[Misc] Add CustomOp interface for device portability (#5255)

tests/kernels/test_activation.py

@@ -44,7 +44,7 @@ def test_act_and_mul(
     elif activation == "gelu_tanh":
         layer = GeluAndMul(approximate="tanh")
     out = layer(x)
-    ref_out = layer._forward(x)
+    ref_out = layer.forward_native(x)
     # The SiLU and GELU implementations are equivalent to the native PyTorch
     # implementations, so we can do exact comparison.
     assert torch.allclose(out, ref_out, atol=0.0, rtol=0.0)
@@ -72,7 +72,7 @@ def test_activation(
     x = torch.randn(num_tokens, d, dtype=dtype)
     layer = activation()
     out = layer(x)
-    ref_out = layer._forward(x)
+    ref_out = layer.forward_native(x)
     assert torch.allclose(out,
                           ref_out,
                           atol=get_default_atol(out),

tests/kernels/test_layernorm.py

@@ -42,7 +42,7 @@ def test_rms_norm(
 
     # NOTE(woosuk): The reference implementation should be executed first
     # because the custom kernel is in-place.
-    ref_out = layer._forward(x, residual)
+    ref_out = layer.forward_native(x, residual)
     out = layer(x, residual)
     # NOTE(woosuk): LayerNorm operators (including RMS) typically have larger
     # numerical errors than other operators because they involve reductions.

tests/kernels/test_pos_encoding.py

@@ -64,7 +64,7 @@ def test_rotary_embedding(
 
     # NOTE(woosuk): The reference implementation should be executed first
     # because the custom kernel is in-place.
-    ref_query, ref_key = rope._forward(positions, query, key)
+    ref_query, ref_key = rope.forward_native(positions, query, key)
     out_query, out_key = rope.forward(positions, query, key)
     # Compare the results.
     assert torch.allclose(out_query,
@@ -121,7 +121,7 @@ def test_batched_rotary_embedding(
 
     # NOTE(woosuk): The reference implementation should be executed first
     # because the custom kernel is in-place.
-    ref_query, ref_key = rope._forward(positions, query, key)
+    ref_query, ref_key = rope.forward_native(positions, query, key)
     out_query, out_key = rope.forward(positions,
                                       query,
                                       key,
@@ -195,7 +195,8 @@ def test_batched_rotary_embedding_multi_lora(
 
     # NOTE(woosuk): The reference implementation should be executed first
     # because the custom kernel is in-place.
-    ref_query, ref_key = rope._forward(positions, query, key, query_offsets)
+    ref_query, ref_key = rope.forward_native(positions, query, key,
+                                             query_offsets)
     out_query, out_key = rope.forward(positions, query, key,
                                       query_offsets.flatten())
     # Compare the results.

vllm/model_executor/custom_op.py

@@ -0,0 +1,60 @@
+import torch.nn as nn
+
+from vllm.utils import is_cpu, is_hip
+
+
+class CustomOp(nn.Module):
+
+    def __init__(self, *args, **kwargs):
+        super().__init__()
+        self._forward_method = self.dispatch_forward()
+
+    def forward(self, *args, **kwargs):
+        return self._forward_method(*args, **kwargs)
+
+    def forward_native(self, *args, **kwargs):
+        """PyTorch-native implementation of the forward method.
+
+        This method is optional. If implemented, it can be used with compilers
+        such as torch.compile or PyTorch XLA. Also, it can be used for testing
+        purposes.
+        """
+        raise NotImplementedError
+
+    def forward_cuda(self, *args, **kwargs):
+        raise NotImplementedError
+
+    def forward_hip(self, *args, **kwargs):
+        # By default, we assume that HIP ops are compatible with CUDA ops.
+        return self.forward_cuda(*args, **kwargs)
+
+    def forward_xpu(self, *args, **kwargs):
+        # By default, we assume that XPU ops are compatible with CUDA ops.
+        # NOTE(woosuk): This is a placeholder for future extensions.
+        return self.forward_cuda(*args, **kwargs)
+
+    def forward_cpu(self, *args, **kwargs):
+        # By default, we assume that CPU ops are compatible with CUDA ops.
+        return self.forward_cuda(*args, **kwargs)
+
+    def forward_tpu(self, *args, **kwargs):
+        # By default, we assume that TPU ops are compatible with the
+        # PyTorch-native implementation.
+        # NOTE(woosuk): This is a placeholder for future extensions.
+        return self.forward_native(*args, **kwargs)
+
+    def forward_gaudi(self, *args, **kwargs):
+        # By default, we assume that Gaudi ops are compatible with the
+        # PyTorch-native implementation.
+        # NOTE(woosuk): This is a placeholder for future extensions.
+        return self.forward_native(*args, **kwargs)
+
+    def dispatch_forward(self):
+        # NOTE(woosuk): Here we assume that vLLM was built for only one
+        # specific backend. Currently, we do not support dynamic dispatching.
+        if is_hip():
+            return self.forward_hip
+        elif is_cpu():
+            return self.forward_cpu
+        else:
+            return self.forward_cuda

