Add versioned trainer graphs and logical-rank APIs - #925
Draft
bradhilton wants to merge 44 commits into
Draft
bradhilton wants to merge 44 commits into
bradhilton wants to merge 44 commits into
Conversation
bradhilton
had a problem deploying
to
trainer-rank-gpu-validation
September 18, 2026 00:36 — with 
GitHub Actions
Failure
GitHub Actions
Failure
bradhilton
had a problem deploying
to
trainer-rank-gpu-validation
September 18, 2026 01:04 — with
GitHub Actions
Error
GitHub Actions
Error
bradhilton
deployed
to
trainer-rank-gpu-validation
September 18, 2026 01:24 — with
GitHub Actions
Active
GitHub Actions
Active
…view-fixes-merge-20260919
bradhilton
deployed
to
trainer-rank-gpu-validation
September 19, 2026 21:25 — with
GitHub Actions
Active
GitHub Actions
Active
bradhilton
deployed
to
trainer-rank-gpu-validation
September 20, 2026 02:55 — with
GitHub Actions
Active
GitHub Actions
Active
bradhilton
deployed
to
trainer-rank-gpu-validation
September 20, 2026 03:35 — with
GitHub Actions
Active
GitHub Actions
Active
bradhilton
deployed
to
trainer-rank-gpu-validation
September 20, 2026 04:28 — with
GitHub Actions
Active
GitHub Actions
Active
bradhilton
deployed
to
trainer-rank-gpu-validation
September 21, 2026 07:26 — with
GitHub Actions
Active
GitHub Actions
Active
bradhilton
deployed
to
trainer-rank-gpu-validation
September 21, 2026 08:33 — with
GitHub Actions
Active
GitHub Actions
Active
bradhilton
deployed
to
trainer-rank-gpu-validation
September 21, 2026 11:24 — with
GitHub Actions
Active
GitHub Actions
Active
bradhilton
deployed
to
trainer-rank-gpu-validation
September 21, 2026 14:53 — with
GitHub Actions
Active
GitHub Actions
Active
bradhilton
deployed
to
trainer-rank-gpu-validation
September 21, 2026 16:58 — with
GitHub Actions
Active
GitHub Actions
Active
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.
Adds logical DP and rank-zero training APIs, versioned backward graphs, and checkpoint-owned live parameters/modules. The planner can retain, offload or replay state; backward commits gradients transactionally. Default staleness is two updates (
0is strict), with eligible importance correction when current probabilities are available. Native retained backward supports BF16 and rejects unsupported quantized Transformer Engine state.Paired with Caladan #459. Fixes delivery-test typing, keeps Tinker client imports from patching global asyncio, and releases the previous logical batch before advancing the physical iterator. Fresh Astra and Fable source reviews approve the fixes. CPU CI and 2×H200 GPU CI pass at
4d9c63111a, including the new lifetime regression. The GPU cluster termination completed. A secondary cleanup-error gap in placement/client export remains a documented follow-up.Earlier revisions passed eight-GPU derivative/memory canaries, learning comparisons, and a configured protocol-8 factory smoke. Qwen3-8B measurements showed 12.1% lower throughput at four 1K inputs (ABBA) and 4.5% lower at four 4K inputs (one pair); these retain their original source scope and make no throughput claim for the latest fixes.