the graph but
    # the PyTorch-side bits for that aren't ready yet either.
    #
    # If this value is None (like in some tests), then we end up baking the string
    # into the graph. Otherwise, the moe custom ops will pop a string from this list.
    all_moe_layers: list[str] | None = None
    moe_layer_index: int = 0

    additional_kwargs: dict[str, Any] = field(default_factory=dict)

    def __post_init__(self):
        assert self.cudagraph_runtime_mode.is_valid_runtime_mode(), (
            f"Invalid cudagraph runtime mode: {self.cudagraph_runtime_mode}"
        )


_forward_context: ForwardContext | None = None


def get_forward_context() -> ForwardContext:
    """Get the current forward context."""
    assert _forward_context is not None, (
        "Forward context is not set. "
        "Please use `set_forward_context` to set the forward context."
    )
    return _forward_context


def is_forward_context_available() -> bool:
    return _forward_context is not None


def create_forward_context(
    attn_metadata: Any,
    vllm_config: VllmConfig,
    dp_metadata: DPMetadata | None = None,
    cudagraph_runtime_mode: CUDAGraphMode = CUDAGraphMode.NONE,
    batch_descriptor: BatchDescriptor | None = None,
    ubatch_slices: UBatchSlices | None = None,
    slot_mapping: dict[str, torch.Tensor] | list[dict[str, torch.Tensor]] | None = None,
    additional_kwargs: dict[str, Any] | None = None,
    skip_compiled: bool = False,
):
    if vllm_config.compilation_config.fast_moe_cold_start:
        all_moe_layers = vllm_config.compilation_config.static_all_moe_layers
    else:
        all_moe_layers = None

    return ForwardContext(
        no_compile_layers=vllm_config.compilation_config.static_forward_context,
        all_moe_layers=all_moe_layers,
        attn_metadata=attn_metadata,
        slot_mapping=slot_mapping or {},
        dp_metadata=dp_metadata,
        cudagraph_runtime_mode=cudagraph_runtime_mode,
        batch_descriptor=batch_descriptor,
        ubatch_slices=ubatch_slices,
        skip_compiled=skip_compiled,
        additional_kwargs=additional_kwargs or {},
    )


@contextmanager
def override_forward_context(forward_context: ForwardContext | None):
    """A context manager that overrides the current forward context.
    This is used to override the forward context for a specific
    forward pass.
    """
    global _forward_context
    prev_context = _forward_context
    _forward_context = forward_context
    try:
        yield
    finally:
        _forward_context = prev_context


@contextmanager
def set_forward_context(
    attn_metadata: Any,
    vllm_config: VllmConfig,
    num_tokens: int | None = None,
    num_tokens_across_dp: torch.Tensor | None = None,
    cudagraph_runtime_mode: CUDAGraphMode = CUDAGraphMode.NONE,
    batch_descriptor: BatchDescriptor | None = None,
    ubatch_slices: UBatchSlices | None = None,
    slot_mapping: dict[str, torch.Tensor] | list[dict[str, torch.Tensor]] | None = None,
    skip_compiled: bool = False,
):
    """A context manager that stores the current forward context,
    can be attention metadata, etc.
    Here we can inject common logic for every model forward pass.
    """
    global forward_start_time
    need_to_track_batchsize = track_batchsize and attn_metadata is not None
    if need_to_track_batchsize:
        forward_start_time = time.perf_counter()

    dp_metadata: DPMetadata | None = None
    if (
        vllm_config.parallel_config.data_parallel_size > 1
        and vllm_config.parallel_config.is_moe_model is not False
        and (attn_metadata is not None or num_tokens is not None)
    ):
        # If num_tokens_across_dp hasn't already been initialized, then
        # initialize it here. Both DP padding and Microbatching will be
        # disabled.
        if num_tokens_across_dp is None:
            assert ubatch_slices is None
            assert num_tokens is not None
            _, num_tokens_across_dp, _ = coordinate_batch_across_dp(
                num_tokens_unpadded=num_tokens,
                parallel_config=vllm_config.parallel_config,
                allow_microbatching=False,
            )
            assert num_tokens_across_dp is not None
        dp_metadata = DPMetadata.make(
            vllm_config.parallel_config, num_tokens or 0, num_tokens_across_dp
        )

    # Convenience: if cudagraph is used and num_tokens is given, we can just
    # create a batch descriptor here if not given (there's no harm since if it
    # doesn't match in the wrapper it'll fall through).
    if cudagraph_runtime_mode != CUDAGraphMode.NONE and num_tokens is not None:
        batch_descriptor = batch_descriptor or BatchDescriptor(num_tokens=num_tokens)

    additional_kwargs = current_platform.set_additional_forward_context(
        attn_metadata=attn_metadata,
        vllm_config=vllm_config,
        dp_metadata=dp_metadata,
        num_tokens=num_tokens,
        num_tokens_across_dp=num_tokens_across_dp,
        cudagraph_runtime_mode=cudagraph_runtime_mode,
        batch_descriptor=batch_descriptor,
        ubatch_slices=ubatch_slices,
    )

    forward_context = create_forward_context(
        attn_metadata,
        vllm_config,
        dp_metadata,
        cudagraph_runtime_mode,
        batch_descriptor,
        ubatch_slices,
        slot_mapping,
        additional_kwargs,
        skip_compiled,
    )

    try:
        with (
            override_forward_context(forward_context),
            vllm_config.kernel_config.ir_op_priority.set_priority(),
            vllm.ir.enable_torch_wrap(
                vllm_config.compilation_config.ir_enable_torch_wrap
            ),
        ):
            yield
    finally:
        global last_logging_time, batchsize_logging_interval
        if need_to_track_batchsize:
            batchsize = num_tokens
            # we use synchronous scheduling right now,
            # adding a sync point here should not affect
            # scheduling of the next batch
            synchronize = current_platform.synchronize
            if synchronize is not None:
                synchronize()
            now = time.perf_counter()
            # time measurement is in milliseconds
            batchsize_forward_time[batchsize].append((now - forward_start_time) * 1000)
            if now - last_logging_time > batchsize_logging_interval:
                last_logging_time = now
                forward_stats = []
                for bs, times in batchsize_forward_time.items():
                    if len(times) <= 1:
                        # can be cudagraph / profiling run
                        continue
                    medium = torch.quantile(torch.tensor(times), q=0.5).item()
                    medium = round(medium, 2)
                    forward_stats.append((bs, len(times), medium))
                forward_stats.sort(key=lambda x: x[1], reverse=True)
                if forward_stats:
                    logger.info(
                        (
                            "Batchsize forward time stats "
                            "(batchsize, count, median_time(ms)): %s"
                        ),
                        forward_stats,
                    )