[torch.compile] Compile CustomOp.forward_native for SiluAndMul and QuantFP8 to avoid raw torch ops inside opaque custom ops (#32806)

Signed-off-by: Luka Govedič <lgovedic@redhat.com>
Signed-off-by: Luka Govedič <ProExpertProg@users.noreply.github.com>
Co-authored-by: Michael Goin <mgoin64@gmail.com>

tests/compile/test_silu_mul_quant_fusion.py

@@ -222,7 +222,7 @@ def test_fusion_silu_and_mul_quant(
     x = torch.rand(num_tokens, hidden_size * 2)
 
     # Reshape pass is needed for the fusion pass to work
-    custom_ops = []
+    custom_ops = ["none"]
     if enable_silu_mul_custom_op:
         custom_ops.append("+silu_and_mul")
     if enable_quant_fp8_custom_op:
@@ -231,6 +231,7 @@ def test_fusion_silu_and_mul_quant(
         compilation_config=CompilationConfig(
             mode=CompilationMode.VLLM_COMPILE,
             custom_ops=custom_ops,
+            backend="eager",  # avoid compilation for SiluAndMul and QuantFP8
             pass_config=PassConfig(fuse_act_quant=True, eliminate_noops=True),
         ),
     )

tests/kernels/core/test_activation.py

@@ -57,7 +57,7 @@ def test_act_and_mul(
     torch.set_default_device(device)
     x = torch.randn(num_tokens, 2 * d, dtype=dtype)
     if activation == "silu_and_mul":
-        layer = SiluAndMul()
+        layer = SiluAndMul(compile_native=False)
         fn = torch.ops._C.silu_and_mul
     if activation == "mul_and_silu":
         layer = MulAndSilu()

vllm/compilation/matcher_utils.py

@@ -336,6 +336,7 @@ class MatcherQuantFP8(MatcherCustomOp):
             quant_key.scale.group_shape,
             column_major_scales=has_col_major_scales,
             use_ue8m0=is_e8m0,
+            compile_native=False,
         )
 
     def forward_rocm_aiter(

vllm/model_executor/custom_op.py

@@ -1,16 +1,15 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-
-
+import torch
 import torch.nn as nn
 
 from vllm.config import get_cached_compilation_config
 from vllm.logger import init_logger
+from vllm.model_executor.utils import maybe_disable_graph_partition
 from vllm.platforms import current_platform
 
 logger = init_logger(__name__)
 
-
 # Dictionary of all custom ops (classes, indexed by registered name).
 # To check if an op with a name is enabled, call .enabled() on the class.
 # Examples:
@@ -118,10 +117,10 @@ class CustomOp(nn.Module):
             )
         return super().__new__(op_cls_to_instantiate)
 
-    def __init__(self, enforce_enable: bool = False):
+    def __init__(self, *, enforce_enable: bool = False, compile_native: bool = False):
         super().__init__()
         self._enforce_enable = enforce_enable
-        self._forward_method = self.dispatch_forward()
+        self._forward_method = self.dispatch_forward(compile_native=compile_native)
 
     def forward(self, *args, **kwargs):
         return self._forward_method(*args, **kwargs)
@@ -162,7 +161,7 @@ class CustomOp(nn.Module):
         # PyTorch-native implementation.
         return self.forward_native(*args, **kwargs)
 
-    def dispatch_forward(self):
+    def dispatch_forward(self, compile_native: bool):
         # NOTE(woosuk): Here we assume that vLLM was built for only one
         # specific backend. Currently, we do not support dynamic dispatching.
         compilation_config = get_cached_compilation_config()
@@ -180,7 +179,9 @@ class CustomOp(nn.Module):
             compilation_config.disabled_custom_ops.update([self.__class__.name])
 
         if not enabled:
-            return self.forward_native
+            # Compile forward_native to avoid eager torch ops if inside
+            # opaque torch custom op (e.g. fused_moe, unified_attention, etc.)
+            return self.maybe_compile(self.forward_native, enable=compile_native)
 
         if current_platform.is_rocm():
             return self.forward_hip
@@ -195,6 +196,40 @@ class CustomOp(nn.Module):
         else:
             return self.forward_cuda
 
+    def maybe_compile(self, fn, *, enable: bool = True):
+        """
+        Compile fn if compilation enabled.
+        Useful for CustomOp instances called from within a torch custom op,
+        meaning the forward call is hidden from the model-level torch.compile.
+
+        NOTE: this does not enable fusion across ops, so opaque custom ops
+        should still be unwrapped wherever possible.
+        """
+        # Do not compile if compilation disabled
+        from vllm.config.compilation import CompilationMode
+
+        if not enable:
+            return fn
+
+        # Do not compile if global compilation disabled
+        compilation_config = get_cached_compilation_config()
+        if compilation_config.mode == CompilationMode.NONE:
+            return fn
+
+        # If eager backend is used, do not compile either
+        if compilation_config.backend == "eager":
+            return fn
+
+        # dynamic=True to avoid recompilations
+        return torch.compile(
+            fn,
+            dynamic=True,
+            backend=current_platform.simple_compile_backend,
+            options=maybe_disable_graph_partition(
+                current_platform.simple_compile_backend
+            ),
+        )
+
     @classmethod
     def enabled(cls) -> bool:
         # if no name, then it was not registered

vllm/model_executor/layers/activation.py

@@ -75,8 +75,8 @@ class SiluAndMul(CustomOp):
 
     # --8<-- [end:silu_and_mul]
 
-    def __init__(self):
-        super().__init__()
+    def __init__(self, *, compile_native: bool = True):
+        super().__init__(compile_native=compile_native)
         if current_platform.is_cuda_alike():
             self.op = torch.ops._C.silu_and_mul
         elif current_platform.is_xpu():

vllm/model_executor/layers/quantization/input_quant_fp8.py

@@ -38,6 +38,7 @@ class QuantFP8(CustomOp):
         column_major_scales: bool = False,
         tma_aligned_scales: bool = False,
         use_ue8m0: bool | None = None,  # for Torch compile
+        compile_native: bool = True,
     ):
         """
         :param static: static or dynamic quantization
@@ -49,8 +50,9 @@ class QuantFP8(CustomOp):
             TMA-aligned layout
         :param column_major_scales: For group quantization, output scales in
             column major format
+        :param compile_native: Manually compile forward_native if compile mode > None
         """
-        super().__init__()
+        super().__init__(compile_native=compile_native)
         self.static = static
         self.group_shape = group_shape
         self.use_per_token_if_dynamic = group_shape == GroupShape.PER_TOKEN

vllm/model_executor/layers/rotary_embedding/common.py

@@ -129,7 +129,7 @@ class ApplyRotaryEmb(CustomOp):
         is_neox_style: bool = True,
         enable_fp32_compute: bool = False,
     ) -> None:
-        super().__init__(enforce_enable)
+        super().__init__(enforce_enable=enforce_enable)
         self.is_neox_style = is_neox_style
         self.enable_fp32_compute = enable_fp32_compute