vllm.compilation.sequence_parallelism ¶

class FirstAllReduceRMSNormPattern(_SequenceParallelPatternHelper):
    def get_inputs(self):
        input = torch.empty([1, 8, 4], device=self.device, dtype=self.dtype)
        permute = torch.empty([1, 8, 4], device=self.device, dtype=self.dtype)
        arg3_1 = torch.empty([4], device=self.device, dtype=self.dtype)

        return [input, permute, arg3_1]

    def register(self, pm_pass: PatternMatcherPass):
        def pattern(
            input: torch.Tensor,
            permute: torch.Tensor,
            arg3_1: torch.Tensor,
        ):
            all_reduce = self._all_reduce(input)
            rmsnorm = self._functional_rmsnorm(permute, all_reduce, arg3_1)

            return rmsnorm[1], all_reduce

        def replacement(
            input: torch.Tensor,
            permute: torch.Tensor,
            arg3_1: torch.Tensor,
        ):
            reduce_scatter = self._reduce_scatter(input)

            rmsnorm_result = torch.empty_like(reduce_scatter)
            rmsnorm = self._functional_rmsnorm(rmsnorm_result, reduce_scatter, arg3_1)

            all_gather = self._all_gather(rmsnorm[1])

            return all_gather, reduce_scatter

        pm.register_replacement(
            pattern, replacement, self.get_inputs(), pm.fwd_only, pm_pass
        )

class FirstAllReduceRMSNormStaticFP8Pattern(_SequenceParallelPatternHelper):
    def __init__(
        self, epsilon: float, dtype: torch.dtype, device: str, op: torch._ops.OpOverload
    ):
        super().__init__(epsilon, dtype, device, quant_op=op)

    def get_inputs(self):
        input = torch.zeros([1, 8, 4], device=self.device, dtype=self.dtype)
        rmsnorm_result = torch.empty([1, 8, 4], device=self.device, dtype=self.dtype)
        quant_result = torch.empty([1, 8, 4], device=self.device, dtype=FP8_DTYPE)
        weight = torch.empty([4], device=self.device, dtype=self.dtype)
        scale = torch.tensor(1.0, device=self.device, dtype=torch.float32)
        return [input, rmsnorm_result, quant_result, weight, scale]

    def register(self, pm_pass: PatternMatcherPass):
        def pattern(
            input: torch.Tensor,
            rmsnorm_result: torch.Tensor,
            quant_result: torch.Tensor,
            weight: torch.Tensor,
            scale: torch.Tensor,
        ):
            all_reduce = self._all_reduce(input)
            static_fp8 = self._functional_rmsnorm_then_quant(
                rmsnorm_result, quant_result, all_reduce, weight, scale
            )
            return static_fp8[1], all_reduce

        def replacement(
            input: torch.Tensor,
            rmsnorm_result: torch.Tensor,
            quant_result: torch.Tensor,
            weight: torch.Tensor,
            scale: torch.Tensor,
        ):
            reduce_scatter = self._reduce_scatter(input)

            rmsnorm_result = torch.empty_like(
                reduce_scatter, dtype=rmsnorm_result.dtype
            )
            quant_result = torch.empty_like(
                rmsnorm_result,  # Output of RMSNorm
                dtype=quant_result.dtype,
            )
            static_fp8 = self._functional_rmsnorm_then_quant(
                rmsnorm_result, quant_result, reduce_scatter, weight, scale
            )
            all_gather = self._all_gather(static_fp8[1])

            return all_gather, reduce_scatter

        pm.register_replacement(
            pattern, replacement, self.get_inputs(), pm.fwd_only, pm_pass
        )

class LastAllReduceRMSNormPattern(_SequenceParallelPatternHelper):
    def get_inputs(self):
        mm_1 = torch.empty([4, 4], device=self.device, dtype=self.dtype)

        residual = torch.empty([4, 4], device=self.device, dtype=self.dtype)
        rms_norm_weights = torch.empty([4, 4], device=self.device, dtype=self.dtype)

        return [
            residual,
            mm_1,
            rms_norm_weights,
        ]

