[AMD][ROCm] MoRI EP: a high-performance all2all backend

vllm/config/parallel.py

@@ -123,6 +123,7 @@ class ParallelConfig:
             "pplx",
             "deepep_high_throughput",
             "deepep_low_latency",
+            "mori",
             "allgather_reducescatter",
             "flashinfer_all2allv",
         ]
@@ -135,6 +136,7 @@ class ParallelConfig:
     - "pplx": Use pplx kernels
     - "deepep_high_throughput": Use deepep high-throughput kernels
     - "deepep_low_latency": Use deepep low-latency kernels
+    - "mori": Use mori kernels
     - "flashinfer_all2allv": Use flashinfer alltoallv kernels for mnnvl"""
     num_redundant_experts: int | None = None
     """`num_redundant_experts` is deprecated and has been replaced with
@@ -370,6 +372,7 @@ def use_sequence_parallel_moe(self) -> bool:
                 "naive",
                 "deepep_high_throughput",
                 "deepep_low_latency",
+                "mori",
             )
             and self.enable_expert_parallel
             and self.tensor_parallel_size > 1

vllm/distributed/device_communicators/all2all.py

@@ -10,7 +10,7 @@
 from vllm.forward_context import get_forward_context
 from vllm.logger import init_logger
 from vllm.utils.flashinfer import has_flashinfer_all2all
-from vllm.utils.import_utils import has_deep_ep, has_pplx
+from vllm.utils.import_utils import has_deep_ep, has_mori, has_pplx
 
 from .base_device_communicator import All2AllManagerBase, Cache
 
@@ -488,3 +488,80 @@ def cleanup(self):
                 self.prepare_workspace_tensor = None
                 self.mapping = None
                 self.initialized = False
+
+
+class MoriAll2AllManager(All2AllManagerBase):
+    def __init__(self, cpu_group):
+        assert has_mori(), (
+            "MoRI kernels not found. Please follow https://github.com/ROCm/mori/blob/main/README.md"
+            " to install MoRI kernels."
+        )  # noqa
+        import mori
+
+        super().__init__(cpu_group)
+        self.handle_cache = Cache()
+
+        torch._C._distributed_c10d._register_process_group("mori", cpu_group)
+        mori.shmem.shmem_torch_process_group_init("mori")
+
+    def _make_all2all_kwargs(
+        self,
+        rank: int,
+        num_ep_ranks: int,
+        input_dtype: torch.dtype,
+        quant_dtype: torch.dtype,
+        token_hidden_size: int,
+        scale_dim: int,
+        scale_type_size: int,
+        max_num_tokens_per_dp_rank: int,
+        num_local_experts: int,
+        num_experts_per_token: int,
+    ):
+        import mori  # type: ignore[import-not-found]
+
+        if not self.internode:
+            # single node
+            kernel_type = mori.ops.EpDispatchCombineKernelType.IntraNode
+            warp_num_per_block = 16
+            block_num = 80
+            rdma_block_num = 0
+        else:
+            # multi node
+            kernel_type = mori.ops.EpDispatchCombineKernelType.InterNodeV1
+            warp_num_per_block = 16
+            block_num = 32
+            rdma_block_num = 16
+
+        return dict(
+            rank=rank,
+            world_size=num_ep_ranks,
+            data_type=quant_dtype,
+            hidden_dim=token_hidden_size,
+            scale_dim=scale_dim,
+            scale_type_size=scale_type_size,
+            max_token_type_size=input_dtype.itemsize,
+            max_num_inp_token_per_rank=max_num_tokens_per_dp_rank,
+            num_experts_per_rank=num_local_experts,
+            num_experts_per_token=num_experts_per_token,
+            warp_num_per_block=warp_num_per_block,
+            block_num=block_num,
+            kernel_type=kernel_type,
+            rdma_block_num=rdma_block_num,
+        )
+
+    def _make_handle(self, **kwargs):
+        import mori  # type: ignore[import-not-found]
+
+        mori_config = mori.ops.EpDispatchCombineConfig(**kwargs)
+        handle = mori.ops.EpDispatchCombineOp(mori_config)
+        return handle
+
+    def get_handle(self, kwargs):
+        import mori  # type: ignore[import-not-found]
+
+        mori_kwargs = self._make_all2all_kwargs(**kwargs)
+        logger.debug("MoRI all2all args %s", mori_kwargs)
+        handle: mori.ops.EpDispatchCombineOp = self.handle_cache.get_or_create(
+            mori_kwargs, self._make_handle
+        )
+        return handle

