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

Fix flash attention bugs with Mistral and Falcon (#27625)

Browse Source 
* fix various bugs with flash attention

* bump

* fix test

* fix mistral

* use skiptest instead of return that may be misleading

* fix on review
This commit is contained in:
fxmarty
2023-11-21 15:20:44 +01:00
committed by GitHub
parent f93c1e9ece
commit 82cc0a79ac
5 changed files with 50 additions and 32 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
tests/models/llama/test_modeling_llama.py
View File

@@ -22,6 +22,7 @@ from parameterized import parameterized
from transformers import LlamaConfig, is_torch_available, set_seed
from transformers.testing_utils import (
require_bitsandbytes,
require_flash_attn,
require_torch,
require_torch_accelerator,
@@ -385,6 +386,7 @@ class LlamaModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixi
@require_flash_attn
@require_torch_gpu
@require_bitsandbytes
@pytest.mark.flash_attn_test
@slow
def test_flash_attn_2_generate_padding_right(self):

54
tests/models/mistral/test_modeling_mistral.py
View File

@@ -375,9 +375,6 @@ class MistralModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMi
import torch
for model_class in self.all_generative_model_classes:
if not model_class._supports_flash_attn_2:
return
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
model = model_class(config)
@@ -405,36 +402,49 @@ class MistralModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMi
@require_torch_gpu
@pytest.mark.flash_attn_test
@slow
def test_flash_attn_2_inference_padding_right(self):
def test_flash_attn_2_generate_use_cache(self):
import torch
for model_class in self.all_model_classes:
if not model_class._supports_flash_attn_2:
return
max_new_tokens = 30
for model_class in self.all_generative_model_classes:
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
dummy_input = inputs_dict[model_class.main_input_name]
if dummy_input.dtype in [torch.float32, torch.bfloat16]:
dummy_input = dummy_input.to(torch.float16)
# make sure that all models have enough positions for generation
if hasattr(config, "max_position_embeddings"):
config.max_position_embeddings = max_new_tokens + dummy_input.shape[1] + 1
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
model = model_class(config)
with tempfile.TemporaryDirectory() as tmpdirname:
model.save_pretrained(tmpdirname)
model_fa = model_class.from_pretrained(
tmpdirname, torch_dtype=torch.bfloat16, use_flash_attention_2=True
)
model_fa.to(torch_device)
dummy_attention_mask = inputs_dict.get("attention_mask", torch.ones_like(dummy_input))
# NOTE: Mistral apparently does not support right padding + use_cache with FA2.
dummy_attention_mask[:, -1] = 1
model = model_class.from_pretrained(
tmpdirname, torch_dtype=torch.bfloat16, use_flash_attention_2=False
tmpdirname,
torch_dtype=torch.float16,
use_flash_attention_2=True,
low_cpu_mem_usage=True,
).to(torch_device)
# Just test that a large cache works as expected
_ = model.generate(
dummy_input, attention_mask=dummy_attention_mask, max_new_tokens=max_new_tokens, do_sample=False
)
model.to(torch_device)
dummy_input = torch.LongTensor([[1, 2, 3, 4, 5]]).to(torch_device)
dummy_attention_mask = torch.LongTensor([[1, 1, 1, 1, 0]]).to(torch_device)
_ = model(dummy_input, output_hidden_states=True).hidden_states[-1]
with self.assertRaises(ValueError):
_ = model_fa(
dummy_input, attention_mask=dummy_attention_mask, output_hidden_states=True
).hidden_states[-1]
@require_flash_attn
@require_torch_gpu
@pytest.mark.flash_attn_test
@slow
def test_flash_attn_2_inference_padding_right(self):
self.skipTest("Mistral flash attention does not support right padding")
@require_torch