Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Restore SDPA in Gemma2 models for transformers > 4.45 #976

Merged
merged 3 commits into from
Oct 28, 2024
Merged
Show file tree
Hide file tree
Changes from 1 commit
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
20 changes: 20 additions & 0 deletions optimum/exporters/openvino/model_patcher.py
Original file line number Diff line number Diff line change
Expand Up @@ -2505,6 +2505,26 @@ def patched_forward(*args, **kwargs):

self.patched_forward = patched_forward

def __enter__(self):
super().__enter__()
if is_transformers_version(">=", "4.45.0"):
from transformers.models.gemma2.modeling_gemma2 import GEMMA2_ATTENTION_CLASSES

sdpa_attn = GEMMA2_ATTENTION_CLASSES["sdpa"]
eager_attn = GEMMA2_ATTENTION_CLASSES["eager"]

for layer in self._model.model.layers:
if isinstance(layer.self_attn, eager_attn):
layer.self_attn._orig_forward = layer.self_attn.forward
layer.self_attn.forward = types.MethodType(sdpa_attn.forward, layer.self_attn)

def __exit__(self, exc_type, exc_value, traceback):
super().__exit__(exc_type, exc_value, traceback)
if is_transformers_version(">=", "4.45.0"):
for layer in self._model.model.layers:
if hasattr(layer.self_attn, "_orig_forward"):
layer.self_attn.forward = layer.self_attn._orig_forward


def _decilm_attn_forward(
self,
Expand Down
8 changes: 8 additions & 0 deletions tests/openvino/test_modeling.py
Original file line number Diff line number Diff line change
Expand Up @@ -863,6 +863,10 @@ def test_compare_to_transformers(self, model_arch):
if model_arch in self.REMOTE_CODE_MODELS:
model_kwargs = {"trust_remote_code": True}

# starting from transformers 4.45.0 gemma2 uses eager attention by default, while ov - sdpa
if model_arch == "gemma2" and is_transformers_version(">=", "4.45.0"):
model_kwargs["attn_implemenation"] = "sdpa"
eaidova marked this conversation as resolved.
Show resolved Hide resolved

ov_model = OVModelForCausalLM.from_pretrained(model_id, export=True, ov_config=F32_CONFIG, **model_kwargs)
self.assertIsInstance(ov_model.config, PretrainedConfig)
self.assertTrue(ov_model.use_cache)
Expand Down Expand Up @@ -1094,6 +1098,10 @@ def test_beam_search(self, model_arch):
"config": AutoConfig.from_pretrained(model_id, trust_remote_code=True),
"trust_remote_code": True,
}

# starting from transformers 4.45.0 gemma2 uses eager attention by default, while ov - sdpa
if model_arch == "gemma2" and is_transformers_version(">=", "4.45.0"):
model_kwargs["attn_implemenation"] = "sdpa"
eaidova marked this conversation as resolved.
Show resolved Hide resolved
# Qwen tokenizer does not support padding, chatglm, glm4 testing models produce nan that incompatible with beam search
if model_arch in ["qwen", "chatglm", "glm4"]:
return
Expand Down
Loading