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

F.scaled_dot_product_attention support (#26572)

Browse Source 
* add sdpa

* wip

* cleaning

* add ref

* yet more cleaning

* and more :)

* wip llama

* working llama

* add output_attentions=True support

* bigcode sdpa support

* fixes

* gpt-bigcode support, require torch>=2.1.1

* add falcon support

* fix conflicts falcon

* style

* fix attention_mask definition

* remove output_attentions from attnmaskconverter

* support whisper without removing any Copied from statement

* fix mbart default to eager renaming

* fix typo in falcon

* fix is_causal in SDPA

* check is_flash_attn_2_available in the models init as well in case the model is not initialized through from_pretrained

* add warnings when falling back on the manual implementation

* precise doc

* wip replace _flash_attn_enabled by config.attn_implementation

* fix typo

* add tests

* style

* add a copy.deepcopy on the config in from_pretrained, as we do not want to modify it inplace

* obey to config.attn_implementation if a config is passed in from_pretrained

* fix is_torch_sdpa_available when torch is not installed

* remove dead code

* Update src/transformers/modeling_attn_mask_utils.py

Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

* Update src/transformers/modeling_attn_mask_utils.py

Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

* Update src/transformers/modeling_attn_mask_utils.py

Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

* Update src/transformers/modeling_attn_mask_utils.py

Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

* Update src/transformers/modeling_attn_mask_utils.py

Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

* Update src/transformers/models/bart/modeling_bart.py

Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

* remove duplicate pretraining_tp code

* add dropout in llama

* precise comment on attn_mask

* add fmt: off for _unmask_unattended docstring

* precise num_masks comment

* nuke pretraining_tp in LlamaSDPAAttention following Arthur's suggestion

* cleanup modeling_utils

* backward compatibility

* fix style as requested

* style

* improve documentation

* test pass

* style

* add _unmask_unattended tests

* skip meaningless tests for idefics

* hard_check SDPA requirements when specifically requested

* standardize the use if XXX_ATTENTION_CLASSES

* fix SDPA bug with mem-efficient backend on CUDA when using fp32

* fix test

* rely on SDPA is_causal parameter to handle the causal mask in some cases

* fix FALCON_ATTENTION_CLASSES

* remove _flash_attn_2_enabled occurences

* fix test

* add OPT to the list of supported flash models

* improve test

* properly test on different SDPA backends, on different dtypes & properly handle separately the pad tokens in the test

* remove remaining _flash_attn_2_enabled occurence

* Update src/transformers/modeling_utils.py

Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

* Update src/transformers/modeling_utils.py

Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

* Update src/transformers/modeling_utils.py

Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

* Update src/transformers/modeling_attn_mask_utils.py

Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

* Update docs/source/en/perf_infer_gpu_one.md

Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

* remove use_attn_implementation

* fix docstring & slight bug

* make attn_implementation internal (_attn_implementation)

* typos

* fix tests

* deprecate use_flash_attention_2=True

* fix test

* add back llama that was removed by mistake

* fix tests

* remove _flash_attn_2_enabled occurences bis

* add check & test that passed attn_implementation is valid

* fix falcon torchscript export

* fix device of mask in tests

* add tip about torch.jit.trace and move bt doc below sdpa

* fix parameterized.expand order

* move tests from test_modeling_attn_mask_utils to test_modeling_utils as a relevant test class is already there

* update sdpaattention class with the new cache

* Update src/transformers/configuration_utils.py

Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

* Update src/transformers/models/bark/modeling_bark.py

* address review comments

* WIP torch.jit.trace fix. left: test both eager & sdpa

* add test for torch.jit.trace for both eager/sdpa

* fix falcon with torch==2.0 that needs to use sdpa

* fix doc

* hopefully last fix

* fix key_value_length that has no default now in mask converter

* is it flacky?

