Add PyTorch-native implementation of custom layers (#1898)

vllm/model_executor/layers/activation.py

@@ -1,8 +1,10 @@
 """Custom activation functions."""
+import math
 from typing import Optional
 
 import torch
 import torch.nn as nn
+import torch.nn.functional as F
 
 from vllm._C import ops
 from vllm.model_executor.layers.quantization import QuantizationConfig
@@ -22,6 +24,11 @@ class SiluAndMul(nn.Module):
         return: (batch_size, seq_len, d) or (num_tokens, d)
     """
 
+    def _forward(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:
         d = x.shape[-1] // 2
         output_shape = (x.shape[:-1] + (d, ))
@@ -32,6 +39,12 @@ class SiluAndMul(nn.Module):
 
 class NewGELU(nn.Module):
 
+    def _forward(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:
         out = torch.empty_like(x)
         ops.gelu_new(out, x)
@@ -40,6 +53,11 @@ class NewGELU(nn.Module):
 
 class FastGELU(nn.Module):
 
+    def _forward(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:
         out = torch.empty_like(x)
         ops.gelu_fast(out, x)

vllm/model_executor/layers/layernorm.py

@@ -23,6 +23,26 @@ class RMSNorm(nn.Module):
         self.weight = nn.Parameter(torch.ones(hidden_size))
         self.variance_epsilon = eps
 
+    def _forward(
+        self,
+        x: torch.Tensor,
+        residual: Optional[torch.Tensor] = None,
+    ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
+        """PyTorch-native implementation equivalent to forward()."""
+        orig_dtype = x.dtype
+        x = x.to(torch.float32)
+        if residual is not None:
+            x = x + residual.to(torch.float32)
+            residual = x.to(orig_dtype)
+
+        variance = x.pow(2).mean(dim=-1, keepdim=True)
+        x = x * torch.rsqrt(variance + self.variance_epsilon)
+        x = x.to(orig_dtype) * self.weight
+        if residual is None:
+            return x
+        else:
+            return x, residual
+
     def forward(
         self,
         x: torch.Tensor,

vllm/model_executor/layers/rotary_embedding.py

@@ -30,6 +30,19 @@ import torch.nn as nn
 from vllm._C import ops
 
 
+def _rotate_neox(x: torch.Tensor) -> torch.Tensor:
+    x1 = x[..., :x.shape[-1] // 2]
+    x2 = x[..., x.shape[-1] // 2:]
+    return torch.cat((-x2, x1), dim=-1)
+
+
+def _rotate_gptj(x: torch.Tensor) -> torch.Tensor:
+    x1 = x[..., ::2]
+    x2 = x[..., 1::2]
+    x = torch.stack((-x2, x1), dim=-1)
+    return x.flatten(-2)
+
+
 class RotaryEmbedding(nn.Module):
     """Original rotary positional embedding."""
 
@@ -81,6 +94,47 @@ class RotaryEmbedding(nn.Module):
         cache = torch.cat((cos, sin), dim=-1)
         return cache
 
+    def _forward(
+        self,
+        positions: torch.Tensor,
+        query: torch.Tensor,
+        key: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """PyTorch-native implementation equivalent to forward()."""
+        query = query.view(*query.shape[:-1], -1, self.head_size)
+        key = key.view(*key.shape[:-1], -1, self.head_size)
+
+        query_rot = query[..., :self.rotary_dim]
+        key_rot = key[..., :self.rotary_dim]
+        if self.rotary_dim < self.head_size:
+            query_pass = query[..., self.rotary_dim:]
+            key_pass = key[..., self.rotary_dim:]
+
+        cos_sin = self.cos_sin_cache[positions]
+        cos, sin = cos_sin.chunk(2, dim=-1)
+        if self.is_neox_style:
+            # NOTE(woosuk): Here we assume that the positions tensor has the
+            # shape [batch_size, seq_len].
+            cos = cos.repeat(1, 1, 2).unsqueeze(-2)
+            sin = sin.repeat(1, 1, 2).unsqueeze(-2)
+        else:
+            cos = cos.repeat_interleave(2, dim=-1).unsqueeze(-2)
+            sin = sin.repeat_interleave(2, dim=-1).unsqueeze(-2)
+
+        rotate_fn = _rotate_neox if self.is_neox_style else _rotate_gptj
+        query_rot = query_rot * cos + rotate_fn(query_rot) * sin
+        key_rot = key_rot * cos + rotate_fn(key_rot) * sin
+
+        if self.rotary_dim < self.head_size:
+            query = torch.cat((query_rot, query_pass), dim=-1)
+            key = torch.cat((key_rot, key_pass), dim=-1)
+        else:
+            query = query_rot
+            key = key_rot
+        query = query.flatten(-2)
+        key = key.flatten(-2)
+        return query, key
+
     def forward(
         self,
         positions: torch.Tensor,