[Kernel] Integrate CUTLASS MoE kernel with PPLX (#18762)

Signed-off-by: ElizaWszola <ewszola@redhat.com>
Signed-off-by: Tyler Michael Smith <tyler@neuralmagic.com>
Co-authored-by: Tyler Michael Smith <tyler@neuralmagic.com>

vllm/model_executor/layers/fused_moe/layer.py

@@ -9,6 +9,9 @@ from typing import Callable, Optional, Union
 
 import torch
 import torch.nn.functional as F
+from compressed_tensors.quantization import (QuantizationArgs,
+                                             QuantizationStrategy,
+                                             QuantizationType)
 from torch.nn.parameter import UninitializedParameter
 
 import vllm.envs as envs
@@ -210,6 +213,7 @@ class MoEConfig:
     moe_parallel_config: FusedMoEParallelConfig
 
     in_dtype: torch.dtype  # The activation type.
+    quant_dtype: torch.dtype = None
 
     # TODO: add more quantization params, blocked, per-token, etc.
     block_size: int = 128
@@ -264,8 +268,22 @@ class FusedMoeWeightScaleSupported(Enum):
     BLOCK = "block"
 
 
+def get_quant_config_input_activations(
+        quant_config: Optional[QuantizationConfig]
+) -> Optional[QuantizationArgs]:
+    if (quant_config is not None and hasattr(quant_config, 'target_scheme_map')
+            and "Linear" in quant_config.target_scheme_map and
+            "input_activations" in quant_config.target_scheme_map["Linear"]):
+        return quant_config.target_scheme_map["Linear"].get(
+            "input_activations")
+    else:
+        return None
+
+
 class FusedMoEMethodBase(QuantizeMethodBase):
 
+    moe: MoEConfig
+
     @abstractmethod
     def create_weights(self, layer: torch.nn.Module, num_experts: int,
                        hidden_size: int, intermediate_size_per_partition: int,
@@ -277,6 +295,7 @@ class FusedMoEMethodBase(QuantizeMethodBase):
         all2all_manager = get_ep_group().device_communicator.all2all_manager
         assert all2all_manager is not None
 
+        self.moe = moe
         quant_dtype = None
         act_quant_block_size = None
         from vllm.model_executor.layers.quantization.fp8 import Fp8Config
@@ -297,13 +316,14 @@ class FusedMoEMethodBase(QuantizeMethodBase):
                 # dp_size actually means tp_size, bug in pplx kernels
                 dp_size=all2all_manager.tp_group.world_size,
                 hidden_dim=moe.hidden_dim,
-                hidden_dim_bytes=moe.hidden_dim * moe.in_dtype.itemsize,
+                hidden_dim_bytes=moe.hidden_dim * moe.quant_dtype.itemsize,
                 # For blocked per token: set to
                 #   ceil_div(hidden_dim, block_size) * sizeof(float32)
                 # For per-token: set to sizeof(float32)
-                hidden_dim_scale_bytes=(0 if moe.in_dtype.itemsize != 1 else (
-                    (moe.hidden_dim + moe.block_size - 1) // moe.block_size *
-                    torch.float32.itemsize)),
+                hidden_dim_scale_bytes=(
+                    0 if moe.quant_dtype.itemsize != 1 else
+                    ((moe.hidden_dim + moe.block_size - 1) // moe.block_size *
+                     torch.float32.itemsize)),
             )
 
             # Intranode pplx a2a takes a group name while internode does not.
@@ -313,6 +333,9 @@ class FusedMoEMethodBase(QuantizeMethodBase):
 
             handle = all2all_manager.get_handle(all_to_all_args)
 
+            input_activations = get_quant_config_input_activations(
+                quant_config)
+
             prepare_finalize = PplxPrepareAndFinalize(
                 handle,
                 max_num_tokens=moe.max_num_tokens,
@@ -320,7 +343,10 @@ class FusedMoEMethodBase(QuantizeMethodBase):
                 rank=all2all_manager.rank,
                 # dp_size actually means tp_size, bug in pplx kernels
                 dp_size=all2all_manager.tp_group.world_size,
-                quant_dtype=moe.in_dtype,
+                quant_dtype=moe.quant_dtype,
+                per_act_token=(input_activations.strategy
+                               == QuantizationStrategy.TOKEN
+                               if input_activations is not None else False),
             )
         elif moe.use_deepep_ht_kernels:
             assert moe.dp_size == all2all_manager.dp_world_size
@@ -365,15 +391,15 @@ class FusedMoEMethodBase(QuantizeMethodBase):
         self.topk_indices_dtype = None
         if prepare_finalize is not None:
             self.topk_indices_dtype = prepare_finalize.topk_indices_dtype()
-            experts = self.select_gemm_impl(prepare_finalize)
+            experts = self.select_gemm_impl(prepare_finalize, moe)
             self.fused_experts = FusedMoEModularKernel(
                 prepare_finalize,
                 experts,
             )
 
     def select_gemm_impl(
-        self, prepare_finalize: FusedMoEPrepareAndFinalize
-    ) -> FusedMoEPermuteExpertsUnpermute:
+            self, prepare_finalize: FusedMoEPrepareAndFinalize,
+            moe: Optional[MoEConfig]) -> FusedMoEPermuteExpertsUnpermute:
         # based on the all2all implementation, select the appropriate
         # gemm implementation
         raise NotImplementedError(
@@ -419,7 +445,8 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
         else:
             self.rocm_aiter_fused_experts = None  # type: ignore
 
