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

use torch.testing.assertclose instead to get more details about error in cis (#35659)

Browse Source 
* use torch.testing.assertclose instead to get more details about error in cis

* fix

* style

* test_all

* revert for I bert

* fixes and updates

* more image processing fixes

* more image processors

* fix mamba and co

* style

* less strick

* ok I won't be strict

* skip and be done

* up
This commit is contained in:
Arthur
2025-01-24 16:55:28 +01:00
committed by GitHub
parent 72d1a4cd53
commit b912f5ee43
255 changed files with 1048 additions and 969 deletions
Download Patch File Download Diff File Expand all files Collapse all files

50
tests/test_modeling_common.py
View File

@@ -1771,7 +1771,7 @@ class ModelTesterMixin:
model.eval()
hidden_states_with_chunk = model(**self._prepare_for_class(inputs_dict, model_class))[0]
self.assertTrue(torch.allclose(hidden_states_no_chunk, hidden_states_with_chunk, atol=1e-3))
torch.testing.assert_close(hidden_states_no_chunk, hidden_states_with_chunk, rtol=1e-3, atol=1e-3)
def test_resize_position_vector_embeddings(self):
if not self.test_resize_position_embeddings:
@@ -1898,7 +1898,7 @@ class ModelTesterMixin:
else:
old_embeddings_mean = torch.mean(model_embed.weight.data[:-10, :], axis=0)
new_embeddings_mean = torch.mean(model_embed.weight.data[-10:, :], axis=0)
torch.testing.assert_close(old_embeddings_mean, new_embeddings_mean, atol=1e-3, rtol=1e-1)
torch.testing.assert_close(old_embeddings_mean, new_embeddings_mean, rtol=1e-3, atol=1e-3)
# Check that the model can still do a forward pass successfully (every parameter should be resized)
if not is_deepspeed_zero3_enabled():
@@ -2006,7 +2006,7 @@ class ModelTesterMixin:
else:
old_embeddings_mean = torch.mean(model_embed.weight.data[:-10, :], axis=0)
new_embeddings_mean = torch.mean(model_embed.weight.data[-10:, :], axis=0)
torch.testing.assert_close(old_embeddings_mean, new_embeddings_mean, atol=1e-3, rtol=1e-1)
torch.testing.assert_close(old_embeddings_mean, new_embeddings_mean, rtol=1e-3, atol=1e-3)
@require_deepspeed
@require_torch_accelerator
@@ -2081,7 +2081,7 @@ class ModelTesterMixin:
else:
old_embeddings_mean = torch.mean(output_embeds.weight.data[:-10, :], axis=0)
new_embeddings_mean = torch.mean(output_embeds.weight.data[-10:, :], axis=0)
torch.testing.assert_close(old_embeddings_mean, new_embeddings_mean, atol=1e-3, rtol=1e-1)
torch.testing.assert_close(old_embeddings_mean, new_embeddings_mean, rtol=1e-3, atol=1e-3)
# check if the old bias mean close to added bias mean.
if output_embeds.bias is not None:
if is_deepspeed_zero3_enabled():
@@ -2092,7 +2092,7 @@ class ModelTesterMixin:
old_bias_mean = torch.mean(output_embeds.bias.data[:-10], axis=0)
new_bias_mean = torch.mean(output_embeds.bias.data[-10:], axis=0)
torch.testing.assert_close(old_bias_mean, new_bias_mean, atol=1e-5, rtol=1e-2)
torch.testing.assert_close(old_bias_mean, new_bias_mean, rtol=1e-5, atol=1e-5)
# Check that resizing the token embeddings with a smaller vocab size decreases the model's vocab size
model.resize_token_embeddings(model_vocab_size - 15)
@@ -3049,7 +3049,7 @@ class ModelTesterMixin:
out_embeds = model(
inputs_embeds=inputs_embeds, decoder_inputs_embeds=decoder_inputs_embeds, **inputs
)[0]
self.assertTrue(torch.allclose(out_embeds, out_ids))
torch.testing.assert_close(out_embeds, out_ids)
@require_non_xpu
@require_torch_multi_gpu
@@ -3170,10 +3170,10 @@ class ModelTesterMixin:
for value, parallel_value in zip(output, parallel_output):
if isinstance(value, torch.Tensor):
self.assertTrue(torch.allclose(value, parallel_value.to("cpu"), atol=1e-7))
torch.testing.assert_close(value, parallel_value.to("cpu"), rtol=1e-7, atol=1e-7)
elif isinstance(value, (Tuple, List)):
