[misc] update torch version (#6206)

* [misc] update torch version

* fix test

* fix test

* fix test

* fix test

.compatibility

@@ -1,3 +1,3 @@
-2.2.2-12.1.0
 2.3.0-12.1.0
 2.4.0-12.4.1
+2.5.1-12.4.1

.cuda_ext.json

@@ -1,11 +1,11 @@
 {
   "build": [
     {
-      "torch_command": "pip install torch==2.1.0 torchvision==0.16.0 torchaudio==2.1.0 --index-url https://download.pytorch.org/whl/cu121",
+      "torch_command": "pip install torch==2.3.0 torchvision==0.18.0 torchaudio==2.3.0 --index-url https://download.pytorch.org/whl/cu121",
       "cuda_image": "hpcaitech/cuda-conda:12.1"
     },
     {
-      "torch_command": "pip install torch==2.4.0 torchvision==0.19.0 torchaudio==2.4.0 --index-url https://download.pytorch.org/whl/cu124",
+      "torch_command": "pip install torch==2.5.1 torchvision==0.20.1 torchaudio==2.5.1 --index-url https://download.pytorch.org/whl/cu124",
       "cuda_image": "hpcaitech/cuda-conda:12.4"
     }
   ]

requirements/requirements.txt

@@ -8,7 +8,7 @@ click
 fabric
 contexttimer
 ninja
-torch>=2.2.0,<=2.4.1
+torch>=2.2.0,<=2.5.1
 safetensors
 einops
 pydantic

tests/test_cluster/test_device_mesh_manager.py

@@ -1,7 +1,7 @@
 from colossalai.cluster.device_mesh_manager import DeviceMeshInfo, DeviceMeshManager
 from colossalai.initialize import launch
 from colossalai.logging import disable_existing_loggers
-from colossalai.testing import spawn
+from colossalai.testing import rerun_if_address_is_in_use, spawn
 
 
 def check_device_mesh_manager(rank, world_size, port):
@@ -24,6 +24,7 @@ def check_device_mesh_manager(rank, world_size, port):
     assert device_mesh_with_shape._logical_mesh_id.tolist() == [[0, 1], [2, 3]]
 
 
+@rerun_if_address_is_in_use()
 def test_device_mesh_manager():
     spawn(check_device_mesh_manager, 4)
 

tests/test_shardformer/test_model/test_shard_t5.py

@@ -51,7 +51,7 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn,
     if test_config["precision"] == "fp32":
         atol, rtol = 1e-5, 1e-3
     else:
-        atol, rtol = 5e-2, 5e-2
+        atol, rtol = 9e-2, 0
     if (stage_manager is None or stage_manager.is_first_stage()) and booster.plugin.zero_stage == 0:
         row_layer_grads = get_grad_tensors_for_check(
             t5, sharded_t5, row_layer_for_check, tp_group, atol=atol, rtol=rtol, dim=0