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
@@ -32,6 +32,7 @@ from transformers.testing_utils import (
|
||||
require_accelerate,
|
||||
require_bitsandbytes,
|
||||
require_torch,
|
||||
require_torch_sdpa,
|
||||
require_vision,
|
||||
slow,
|
||||
torch_device,
|
||||
@@ -481,6 +482,7 @@ class InstructBlipVideoForConditionalGenerationDecoderOnlyTest(
|
||||
test_resize_embeddings = False
|
||||
test_attention_outputs = False
|
||||
test_torchscript = False
|
||||
_is_composite = True
|
||||
|
||||
def setUp(self):
|
||||
self.model_tester = InstructBlipVideoForConditionalGenerationDecoderOnlyModelTester(self)
|
||||
@@ -550,6 +552,66 @@ class InstructBlipVideoForConditionalGenerationDecoderOnlyTest(
|
||||
model = InstructBlipVideoForConditionalGeneration.from_pretrained(model_name)
|
||||
self.assertIsNotNone(model)
|
||||
|
||||
@require_torch_sdpa
|
||||
def test_sdpa_can_dispatch_composite_models(self):
|
||||
"""
|
||||
Tests if composite models dispatch correctly on SDPA/eager when requested so when loading the model.
|
||||
This tests only by looking at layer names, as usually SDPA layers are calles "SDPAAttention".
|
||||
In contrast to the above test, this one checks if the "config._attn_implamentation" is a dict after the model
|
||||
is loaded, because we manually replicate requested attn implementation on each sub-config when loading.
|
||||
See https://github.com/huggingface/transformers/pull/32238 for more info
|
||||
|
||||
The test tries to cover most general cases of composite models, VLMs with vision and text configs. Any model
|
||||
that has a different set of sub-configs has to overwrite this test.
|
||||
"""
|
||||
if not self.has_attentions:
|
||||
self.skipTest(reason="Model architecture does not support attentions")
|
||||
|
||||
if not self._is_composite:
|
||||
self.skipTest(f"{self.all_model_classes[0].__name__} does not support SDPA")
|
||||
|
||||
for model_class in self.all_model_classes:
|
||||
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
||||
model = model_class(config)
|
||||
|
||||
with tempfile.TemporaryDirectory() as tmpdirname:
|
||||
model.save_pretrained(tmpdirname)
|
||||
model_sdpa = model_class.from_pretrained(tmpdirname)
|
||||
model_sdpa = model_sdpa.eval().to(torch_device)
|
||||
|
||||
text_attn = "sdpa" if model.language_model._supports_sdpa else "eager"
|
||||
vision_attn = "sdpa" if model.vision_model._supports_sdpa else "eager"
|
||||
qformer_attn = "sdpa" if model.qformer._supports_sdpa else "eager"
|
||||
|
||||
# `None` as it is the requested one which will be assigned to each sub-config
|
||||
# Sub-model will dispatch to SDPA if it can (checked below that `SDPA` layers are present)
|
||||
self.assertTrue(model.language_model.config._attn_implementation == text_attn)
|
||||
self.assertTrue(model.vision_model.config._attn_implementation == vision_attn)
|
||||
self.assertTrue(model.qformer.config._attn_implementation == qformer_attn)
|
||||
|
||||
model_eager = model_class.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.language_model.config._attn_implementation == "eager")
|
||||
self.assertTrue(model_eager.vision_model.config._attn_implementation == "eager")
|
||||
self.assertTrue(model_eager.qformer.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 and any(
|
||||
module_attn == "sdpa" for module_attn in [text_attn, vision_attn, qformer_attn]
|
||||
):
|
||||
raise ValueError("The SDPA model should have SDPA attention layers")
|
||||
|
||||
|
||||
# We will verify our results on an image of cute cats
|
||||
def prepare_video():
|
||||
|
||||
Reference in New Issue
Block a user