Export MoE

torchtitan/experiments/deepseek_v3/tracing.py

@@ -0,0 +1,105 @@
+# torchrun --nproc-per-node 4 --standalone tracing.py
+
+import torch
+import torch.distributed as dist
+
+from model import MoE
+from model_config import deepseek_config_registry
+
+
+def print0(*args, **kwargs):
+    if dist.get_rank() == 0:
+        print("\n")
+        print(*args, **kwargs)
+
+def setup_mesh():
+    ep_size = dist.get_world_size()
+    mesh_shape = (ep_size,)
+    mesh = dist.init_device_mesh("cuda", mesh_shape, mesh_dim_names=("ep",))
+    return mesh
+
+def setup_model(mesh):
+    group_size = mesh["ep"].size()
+    rank = mesh["ep"].get_local_rank()
+
+    model_id = "deepseek-ai/DeepSeek-V2-Lite-Chat"
+    config = deepseek_config_registry[model_id]
+    config.ep_size = group_size
+
+    device = torch.device("cuda", rank % torch.cuda.device_count())
+    dtype = torch.bfloat16
+    torch.set_default_dtype(dtype)
+
+    # Initialize the model
+    print0("Initializing MoE model...")
+    with mesh, torch.device(device):
+        moe = MoE(config)
+
+    print0("Setting up Symmetric Memory ...")
+    moe.setup_symm_mem(torch.bfloat16, device)
+
+    return moe
+
+def test_export(moe, mesh):
+    seqlen = 256
+    batch_size = 1
+    config = moe.config
+
+    rank = mesh["ep"].get_local_rank()
+    device = torch.device("cuda", rank % torch.cuda.device_count())
+
+    x = torch.randn(
+        batch_size, seqlen, config.hidden_size, dtype=torch.bfloat16, device=device
+    )
+    y = moe(x)
+    # print(y.shape)
+
+    # Let's export the model
+    print0("Exporting MoE model using torch.export...")
+
+    # Put model in eval mode for export
+    moe.eval()
+
+    # Create example input for export
+    example_input = (
+        torch.randn(
+            batch_size, seqlen, config.hidden_size, dtype=torch.bfloat16, device=device
+        ),
+    )
+
+    # Export using torch.export.export
+    exported_model = torch.export.export(moe, example_input)
+    print0("Successfully exported the MoE model using torch.export")
+
+    # Save the exported model
+    # export_path = "exported_moe_model.pt"
+    # torch.export.save(exported_model, export_path)
+    # print(f"Exported model saved to: {export_path}")
+
+    # Test the exported model
+    print0("Testing exported model...")
+    with torch.no_grad():
+        original_output = moe(*example_input)
+        exported_output = exported_model.module()(*example_input)
+
+        # Check if outputs are close
+        if torch.allclose(original_output, exported_output, rtol=1e-3, atol=1e-3):
+            print0("✓ Exported model outputs match original model outputs")
+        else:
+            print0("⚠ Exported model outputs differ from original model")
+            print0(
+                f"Max difference: {torch.max(torch.abs(original_output - exported_output))}"
+            )
+
+    print0("Model export completed!\n")
+
+    if rank == 0:
+        exported_model.graph_module.print_readable()
+
+
+if __name__ == "__main__":
+    torch.distributed.init_process_group(backend="nccl")
+    mesh = setup_mesh()
+    moe = setup_model(mesh)
+    test_export(moe, mesh)
+    torch.distributed.destroy_process_group()