Skip to content

fix(dflash): gate flex_attention BACKEND=TRITON on torch>=2.11 (fixes ROCm) #773

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
jiapingW merged 2 commits into from
Aug 27, 2026
Merged

fix(dflash): gate flex_attention BACKEND=TRITON on torch>=2.11 (fixes ROCm) #773

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
50c3407
fix(dflash): gate flex_attention BACKEND=TRITON on torch>=2.11 (fixes…
Aug 20, 2026
cb1b5e3
Merge branch 'main' into fix/rocm-dflash-flex-backend
zhentaocc Aug 25, 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
16 changes: 13 additions & 3 deletions specforge/algorithms/common/dflash_family_model.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -25,7 +25,6 @@
if hasattr(torch, "npu") and torch.npu.is_available():
FLEX_ATTENTION_AVAILABLE = False


_VALID_LOSS_TYPES = {
"dflash",
"dpace",
Expand Down Expand Up @@ -363,9 +362,20 @@ def _forward_draft_blocks(
# DFlash's dynamic short-query batches are training/prefill shaped,
# not autoregressive decoding. AUTO may route q_len < 128 to the
# more restrictive flex-decoding kernel, whose config set can be
# empty for DFlash's sparse BlockMask. Keep the general Triton
# empty for DFlash's sparse BlockMask. Force the general Triton
# Flex Attention kernel for every DFlash-family batch.
draft_kwargs["kernel_options"] = {"BACKEND": "TRITON"}
#
# The "BACKEND" kernel_option only exists on torch >= 2.11, where
# the inductor lowering sanitizes it out of the generated Triton
# constexprs. On older builds (including current torch ROCm wheels)
# the string leaks into the kernel as a bare identifier and fails to
# compile (NameError: 'TRITON' is not defined), so we fall back to
# FORCE_USE_FLEX_ATTENTION, which selects the same kernel and has
# been supported since torch 2.5.
if torch.__version__ >= "2.11":
draft_kwargs["kernel_options"] = {"BACKEND": "TRITON"}
else:
draft_kwargs["kernel_options"] = {"FORCE_USE_FLEX_ATTENTION": True}
output_hidden = self.draft_model(
position_ids=full_position_ids,
noise_embedding=noise_embedding,
Expand Down
7 changes: 6 additions & 1 deletion tests/test_modeling/test_dflash_sliding.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -169,7 +169,12 @@ def test_online_wrapper_forces_standard_triton_flex_backend(self):
loss_mask=torch.ones(1, 4),
)

self.assertEqual(layers[0].kernel_options, {"BACKEND": "TRITON"})
expected_kernel_options = (
{"BACKEND": "TRITON"}
if torch.__version__ >= "2.11"
else {"FORCE_USE_FLEX_ATTENTION": True}
)
self.assertEqual(layers[0].kernel_options, expected_kernel_options)


class TestDFlashSlidingConfig(unittest.TestCase):
Expand Down
Loading