vllm/distributed/device_communicators/cuda_communicator.py

@@ -110,6 +110,10 @@ def __init__(
                 from .all2all import DeepEPLLAll2AllManager
 
                 self.all2all_manager = DeepEPLLAll2AllManager(self.cpu_group)
+            elif self.all2all_backend == "mori":
+                from .all2all import MoriAll2AllManager
+
+                self.all2all_manager = MoriAll2AllManager(self.cpu_group)
             elif self.all2all_backend == "flashinfer_all2allv":
                 from .all2all import FlashInferAllToAllManager
 

vllm/envs.py

@@ -111,7 +111,7 @@
     VLLM_ROCM_USE_TRITON_ROPE: bool = False
     VLLM_ROCM_USE_AITER_FP8BMM: bool = True
     VLLM_ROCM_USE_AITER_UNIFIED_ATTENTION: bool = False
-    VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS: bool = True
+    VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS: bool = False
     VLLM_ROCM_USE_SKINNY_GEMM: bool = True
     VLLM_ROCM_FP8_PADDING: bool = True
     VLLM_ROCM_MOE_PADDING: bool = True
@@ -166,6 +166,7 @@
         "pplx",
         "deepep_high_throughput",
         "deepep_low_latency",
+        "mori",
         "allgather_reducescatter",
         "flashinfer_all2allv",
     ] = "allgather_reducescatter"
@@ -935,7 +936,7 @@ def get_vllm_port() -> int | None:
     # Whether to use aiter fusion shared experts ops.
     # By default is enabled.
     "VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS": lambda: (
-        os.getenv("VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS", "True").lower()
+        os.getenv("VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS", "False").lower()
         in ("true", "1")
     ),
     # use rocm skinny gemms
@@ -1197,6 +1198,7 @@ def get_vllm_port() -> int | None:
     # - "pplx": use pplx kernels
     # - "deepep_high_throughput", use deepep high-throughput kernels
     # - "deepep_low_latency", use deepep low-latency kernels
+    # - "mori", use MoRI kernels
     # - "flashinfer_all2allv", use flashinfer alltoallv kernels for mnnvl
     "VLLM_ALL2ALL_BACKEND": env_with_choices(
         "VLLM_ALL2ALL_BACKEND",
@@ -1206,6 +1208,7 @@ def get_vllm_port() -> int | None:
             "pplx",
             "deepep_high_throughput",
             "deepep_low_latency",
+            "mori",
             "allgather_reducescatter",
             "flashinfer_all2allv",
         ],

vllm/model_executor/layers/fused_moe/__init__.py

@@ -64,6 +64,9 @@ def get_config() -> dict[str, Any] | None:
         cutlass_moe_fp8,
     )
     from vllm.model_executor.layers.fused_moe.deep_gemm_moe import DeepGemmExperts
+    from vllm.model_executor.layers.fused_moe.fused_aiter_moe import (
+        AiterExperts,
+    )
     from vllm.model_executor.layers.fused_moe.fused_batched_moe import (
         BatchedTritonExperts,
     )
@@ -93,6 +96,7 @@ def get_config() -> dict[str, Any] | None:
         "BatchedDeepGemmExperts",
         "TritonOrDeepGemmExperts",
         "BatchedTritonOrDeepGemmExperts",
+        "AiterExperts",
     ]
 else:
     # Some model classes directly use the custom ops. Add placeholders

vllm/model_executor/layers/fused_moe/config.py

@@ -691,6 +691,10 @@ def use_deepep_ht_kernels(self):
     def use_deepep_ll_kernels(self):
         return self.use_all2all_kernels and self.all2all_backend == "deepep_low_latency"
 
