[torch.compile] expanding support and fix allgather compilation (vllm…

…-project#9637)

Signed-off-by: youkaichao <[email protected]>
Co-authored-by: youkaichao <[email protected]>
Signed-off-by: Sumit Dubey <[email protected]>

vllm/distributed/parallel_state.py

@@ -392,15 +392,20 @@ def all_gather(self, input_: torch.Tensor, dim: int = -1) -> torch.Tensor:
             # Convert negative dim to positive.
             dim += input_.dim()
         input_size = input_.size()
+        # NOTE: we have to use concat-style all-gather here,
+        # stack-style all-gather has compatibility issues with
+        # torch.compile . see https://github.com/pytorch/pytorch/issues/138795
+        output_size = (input_size[0] * world_size, ) + input_size[1:]
         # Allocate output tensor.
-        output_tensor = torch.empty((world_size, ) + input_size,
+        output_tensor = torch.empty(output_size,
                                     dtype=input_.dtype,
                                     device=input_.device)
         # All-gather.
         torch.distributed.all_gather_into_tensor(output_tensor,
                                                  input_,
                                                  group=self.device_group)
         # Reshape
+        output_tensor = output_tensor.reshape((world_size, ) + input_size)
         output_tensor = output_tensor.movedim(0, dim)
         output_tensor = output_tensor.reshape(input_size[:dim] +
                                               (world_size *

vllm/model_executor/models/gpt_bigcode.py

@@ -25,6 +25,7 @@
 from transformers import GPTBigCodeConfig
 
 from vllm.attention import Attention, AttentionMetadata
+from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, LoRAConfig
 from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
 from vllm.model_executor.layers.activation import get_act_fn
@@ -187,6 +188,7 @@ def forward(
         return hidden_states
 
 
+@support_torch_compile
 class GPTBigCodeModel(nn.Module):
 
     def __init__(

vllm/model_executor/models/gpt_j.py

@@ -23,6 +23,7 @@
 from transformers import GPTJConfig
 
 from vllm.attention import Attention, AttentionMetadata
+from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig
 from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
 from vllm.model_executor.layers.activation import get_act_fn
@@ -174,6 +175,7 @@ def forward(
         return hidden_states
 
 
+@support_torch_compile
 class GPTJModel(nn.Module):
 
     def __init__(

vllm/model_executor/models/gpt_neox.py

@@ -23,6 +23,7 @@
 from transformers import GPTNeoXConfig
 
 from vllm.attention import Attention, AttentionMetadata
+from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig
 from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
 from vllm.model_executor.layers.activation import get_act_fn
@@ -187,6 +188,7 @@ def forward(
         return hidden_states
 
 
+@support_torch_compile
 class GPTNeoXModel(nn.Module):
 
     def __init__(

vllm/model_executor/models/granite.py

@@ -28,6 +28,7 @@
 from transformers import GraniteConfig
 
 from vllm.attention import Attention, AttentionMetadata
+from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, LoRAConfig
 from vllm.distributed import (get_pp_group, get_tensor_model_parallel_rank,
                               get_tensor_model_parallel_world_size)
@@ -254,6 +255,7 @@ def forward(
         return hidden_states
 
 
+@support_torch_compile
 class GraniteModel(nn.Module):
 
     def __init__(

vllm/model_executor/models/internlm2.py

@@ -7,6 +7,7 @@
 from transformers import PretrainedConfig
 
 from vllm.attention import Attention, AttentionMetadata
+from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig
 from vllm.distributed import (get_pp_group, get_tensor_model_parallel_rank,
                               get_tensor_model_parallel_world_size,
@@ -230,6 +231,7 @@ def forward(
         return hidden_states, residual
 
 
+@support_torch_compile
 class InternLM2Model(nn.Module):
 
     def __init__(