    def register(self, pm_pass: PatternMatcherPass):
        def pattern(
            residual: torch.Tensor,
            mm_1: torch.Tensor,
            rms_norm_weights: torch.Tensor,
        ) -> tuple[torch.Tensor, torch.Tensor]:
            all_reduce = self._all_reduce(mm_1)
            rmsnorm = self._functional_fused_add_rmsnorm(
                all_reduce, residual, rms_norm_weights
            )
            return rmsnorm[1]

        def replacement(
            residual: torch.Tensor,
            mm_1: torch.Tensor,
            rms_norm_weights: torch.Tensor,
        ) -> tuple[torch.Tensor, torch.Tensor]:
            reduce_scatter = self._reduce_scatter(mm_1)
            rmsnorm = self._functional_fused_add_rmsnorm(
                reduce_scatter, residual, rms_norm_weights
            )
            normalized = self._all_gather(rmsnorm[1])
            return normalized

        pm.register_replacement(
            pattern, replacement, self.get_inputs(), pm.fwd_only, pm_pass
        )

class LastAllReduceRMSNormStaticFP8Pattern(_SequenceParallelPatternHelper):
    def __init__(
        self, epsilon: float, dtype: torch.dtype, device: str, op: torch._ops.OpOverload
    ):
        super().__init__(epsilon, dtype, device, quant_op=op)

    def get_inputs(self):
        mm_1 = torch.empty([4, 4], device=self.device, dtype=self.dtype)

        residual = torch.empty([4, 4], device=self.device, dtype=self.dtype)
        rms_norm_weights = torch.empty([4, 4], device=self.device, dtype=self.dtype)
        result = torch.empty([4, 4], device=self.device, dtype=FP8_DTYPE)
        scale = torch.empty([1, 1], device=self.device, dtype=torch.float32)

        return [
            result,
            residual,
            mm_1,
            rms_norm_weights,
            scale,
        ]

    def register(self, pm_pass: PatternMatcherPass):
        def pattern(
            result: torch.Tensor,
            residual: torch.Tensor,
            mm_1: torch.Tensor,
            rms_norm_weights: torch.Tensor,
            scale: torch.Tensor,
        ) -> tuple[torch.Tensor, torch.Tensor]:
            all_reduce = self._all_reduce(mm_1)
            static_fp8, _ = self._functional_fused_add_rmsnorm_then_quant(
                result, all_reduce, residual, rms_norm_weights, scale
            )
            return static_fp8[1]

        def replacement(
            result: torch.Tensor,
            residual: torch.Tensor,
            mm_1: torch.Tensor,
            rms_norm_weights: torch.Tensor,
            scale: torch.Tensor,
        ) -> tuple[torch.Tensor, torch.Tensor]:
            reduce_scatter = self._reduce_scatter(mm_1)
            quant_result_buf = torch.empty_like(reduce_scatter, dtype=result.dtype)
            static_fp8, _ = self._functional_fused_add_rmsnorm_then_quant(
                quant_result_buf, reduce_scatter, residual, rms_norm_weights, scale
            )
            normalized = self._all_gather(static_fp8[1])
            return normalized

        pm.register_replacement(
            pattern, replacement, self.get_inputs(), pm.fwd_only, pm_pass
        )

class MiddleAllReduceRMSNormPattern(_SequenceParallelPatternHelper):
    def get_inputs(self):
        mm_1 = torch.empty([4, 4], device=self.device, dtype=self.dtype)

        residual = torch.empty([4, 4], device=self.device, dtype=self.dtype)
        rms_norm_weights = torch.empty([4, 4], device=self.device, dtype=self.dtype)

        return [
            residual,
            mm_1,
            rms_norm_weights,
        ]

    def register(self, pm_pass: PatternMatcherPass):
        def pattern(
            residual: torch.Tensor,
            mm_1: torch.Tensor,
            rms_norm_weights: torch.Tensor,
        ) -> tuple[torch.Tensor, torch.Tensor]:
            all_reduce = self._all_reduce(mm_1)
            rmsnorm = self._functional_fused_add_rmsnorm(
                all_reduce, residual, rms_norm_weights
            )
            return rmsnorm[1], rmsnorm[2]