+    @property
+    def use_mori_kernels(self):
+        return self.use_all2all_kernels and self.all2all_backend == "mori"
+
     @staticmethod
     def flatten_tp_across_dp(
         tp_size: int, dp_size: int, dp_rank: int
@@ -883,6 +887,10 @@ def use_deepep_ht_kernels(self):
     def use_deepep_ll_kernels(self):
         return self.moe_parallel_config.use_deepep_ll_kernels
 
+    @property
+    def use_mori_kernels(self):
+        return self.moe_parallel_config.use_mori_kernels
+
     @property
     def use_flashinfer_cutlass_kernels(self):
         """

vllm/model_executor/layers/fused_moe/fused_aiter_moe.py

@@ -0,0 +1,90 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import torch
+
+import vllm.model_executor.layers.fused_moe.modular_kernel as mk
+from vllm.model_executor.layers.fused_moe.config import FusedMoEQuantConfig
+from vllm.model_executor.layers.fused_moe.rocm_aiter_fused_moe import (
+    rocm_aiter_fused_experts,
+)
+from vllm.model_executor.layers.fused_moe.topk_weight_and_reduce import (
+    TopKWeightAndReduceNoOP,
+)
+
+
+class AiterExperts(mk.FusedMoEPermuteExpertsUnpermute):
+    def __init__(self, quant_config: FusedMoEQuantConfig):
+        super().__init__(quant_config)
+
+    @property
+    def activation_formats(
+        self,
+    ) -> tuple[mk.FusedMoEActivationFormat, mk.FusedMoEActivationFormat]:
+        return (
+            mk.FusedMoEActivationFormat.Standard,
+            mk.FusedMoEActivationFormat.Standard,
+        )
+
+    def supports_chunking(self) -> bool:
+        return True
+
+    def supports_expert_map(self) -> bool:
+        return True
+
+    def finalize_weight_and_reduce_impl(self) -> mk.TopKWeightAndReduce:
+        return TopKWeightAndReduceNoOP()
+
+    def workspace_shapes(
+        self,
+        M: int,
+        N: int,
+        K: int,
+        topk: int,
+        global_num_experts: int,
+        local_num_experts: int,
+        expert_tokens_meta: mk.ExpertTokensMetadata | None,
+    ) -> tuple[tuple[int, ...], tuple[int, ...], tuple[int, ...]]:
+        workspace13 = (M, K)
+        workspace2 = (0,)
+        output = (M, K)
+        return (workspace13, workspace2, output)
+
+    def apply(
+        self,
+        output: torch.Tensor,
+        hidden_states: torch.Tensor,
+        w1: torch.Tensor,
+        w2: torch.Tensor,
+        topk_weights: torch.Tensor,
+        topk_ids: torch.Tensor,
+        activation: str,
+        global_num_experts: int,
+        expert_map: torch.Tensor | None,
+        a1q_scale: torch.Tensor | None,
+        a2_scale: torch.Tensor | None,
+        workspace13: torch.Tensor,
+        workspace2: torch.Tensor,
+        expert_tokens_meta: mk.ExpertTokensMetadata | None,
+        apply_router_weight_on_input: bool,
+    ) -> None:
+        if expert_tokens_meta is not None:
+            num_local_tokens = expert_tokens_meta.expert_num_tokens
+        else:
+            num_local_tokens = None
+
+        result = rocm_aiter_fused_experts(
+            hidden_states,
+            w1,
+            w2,
+            topk_weights=topk_weights,
+            topk_ids=topk_ids,
+            activation=activation,
+            apply_router_weight_on_input=apply_router_weight_on_input,
+            expert_map=expert_map,
+            quant_config=self.quant_config,
+            a1q_scale=a1q_scale,
+            num_local_tokens=num_local_tokens,
+            output_dtype=output.dtype,
+        )
+        output.copy_(result)