-    def select_gemm_impl(self, prepare_finalize: FusedMoEPrepareAndFinalize):
+    def select_gemm_impl(self, prepare_finalize: FusedMoEPrepareAndFinalize,
+                         moe: Optional[MoEConfig]):
 
         assert self.fused_experts == fused_experts
 
@@ -809,7 +836,6 @@ class FusedMoE(torch.nn.Module):
         activation: str = "silu",
     ):
         super().__init__()
-
         if params_dtype is None:
             params_dtype = torch.get_default_dtype()
         self.params_dtype = params_dtype
@@ -869,14 +895,24 @@ class FusedMoE(torch.nn.Module):
             from vllm_hpu_extension.ops import DynamicFusedMOE
             self.hpu_fused_moe = DynamicFusedMOE(self.global_num_experts)
 
+        # Only support float8 for now.
+        quant_dtype = params_dtype
+        if quant_config is not None:
+            input_activations = get_quant_config_input_activations(
+                quant_config)
+            if (input_activations is not None
+                    and input_activations.num_bits == 8
+                    and input_activations.type == QuantizationType.FLOAT):
+                quant_dtype = torch.float8_e4m3fn
+
         moe = MoEConfig(
             num_experts=self.global_num_experts,
             experts_per_token=top_k,
             hidden_dim=hidden_size,
             num_local_experts=self.local_num_experts,
             moe_parallel_config=self.moe_parallel_config,
-            # TODO (bnell): this needs to be fixed for quantized types.
             in_dtype=params_dtype,
+            quant_dtype=quant_dtype,
             max_num_tokens=MOE_DP_CHUNK_SIZE,
         )
         self.moe_config = moe

vllm/model_executor/layers/fused_moe/modular_kernel.py

@@ -175,6 +175,7 @@ class FusedMoEPermuteExpertsUnpermute(ABC):
     def workspace_shapes(
         self,
         a: torch.Tensor,
+        aq: torch.Tensor,
         M: int,
         N: int,
         K: int,
@@ -309,7 +310,7 @@ class FusedMoEModularKernel(torch.nn.Module):
 
         # Use a1 here to decipher the correct workspace datatype
         workspace13_shape, workspace2_shape, workspace_dtype = (
-            self.fused_experts.workspace_shapes(a1, M, N, K, top_k,
+            self.fused_experts.workspace_shapes(a1, a1q, M, N, K, top_k,
                                                 global_num_experts))
 
         # We can reuse the memory between cache1 and cache3 because by the time

vllm/model_executor/layers/fused_moe/pplx_prepare_finalize.py

@@ -21,7 +21,8 @@ class PplxPrepareAndFinalize(mk.FusedMoEPrepareAndFinalize):
                  rank: int,
                  dp_size: int,
                  quant_dtype: Optional[torch.dtype] = None,
-                 block_shape: Optional[list[int]] = None):
+                 block_shape: Optional[list[int]] = None,
+                 per_act_token: bool = False):
         super().__init__()
         assert max_num_tokens > 0
         self.a2a = a2a
@@ -31,6 +32,7 @@ class PplxPrepareAndFinalize(mk.FusedMoEPrepareAndFinalize):
         self.rank = rank
         self.dp_size = dp_size
         self.quant_dtype = quant_dtype
+        self.per_act_token = per_act_token
 
     def max_num_tokens_per_rank(self) -> Optional[int]:
         return self.max_num_tokens
@@ -66,13 +68,14 @@ class PplxPrepareAndFinalize(mk.FusedMoEPrepareAndFinalize):
                 "apply_router_weight_on_input is only implemented for topk=1")
             a1 = a1 * rank_topk_weights.to(a1.dtype)
 
-        per_act_token = a1_scale.numel() != 1 if a1_scale is not None else (
-            a2_scale.numel() != 1 if a2_scale is not None else False)
+        repeat_cols = 4
+        repeat_rows = 1 if self.per_act_token else a1.shape[0]
+        a1q, a1q_scale = moe_kernel_quantize_input(
+            a1, (None if self.per_act_token else a1_scale), self.quant_dtype,
+            self.per_act_token, self.block_shape)
 
