[LoRA] Support FusedMoE LoRA Triton kernel for mxfp4 (#28971)

Signed-off-by: Xin Yang <xyangx@amazon.com>
Co-authored-by: Jee Jee Li <pandaleefree@gmail.com>

tests/kernels/moe/test_modular_oai_triton_moe.py

+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""
+Test modular OAI Triton MoE
+"""
+
+import pytest
+import torch
+
+from vllm.utils.import_utils import has_triton_kernels
+
+if not has_triton_kernels():
+    pytest.skip(
+        "triton_kernels not found, skipping all related tests",
+        allow_module_level=True,
+    )
+
+from triton_kernels.matmul_ogs import FlexCtx, PrecisionConfig
+from triton_kernels.numerics import InFlexData
+from triton_kernels.numerics_details.mxfp import downcast_to_mxfp, upcast_from_mxfp
+from triton_kernels.tensor import FP4, convert_layout, wrap_torch_tensor
+from triton_kernels.tensor_details import layout
+from triton_kernels.testing import assert_close
+
+from vllm.config import VllmConfig, set_current_vllm_config
+from vllm.model_executor.layers.fused_moe.config import mxfp4_w4a16_moe_quant_config
+from vllm.model_executor.layers.fused_moe.gpt_oss_triton_kernels_moe import (
+    OAITritonExperts,
+    UnfusedOAITritonExperts,
+)
+from vllm.model_executor.layers.fused_moe.modular_kernel import FusedMoEModularKernel
+from vllm.model_executor.layers.fused_moe.prepare_finalize import (
+    MoEPrepareAndFinalizeNoEP,
+)
+from vllm.model_executor.layers.utils import shuffle_weight
+from vllm.platforms import current_platform
+
+MNK = [
+    (1, 512, 384),
+    (1, 2880, 2880),
+    (2, 512, 384),
+    (2, 2880, 2880),
+    (32, 2880, 2880),
+    (64, 2880, 2880),
+]
+
+
+def unshuffle_weight(w: torch.Tensor):
+    first = w[..., ::2]
+    second = w[..., 1::2]
+    return torch.concat((first, second), dim=-1)
+
+
+def make_weights(dtype, k, n, e):
+    w1 = torch.randn((e, k, 2 * n), dtype=dtype, device="cuda")
+    w1_bias = torch.randn((e, 2 * n), dtype=dtype, device="cuda")
+
+    w2 = torch.randn((e, n, k), dtype=dtype, device="cuda")
+    w2_bias = torch.randn((e, k), dtype=dtype, device="cuda")
+
+    w1_tri = w1.clone()
+    w2_tri = w2.clone()
+
+    w1_bias_tri = w1_bias.clone()
+    w2_bias_tri = w2_bias.clone()
+    w1_bias_tri = w1_bias_tri.to(torch.float32)
+    w2_bias_tri = w2_bias_tri.to(torch.float32)
+
+    # shuffle weights
+    w1_tri = shuffle_weight(w1_tri)
+    w1_bias_tri = shuffle_weight(w1_bias_tri)
+
+    # quant triton_weights
+    w1_tri, w1_scale_tri = downcast_to_mxfp(w1_tri, torch.uint8, axis=1)
+    w1 = upcast_from_mxfp(w1_tri, w1_scale_tri, dtype, axis=1)
+    w1 = unshuffle_weight(w1)
+
+    w2_tri, w2_scale_tri = downcast_to_mxfp(w2_tri, torch.uint8, axis=1)
+    w2 = upcast_from_mxfp(w2_tri, w2_scale_tri, dtype, axis=1)
+
+    num_warps = 8
+    w_layout, w_layout_opts = layout.make_default_matmul_mxfp4_w_layout(mx_axis=1)
+    w_scale_layout, w_scale_layout_opts = (
+        layout.make_default_matmul_mxfp4_w_scale_layout(mx_axis=1, num_warps=num_warps)
+    )
+
+    w1_tri = convert_layout(wrap_torch_tensor(w1_tri, FP4), w_layout, **w_layout_opts)
+    w1_scale_tri = convert_layout(
+        wrap_torch_tensor(w1_scale_tri),
+        w_scale_layout,
+        **w_scale_layout_opts,
+    )
+
+    w2_tri = convert_layout(wrap_torch_tensor(w2_tri, FP4), w_layout, **w_layout_opts)
+    w2_scale_tri = convert_layout(
+        wrap_torch_tensor(w2_scale_tri),
+        w_scale_layout,
+        **w_scale_layout_opts,
+    )
+
+    w1_precision_config = PrecisionConfig(
+        weight_scale=w1_scale_tri, flex_ctx=FlexCtx(rhs_data=InFlexData())
+    )
+    w2_precision_config = PrecisionConfig(
+        weight_scale=w2_scale_tri, flex_ctx=FlexCtx(rhs_data=InFlexData())
+    )
+
+    return (
+        w1,
+        w2,
+        w1_bias,
+        w2_bias,
+        w1_tri,
+        w2_tri,
+        w1_bias_tri,
+        w2_bias_tri,
+        w1_precision_config,
+        w2_precision_config,