vllm/model_executor/layers/fused_moe/layer.py

@@ -56,7 +56,7 @@
 from vllm.platforms import current_platform
 from vllm.platforms.interface import CpuArchEnum
 from vllm.utils.flashinfer import has_flashinfer_cutlass_fused_moe
-from vllm.utils.import_utils import has_deep_ep, has_pplx
+from vllm.utils.import_utils import has_deep_ep, has_mori, has_pplx
 from vllm.utils.math_utils import cdiv, round_up
 from vllm.utils.torch_utils import current_stream, direct_register_custom_op
 from vllm.v1.worker.ubatching import dbo_current_ubatch_id
@@ -76,6 +76,8 @@
             DEEPEP_QUANT_BLOCK_SHAPE,
             DeepEPLLPrepareAndFinalize,
         )
+    if has_mori():
+        from .mori_prepare_finalize import MoriPrepareAndFinalize
 else:
     fused_experts = None  # type: ignore
     FusedMoEPermuteExpertsUnpermute = object  # type: ignore
@@ -233,6 +235,36 @@ def _maybe_make_prepare_finalize(
                 use_fp8_dispatch=use_fp8_dispatch,
             )
 
+        elif moe.use_mori_kernels:
+            assert quant_config is not None
+            # For PTPC (per token per channel) quant, the scale dim for each token is 1
+            # For 1x128 quant, the scale dim for each token is hidden_dim // 128
+            scale_dim = 1 if quant_config.is_per_act_token else moe.hidden_dim // 128
+            all_to_all_args = dict(
+                rank=all2all_manager.rank,
+                num_ep_ranks=all2all_manager.world_size,
+                quant_dtype=quant_config.quant_dtype,
+                token_hidden_size=moe.hidden_dim,
+                scale_dim=scale_dim,
+                scale_type_size=torch.float32.itemsize,
+                max_num_tokens_per_dp_rank=moe.max_num_tokens,
+                input_dtype=moe.in_dtype,
+                num_local_experts=moe.num_experts // all2all_manager.world_size,
+                num_experts_per_token=moe.experts_per_token,
+            )
+            handle = all2all_manager.get_handle(all_to_all_args)
+
+            # Note: We may want to use FP8 dispatch just to reduce
+            # data movement.
+            use_fp8_dispatch = is_rocm_aiter_moe_enabled()
+
+            prepare_finalize = MoriPrepareAndFinalize(
+                handle,
+                max_tokens_per_rank=moe.max_num_tokens,
+                num_dispatchers=all2all_manager.world_size,
+                use_fp8_dispatch=use_fp8_dispatch,
+            )
+
         return prepare_finalize
 
     def maybe_make_prepare_finalize(self) -> FusedMoEPrepareAndFinalize | None:
@@ -1407,6 +1439,11 @@ def __init__(
             is_act_and_mul=is_act_and_mul,
             is_lora_enabled=vllm_config.lora_config is not None,
         )
+        if self.use_mori_kernels:
+            assert not is_rocm_aiter_fusion_shared_expert_enabled(), (
+                "Mori does not support fusion shared expert now. "
+                "Turn it off by setting VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS=0"
+            )
 
         self.moe_quant_config: FusedMoEQuantConfig | None = None
         self.quant_config = quant_config
@@ -1538,6 +1575,10 @@ def use_deepep_ht_kernels(self):
     def use_deepep_ll_kernels(self):
         return self.moe_parallel_config.use_deepep_ll_kernels
 
+    @property
+    def use_mori_kernels(self):
+        return self.moe_parallel_config.use_mori_kernels
+
     @property
     def use_flashinfer_cutlass_kernels(self):
         return (
@@ -1551,6 +1592,7 @@ def use_dp_chunking(self) -> bool:
         return (
             self.moe_parallel_config.use_pplx_kernels
             or self.moe_parallel_config.use_deepep_ll_kernels
+            or self.moe_parallel_config.use_mori_kernels
             or (self.dp_size > 1 and self.use_flashinfer_cutlass_kernels)
         )