[CustomOp] CustomOp FusedRMSNormGated (#35877)

Signed-off-by: Elias Ellison <elias.ellison@gmail.com>
Signed-off-by: eellison <elias.ellison@gmail.com>

tests/kernels/core/test_fused_rms_norm_gated.py

@@ -0,0 +1,103 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+"""Tests that FusedRMSNormGated decomposes correctly under torch.compile,
+matching the eager triton kernel output."""
+
+import pytest
+import torch
+
+from vllm.model_executor.layers.fla.ops.kda import FusedRMSNormGated
+from vllm.utils.torch_utils import set_random_seed
+
+DTYPES = [torch.bfloat16]
+HIDDEN_SIZES = [128, 512]
+NUM_TOKENS = [64, 128]
+ACTIVATIONS = ["swish", "sigmoid"]
+ELEMENTWISE_AFFINE = [True, False]
+SEEDS = [0]
+
+
+@pytest.mark.parametrize("num_tokens", NUM_TOKENS)
+@pytest.mark.parametrize("hidden_size", HIDDEN_SIZES)
+@pytest.mark.parametrize("activation", ACTIVATIONS)
+@pytest.mark.parametrize("elementwise_affine", ELEMENTWISE_AFFINE)
+@pytest.mark.parametrize("dtype", DTYPES)
+@pytest.mark.parametrize("seed", SEEDS)
+@torch.inference_mode()
+def test_compiled_vs_eager(
+    default_vllm_config,
+    num_tokens: int,
+    hidden_size: int,
+    activation: str,
+    elementwise_affine: bool,
+    dtype: torch.dtype,
+    seed: int,
+) -> None:
+    """forward_native decomposition matches forward_cuda triton kernel."""
+    torch._dynamo.reset()
+    set_random_seed(seed)
+    device = torch.device("cuda:0")
+
+    module = FusedRMSNormGated(
+        hidden_size,
+        elementwise_affine=elementwise_affine,
+        eps=1e-5,
+        activation=activation,
+        device=device,
+        dtype=dtype,
+    )
+    x = torch.randn(num_tokens, hidden_size, dtype=dtype, device=device)
+    g = torch.randn(num_tokens, hidden_size, dtype=dtype, device=device)
+
+    # forward_cuda may modify x in-place, so clone inputs
+    cuda_out = module.forward_cuda(x.clone(), g.clone())
+    compiled_native = torch.compile(module.forward_native, fullgraph=True)
+    native_out = compiled_native(x.clone(), g.clone())
+
+    torch.testing.assert_close(native_out, cuda_out, atol=1e-3, rtol=1e-2)
+
+
+@pytest.mark.parametrize(
+    "shape",
+    [
+        (1, 16, 32, 128),
+        (2, 8, 16, 64),
+    ],
+)
+@pytest.mark.parametrize("activation", ACTIVATIONS)
+@pytest.mark.parametrize("elementwise_affine", ELEMENTWISE_AFFINE)
+@pytest.mark.parametrize("dtype", DTYPES)
+@pytest.mark.parametrize("seed", SEEDS)
+@torch.inference_mode()
+def test_compiled_vs_eager_multidim(
+    default_vllm_config,
+    shape: tuple,
+    activation: str,
+    elementwise_affine: bool,
+    dtype: torch.dtype,
+    seed: int,
+) -> None:
+    """forward_native decomposition handles multi-dimensional inputs."""
+    torch._dynamo.reset()
+    set_random_seed(seed)
+    device = torch.device("cuda:0")
+    head_dim = shape[-1]
+
+    module = FusedRMSNormGated(
+        head_dim,
+        elementwise_affine=elementwise_affine,
+        eps=1e-5,
+        activation=activation,
+        device=device,
+        dtype=dtype,
+    )
+    x = torch.randn(*shape, dtype=dtype, device=device)
+    g = torch.randn(*shape, dtype=dtype, device=device)
+
+    # forward_cuda may modify x in-place, so clone inputs
+    cuda_out = module.forward_cuda(x.clone(), g.clone())
+    compiled_native = torch.compile(module.forward_native, fullgraph=True)
+    native_out = compiled_native(x.clone(), g.clone())
+
+    torch.testing.assert_close(native_out, cuda_out, atol=1e-3, rtol=1e-2)

vllm/model_executor/layers/fla/ops/kda.py

@@ -12,6 +12,7 @@
 import torch
 import torch.nn as nn
 
+from vllm.model_executor.custom_op import CustomOp
 from vllm.triton_utils import tl, triton
 from vllm.utils.math_utils import cdiv, next_power_of_2
 
@@ -431,7 +432,8 @@ def rms_norm_gated(
     return y if not prenorm else (y, residual_out.reshape(x_shape_og))
 
 
-class FusedRMSNormGated(nn.Module):
+@CustomOp.register("fused_rms_norm_gated")
+class FusedRMSNormGated(CustomOp):
     def __init__(
         self,
         hidden_size: int,
@@ -458,7 +460,33 @@ class FusedRMSNormGated(nn.Module):
             self.register_parameter("weight", None)
         self.register_parameter("bias", None)
 
-    def forward(
+    def forward_native(
+        self,
+        x: torch.Tensor,
+        g: torch.Tensor,
+        residual: torch.Tensor | None = None,
+        prenorm: bool = False,
+        residual_in_fp32: bool = False,
+    ) -> torch.Tensor:
+        """Decomposed PyTorch ops for torch.compile/inductor fusion."""
+        # TODO(https://github.com/vllm-project/vllm/issues/36175): implement
+        # native residual/prenorm path and unify with RMSNormGated.
+        # For now, fall back to the triton kernel.
+        if residual is not None or prenorm:
+            return self.forward_cuda(x, g, residual, prenorm, residual_in_fp32)
+        x_float = x.float()
+        variance = x_float.pow(2).mean(dim=-1, keepdim=True)
+        x_normed = x_float * torch.rsqrt(variance + self.eps)
+        if self.weight is not None:
+            x_normed = x_normed * self.weight.float()
+        g_float = g.float()
+        if self.activation in ("swish", "silu"):
+            out = x_normed * g_float * torch.sigmoid(g_float)
+        else:  # sigmoid
+            out = x_normed * torch.sigmoid(g_float)
+        return out.to(x.dtype)
+
+    def forward_cuda(
         self,
         x: torch.Tensor,
         g: torch.Tensor,