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

switch to device agnostic device calling for test cases (#38247)

Browse Source 
* use device agnostic APIs in test cases

Signed-off-by: Matrix Yao <matrix.yao@intel.com>

* fix style

Signed-off-by: Matrix Yao <matrix.yao@intel.com>

* add one more

Signed-off-by: YAO Matrix <matrix.yao@intel.com>

* xpu now supports integer device id, aligning to CUDA behaviors

Signed-off-by: Matrix Yao <matrix.yao@intel.com>

* update to use device_properties

Signed-off-by: Matrix Yao <matrix.yao@intel.com>

* fix style

Signed-off-by: Matrix Yao <matrix.yao@intel.com>

* update comment

Signed-off-by: Matrix Yao <matrix.yao@intel.com>

* fix comments

Signed-off-by: Matrix Yao <matrix.yao@intel.com>

* fix style

Signed-off-by: Matrix Yao <matrix.yao@intel.com>

---------

Signed-off-by: Matrix Yao <matrix.yao@intel.com>
Signed-off-by: YAO Matrix <matrix.yao@intel.com>
Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>
This commit is contained in:
Yao Matrix
2025-05-26 16:18:53 +08:00
committed by GitHub
parent cba279f46c
commit a5a0c7b888
39 changed files with 259 additions and 389 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
tests/models/bloom/test_modeling_bloom.py
View File

@@ -390,7 +390,7 @@ class BloomModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixi
def test_simple_generation(self):
# This test is a bit flaky. For some GPU architectures, pytorch sets by default allow_fp16_reduced_precision_reduction = True and some operations
# do not give the same results under this configuration, especially torch.baddmm and torch.bmm. https://pytorch.org/docs/stable/notes/numerical_accuracy.html#fp16-on-mi200
# As we leave the default value (True) for allow_fp16_reduced_precision_reduction , the tests failed when running in half-precision with smaller models (560m)
# As we leave the default value (True) for allow_fp16_reduced_precision_reduction, the tests failed when running in half-precision with smaller models (560m)
# Please see: https://pytorch.org/docs/stable/notes/cuda.html#reduced-precision-reduction-in-fp16-gemms
# This discrepancy is observed only when using small models and seems to be stable for larger models.
# Our conclusion is that these operations are flaky for small inputs but seems to be stable for larger inputs (for the functions `baddmm` and `bmm`), and therefore for larger models.
@@ -763,7 +763,6 @@ class BloomEmbeddingTest(unittest.TestCase):
@require_torch
def test_hidden_states_transformers(self):
cuda_available = torch.cuda.is_available()
model = BloomModel.from_pretrained(self.path_bigscience_model, use_cache=False, torch_dtype="auto").to(
torch_device
)
@@ -782,7 +781,7 @@ class BloomEmbeddingTest(unittest.TestCase):
"max": logits.last_hidden_state.max(dim=-1).values[0][0].item(),
}
if cuda_available:
if torch_device == "cuda":
self.assertAlmostEqual(MEAN_VALUE_LAST_LM, logits.last_hidden_state.mean().item(), places=4)
else:
self.assertAlmostEqual(MEAN_VALUE_LAST_LM, logits.last_hidden_state.mean().item(), places=3)
@@ -791,7 +790,6 @@ class BloomEmbeddingTest(unittest.TestCase):
@require_torch
def test_logits(self):
cuda_available = torch.cuda.is_available()
model = BloomForCausalLM.from_pretrained(self.path_bigscience_model, use_cache=False, torch_dtype="auto").to(
torch_device
) # load in bf16
@@ -807,9 +805,5 @@ class BloomEmbeddingTest(unittest.TestCase):
output = model(tensor_ids).logits
output_gpu_1, output_gpu_2 = output.split(125440, dim=-1)
if cuda_available:
self.assertAlmostEqual(output_gpu_1.mean().item(), MEAN_LOGITS_GPU_1, places=6)
self.assertAlmostEqual(output_gpu_2.mean().item(), MEAN_LOGITS_GPU_2, places=6)
else:
self.assertAlmostEqual(output_gpu_1.mean().item(), MEAN_LOGITS_GPU_1, places=6) # 1e-06 precision!!
self.assertAlmostEqual(output_gpu_2.mean().item(), MEAN_LOGITS_GPU_2, places=6)
self.assertAlmostEqual(output_gpu_1.mean().item(), MEAN_LOGITS_GPU_1, places=6)
self.assertAlmostEqual(output_gpu_2.mean().item(), MEAN_LOGITS_GPU_2, places=6)