+    )
+
+
+def swiglu(x, alpha: float = 1.702, limit: float = 1.0):
+    # Note we add an extra bias of 1 to the linear layer
+    x_glu, x_linear = torch.chunk(x, 2, dim=-1)
+    if limit is not None:
+        x_glu = x_glu.clamp(max=limit)
+    out_glu = x_glu * torch.sigmoid(alpha * x_glu)
+    if limit is not None:
+        x_linear = x_linear.clamp(min=-limit, max=limit)
+    return out_glu * (x_linear + 1)
+
+
+def torch_moe_impl(
+    hidden_states: torch.Tensor,  # (M, K)
+    w1: torch.Tensor,  # (E, K, 2N)
+    w2: torch.Tensor,  # (E, N, K)
+    w1_bias: torch.Tensor,  # (E, 2N)
+    w2_bias: torch.Tensor,  # (E, K)
+    topk_weights: torch.Tensor,  # (M, topk)
+    topk_ids: torch.Tensor,  # (M, topk)
+):
+    w1 = w1[topk_ids, ...]
+    w1_bias = w1_bias[topk_ids, ...]
+    hidden_states = torch.einsum("bekc,bk->bec", w1, hidden_states) + w1_bias
+    hidden_states = swiglu(hidden_states, limit=7)
+
+    w2 = w2[topk_ids, ...]
+    w2_bias = w2_bias[topk_ids, ...]
+    hidden_states = torch.einsum("bekc,bek->bec", w2, hidden_states) + w2_bias
+
+    # Weighted sum of experts
+    hidden_states = torch.einsum("bec,be->bc", hidden_states, topk_weights)
+    return hidden_states
+
+
+def oai_triton_moe_impl(
+    x: torch.Tensor,
+    w1: torch.Tensor,
+    w2: torch.Tensor,
+    w1_scale: "PrecisionConfig",
+    w2_scale: "PrecisionConfig",
+    w1_bias: torch.Tensor | None,
+    w2_bias: torch.Tensor | None,
+    num_experts: int,
+    topk_weights: torch.Tensor,
+    topk_ids: torch.Tensor,
+    unfused: bool = False,
+) -> torch.Tensor:
+    quant_config = mxfp4_w4a16_moe_quant_config(
+        w1_bias=w1_bias,
+        w2_bias=w2_bias,
+        w1_scale=w1_scale,
+        w2_scale=w2_scale,
+    )
+
+    if unfused:
+        fused_experts = UnfusedOAITritonExperts(quant_config)
+    else:
+        fused_experts = OAITritonExperts(quant_config)
+
+    mk = FusedMoEModularKernel(MoEPrepareAndFinalizeNoEP(), fused_experts)
+
+    return mk.forward(
+        hidden_states=x,
+        w1=w1,
+        w2=w2,
+        topk_weights=topk_weights,
+        topk_ids=topk_ids,
+        inplace=True,
+        activation="swigluoai",
+        global_num_experts=num_experts,
+        expert_map=None,
+        apply_router_weight_on_input=False,
+    )
+
+
+@pytest.mark.skipif(
+    not current_platform.is_cuda(), reason="This test is skipped on non-CUDA platform."
+)
+@pytest.mark.parametrize("dtype", [torch.bfloat16])
+@pytest.mark.parametrize("m,n,k", MNK)
+@pytest.mark.parametrize("num_experts", [32, 128])
+@pytest.mark.parametrize("topk", [4])
+@pytest.mark.parametrize("unfused", [True, False])
+def test_oai_triton_moe(
+    dtype: torch.dtype,
+    m: int,
+    n: int,
+    k: int,
+    num_experts: int,
+    topk: int,
+    unfused: bool,
+):
+    current_platform.seed_everything(0)
+    (
+        w1,
+        w2,
+        w1_bias,
+        w2_bias,
+        w1_tri,
+        w2_tri,
+        w1_bias_tri,
+        w2_bias_tri,
+        w1_precision_config,
+        w2_precision_config,
+    ) = make_weights(dtype, k, n, num_experts)
+
+    x = torch.randn((m, k), dtype=dtype, device="cuda")
+    router_logits = torch.randn(m, num_experts, device="cuda", dtype=dtype)
+    topk_weights, topk_ids = torch.topk(router_logits, k=topk, dim=-1, sorted=True)
+    topk_weights = torch.nn.functional.softmax(topk_weights, dim=-1)
+
+    with set_current_vllm_config(VllmConfig()):
+        out_ref = torch_moe_impl(x, w1, w2, w1_bias, w2_bias, topk_weights, topk_ids)
+
+        out = oai_triton_moe_impl(
+            x,
+            w1_tri,
+            w2_tri,
+            w1_precision_config,
+            w2_precision_config,
+            w1_bias_tri,
+            w2_bias_tri,
+            num_experts,
+            topk_weights,
+            topk_ids,
+            unfused,
+        )
+
+    assert_close(ref=out_ref, tri=out, maxtol=0.025, rmstol=0.005)