        def replacement(
            residual: torch.Tensor,
            mm_1: torch.Tensor,
            rms_norm_weights: torch.Tensor,
        ) -> tuple[torch.Tensor, torch.Tensor]:
            reduce_scatter = self._reduce_scatter(mm_1)
            rmsnorm = self._functional_fused_add_rmsnorm(
                reduce_scatter, residual, rms_norm_weights
            )
            all_gather = self._all_gather(rmsnorm[1])
            return all_gather, rmsnorm[2]

        pm.register_replacement(
            pattern, replacement, self.get_inputs(), pm.fwd_only, pm_pass
        )

class MiddleAllReduceRMSNormStaticFP8Pattern(_SequenceParallelPatternHelper):
    def __init__(
        self, epsilon: float, dtype: torch.dtype, device: str, op: torch._ops.OpOverload
    ):
        super().__init__(epsilon, dtype, device, quant_op=op)

    def get_inputs(self):
        mm_1 = torch.empty([4, 4], device=self.device, dtype=self.dtype)

        residual = torch.empty([4, 4], device=self.device, dtype=self.dtype)
        rms_norm_weights = torch.empty([4, 4], device=self.device, dtype=self.dtype)
        result = torch.empty([4, 4], device=self.device, dtype=FP8_DTYPE)
        scale = torch.empty([1, 1], device=self.device, dtype=torch.float32)

        return [
            result,
            residual,
            mm_1,
            rms_norm_weights,
            scale,
        ]

    def register(self, pm_pass: PatternMatcherPass):
        def pattern(
            result: torch.Tensor,
            residual: torch.Tensor,
            mm_1: torch.Tensor,
            rms_norm_weights: torch.Tensor,
            scale: torch.Tensor,
        ) -> tuple[torch.Tensor, torch.Tensor]:
            all_reduce = self._all_reduce(mm_1)
            static_fp8, rmsnorm_residual_out = (
                self._functional_fused_add_rmsnorm_then_quant(  # noqa: E501
                    result, all_reduce, residual, rms_norm_weights, scale
                )
            )
            return static_fp8[1], rmsnorm_residual_out

        def replacement(
            result: torch.Tensor,
            residual: torch.Tensor,
            mm_1: torch.Tensor,
            rms_norm_weights: torch.Tensor,
            scale: torch.Tensor,
        ) -> tuple[torch.Tensor, torch.Tensor]:
            reduce_scatter = self._reduce_scatter(mm_1)
            quant_result_buf = torch.empty_like(reduce_scatter, dtype=result.dtype)
            static_fp8, rmsnorm_residual_out = (
                self._functional_fused_add_rmsnorm_then_quant(  # noqa: E501
                    quant_result_buf, reduce_scatter, residual, rms_norm_weights, scale
                )
            )
            all_gather = self._all_gather(static_fp8[1])
            return all_gather, rmsnorm_residual_out

        pm.register_replacement(
            pattern, replacement, self.get_inputs(), pm.fwd_only, pm_pass
        )

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.
    """

    @enable_fake_mode
    def __init__(self, config: VllmConfig):
        super().__init__(config)

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

        for epsilon in [1e-5, 1e-6]:
            # RMSNorm + Static FP8 quantization patterns
            fp8_quant_op = torch.ops._C.static_scaled_fp8_quant.default
            FirstAllReduceRMSNormStaticFP8Pattern(
                epsilon, self.model_dtype, self.device, fp8_quant_op
            ).register(self.patterns)
            MiddleAllReduceRMSNormStaticFP8Pattern(
                epsilon, self.model_dtype, self.device, fp8_quant_op
            ).register(self.patterns)
            LastAllReduceRMSNormStaticFP8Pattern(
                epsilon, self.model_dtype, self.device, fp8_quant_op
            ).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)

            LastAllReduceRMSNormPattern(
                epsilon, self.model_dtype, self.device
            ).register(self.patterns)
        self.dump_patterns(config, self.patterns)

    def is_applicable(self, shape: int | None) -> 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()
        return shape is not None and shape % tp_size == 0

    @VllmInductorPass.time_and_log
    def __call__(self, graph: fx.Graph):
        self.matched_count = self.patterns.apply(graph)
        logger.debug("Replaced %s patterns", self.matched_count)