for value_, parallel_value_ in zip(value, parallel_value):
self.assertTrue(torch.allclose(value_, parallel_value_.to("cpu"), atol=1e-7))
torch.testing.assert_close(value_, parallel_value_.to("cpu"), rtol=1e-7, atol=1e-7)
def check_device_map_is_respected(self, model, device_map):
for param_name, param in model.named_parameters():
@@ -3229,9 +3229,12 @@ class ModelTesterMixin:
new_output = new_model(**inputs_dict_class)
if isinstance(base_output[0], tuple) and isinstance(new_output[0], tuple):
self.assertTrue(torch.allclose(a, b, atol=1e-5) for a, b in zip(base_output[0], new_output[0]))
[
torch.testing.assert_close(a, b, rtol=1e-5, atol=1e-5)
for a, b in zip(base_output[0], new_output[0])
]
else:
self.assertTrue(torch.allclose(base_output[0], new_output[0], atol=1e-5))
torch.testing.assert_close(base_output[0], new_output[0], rtol=1e-5, atol=1e-5)
@require_accelerate
@mark.accelerate_tests
@@ -3264,9 +3267,12 @@ class ModelTesterMixin:
new_output = new_model(**inputs_dict_class)
if isinstance(base_output[0], tuple) and isinstance(new_output[0], tuple):
self.assertTrue(torch.allclose(a, b, atol=1e-5) for a, b in zip(base_output[0], new_output[0]))
[
torch.testing.assert_close(a, b, rtol=1e-5, atol=1e-5)
for a, b in zip(base_output[0], new_output[0])
]
else:
self.assertTrue(torch.allclose(base_output[0], new_output[0], atol=1e-5))
torch.testing.assert_close(base_output[0], new_output[0], rtol=1e-5, atol=1e-5)
@require_accelerate
@mark.accelerate_tests
@@ -3303,9 +3309,12 @@ class ModelTesterMixin:
new_output = new_model(**inputs_dict_class)
if isinstance(base_output[0], tuple) and isinstance(new_output[0], tuple):
self.assertTrue(torch.allclose(a, b, atol=1e-5) for a, b in zip(base_output[0], new_output[0]))
[
torch.testing.assert_close(a, b, rtol=1e-5, atol=1e-5)
for a, b in zip(base_output[0], new_output[0])
]
else:
self.assertTrue(torch.allclose(base_output[0], new_output[0], atol=1e-5))
torch.testing.assert_close(base_output[0], new_output[0], rtol=1e-5, atol=1e-5)
@require_accelerate
@mark.accelerate_tests
@@ -3341,9 +3350,12 @@ class ModelTesterMixin:
new_output = new_model(**inputs_dict_class)
if isinstance(base_output[0], tuple) and isinstance(new_output[0], tuple):
self.assertTrue(torch.allclose(a, b, atol=1e-5) for a, b in zip(base_output[0], new_output[0]))
[
torch.testing.assert_close(a, b, rtol=1e-5, atol=1e-5)
for a, b in zip(base_output[0], new_output[0])
]
else:
self.assertTrue(torch.allclose(base_output[0], new_output[0], atol=1e-5))
torch.testing.assert_close(base_output[0], new_output[0], rtol=1e-5, atol=1e-5)
def test_problem_types(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
@@ -4555,10 +4567,10 @@ class ModelTesterMixin:
logits_padded = res_padded.logits[inputs_dict["attention_mask"].bool()]
logits_padfree = res_padfree.logits[0]
torch.testing.assert_close(logits_padded.argmax(-1), logits_padfree.argmax(-1), atol=0, rtol=0)
torch.testing.assert_close(logits_padded.argmax(-1), logits_padfree.argmax(-1), rtol=0, atol=0)
# acceptable numerical instability
tol = torch.finfo(torch.float16).eps
torch.testing.assert_close(logits_padded, logits_padfree, atol=tol, rtol=tol)
torch.testing.assert_close(logits_padded, logits_padfree, rtol=tol, atol=tol)
@is_pt_tf_cross_test
def test_tf_from_pt_safetensors(self):
@@ -4780,7 +4792,7 @@ class ModelTesterMixin:
self.assertEqual(tuple(last_token_logits.shape), (batch_size, 1, vocab_size))
# Assert the last tokens are actually the same (except for the natural fluctuation due to order of FP ops)
self.assertTrue(torch.allclose(all_logits[:, -1:, :], last_token_logits, atol=1e-5))
torch.testing.assert_close(all_logits[:, -1:, :], last_token_logits, rtol=1e-5, atol=1e-5)
@require_torch_gpu
def test_flex_attention_with_grads(self):