runtime (7/7): integration — launcher + end-to-end equivalence gates#8
Closed
maocheng23 wants to merge 3 commits into
Closed
runtime (7/7): integration — launcher + end-to-end equivalence gates#8maocheng23 wants to merge 3 commits into
maocheng23 wants to merge 3 commits into
Conversation
Wire the offline-EAGLE3 dataflow (launch.py + thin train_eagle3_dataflow.py), README, and the cross-layer GPU gates: offline/online/trainer-split/checkpoint equivalence, extraction-vs-HF, and the FSDP launcher path.
maocheng23
force-pushed
the
stack/7-integration
branch
from
dfaa8b1 to
58ee2e5
Compare
maocheng23
force-pushed
the
stack/6-training
branch
from
1cc8520 to
41824dc
Compare
Whole-system map (autonomous compute loops, a passive metadata-only control plane, a tensor-only data plane), an endpoint reference table, and the autonomy model. Endpoints verified against source. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…stroy_distributed - train_eagle3_dataflow.py: parse_args() does not set args.target_batch_size (train_eagle3.main derives it inline), so the offline dataflow launcher crashed with AttributeError in build_dataloaders before training. Derive it right after parse_args, mirroring train_eagle3. - distributed.py: destroy_distributed() raised "Invalid process group specified" during teardown when a group is None or already destroyed, making a successful run exit non-zero. Destroy each group defensively so cleanup never crashes. Both surfaced by a real Qwen2.5-7B offline old-vs-new e2e training run. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Owner
Author
|
Superseded: merged upstream as sgl-project#600. Closing this fork-internal review PR. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Part 7/7 of the DataFlow runtime stack (M1–M4). Review bottom-up.
Wires the offline-EAGLE3 dataflow (launch.py + thin train_eagle3_dataflow.py) + README + the cross-layer GPU gates: offline/online/trainer-split/checkpoint equivalence, extraction-vs-HF, and the FSDP launcher path. Validated 62/62 on H200.
stack/6-training→ Head:stack/7-integrationunittest discover tests/test_runtime, GPU testsskipUnless(cuda)).Each layer compiles and passes its own CPU tests; the cross-layer GPU equivalence gates land in 7/7. Split out of the monolithic PR for reviewability.
🤖 Generated with Claude Code