vllm/lora/layers/fused_moe.py

@@ -20,15 +20,24 @@ from vllm.model_executor.layers.fused_moe.config import (
     _get_config_dtype_str,
 )
 from vllm.model_executor.layers.fused_moe.fused_marlin_moe import (
-    modular_marlin_fused_moe,
+    MarlinExperts,
 )
 from vllm.model_executor.layers.fused_moe.fused_moe import (
-    modular_triton_fused_moe,
+    TritonExperts,
     try_get_optimal_moe_config,
 )
 from vllm.model_executor.layers.fused_moe.fused_moe_modular_method import (
     FusedMoEModularMethod,
 )
+from vllm.model_executor.layers.fused_moe.gpt_oss_triton_kernels_moe import (
+    UnfusedOAITritonExperts,
+)
+from vllm.model_executor.layers.fused_moe.modular_kernel import (
+    FusedMoEModularKernel,
+)
+from vllm.model_executor.layers.fused_moe.prepare_finalize import (
+    MoEPrepareAndFinalizeNoEP,
+)
 
 from .utils import _get_lora_device
 
@@ -114,14 +123,22 @@ class FusedMoEWithLoRA(BaseLayerWithLoRA):
         self.base_layer.ensure_moe_quant_config_init()
         quant_config = self.base_layer.quant_method.moe_quant_config
 
-        m_fused_moe_fn = (
-            modular_triton_fused_moe(
-                quant_config, shared_experts=self.base_layer.shared_experts
+        prepare_finalize = MoEPrepareAndFinalizeNoEP()
+        m_fused_moe_fn = FusedMoEModularKernel(
+            prepare_finalize,
+            self.base_layer.quant_method.select_gemm_impl(
+                prepare_finalize, self.base_layer
+            ),
+            self.base_layer.shared_experts,
+            getattr(self.base_layer, "shared_experts_stream", None),
         )
-            if not quant_config.use_mxfp4_w4a16
-            else modular_marlin_fused_moe(
-                quant_config, shared_experts=self.base_layer.shared_experts
+        if quant_config.use_mxfp4_w4a16:
+            assert isinstance(
+                m_fused_moe_fn.fused_experts, (MarlinExperts, UnfusedOAITritonExperts)
             )
+        else:
+            assert isinstance(
+                m_fused_moe_fn.fused_experts, (MarlinExperts, TritonExperts)
             )
 
         def fwd_decorator(layer, func):