vllm/model_executor/layers/activation.py

@@ -6,14 +6,14 @@ import torch
 import torch.nn as nn
 import torch.nn.functional as F
 
-from vllm import _custom_ops as ops
 from vllm.distributed import (divide, get_tensor_model_parallel_rank,
                               get_tensor_model_parallel_world_size)
+from vllm.model_executor.custom_op import CustomOp
 from vllm.model_executor.layers.quantization import QuantizationConfig
 from vllm.model_executor.utils import set_weight_attrs
 
 
-class SiluAndMul(nn.Module):
+class SiluAndMul(CustomOp):
     """An activation function for SwiGLU.
 
     The function computes x -> silu(x[:d]) * x[d:] where d = x.shape[-1] // 2.
@@ -23,12 +23,14 @@ class SiluAndMul(nn.Module):
         return: (num_tokens, d) or (batch_size, seq_len, d)
     """
 
-    def _forward(self, x: torch.Tensor) -> torch.Tensor:
+    def forward_native(self, x: torch.Tensor) -> torch.Tensor:
         """PyTorch-native implementation equivalent to forward()."""
         d = x.shape[-1] // 2
         return F.silu(x[..., :d]) * x[..., d:]
 
-    def forward(self, x: torch.Tensor) -> torch.Tensor:
+    def forward_cuda(self, x: torch.Tensor) -> torch.Tensor:
+        from vllm import _custom_ops as ops
+
         d = x.shape[-1] // 2
         output_shape = (x.shape[:-1] + (d, ))
         out = torch.empty(output_shape, dtype=x.dtype, device=x.device)
@@ -36,7 +38,7 @@ class SiluAndMul(nn.Module):
         return out
 
 
-class GeluAndMul(nn.Module):
+class GeluAndMul(CustomOp):
     """An activation function for GeGLU.
 
     The function computes x -> GELU(x[:d]) * x[d:] where d = x.shape[-1] // 2.
@@ -52,12 +54,14 @@ class GeluAndMul(nn.Module):
         if approximate not in ("none", "tanh"):
             raise ValueError(f"Unknown approximate mode: {approximate}")
 
-    def _forward(self, x: torch.Tensor) -> torch.Tensor:
+    def forward_native(self, x: torch.Tensor) -> torch.Tensor:
         """PyTorch-native implementation equivalent to forward()."""
         d = x.shape[-1] // 2
         return F.gelu(x[..., :d], approximate=self.approximate) * x[..., d:]
 
-    def forward(self, x: torch.Tensor) -> torch.Tensor:
+    def forward_cuda(self, x: torch.Tensor) -> torch.Tensor:
+        from vllm import _custom_ops as ops
+
         d = x.shape[-1] // 2
         output_shape = (x.shape[:-1] + (d, ))
         out = torch.empty(output_shape, dtype=x.dtype, device=x.device)
@@ -71,28 +75,32 @@ class GeluAndMul(nn.Module):
         return f'approximate={repr(self.approximate)}'
 
 
-class NewGELU(nn.Module):
+class NewGELU(CustomOp):
 
