Skip to content

Fix THD P2P pad detection and tail-zero gating #3506

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
RPalmr wants to merge 13 commits into from
Closed

Fix THD P2P pad detection and tail-zero gating #3506

Show file tree
Hide file tree
Changes from 9 commits
Commits
Show all changes
13 commits
Select commit Hold shift + click to select a range
64c6e63
Fix THD P2P pad detection and tail-zero gating
RPalmr Sep 10, 2026
0b194aa
Merge branch 'NVIDIA:main' into fix/thd-p2p-pad-detect
RPalmr Sep 10, 2026
b5063c6
[pre-commit.ci] auto fixes from pre-commit.com hooks
pre-commit-ci[bot] Sep 10, 2026
6e05fa0
Merge branch 'NVIDIA:main' into fix/thd-p2p-pad-detect
RPalmr Sep 10, 2026
9c7b1e6
Reuse eager pad decision during graph capture
RPalmr Sep 10, 2026
0c3295d
[pre-commit.ci] auto fixes from pre-commit.com hooks
pre-commit-ci[bot] Sep 10, 2026
998ea38
Include tensor version in pad cache key
RPalmr Sep 10, 2026
83d0411
[pre-commit.ci] auto fixes from pre-commit.com hooks
pre-commit-ci[bot] Sep 10, 2026
f958198
Merge branch 'main' into fix/thd-p2p-pad-detect
RPalmr Sep 12, 2026
b29514b
Use id-version key with lru_cache for pad decision
RPalmr Sep 15, 2026
f3325ff
Merge origin/main into fix/thd-p2p-pad-detect
RPalmr Sep 15, 2026
38bbb40
[pre-commit.ci] auto fixes from pre-commit.com hooks
pre-commit-ci[bot] Sep 15, 2026
d42dfab
Merge branch 'main' into fix/thd-p2p-pad-detect
RPalmr Sep 16, 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
58 changes: 37 additions & 21 deletions transformer_engine/pytorch/attention/dot_product_attention/backends.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1868,29 +1868,45 @@ def backward(ctx, d_out, *_args):
dq = dq[..., : d_out.shape[-1]]
dk = dk[..., : d_out.shape[-1]]
dv = dv[..., : d_out.shape[-1]]
# Zero-fill positions beyond cu_seqlens_*_padded[-1] in dQ/dK/dV for THD.
# Use Q's padded boundary for dQ and KV's padded boundary for dK/dV.
# Sync-free `arange + mask` so capture and eager paths run the same code.
_qkv_format = ctx.qkv_layout.split("_")[0].replace("3", "").replace("2", "")
if _qkv_format == "thd":
if (
cu_seqlens_q_padded is not None
and isinstance(dq, torch.Tensor)
and dq.shape[0] > 0
):
q_pad_mask = (
torch.arange(dq.shape[0], device=dq.device) >= cu_seqlens_q_padded[-1]
)
dq[q_pad_mask] = 0
if cu_seqlens_kv_padded is not None:
kv_actual_t = cu_seqlens_kv_padded[-1]
for d_tensor in (dk, dv):
if isinstance(d_tensor, torch.Tensor) and d_tensor.shape[0] > 0:
kv_pad_mask = (
torch.arange(d_tensor.shape[0], device=d_tensor.device)
>= kv_actual_t
)
d_tensor[kv_pad_mask] = 0
if is_graph_capturing():
if (
cu_seqlens_q_padded is not None
and isinstance(dq, torch.Tensor)
and dq.shape[0] > 0
):
q_pad_mask = (
torch.arange(dq.shape[0], device=dq.device) >= cu_seqlens_q_padded[-1]
)
dq[q_pad_mask] = 0
if cu_seqlens_kv_padded is not None:
kv_actual_t = cu_seqlens_kv_padded[-1]
for d_tensor in (dk, dv):
if isinstance(d_tensor, torch.Tensor) and d_tensor.shape[0] > 0:
kv_pad_mask = (
torch.arange(d_tensor.shape[0], device=d_tensor.device)
>= kv_actual_t
)
d_tensor[kv_pad_mask] = 0
else:
if (
cu_seqlens_q_padded is not None
and isinstance(dq, torch.Tensor)
and dq.shape[0] > 0
):
q_end = int(cu_seqlens_q_padded[-1])
if q_end < dq.shape[0]:
dq[q_end:].fill_(0)
if cu_seqlens_kv_padded is not None:
kv_end = int(cu_seqlens_kv_padded[-1])
for d_tensor in (dk, dv):
if (
isinstance(d_tensor, torch.Tensor)
and d_tensor.shape[0] > 0
and kv_end < d_tensor.shape[0]
):
d_tensor[kv_end:].fill_(0)
else:
with get_nvtx_range_context("FusedAttnFunc.backward"):
# get nominal data type of dq, dk, dv
Expand Down
18 changes: 9 additions & 9 deletions transformer_engine/pytorch/attention/dot_product_attention/context_parallel.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3109,8 +3109,6 @@ def backward(ctx, dout, *_args):
and cu_seqlens_kv_padded is not None
):
if is_graph_capturing():
# arange+mask under capture: `tensor[scalar_tensor:]` slicing would
# force a GPU->CPU sync that is forbidden during CUDA graph capture.
q_pad_mask = torch.arange(dq.shape[0], device=dq.device) >= cu_seqlens_q_padded[-1]
kv_pad_mask = (
torch.arange(dk.shape[0], device=dk.device) >= cu_seqlens_kv_padded[-1]
Expand All @@ -3119,10 +3117,14 @@ def backward(ctx, dout, *_args):
dk[kv_pad_mask] = 0
dv[kv_pad_mask] = 0
else:
# Pre-existing TE eager-mode behaviour.
dq[cu_seqlens_q_padded[-1] :].fill_(0)
dk[cu_seqlens_kv_padded[-1] :].fill_(0)
dv[cu_seqlens_kv_padded[-1] :].fill_(0)
q_end = int(cu_seqlens_q_padded[-1])
kv_end = int(cu_seqlens_kv_padded[-1])
if q_end < dq.shape[0]:
dq[q_end:].fill_(0)
if kv_end < dk.shape[0]:
dk[kv_end:].fill_(0)
if kv_end < dv.shape[0]:
dv[kv_end:].fill_(0)