vllm/model_executor/layers/fused_moe/gpt_oss_triton_kernels_moe.py

@@ -5,6 +5,7 @@
 import torch
 
 import vllm.model_executor.layers.fused_moe.modular_kernel as mk
+from vllm import _custom_ops as ops
 from vllm.logger import init_logger
 from vllm.model_executor.layers.fused_moe.config import (
     FUSED_MOE_UNQUANTIZED_CONFIG,
@@ -376,3 +377,148 @@ class OAITritonExperts(BaseOAITritonExperts):
             intermediate_cache=workspace2,
             a1q_scale=a1q_scale,
         )
+
+
+class UnfusedOAITritonExperts(BaseOAITritonExperts):
+    """
+    A Triton based MoE expert class that operates on expert standard
+    format and explicitly keeps the activation and reduction (moe_sum) steps
+    unfused from the matmul_ogs kernel. This exposes injection points
+    for activation and moe_sum.
+
+    One use case for it is to inject LoRA modules on the activation and moe_sum.
+    """
+
+    def __init__(self, quant_config: FusedMoEQuantConfig):
+        # TODO (varun) : Enable activation quantization
+        assert quant_config.use_mxfp4_w4a16, "Supports only mxfp4_w4a16"
+        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 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, ...]]:
+        # workspace are allocated inside the kernel
+        workspace1 = (M * topk, N // 2)
+        workspace2 = (M * topk, max(N, K))
+        output = (M, K)
+        return (workspace1, workspace2, output)
+
+    def moe_sum(self, input: torch.Tensor, output: torch.Tensor):
+        ops.moe_sum(input, 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,
+    ):
+        if self.quant_config is None:
+            self.quant_config = FUSED_MOE_UNQUANTIZED_CONFIG
+
+        if expert_map is not None:
+            topk_ids = expert_map[topk_ids]
+
+        local_num_experts = w1.size(0)
+        if global_num_experts == -1:
+            global_num_experts = local_num_experts
+
+        routing_data, gather_indx, scatter_indx = self._make_routing_data(
+            topk_ids, topk_weights, local_num_experts
+        )
+
+        topk = topk_ids.size(1)
+
+        # type check, uint8 means mxfp4
+        assert hidden_states.dtype == torch.bfloat16
+        assert (
+            self.quant_config.w1_bias is None
+            or self.quant_config.w1_bias.dtype == torch.float32
+        )
+        assert (
+            self.quant_config.w2_bias is None
+            or self.quant_config.w2_bias.dtype == torch.float32
+        )
+
+        # Shape check, only check non-mxfp4
+        assert hidden_states.ndim == 2
+        assert hidden_states.shape[-1] == w1.shape[-2]
+        assert w2.shape[-1] == w1.shape[1]
+
+        batch_dim = 1
+        M, K = hidden_states.shape
+        E, _, N = w1.shape
+
+        if global_num_experts == -1:
+            global_num_experts = E
+
+        # Note that the output tensor might be in workspace13
+        intermediate_cache1 = _resize_cache(workspace2, (batch_dim, M * topk, N))
+        intermediate_cache3 = _resize_cache(workspace2, (batch_dim, M * topk, K))
+        intermediate_cache2 = _resize_cache(workspace13, (M * topk, N // 2))
+
+        gammas = routing_data.gate_scal if routing_data else None
+
+        matmul_ogs(
+            hidden_states,
+            w1,
+            self.quant_config.w1_bias,
+            routing_data,
+            gather_indx=gather_indx,
+            precision_config=self.quant_config.w1_precision,
+            gammas=gammas if apply_router_weight_on_input else None,
+            fused_activation=None,
+            y=intermediate_cache1,
+        )
+
+        self.activation(
+            activation, intermediate_cache2, intermediate_cache1.view(-1, N)
+        )
+
+        # matmul_ogs grouped reduction fuse sum across multiple experts:
+        # y[dst_ind // n_expts_act, :] += x[src_ind, :]
+        # Need to set n_expts_act to 1 to unfuse moe_sum
+        routing_data.n_expts_act = 1
+
+        matmul_ogs(
+            intermediate_cache2,
+            w2,
+            self.quant_config.w2_bias,
+            routing_data,
+            scatter_indx=scatter_indx,
+            precision_config=self.quant_config.w2_precision,
+            gammas=None if apply_router_weight_on_input else gammas,
+            y=intermediate_cache3,
+        )
+
+        self.moe_sum(intermediate_cache3.view(-1, topk, K), output)

vllm/model_executor/layers/quantization/mxfp4.py

@@ -30,6 +30,7 @@ from vllm.model_executor.layers.fused_moe.fused_marlin_moe import (
 )
 from vllm.model_executor.layers.fused_moe.gpt_oss_triton_kernels_moe import (
     OAITritonExperts,
+    UnfusedOAITritonExperts,
 )
 from vllm.model_executor.layers.fused_moe.trtllm_moe import TrtLlmGenExperts
 from vllm.model_executor.layers.linear import LinearBase, UnquantizedLinearMethod
@@ -83,9 +84,22 @@ def get_mxfp4_backend_with_lora() -> Mxfp4Backend:
     if not current_platform.is_cuda():
         return Mxfp4Backend.NONE
 
+    # If FlashInfer is not available, try either Marlin or Triton
+    triton_kernels_supported = (
+        has_triton_kernels()
+        and is_torch_equal_or_newer("2.8.0")
+        # NOTE: triton_kernels are only confirmed to work on SM90 and SM100
+        # SM110 fails with this error: https://github.com/vllm-project/vllm/issues/29317
+        # SM120 needs this fix: https://github.com/triton-lang/triton/pull/8498
+        and (9, 0) <= current_platform.get_device_capability() < (11, 0)
+    )
+    if envs.VLLM_MXFP4_USE_MARLIN or not triton_kernels_supported:
         logger.info_once("[get_mxfp4_backend_with_lora] Using Marlin backend")
         return Mxfp4Backend.MARLIN
 
+    logger.info_once("[get_mxfp4_backend_with_lora] Using Triton backend")
+    return Mxfp4Backend.TRITON
+
 
 def get_mxfp4_backend(with_lora_support: bool) -> Mxfp4Backend:
     # Backend Selection
@@ -854,6 +868,8 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
             elif self.mxfp4_backend == Mxfp4Backend.MARLIN:
                 return MarlinExperts(self.moe_quant_config)
             elif self.mxfp4_backend == Mxfp4Backend.TRITON:
+                if self.moe.is_lora_enabled:
+                    return UnfusedOAITritonExperts(self.moe_quant_config)
                 return OAITritonExperts(self.moe_quant_config)
             else:
                 raise NotImplementedError(