Skip to content

Gemma attn #5346

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.

Sign up for GitHub

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

Merged
danielhanchen merged 3 commits into from
May 13, 2026
Merged

Gemma attn #5346

Changes from 2 commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
ef52526
Improve attn refactor
Datta0 May 8, 2026
426bf74
refactor attention impl determine
Datta0 May 9, 2026
6640371
Merge branch 'main' into gemma_attn
Datta0 May 13, 2026
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
67 changes: 44 additions & 23 deletions unsloth/models/_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -234,6 +234,7 @@ def apply_unsloth_gradient_checkpointing(
# ModernBERT: create_block_mask with _compile=True hits CUDA illegal memory
# access on some GPU architectures (B200). Falls back to eager safely.
_FLEX_EXCLUDED_MODELS = ("gpt_oss", "mllama", "nemotron_h", "modernbert")
_FLEX_PREFERRED_MODELS = ("gemma3", "gemma3_text", "shieldgemma2")
_EAGER_ONLY_PREFIXES = ("gemma3n",)
_FLASH_ATTENTION_MAX_HEAD_DIM = 256
_FLASH_ATTENTION_DISABLED_WARNED = set()
Expand All @@ -243,10 +244,37 @@ def _is_flex_excluded(model_type):
return model_type in _FLEX_EXCLUDED_MODELS


def _config_prefers_flex_attention(config):
return any(
_config_get(attention_config, "model_type", "").lower()
in _FLEX_PREFERRED_MODELS
for attention_config in _iter_attention_configs(config)
)


def _is_eager_only(model_type):
return any(model_type.startswith(p) for p in _EAGER_ONLY_PREFIXES)


def _supports_flex_attention(model_class, config, model_type):
if os.environ.get("UNSLOTH_ENABLE_FLEX_ATTENTION", "1") == "0":
return False
if not getattr(model_class, "_supports_flex_attn", False):
return False
if _is_flex_excluded(model_type):
return False
for attention_config in _iter_attention_configs(config):
attention_dropout = _config_get(attention_config, "attention_dropout", 0) or 0
if attention_dropout != 0:
return False
try:
from transformers.utils.import_utils import is_torch_flex_attn_available

return is_torch_flex_attn_available()
except Exception:
return False


def _config_items(config):
if isinstance(config, dict):
return config.items()
Expand Down Expand Up @@ -465,8 +493,23 @@ def resolve_attention_implementation(
if model_class is None:
attn_impl = _set_attn_impl(config, "sdpa" if supports_sdpa else "eager")
else:
supports_flex_attention = _supports_flex_attention(
model_class, config, model_type
)
prefers_flex_attention = _config_prefers_flex_attention(config)
if _is_eager_only(model_type):
attn_impl = _set_attn_impl(config, "eager")
elif prefers_flex_attention and supports_flex_attention:
attn_impl = _set_attn_impl(config, "flex_attention")
elif (
not prefers_flex_attention
and not flash_attention_disabled
Comment on lines +505 to +506

@chatgpt-codex-connector chatgpt-codex-connector Bot May 13, 2026

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

P1 Badge Allow flash fallback when preferred flex backend is unavailable

The new not prefers_flex_attention guard in resolve_attention_implementation prevents all preferred models (e.g. gemma3) from ever selecting flash_attention_2 unless flex attention is already available. In environments where is_torch_flex_attn_available() is false but Flash Attention is installed, this now falls through to SDPA/eager instead of using flash, which is a regression from the previous selection order and can reintroduce the SDPA-path failures this change is trying to avoid.

Useful? React with 👍 / 👎.

and HAS_FLASH_ATTENTION
and supports_flash_attention
):
attn_impl = _set_attn_impl(config, "flash_attention_2")
Comment on lines +504 to +510

@gemini-code-assist gemini-code-assist Bot May 9, 2026

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

high

The condition not prefers_flex_attention at line 505 creates a logic bug where models that prefer Flex Attention but cannot use it (e.g., due to an older PyTorch version or environment settings) will skip Flash Attention 2 even if it is available and supported. This forces a fallback to SDPA or Eager, which is significantly less efficient. Since the previous elif block already handles the case where Flex is both preferred and supported, this check is unnecessary and harmful for fallback scenarios.

Suggested change
elif (
not prefers_flex_attention
and not flash_attention_disabled
and HAS_FLASH_ATTENTION
and supports_flash_attention
):
attn_impl = _set_attn_impl(config, "flash_attention_2")
elif (
not flash_attention_disabled
and HAS_FLASH_ATTENTION
and supports_flash_attention
):
attn_impl = _set_attn_impl(config, "flash_attention_2")

elif supports_flex_attention:
attn_impl = _set_attn_impl(config, "flex_attention")
elif flash_attention_disabled:
attn_impl = _disable_flash_attention_if_needed(
config,
Expand All @@ -476,32 +519,10 @@ def resolve_attention_implementation(
),
disable_reason = disable_reason,
)
elif HAS_FLASH_ATTENTION and supports_flash_attention:
attn_impl = _set_attn_impl(config, "flash_attention_2")
elif supports_sdpa:
attn_impl = _set_attn_impl(config, "sdpa")
else:
attn_impl = "eager"
if os.environ.get("UNSLOTH_ENABLE_FLEX_ATTENTION", "1") != "0":
try:
from transformers.utils.import_utils import (
is_torch_flex_attn_available,
)

if (
is_torch_flex_attn_available()
and getattr(model_class, "_supports_flex_attn", False)
and not _is_flex_excluded(model_type)
):
attention_dropout = (
_config_get(config, "attention_dropout", 0) or 0
)
if attention_dropout == 0:
attn_impl = _set_attn_impl(config, "flex_attention")
except Exception:
pass
if attn_impl == "eager":
attn_impl = _set_attn_impl(config, "eager")
attn_impl = _set_attn_impl(config, "eager")

if requested_attn_implementation is None:
final_attn_impl = attn_impl
Expand Down
Loading