feat(testing): add TP-invariant reduction references - #103
Conversation
|
No actionable comments were generated in the recent review. 🎉 ℹ️ Recent review info⚙️ Run configurationConfiguration used: defaults Review profile: CHILL Plan: Pro Run ID: 📒 Files selected for processing (4)
✅ Files skipped from review due to trivial changes (1)
🚧 Files skipped from review as they are similar to previous changes (2)
📝 WalkthroughWalkthroughAdds a TP-invariant reduction contract, reference implementations, shard-aware masked reductions, drift diagnostics, public re-exports, and an extensive test suite (unit, distributed Gloo smoke, large-vocab, and CUDA fp16 gated tests) to validate vocab-sharded selected-logprobs and masked statistics. ChangesTP-Invariant Reduction Contract and Implementation
Estimated code review effort🎯 4 (Complex) | ⏱️ ~60 minutes Suggested reviewers
Poem
🚥 Pre-merge checks | ✅ 4 | ❌ 1❌ Failed checks (1 warning)
✅ Passed checks (4 passed)
✏️ Tip: You can configure your own custom pre-merge checks in the settings. ✨ Finishing Touches📝 Generate docstrings
🧪 Generate unit tests (beta)
Warning Tools execution failed with the following error: Failed to run tools: Ping-pong health check failed Comment |
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 2
🤖 Prompt for all review comments with AI agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
Inline comments:
In `@rl_engine/testing/reference_ops.py`:
- Around line 124-126: owner_ranks_for_token_ids currently overwrites owners for
overlapping shard_ranges which can violate the “exactly one owner” contract used
by summarize_tp_logprob_drift; add validation to detect overlaps and fail fast:
before iterating shard_ranges verify the ranges are sorted and non-overlapping
(each start >= previous end) or, inside the loop, check where owns is true and
owners != -1 and raise an error identifying the overlapping shard and token_ids;
reference the function owner_ranks_for_token_ids and the variables shard_ranges,
token_ids, owns, and owners when adding the check so overlaps are rejected
rather than silently taking the last match.
- Around line 362-367: The maskless branch in sharded_active_token_count trusts
value_shards[0].device but doesn't validate the rest, causing cross-device
tensors; mirror the device validation from sharded_masked_sum by iterating all
entries in value_shards, check that each.values.device equals the chosen device,
and raise a ValueError (or similar) if any shard is on a different device before
computing count and returning the tensor; reference mask_shards, value_shards,
sharded_active_token_count, sharded_masked_sum, sharded_masked_mean, device, and
count when making the change.
🪄 Autofix (Beta)
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: defaults
Review profile: CHILL
Plan: Pro
Run ID: 3e77c135-7d94-4528-b2a0-3054de221438
📒 Files selected for processing (4)
docs/design/tp-invariant-reductions.mdrl_engine/testing/__init__.pyrl_engine/testing/reference_ops.pytests/test_tp_invariant_reductions.py
| for rank, (start, end) in enumerate(shard_ranges): | ||
| owns = active & (token_ids >= int(start)) & (token_ids < int(end)) | ||
| owners = torch.where(owns, torch.full_like(owners, rank), owners) |
There was a problem hiding this comment.
Reject overlapping shard ranges instead of silently taking the last match.
owner_ranks_for_token_ids() overwrites owners on every matching range, so an overlapping shard_ranges input silently reports the last rank as the owner. That breaks the “exactly one owner” contract the TP references rely on and can misattribute diagnostics downstream in summarize_tp_logprob_drift(). Add a one-time validation that shard_ranges are sorted and non-overlapping before the loop, or detect owners != -1 when owns is true and raise.
🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
In `@rl_engine/testing/reference_ops.py` around lines 124 - 126,
owner_ranks_for_token_ids currently overwrites owners for overlapping
shard_ranges which can violate the “exactly one owner” contract used by
summarize_tp_logprob_drift; add validation to detect overlaps and fail fast:
before iterating shard_ranges verify the ranges are sorted and non-overlapping
(each start >= previous end) or, inside the loop, check where owns is true and
owners != -1 and raise an error identifying the overlapping shard and token_ids;
reference the function owner_ranks_for_token_ids and the variables shard_ranges,
token_ids, owns, and owners when adding the check so overlaps are rejected
rather than silently taking the last match.
| if mask_shards is None: | ||
| if not value_shards: | ||
| raise ValueError("value_shards must be provided when mask_shards is None") | ||
| device = value_shards[0].device | ||
| count = sum(int(values.numel()) for values in value_shards) | ||
| return torch.tensor(count, device=device, dtype=reduction_dtype) |
There was a problem hiding this comment.
Validate value_shards devices on the maskless count path.
When mask_shards is None, sharded_active_token_count() trusts value_shards[0].device and never checks the remaining shards. A mixed-device input then returns a count tensor on an arbitrary device, and sharded_masked_mean() can later fail when dividing by a numerator produced on a different device. Mirror the same all-shards device validation used in sharded_masked_sum() before computing count.
🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
In `@rl_engine/testing/reference_ops.py` around lines 362 - 367, The maskless
branch in sharded_active_token_count trusts value_shards[0].device but doesn't
validate the rest, causing cross-device tensors; mirror the device validation
from sharded_masked_sum by iterating all entries in value_shards, check that
each.values.device equals the chosen device, and raise a ValueError (or similar)
if any shard is on a different device before computing count and returning the
tensor; reference mask_shards, value_shards, sharded_active_token_count,
sharded_masked_sum, sharded_masked_mean, device, and count when making the
change.
Signed-off-by: inaniloquentee <3051000145@qq.com>
8781ab6 to
8334615
Compare
|
This pull request has been inactive for 30 days. To keep things moving, please update it or let us know if it's still in progress. Otherwise it will be automatically closed in 30 days. |
1 participant
Summary
Closes #102
Validation
py -3.13 -m pre_commit run --all-filespy -3.13 -m mypy --ignore-missing-imports rl_engine/py -3.13 -m pytest rl_engine/tests/test_dispatch.py -vpy -3.13 -m pytest tests/test_tp_invariant_reductions.py tests/test_reference_ops.py -q -rspy -3.13 -m mkdocs build --strict -f mkdocs.yamlpy -3.13 -m pytest -qDCO
Signed-off-by: inaniloquentee <3051000145@qq.com>Summary by CodeRabbit
Documentation
New Features
Tests