only override forward if using cuda-graph (#2291)

.github/workflows/nv-inference.yml

@@ -40,8 +40,6 @@ jobs:
         run: |
           git clone https://github.com/huggingface/transformers
           cd transformers
-          # if needed switch to the last known good SHA until transformers@master is fixed
-          git checkout v4.21.2
           git rev-parse --short HEAD
           pip uninstall --yes transformers
           pip install .

deepspeed/inference/engine.py

@@ -162,10 +162,7 @@ class InferenceEngine(Module):
             torch.cuda.set_rng_state(_rng_state.cpu())
 
         if self.mp_world_size > 1:
-            self.model_orig_fwd = self.module.forward
-            self.module.forward = self.forward
-        else:
-            self.module.register_forward_pre_hook(self._pre_forward_hook)
+            assert not self.enable_cuda_graph, "Cuda graph is not supported for model parallelism"
 
     def _get_model_config_generate(self, config):
         self.config = getattr(self.module, 'config', None) if config is None else config
@@ -475,14 +472,6 @@ class InferenceEngine(Module):
         elif self.dtype == torch.float:
             self.module.float()
 
-    def _pre_forward_hook(self, module, *inputs, **kwargs):
-        for input in inputs:
-            if torch.is_tensor(input):
-                input = input.to(torch.cuda.current_device())
-        for k in kwargs:
-            if torch.is_tensor(kwargs[k]):
-                kwargs[k] = kwargs[k].to(torch.cuda.current_device())
-
     def _create_cuda_graph(self, *inputs, **kwargs):
         # warmup to create the workspace and cublas handle
         cuda_stream = torch.cuda.Stream()
@@ -519,23 +508,6 @@ class InferenceEngine(Module):
             *inputs: Variable length input list
             **kwargs: variable length keyword arguments
         """
-
-        if self.mp_world_size > 1:
-            if self.mpu is None:
-                for input in inputs:
-                    if torch.is_tensor(input):
-                        input = input.to(torch.cuda.current_device())
-                        if not input.is_contiguous():
-                            input = input.contiguous()
-                        dist.broadcast(input, 0)
-                for k in kwargs:
-                    if torch.is_tensor(kwargs[k]):
-                        kwargs[k] = kwargs[k].to(torch.cuda.current_device())
-                        if not kwargs[k].is_contiguous():
-                            kwargs[k] = kwargs[k].contiguous()
-                        dist.broadcast(kwargs[k], 0)
-            outputs = self.model_orig_fwd(*inputs, **kwargs)
-        else:
         if self.enable_cuda_graph:
             if self.cuda_graph_created:
                 outputs = self._graph_replay(*inputs, **kwargs)
@@ -544,5 +516,5 @@ class InferenceEngine(Module):
                 outputs = self._graph_replay(*inputs, **kwargs)
         else:
             outputs = self.module(*inputs, **kwargs)
-            #outputs = self.module(*inputs, **kwargs)
+
         return outputs

tests/unit/inference/test_inference.py

@@ -292,13 +292,13 @@ class TestModelTask(DistributedTest):
 
 @pytest.mark.seq_inference
 @pytest.mark.parametrize("model_w_task",
-                         [("gpt2",
+                         [("EleutherAI/gpt-neo-1.3B",
                            "text-generation"),
                           ("EleutherAI/gpt-neox-20b",
                            "text-generation"),
                           ("bigscience/bloom-3b",
                            "text-generation")],
-                         ids=["gpt2",
+                         ids=["gpt-neo",
                               "gpt-neox",
                               "bloom"])
 class TestMPSize(DistributedTest):
@@ -308,7 +308,6 @@ class TestMPSize(DistributedTest):
         self,
         model_w_task,
         dtype,
-        enable_cuda_graph,
         query,
         inf_kwargs,
         assert_fn,
@@ -325,14 +324,11 @@ class TestMPSize(DistributedTest):
         pipe = pipeline(task, model=model, device=-1, framework="pt")
         bs_output = pipe(query, **inf_kwargs)
 
-        pipe.model = deepspeed.init_inference(
-            pipe.model,
+        pipe.model = deepspeed.init_inference(pipe.model,
                                               mp_size=self.world_size,
                                               dtype=dtype,
                                               replace_method="auto",
-            replace_with_kernel_inject=True,
-            enable_cuda_graph=enable_cuda_graph,
-        )
+                                              replace_with_kernel_inject=True)
         # Switch device to GPU so that input tensors are not on CPU
         pipe.device = torch.device(f"cuda:{local_rank}")
         ds_output = pipe(query, **inf_kwargs)