vllm.compilation.passes.fusion.sequence_parallelism ¶

class SequenceParallelismPass(VllmPatternMatcherPass):
    """
    This pass enables sequence parallelism for models.
    It identifies patterns where an AllReduce operation is followed by
    an RMSNorm (or RMSNorm and then Quantization) operation.
    These patterns are replaced with a ReduceScatter operation, followed by
    a local RMSNorm/Quantization, and then an AllGather operation.

    The general transformation is:
    Input -> AllReduce -> RMSNorm -> Output
    becomes
    Input -> ReduceScatter -> RMSNorm -> AllGather -> Output

    While this pass itself does not directly yield performance improvements,
    it lays the groundwork for subsequent fusion passes, such as
    GEMM + ReduceScatter and AllGather + GEMM fusions. These fusions can
    significantly reduce communication overhead and improve overall model
    performance.


    This pass splits up the residual tensor across TP ranks and hence divides its size.
    Because the pattern matcher starts at the end of the graph, the replacement
    contains a slice that temporarily conforms the input residual to the correct size.
    After all patterns have been matched, we use a NoOpEliminationPass to clean up
    what have now become no-op slices.

    Note that an older version of the pass did not need this as it operated only on
    custom rms_norm and fused_rms_norm_add custom ops which did not complain about
    mismatched shapes during replacement. So this approach has the same assumption that
    correctness is only maintained if all rms_norm operations are split across ranks.

    Correctness-wise, this is approach strictly better than before - before,
    the graph was incorrect semantically and shape-wise during the pass.
    With this approach there's only semantic incorrectness during the pass.
    Both approaches restore a correct graph once all patterns are matched.
    """

    @enable_fake_mode
    def __init__(self, config: VllmConfig) -> None:
        super().__init__(config)

        # Used to clean up redundant views created temporarily
        # to circumvent residual shape change issues
        self.noop_cleanup = NoOpEliminationPass(config)
        self.noop_cleanup.pass_name = f"{self.pass_name}.{self.noop_cleanup.pass_name}"

        self.patterns: PatternMatcherPass = PatternMatcherPass(
            pass_name="sequence_parallelism_pass"
        )

        for epsilon in [1e-5, 1e-6]:
            # RMSNorm + Static FP8 quantization patterns
            FirstAllReduceRMSNormStaticFP8Pattern(
                epsilon, self.model_dtype, self.device
            ).register(self.patterns)
            MiddleAllReduceRMSNormStaticFP8Pattern(
                epsilon, self.model_dtype, self.device
            ).register(self.patterns)

            # Normal RMSNorm patterns
            FirstAllReduceRMSNormPattern(
                epsilon, self.model_dtype, self.device
            ).register(self.patterns)

            MiddleAllReduceRMSNormPattern(
                epsilon, self.model_dtype, self.device
            ).register(self.patterns)

        self.dump_patterns(config, self.patterns)

    def is_applicable_for_range(self, compile_range: Range) -> bool:
        # When sequence parallelism is enabled, the residual tensor from RMSNorm
        # needs to be split along the sequence dimension. However, this dimension
        # is symbolic during piecewise compilation, and splitting symbolic shapes
        # is not supported.
        #
        # This pass is therefore only applied when the sequence dimension is
        # concrete:
        # 1. In full-graph compilation mode (no Dynamo splitting ops are used).
        #   For this case we always pad num_tokens to be a multiple of
        #   tensor_parallel_size, so there's no need to check shape % tp_size == 0.
        # 2. For specific shape provided during compilation (e.g., from
        #    `compile_sizes`), which must be divisible by the tensor-parallel
        #    size.
        if (
            not self.compilation_config.splitting_ops
            or self.compilation_config.use_inductor_graph_partition
        ):
            return True
        tp_size = get_tensor_model_parallel_world_size()
        result: bool = (compile_range.is_single_size()) and (
            compile_range.end % tp_size == 0
        )
        return result

    @VllmInductorPass.time_and_log
    def __call__(self, graph: fx.Graph) -> None:
        self.matched_count = self.patterns.apply(graph)
        logger.debug("Replaced %s patterns", self.matched_count)
        # Clean up reshape nodes
        self.noop_cleanup(graph)

class _SequenceParallelPatternHelper:
    """Helper for sequence parallelism patterns."""

    def __init__(
        self,
        epsilon: float,
        dtype: torch.dtype,
        device: str | None,
    ) -> None:
        self.epsilon = epsilon
        self.dtype = dtype
        self.device = device
        self.tp_group = get_tp_group()
        self.tp_size = get_tensor_model_parallel_world_size()

    def _all_reduce(self, x: torch.Tensor) -> torch.Tensor:
        return tensor_model_parallel_all_reduce(x)

    def _reduce_scatter(self, x: torch.Tensor) -> torch.Tensor:
        return torch.ops.vllm.reduce_scatter.default(
            x, dim=0, world_size=self.tp_size, group_name=self.tp_group.unique_name
        )

    def _all_gather(self, x: torch.Tensor) -> torch.Tensor:
        return torch.ops.vllm.all_gather.default(
            x, dim=0, world_size=self.tp_size, group_name=self.tp_group.unique_name
        )