Attn implementation for composite models (#32238)
* first try * codestyle * idefics2 is happy * [run-slow] llava, llava_next, video_llava, vipllava, llava_next_video, idefics, idefics2, kosmos2, fuyu, blip, blip_2, instructblip, instructblipvideo, paligemma * fix-copies * [run-slow] llava, llava_next, video_llava, vipllava, llava_next_video, idefics, idefics2, kosmos2, fuyu, blip, blip_2, instructblip, instructblipvideo * blip-2 needs to init vision from config * when was this removed O_o * minor fix * tests * this way? * tests * model-agnostic code * codestyle * add tests for idefics * modify general test for VLMs * no generation test for vlm yet! * no generation test here also * wanr in VIT-SDPA if output attn * add more tests * user can pass dict as attn impl * repo consistency * update * muicgen * no prints * forgot speech enc-dec and clip * how many composite models we have? * musicgen meelody is same as mudicgen * +siglip * fix tests + add some more * remove idefics custom overriden code * make idefics2 automappable * nits * skip tests * doctests * Update src/transformers/models/idefics2/configuration_idefics2.py Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com> * Update tests/models/clip/test_modeling_clip.py Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com> * Update tests/models/idefics2/test_modeling_idefics2.py Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com> * Update tests/models/idefics2/test_modeling_idefics2.py Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com> * Update src/transformers/configuration_utils.py Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com> * major update, no need for automap * clean up * add FA2 test * more tests * style * skip tests * why did these started failing now? * no attributes for FA2 needed * one tiny test * address comment about FA2 false warning * style * add new models and resolve conflicts * fix copies * let it be this way for now, come back tomorrow to review * some more fixes * update * more updates * update * fix copies * style and tests * another big update * fix tests * fix tests * update * another update * fix tests * fix copies * fix tests --------- Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>
This commit is contained in:
Raushan Turganbay
committed by
GitHub
GitHub
parent
32590b5ecb
commit
21d5025826
@@ -18,7 +18,13 @@ import tempfile
|
||||
import unittest
|
||||
|
||||
from transformers import is_torch_available
|
||||
from transformers.testing_utils import require_deterministic_for_xpu, require_torch, slow, torch_device
|
||||
from transformers.testing_utils import (
|
||||
require_deterministic_for_xpu,
|
||||
require_torch,
|
||||
require_torch_sdpa,
|
||||
slow,
|
||||
torch_device,
|
||||
)
|
||||
|
||||
from ...test_modeling_common import floats_tensor, ids_tensor, random_attention_mask
|
||||
from ..bert.test_modeling_bert import BertModelTester
|
||||
@@ -441,6 +447,66 @@ class EncoderDecoderMixin:
|
||||
max_diff = np.amax(np.abs(out_1 - out_2))
|
||||
self.assertLessEqual(max_diff, 1e-5)
|
||||
|
||||
@require_torch_sdpa
|
||||
def test_sdpa_can_dispatch_composite_models(self):
|
||||
inputs_dict = self.prepare_config_and_inputs()
|
||||
encoder_config, decoder_config = inputs_dict["config"], inputs_dict["decoder_config"]
|
||||
config = SpeechEncoderDecoderConfig.from_encoder_decoder_configs(
|
||||
encoder_config=encoder_config, decoder_config=decoder_config
|
||||
)
|
||||
model = SpeechEncoderDecoderModel(config=config)
|
||||
|
||||
with tempfile.TemporaryDirectory() as tmpdirname:
|
||||
model.save_pretrained(tmpdirname)
|
||||
model_sdpa = SpeechEncoderDecoderModel.from_pretrained(tmpdirname)
|
||||
model_sdpa = model_sdpa.eval().to(torch_device)
|
||||
|
||||
# see https://github.com/huggingface/transformers/pull/32238
|
||||
# Sub-model will dispatch to SDPA if it can (checked below that `SDPA` layers are present)
|
||||
encoder_attn = "sdpa" if model.encoder._supports_sdpa else "eager"
|
||||
decoder_attn = "sdpa" if model.decoder._supports_sdpa else "eager"
|
||||
self.assertTrue(model_sdpa.config._attn_implementation == "sdpa")
|
||||
self.assertTrue(model_sdpa.encoder.config._attn_implementation == encoder_attn)
|
||||
self.assertTrue(model_sdpa.decoder.config._attn_implementation == decoder_attn)
|
||||
|
||||
# Also test that nothing break if we request SDPA explicitly, when both sub-parts support it.
|
||||
# If the model supports sdpa (i.e. all of sub-models supports it) we'll dispatch safely
|
||||
# Otherwise we should raise error that SDPA is not supported, as some of the sub-models doesn't support
|
||||
if encoder_attn == "sdpa" and decoder_attn == "sdpa":
|
||||
model_sdpa_explicit = SpeechEncoderDecoderModel.from_pretrained(tmpdirname, attn_implementation="sdpa")
|
||||
model_sdpa_explicit = model_sdpa_explicit.eval().to(torch_device)
|
||||
|
||||
self.assertTrue(model_sdpa_explicit.config._attn_implementation == "sdpa")
|
||||
else:
|
||||
with self.assertRaises(ValueError):
|
||||
model_sdpa_explicit = SpeechEncoderDecoderModel.from_pretrained(
|
||||
tmpdirname, attn_implementation="sdpa"
|
||||
)
|
||||
|
||||
model_eager = SpeechEncoderDecoderModel.from_pretrained(
|
||||
tmpdirname,
|
||||
attn_implementation="eager",
|
||||
)
|
||||
model_eager = model_eager.eval().to(torch_device)
|
||||
|
||||
self.assertTrue(model_eager.config._attn_implementation == "eager")
|
||||
self.assertTrue(model_eager.encoder.config._attn_implementation == "eager")
|
||||
self.assertTrue(model_eager.decoder.config._attn_implementation == "eager")
|
||||
|
||||
for name, submodule in model_eager.named_modules():
|
||||
class_name = submodule.__class__.__name__
|
||||
if "SdpaAttention" in class_name or "SdpaSelfAttention" in class_name:
|
||||
raise ValueError("The eager model should not have SDPA attention layers")
|
||||
|
||||
has_sdpa = False
|
||||
for name, submodule in model_sdpa.named_modules():
|
||||
class_name = submodule.__class__.__name__
|
||||
if "SdpaAttention" in class_name or "SdpaSelfAttention" in class_name:
|
||||
has_sdpa = True
|
||||
break
|
||||
if not has_sdpa:
|
||||
raise ValueError("The SDPA model should have SDPA attention layers")
|
||||
|
||||
|
||||
@require_torch
|
||||
class Wav2Vec2BertModelTest(EncoderDecoderMixin, unittest.TestCase):
|
||||
|
||||
Reference in New Issue
Block a user