SUMIN/HuggingFace_transformer
1
0
Fork 0
Code Issues Pull Requests Actions 7 Packages Projects Releases Wiki Activity

Remove old code for PyTorch, Accelerator and tokenizers (#37234)

Browse Source 
* Remove unneeded library version checks

Signed-off-by: cyy <cyyever@outlook.com>

* Remove PyTorch condition

Signed-off-by: cyy <cyyever@outlook.com>

* Remove PyTorch condition

Signed-off-by: cyy <cyyever@outlook.com>

* Fix ROCm get_device_capability

Signed-off-by: cyy <cyyever@outlook.com>

* Revert "Fix ROCm get_device_capability"

This reverts commit 0e756434bd7e74ffd73de5500476072b096570a6.

* Remove unnecessary check

Signed-off-by: cyy <cyyever@outlook.com>

* Revert changes

Signed-off-by: cyy <cyyever@outlook.com>

---------

Signed-off-by: cyy <cyyever@outlook.com>
This commit is contained in:
cyyever
2025-04-11 02:54:21 +08:00
committed by GitHub
parent 7ff896c0f2
commit 371c44d0ef
42 changed files with 53 additions and 178 deletions
Download Patch File Download Diff File Expand all files Collapse all files

23
tests/trainer/test_trainer.py
View File

@@ -99,7 +99,6 @@ from transformers.testing_utils import (
require_torch_tensorrt_fx,
require_torch_tf32,
require_torch_up_to_2_accelerators,
require_torchdynamo,
require_vision,
require_wandb,
run_first,
@@ -3994,10 +3993,9 @@ class TrainerIntegrationTest(TestCasePlus, TrainerIntegrationCommon):
@require_non_xpu
@require_torch_non_multi_gpu
@require_torchdynamo
@require_torch_tensorrt_fx
def test_torchdynamo_full_eval(self):
import torchdynamo
from torch import _dynamo as torchdynamo
# torchdynamo at the moment doesn't support DP/DDP, therefore require a single gpu
n_gpus = get_gpu_count()
@@ -4017,30 +4015,35 @@ class TrainerIntegrationTest(TestCasePlus, TrainerIntegrationCommon):
del trainer
# 2. TorchDynamo eager
trainer = get_regression_trainer(a=a, b=b, eval_len=eval_len, torchdynamo="eager", output_dir=tmp_dir)
trainer = get_regression_trainer(
a=a, b=b, eval_len=eval_len, torch_compile_backend="eager", output_dir=tmp_dir
)
metrics = trainer.evaluate()
self.assertAlmostEqual(metrics["eval_loss"], original_eval_loss)
del trainer
torchdynamo.reset()
# 3. TorchDynamo nvfuser
trainer = get_regression_trainer(a=a, b=b, eval_len=eval_len, torchdynamo="nvfuser", output_dir=tmp_dir)
trainer = get_regression_trainer(
a=a, b=b, eval_len=eval_len, torch_compile_backend="nvfuser", output_dir=tmp_dir
)
metrics = trainer.evaluate()
self.assertAlmostEqual(metrics["eval_loss"], original_eval_loss)
torchdynamo.reset()
# 4. TorchDynamo fx2trt
trainer = get_regression_trainer(a=a, b=b, eval_len=eval_len, torchdynamo="fx2trt", output_dir=tmp_dir)
trainer = get_regression_trainer(
a=a, b=b, eval_len=eval_len, torch_compile_backend="fx2trt", output_dir=tmp_dir
)
metrics = trainer.evaluate()
self.assertAlmostEqual(metrics["eval_loss"], original_eval_loss)
torchdynamo.reset()
@unittest.skip(reason="torch 2.0.0 gives `ModuleNotFoundError: No module named 'torchdynamo'`.")
@require_torch_non_multi_gpu
@require_torchdynamo
@require_torch_gpu
def test_torchdynamo_memory(self):
# torchdynamo at the moment doesn't support DP/DDP, therefore require a single gpu
import torchdynamo
from torch import _dynamo as torchdynamo
class CustomTrainer(Trainer):
def compute_loss(self, model, inputs, return_outputs=False):
@@ -4085,7 +4088,7 @@ class TrainerIntegrationTest(TestCasePlus, TrainerIntegrationCommon):
with tempfile.TemporaryDirectory() as tmp_dir:
a = torch.ones(1024, 1024, device="cuda", requires_grad=True)
a.grad = None
args = TrainingArguments(output_dir=tmp_dir, torchdynamo="nvfuser")
args = TrainingArguments(output_dir=tmp_dir, torch_compile_backend="nvfuser")
trainer = CustomTrainer(model=mod, args=args)
# warmup
for _ in range(10):