-        a1q, a1q_scale = moe_kernel_quantize_input(a1, a1_scale,
-                                                   self.quant_dtype,
-                                                   per_act_token,
-                                                   self.block_shape)
+        if a1q_scale is not None:
+            a1q_scale = a1q_scale.repeat(repeat_rows, repeat_cols)
 
         # rem_experts need to be 0 for pplx to work properly.
         rem_experts = num_experts % self.world_size
@@ -100,7 +103,7 @@ class PplxPrepareAndFinalize(mk.FusedMoEPrepareAndFinalize):
                           else 1) * float32_size
             expert_x_scale = torch.empty(
                 (
-                    num_experts,
+                    num_local_experts,
                     expert_x.size(1),
                     (expert_x.size(2) + block_size - 1) // block_size,
                 ),
@@ -121,6 +124,8 @@ class PplxPrepareAndFinalize(mk.FusedMoEPrepareAndFinalize):
             indices=rank_topk_ids,
             bound_m=bound_m,
         )
+        if expert_x_scale is not None:
+            expert_x_scale = expert_x_scale[:, :, 0:1]
 
         return expert_x, expert_x_scale, expert_num_tokens, None, None
 

vllm/model_executor/layers/fused_moe/triton_deep_gemm_moe.py

@@ -37,6 +37,7 @@ class TritonOrDeepGemmExperts(mk.FusedMoEPermuteExpertsUnpermute):
     def workspace_shapes(
         self,
         a: torch.Tensor,
+        aq: torch.Tensor,
         M: int,
         N: int,
         K: int,
@@ -49,9 +50,9 @@ class TritonOrDeepGemmExperts(mk.FusedMoEPermuteExpertsUnpermute):
         if self.allow_deep_gemm and _valid_deep_gemm_shape(M, N, K):
             assert self.deep_gemm_expert is not None
             return self.deep_gemm_expert.workspace_shapes(
-                a, M, N, K, topk, num_experts)
+                a, aq, M, N, K, topk, num_experts)
         else:
-            return self.triton_expert.workspace_shapes(a, M, N, K, topk,
+            return self.triton_expert.workspace_shapes(a, aq, M, N, K, topk,
                                                        num_experts)
 
     def apply(

vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py

@@ -2,6 +2,7 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import enum
+import importlib
 from enum import Enum
 from typing import Callable, Optional
 
@@ -11,7 +12,6 @@ from compressed_tensors.quantization import (ActivationOrdering,
                                              QuantizationStrategy)
 
 import vllm.envs as envs
-import vllm.model_executor.layers.fused_moe  # noqa
 from vllm import _custom_ops as ops
 from vllm.logger import init_logger
 from vllm.model_executor.layers.fused_moe import (FusedMoE, FusedMoEMethodBase,
@@ -30,6 +30,15 @@ from vllm.model_executor.utils import set_weight_attrs
 from vllm.platforms import current_platform
 from vllm.scalar_type import scalar_types
 
+has_pplx = importlib.util.find_spec("pplx_kernels") is not None
+
+if current_platform.is_cuda_alike():
+    from vllm.model_executor.layers.fused_moe.fused_batched_moe import (
+        BatchedPrepareAndFinalize)
+    if has_pplx:
+        from vllm.model_executor.layers.fused_moe.pplx_prepare_finalize import (
+            PplxPrepareAndFinalize)
+
 logger = init_logger(__name__)
 
 
@@ -77,8 +86,7 @@ class CompressedTensorsMoEMethod(FusedMoEMethodBase):
             else:
                 logger.info_once("Using CompressedTensorsWNA16MarlinMoEMethod")
                 return CompressedTensorsWNA16MarlinMoEMethod(quant_config)
-        elif (quant_config._is_fp8_w8a8_sm90(weight_quant, input_quant)
-              and layer.activation == "silu"):
+        elif quant_config._is_fp8_w8a8_sm90(weight_quant, input_quant):
             return CompressedTensorsW8A8Fp8MoECutlassMethod(quant_config)
         elif quant_config._is_fp8_w8a8(weight_quant, input_quant):
             return CompressedTensorsW8A8Fp8MoEMethod(quant_config)
@@ -421,6 +429,11 @@ class CompressedTensorsW8A8Fp8MoECutlassMethod(CompressedTensorsMoEMethod):
                 "For FP8 Fused MoE layer, we require either per tensor or "
                 "channelwise, dynamic per token quantization.")
 
