Skip to content

FA4 TEDotProductAttention #4375

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

Closed
bbuschkaemper wants to merge 4 commits into from
Closed

FA4 TEDotProductAttention #4375

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
065e1f0
Pin TE to FA4-support commit. Allow unit tests to use FA4 instead of …
bbuschkaemper Apr 18, 2026
a5392e0
Merge branch 'main' into fa4-training
bbuschkaemper Apr 22, 2026
fe9ef44
Bump TE version to FA4 fix.
bbuschkaemper Apr 24, 2026
af19ad4
Merge branch 'main' into fa4-training
bbuschkaemper Apr 24, 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
2 changes: 1 addition & 1 deletion pyproject.toml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -200,7 +200,7 @@ override-dependencies = [
flash_mla = [
{ git = "https://github.com/deepseek-ai/FlashMLA", rev = "9edee0c022cd0938148a18e334203b0aab43aa19" },
]
transformer-engine = { git = "https://github.com/NVIDIA/TransformerEngine.git", rev = "f031cf87bd054c7558b887df7bed93975456667f" }
transformer-engine = { git = "https://github.com/NVIDIA/TransformerEngine.git", rev = "9e55a255dd2d63bbf6d2c6ec788d0fd27965b42b" }
nemo-run = { git = "https://github.com/NVIDIA-NeMo/Run.git", rev = "17ae86b64d7f75653351664f5d8c9e466faede00" }
emerging_optimizers = { git = "https://github.com/NVIDIA-NeMo/Emerging-Optimizers.git", rev = "v0.2.0" }
nvidia-resiliency-ext = { git = "https://github.com/NVIDIA/nvidia-resiliency-ext.git", rev = "b2bb3d728a18795807d9f76c535e005a609a1b01" }
Expand Down
11 changes: 9 additions & 2 deletions tests/unit_tests/models/test_gpt_model_batch_invariant.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -36,6 +36,13 @@
except ImportError:
HAVE_FA3 = False

try:
from flash_attn.cute.interface import flash_attn_func

HAVE_FA4 = True
except ImportError:
HAVE_FA4 = False


class DummyTokenizer:
def __init__(self, vocab_size: int, bos: int | None = None, eod: int = 0, pad: int = 0):
Expand Down Expand Up @@ -118,8 +125,8 @@ def _train_forward_logprobs(model: torch.nn.Module, tokens: torch.Tensor) -> tor


@pytest.mark.skipif(
not (is_te_min_version("2.10.0") and HAVE_FA3),
reason="TestGPTModelBatchInvariant requires TE >= 2.10.0 and FlashAttention-3",
not (is_te_min_version("2.10.0") and (HAVE_FA3 or HAVE_FA4)),
reason="TestGPTModelBatchInvariant requires TE >= 2.10.0 and FlashAttention-3 or -4",
)
class TestGPTModelBatchInvariant:
"""End-to-end batch-invariance tests for GPT."""
Expand Down
11 changes: 9 additions & 2 deletions tests/unit_tests/transformer/test_te_layers_batch_invariant.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -29,6 +29,13 @@
except ImportError:
HAVE_FA3 = False

try:
from flash_attn.cute.interface import flash_attn_func

HAVE_FA4 = True
except ImportError:
HAVE_FA4 = False


# ============================================================================
# Batch-Invariant test helpers
Expand Down Expand Up @@ -312,8 +319,8 @@ def test_column_parallel_linear_batch_invariant_randomized():


@pytest.mark.skipif(
not (is_te_min_version("2.10.0") and HAVE_FA3),
reason="TE attention BIK tests require TE >= 2.10.0 and FlashAttention-3",
not (is_te_min_version("2.10.0") and (HAVE_FA3 or HAVE_FA4)),
reason="TE attention BIK tests require TE >= 2.10.0 and FlashAttention-3 or -4",
)
def test_te_attention_layer_batch_invariant_randomized():
torch.backends.cuda.matmul.allow_tf32 = False
Expand Down
Loading
Loading