* fix speculative decoding bug

* tests do pass

* fix following #27907

---------

Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>
This commit is contained in:
fxmarty
2023-12-08 21:38:14 +01:00
committed by GitHub
parent ce0bbd5101
commit 80377eb018
54 changed files with 2227 additions and 454 deletions
Download Patch File Download Diff File Expand all files Collapse all files

83
tests/models/llama/test_modeling_llama.py
View File

@@ -14,6 +14,7 @@
# limitations under the License.
""" Testing suite for the PyTorch LLaMA model. """
import tempfile
import unittest
import pytest
@@ -26,6 +27,7 @@ from transformers.testing_utils import (
require_torch,
require_torch_accelerator,
require_torch_gpu,
require_torch_sdpa,
slow,
torch_device,
)
@@ -411,7 +413,7 @@ class LlamaModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixi
output_native = tokenizer.batch_decode(output_native)
model = LlamaForCausalLM.from_pretrained(
"meta-llama/Llama-2-7b-hf", load_in_4bit=True, device_map={"": 0}, use_flash_attention_2=True
"meta-llama/Llama-2-7b-hf", load_in_4bit=True, device_map={"": 0}, attn_implementation="flash_attention_2"
)
output_fa_2 = model.generate(**inputs, max_new_tokens=20, do_sample=False)
@@ -419,6 +421,85 @@ class LlamaModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixi
self.assertListEqual(output_native, output_fa_2)
@require_flash_attn
@require_torch_gpu
@slow
def test_use_flash_attention_2_true(self):
"""
NOTE: this is the only test testing that the legacy `use_flash_attention=2` argument still works as intended.
"""
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
with tempfile.TemporaryDirectory() as tmp_dir:
model = model_class(config)
model.save_pretrained(tmp_dir)
new_model = LlamaForCausalLM.from_pretrained(
tmp_dir, use_flash_attention_2=True, torch_dtype=torch.float16
).to("cuda")
self.assertTrue(new_model.config._attn_implementation == "flash_attention_2")
has_flash = False
for name, submodule in new_model.named_modules():
if "FlashAttention" in submodule.__class__.__name__:
has_flash = True
break
if not has_flash:
raise ValueError("The flash model should have flash attention layers")
@require_torch_sdpa
@slow
def test_eager_matches_sdpa_generate(self):
"""
Overwritting the common test as the test is flaky on tiny models
"""
max_new_tokens = 30
tokenizer = LlamaTokenizer.from_pretrained("meta-llama/Llama-2-7b-hf")
model_sdpa = LlamaForCausalLM.from_pretrained(
"meta-llama/Llama-2-7b-hf",
torch_dtype=torch.float16,
low_cpu_mem_usage=True,
).to(torch_device)
self.assertTrue(model_sdpa.config._attn_implementation == "sdpa")
model_eager = LlamaForCausalLM.from_pretrained(
"meta-llama/Llama-2-7b-hf",
torch_dtype=torch.float16,
low_cpu_mem_usage=True,
attn_implementation="eager",
).to(torch_device)
self.assertTrue(model_eager.config._attn_implementation == "eager")
for name, submodule in model_eager.named_modules():
if "SdpaAttention" in submodule.__class__.__name__:
raise ValueError("The eager model should not have SDPA attention layers")
has_sdpa = False
for name, submodule in model_sdpa.named_modules():
if "SdpaAttention" in submodule.__class__.__name__:
has_sdpa = True
break
if not has_sdpa:
raise ValueError("The SDPA model should have SDPA attention layers")
texts = ["hi", "Hello this is a very long sentence my friend", "Today I am in Paris and"]
for padding_side in ["left", "right"]:
tokenizer.padding_side = padding_side
tokenizer.pad_token = tokenizer.eos_token
inputs = tokenizer(texts, return_tensors="pt", padding=True).to(torch_device)
res_eager = model_eager.generate(**inputs, max_new_tokens=max_new_tokens, do_sample=False)
res_sdpa = model_sdpa.generate(**inputs, max_new_tokens=max_new_tokens, do_sample=False)
self.assertTrue(torch.allclose(res_eager, res_sdpa))
@require_torch
class LlamaIntegrationTest(unittest.TestCase):