+        from vllm.model_executor.layers.fused_moe.cutlass_moe import (
+            cutlass_moe_fp8)
+        self.fused_experts = cutlass_moe_fp8  # type: ignore
+        self.disable_expert_map = False
+
     def create_weights(self, layer: torch.nn.Module, num_experts: int,
                        hidden_size: int, intermediate_size_per_partition: int,
                        params_dtype: torch.dtype, **extra_weight_attrs):
@@ -499,25 +512,6 @@ class CompressedTensorsW8A8Fp8MoECutlassMethod(CompressedTensorsMoEMethod):
             layer.w13_input_scale = None
             layer.w2_input_scale = None
 
-        device = w13_weight.device
-        # TODO strides can be shared across multiple layers
-        self.ab_strides1 = torch.full((num_experts, ),
-                                      hidden_size,
-                                      device=device,
-                                      dtype=torch.int64)
-        self.c_strides1 = torch.full((num_experts, ),
-                                     2 * intermediate_size_per_partition,
-                                     device=device,
-                                     dtype=torch.int64)
-        self.ab_strides2 = torch.full((num_experts, ),
-                                      intermediate_size_per_partition,
-                                      device=device,
-                                      dtype=torch.int64)
-        self.c_strides2 = torch.full((num_experts, ),
-                                     hidden_size,
-                                     device=device,
-                                     dtype=torch.int64)
-
     def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
         # Fp8 moe kernels require a single activation scale.
         # We take the max of all the scales in case they differ.
@@ -558,6 +552,27 @@ class CompressedTensorsW8A8Fp8MoECutlassMethod(CompressedTensorsMoEMethod):
             layer.w13_weight_scale = torch.nn.Parameter(max_w13_scales,
                                                         requires_grad=False)
 
+    def select_gemm_impl(self, prepare_finalize, moe):
+        from vllm.model_executor.layers.fused_moe.cutlass_moe import (
+            CutlassExpertsFp8)
+
+        assert moe is not None
+
+        max_experts_per_worker = (
+            (moe.num_experts + prepare_finalize.world_size - 1) //
+            prepare_finalize.world_size)
+        experts = CutlassExpertsFp8(
+            max_experts_per_worker, moe.in_dtype,
+            self.input_quant.strategy == QuantizationStrategy.TOKEN,
+            self.weight_quant.strategy == QuantizationStrategy.CHANNEL)
+
+        if has_pplx and isinstance(
+                prepare_finalize,
+            (BatchedPrepareAndFinalize, PplxPrepareAndFinalize)):
+            # no expert_map support in this case
+            self.disable_expert_map = True
+        return experts
+
     def apply(
         self,
         layer: torch.nn.Module,
@@ -577,9 +592,6 @@ class CompressedTensorsW8A8Fp8MoECutlassMethod(CompressedTensorsMoEMethod):
         activation: str = "silu",
     ) -> torch.Tensor:
 
-        assert activation == "silu", (
-            f"{activation} not supported for Cutlass MoE.")
-
         topk_weights, topk_ids = FusedMoE.select_experts(
             hidden_states=x,
             router_logits=router_logits,
@@ -590,27 +602,22 @@ class CompressedTensorsW8A8Fp8MoECutlassMethod(CompressedTensorsMoEMethod):
             num_expert_group=num_expert_group,
             custom_routing_function=custom_routing_function,
             scoring_func=scoring_func,
-            e_score_correction_bias=e_score_correction_bias)
-
-        from vllm.model_executor.layers.fused_moe import cutlass_moe_fp8
+            e_score_correction_bias=e_score_correction_bias,
+            indices_type=torch.uint32)
 
-        return cutlass_moe_fp8(
+        return self.fused_experts(
             x,
-            layer.w13_weight.transpose(1, 2),
-            layer.w2_weight.transpose(1, 2),
-            layer.w13_weight_scale,
-            layer.w2_weight_scale,
+            layer.w13_weight,
+            layer.w2_weight,
             topk_weights,
             topk_ids,
-            self.ab_strides1,
-            self.c_strides1,
-            self.ab_strides2,
-            self.c_strides2,
+            activation=activation,
+            global_num_experts=global_num_experts,
+            expert_map=None if self.disable_expert_map else expert_map,
+            w1_scale=layer.w13_weight_scale,
+            w2_scale=layer.w2_weight_scale,
             a1_scale=layer.w13_input_scale,
             a2_scale=layer.w2_input_scale,
-            out_dtype=x.dtype,
-            expert_map=expert_map,
-            apply_router_weight_on_input=apply_router_weight_on_input,
         )
 
 

vllm/model_executor/layers/quantization/fp8.py

@@ -769,7 +769,7 @@ class Fp8MoEMethod(FusedMoEMethodBase):
             del layer.w13_input_scale
             del layer.w2_input_scale
 
-    def select_gemm_impl(self, prepare_finalize):
+    def select_gemm_impl(self, prepare_finalize, moe):
 
         from vllm.model_executor.layers.fused_moe.batched_triton_or_deep_gemm_moe import (  # noqa: E501
             BatchedTritonOrDeepGemmExperts)