-    def _forward(self, x: torch.Tensor) -> torch.Tensor:
+    def forward_native(self, x: torch.Tensor) -> torch.Tensor:
         """PyTorch-native implementation equivalent to forward()."""
         c = math.sqrt(2.0 / math.pi)
         return 0.5 * x * (1.0 + torch.tanh(c *
                                            (x + 0.044715 * torch.pow(x, 3.0))))
 
-    def forward(self, x: torch.Tensor) -> torch.Tensor:
+    def forward_cuda(self, x: torch.Tensor) -> torch.Tensor:
+        from vllm import _custom_ops as ops
+
         out = torch.empty_like(x)
         ops.gelu_new(out, x)
         return out
 
 
-class FastGELU(nn.Module):
+class FastGELU(CustomOp):
 
-    def _forward(self, x: torch.Tensor) -> torch.Tensor:
+    def forward_native(self, x: torch.Tensor) -> torch.Tensor:
         """PyTorch-native implementation equivalent to forward()."""
         return 0.5 * x * (1.0 + torch.tanh(x * 0.7978845608 *
                                            (1.0 + 0.044715 * x * x)))
 
-    def forward(self, x: torch.Tensor) -> torch.Tensor:
+    def forward_cuda(self, x: torch.Tensor) -> torch.Tensor:
+        from vllm import _custom_ops as ops
+
         out = torch.empty_like(x)
         ops.gelu_fast(out, x)
         return out

vllm/model_executor/layers/layernorm.py

@@ -4,10 +4,10 @@ from typing import Optional, Tuple, Union
 import torch
 import torch.nn as nn
 
-from vllm import _custom_ops as ops
+from vllm.model_executor.custom_op import CustomOp
 
 
-class RMSNorm(nn.Module):
+class RMSNorm(CustomOp):
     """Root mean square normalization.
 
     Computes x -> w * x / sqrt(E[x^2] + eps) where w is the learned weight.
@@ -23,7 +23,7 @@ class RMSNorm(nn.Module):
         self.weight = nn.Parameter(torch.ones(hidden_size))
         self.variance_epsilon = eps
 
-    def _forward(
+    def forward_native(
         self,
         x: torch.Tensor,
         residual: Optional[torch.Tensor] = None,
@@ -43,11 +43,13 @@ class RMSNorm(nn.Module):
         else:
             return x, residual
 
-    def forward(
+    def forward_cuda(
         self,
         x: torch.Tensor,
         residual: Optional[torch.Tensor] = None,
     ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
+        from vllm import _custom_ops as ops
+
         if residual is not None:
             ops.fused_add_rms_norm(
                 x,

vllm/model_executor/layers/rotary_embedding.py

@@ -27,7 +27,7 @@ from typing import Any, Dict, List, Optional, Tuple, Union
 import torch
 import torch.nn as nn
 
-from vllm import _custom_ops as ops
+from vllm.model_executor.custom_op import CustomOp
 
 
 def _rotate_neox(x: torch.Tensor) -> torch.Tensor:
@@ -43,7 +43,7 @@ def _rotate_gptj(x: torch.Tensor) -> torch.Tensor:
     return x.flatten(-2)
 
 
-class RotaryEmbedding(nn.Module):
+class RotaryEmbedding(CustomOp):
     """Original rotary positional embedding."""
 
     def __init__(
@@ -93,7 +93,7 @@ class RotaryEmbedding(nn.Module):
         cache = torch.cat((cos, sin), dim=-1)
         return cache
 
-    def _forward(
+    def forward_native(
         self,
         positions: torch.Tensor,
         query: torch.Tensor,
@@ -138,13 +138,15 @@ class RotaryEmbedding(nn.Module):
         key = key.flatten(-2)
         return query, key
 
-    def forward(
+    def forward_cuda(
         self,
         positions: torch.Tensor,
         query: torch.Tensor,
         key: torch.Tensor,
         offsets: Optional[torch.Tensor] = None,
     ) -> Tuple[torch.Tensor, torch.Tensor]:
+        from vllm import _custom_ops as ops
+
         self.cos_sin_cache = self.cos_sin_cache.to(positions.device,
                                                    dtype=query.dtype)
         # ops.rotary_embedding()/batched_rotary_embedding()