Skip to content

perf: shard concat overhead #2002

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

Open
pjo256 wants to merge 6 commits into
base: main
Choose a base branch
from
Open

perf: shard concat overhead #2002

Changes from 4 commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
46d2ffe
fix: avoid repeated concat copies in shard_by_batch_size
pjo256 Feb 21, 2026
fbef9a3
fix: preallocate shard tensors to remove chunk concat copy overhead
pjo256 Feb 21, 2026
6b7d11f
fix: simplify with per-shard range aggregation
pjo256 Feb 21, 2026
950c1c9
docs: cleanup comments
pjo256 Feb 21, 2026
aca061d
tests: fix missing empty PackedTensor, add a unit test for it
pjo256 Feb 21, 2026
534d1c5
fix: lint
pjo256 Feb 21, 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
74 changes: 38 additions & 36 deletions nemo_rl/distributed/batched_data_dict.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -537,8 +537,9 @@ def _get_padded_seqlen(seqlen: int) -> int:

aggregated_shards = [SlicedDataDict() for _ in range(shards)]

# Group data by shard position across all chunks
# Group data by shard position across all chunks.
for shard_idx in range(shards):
shard_ranges: list[tuple[int, int]] = []
for chunk_idx in range(num_chunks):
# Calculate indices for this particular sub-shard within the chunk
chunk_start = chunk_idx * batch_size
Expand All @@ -549,41 +550,42 @@ def _get_padded_seqlen(seqlen: int) -> int:
# or if shard_end calculation goes beyond total_batch_size
shard_start = min(shard_start, total_batch_size)
shard_end = min(shard_end, total_batch_size)
indices = torch.arange(shard_start, shard_end)

for k in data:
if k not in aggregated_shards[shard_idx]:
# First time seeing this key for this shard, initialize it
if torch.is_tensor(data[k]):
aggregated_shards[shard_idx][k] = data[k][indices].clone()
elif isinstance(data[k], PackedTensor):
aggregated_shards[shard_idx][k] = data[k].slice(
indices.tolist()
)
else:
aggregated_shards[shard_idx][k] = [
data[k][i] for i in indices
]
else:
# Append to existing data - concatenate tensors or extend lists
if torch.is_tensor(data[k]):
aggregated_shards[shard_idx][k] = torch.cat(
[
aggregated_shards[shard_idx][k],
data[k][indices].clone(),
]
)
elif isinstance(data[k], PackedTensor):
aggregated_shards[shard_idx][k] = PackedTensor.concat(
[
aggregated_shards[shard_idx][k],
data[k].slice(indices.tolist()),
]
)
else:
aggregated_shards[shard_idx][k].extend(
[data[k][i] for i in indices]
)

if shard_start < shard_end:
shard_ranges.append((shard_start, shard_end))

for k, v in data.items():
if torch.is_tensor(v):
# Pre-allocate and copy each chunk exactly once.
rows = sum(end - start for start, end in shard_ranges)
shard_tensor = torch.empty(
(rows, *v.shape[1:]),
dtype=v.dtype,
device=v.device,
)
offset = 0
for start, end in shard_ranges:
span = end - start
shard_tensor[offset : offset + span].copy_(v[start:end])
offset += span
aggregated_shards[shard_idx][k] = shard_tensor
elif isinstance(v, PackedTensor):
# PackedTensor is collected per chunk then concatenated once.
packed_slices = [
v.slice(list(range(start, end))) for start, end in shard_ranges
]
if packed_slices:
aggregated_shards[shard_idx][k] = (
PackedTensor.concat(packed_slices)
if len(packed_slices) > 1
else packed_slices[0]
)
else:
shard_values = []
for start, end in shard_ranges:
shard_values.extend([v[i] for i in range(start, end)])

aggregated_shards[shard_idx][k] = shard_values

# map inputs to microbatches such that the total number tokens in
# a microbatch is as close to (including padding tokens) 'max_tokens_per_microbatch'
Expand Down
Loading