if ctx.fp8 and ctx.is_input_fp8:
dq, dk, dv, _, _ = combine_and_quantize(ctx.qkv_layout, dq, dk, dv, ctx.dQKV_quantizer)
Expand All @@ -3140,9 +3142,7 @@ def backward(ctx, dout, *_args):
ctx.dP_quantizer,
)

# Partial-gradient reduction can write THD inter-sequence padding.
# Clean it while gradients and per-step sequence metadata share sequence order.
if ctx.qkv_format == "thd":
if ctx.qkv_format == "thd" and ctx.pad_between_seqs:
_zero_thd_padding((dq,), cu_seqlens_q_per_step[0], cu_seqlens_q_padded)
_zero_thd_padding((dk, dv), cu_seqlens_kv_per_step[0], cu_seqlens_kv_padded)

Expand Down
72 changes: 59 additions & 13 deletions transformer_engine/pytorch/attention/dot_product_attention/dot_product_attention.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -113,6 +113,29 @@
_THD_POLICY_VALIDATION_CACHE_LIMIT = 64
_thd_policy_validation_cache = []

_PAD_BETWEEN_SEQS_EAGER_CACHE_LIMIT = 64
_pad_between_seqs_eager_cache = {}

@timmoon10 timmoon10 Sep 14, 2026

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

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

functools.lru_cache would be simpler and less error-prone than implementing the cache from scratch. That said, we already follow that anti-pattern with _thd_policy_validation_cache, so there's something to be said about maintaining consistency.



def _pad_between_seqs_cache_key(
cu_seqlens_q, cu_seqlens_kv, cu_seqlens_q_padded, cu_seqlens_kv_padded
):
def _key(t):
if t is None:
return (0, 0, 0, 0)
try:
version = t._version
except AttributeError:
version = 0
return (t.data_ptr(), tuple(t.shape), t.dtype, version)

@timmoon10 timmoon10 Sep 14, 2026

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

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

Is it safe to assume that two tensors with the same buffer will have the same padding requirements? If cu_seqlens_* are not persistent, the same buffers may be reused in different steps or different layers. Perhaps it would be safer to do something like:

return (id(t), version)


return (
_key(cu_seqlens_q),
_key(cu_seqlens_kv),
_key(cu_seqlens_q_padded),
_key(cu_seqlens_kv_padded),
)


def _get_thd_policy_attention_backend(
policy: Dict[str, Any],
Expand Down Expand Up @@ -2706,24 +2729,47 @@ def forward(
else None
)

# Default pad_between_seqs auto-detect. For THD, infer presence of
# inter-sequence padding from whether padded cu_seqlens were supplied --
# sync-free, and stable across eager and CUDA graph capture (the auto-detect
# must return the same value in both modes for backend selection to match).
# If padded cu_seqlens are the *same object* as the unpadded ones, no real
# inter-sequence padding exists (only THD tail padding) -- treat as False so
# FlashAttention v2/v4 remain eligible.
if pad_between_seqs is None:
if qkv_format == "thd":
if (
cu_seqlens_q_padded is cu_seqlens_q
and cu_seqlens_kv_padded is cu_seqlens_kv
):
pad_between_seqs = False
if is_graph_capturing():
cache_key = _pad_between_seqs_cache_key(
cu_seqlens_q,
cu_seqlens_kv,
cu_seqlens_q_padded,
cu_seqlens_kv_padded,
)
if cache_key in _pad_between_seqs_eager_cache:
pad_between_seqs = _pad_between_seqs_eager_cache[cache_key]

@timmoon10 timmoon10 Sep 14, 2026

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

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

Noting that the cache is not populated during graph capture. In order to take advantage of the cache, we assume that the cache is populated in the warmup steps.

elif (
cu_seqlens_q_padded is cu_seqlens_q
and cu_seqlens_kv_padded is cu_seqlens_kv
):
pad_between_seqs = False
else:
pad_between_seqs = (
cu_seqlens_q_padded is not None or cu_seqlens_kv_padded is not None
)
else:
Comment thread
greptile-apps[bot] marked this conversation as resolved.
pad_between_seqs = (
cu_seqlens_q_padded is not None or cu_seqlens_kv_padded is not None
cu_seqlens_q_padded is not None
and not torch.equal(cu_seqlens_q_padded[:-1], cu_seqlens_q[:-1])
) or (
cu_seqlens_kv_padded is not None
and not torch.equal(cu_seqlens_kv_padded[:-1], cu_seqlens_kv[:-1])
)
cache_key = _pad_between_seqs_cache_key(
cu_seqlens_q,
cu_seqlens_kv,
cu_seqlens_q_padded,
cu_seqlens_kv_padded,
)
_pad_between_seqs_eager_cache[cache_key] = pad_between_seqs
while (
len(_pad_between_seqs_eager_cache) > _PAD_BETWEEN_SEQS_EAGER_CACHE_LIMIT
):
_pad_between_seqs_eager_cache.pop(
next(iter(_pad_between_seqs_eager_cache))
)
else:
pad_between_seqs = False

Expand Down