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

Fix more inefficient PT operations (#37060)

Browse Source 
* Fix inefficient operations

* Remove cpu() call

* Reorder detach()

* Reorder detach()

* tolist without detach

* item without detach

* Update src/transformers/models/rag/modeling_rag.py

Co-authored-by: Joao Gante <joaofranciscocardosogante@gmail.com>

* Update tests/models/encodec/test_modeling_encodec.py

Co-authored-by: Joao Gante <joaofranciscocardosogante@gmail.com>

* Use detach().cpu().numpy

* Revert some numpy operations

* More fixes

---------

Co-authored-by: Joao Gante <joaofranciscocardosogante@gmail.com>
This commit is contained in:
cyyever
2025-03-31 23:31:24 +08:00
committed by GitHub
parent a1e389e637
commit 786d9c5ed9
54 changed files with 106 additions and 104 deletions
Download Patch File Download Diff File Expand all files Collapse all files

18
tests/models/moshi/test_modeling_moshi.py
View File

@@ -951,8 +951,8 @@ class MoshiIntegrationTests(unittest.TestCase):
expected_text_token = 452
expected_audio_tokens = [916, 1396, 1238, 579, 1105, 914, 1257, 810] # fmt: skip
self.assertTrue(expected_text_token == model_outputs.sequences[0, -2].cpu().item())
self.assertTrue(expected_audio_tokens == model_outputs.audio_codes[0, :, -1].cpu().tolist())
self.assertTrue(expected_text_token == model_outputs.sequences[0, -2].item())
self.assertTrue(expected_audio_tokens == model_outputs.audio_codes[0, :, -1].tolist())
@slow
def test_moshiko_greedy_unconditional_fp16_eager(self):
@@ -966,7 +966,7 @@ class MoshiIntegrationTests(unittest.TestCase):
)
# eager equivalence is not as strict as sdpa.
self.assertTrue(some_expected_audio_tokens == model_outputs.audio_codes[0, :, :2].cpu().tolist())
self.assertTrue(some_expected_audio_tokens == model_outputs.audio_codes[0, :, :2].tolist())
@slow
def test_moshiko_greedy_unconditional_fp32(self):
@@ -986,8 +986,8 @@ class MoshiIntegrationTests(unittest.TestCase):
audio_code_sums = model_outputs.audio_codes.sum().item()
self.assertTrue(np.abs(audio_code_sums - expected_audio_codesum) <= (3e-3 * audio_code_sums))
self.assertTrue(expected_text_tokens == model_outputs.sequences[0, 1:].cpu().tolist())
self.assertTrue(some_expected_audio_tokens == model_outputs.audio_codes[0, :, :2].cpu().tolist())
self.assertTrue(expected_text_tokens == model_outputs.sequences[0, 1:].tolist())
self.assertTrue(some_expected_audio_tokens == model_outputs.audio_codes[0, :, :2].tolist())
@slow
@require_torch_fp16
@@ -1008,8 +1008,8 @@ class MoshiIntegrationTests(unittest.TestCase):
audio_code_sums = model_outputs.audio_codes.sum().item()
self.assertTrue(np.abs(audio_code_sums - expected_audio_codesum) <= (3e-3 * audio_code_sums))
self.assertTrue(expected_text_tokens == model_outputs.sequences[0, 1:].cpu().tolist())
self.assertTrue(some_expected_audio_tokens == model_outputs.audio_codes[0, :, :2].cpu().tolist())
self.assertTrue(expected_text_tokens == model_outputs.sequences[0, 1:].tolist())
self.assertTrue(some_expected_audio_tokens == model_outputs.audio_codes[0, :, :2].tolist())
@slow
@require_torch_fp16
@@ -1030,5 +1030,5 @@ class MoshiIntegrationTests(unittest.TestCase):
audio_code_sums = model_outputs.audio_codes.sum().item()
self.assertTrue(np.abs(audio_code_sums - expected_audio_codesum) <= 2048)
self.assertTrue(expected_text_tokens == model_outputs.sequences[0, 1:].cpu().tolist())
self.assertTrue(some_expected_audio_tokens == model_outputs.audio_codes[0, :, :2].cpu().tolist())
self.assertTrue(expected_text_tokens == model_outputs.sequences[0, 1:].tolist())
self.assertTrue(some_expected_audio_tokens == model_outputs.audio_codes[0, :, :2].tolist())