miles: submission-order execution + CP-safe parallelism default; seam gate harnesses - #18
Merged
Merged
Conversation
The task manager runs request handlers as concurrent asyncio tasks, so pipelined fb/optim_step broadcasts interleaved on the shared TrainGroup: optim_step could consume a later fb's gradients, and inconsistent per-actor mailbox order mispairs DP collectives. Per-handle FIFO lock restores the client's submission-order contract; delete_model holds it too (delete-during-optim_step crash class). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
At CP>1 the miles seam returns per-CP-rank logprob chunks (no full-sequence reassembly yet), so clients see half-length per-sample logprobs. 8K fits without CP for supported model sizes; raise the auto-parallelism CP threshold until the seam stitches CP chunks. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Durable copies of the G1/G2/G3 gate scripts (previously pod-local /tmp, lost on pod recreate). Miles profile default image bumped to the 20260730 multi-LoRA-era bake. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Fixes surfaced by re-validating the miles seam on the
miles_tinker_seam:20260730base (upstream multi-LoRA era). All three gates green after these changes: G1 grad-accum split-invariance bit-identical (grad_norm 178.9398358212894 for 1xfb(8) vs 2xfb(4)+step, and across reruns), G2 client-order exact (8/128 datums, incl. pipelined submission), G3 sl_basic 30-step SFT NLL 2.80->2.28 on Qwen2.5-0.5B r32.1.
dc470bd— serialize per-model backend ops. The task manager runs request handlers as concurrent asyncio tasks, so pipelined fb/optim_step broadcasts executed overlapped on one TrainGroup: optim_step consumed a later fb's gradients, and concurrent fan-outs reach the DP actors' mailboxes in inconsistent per-rank order, mispairing collectives (observed: per-sample logprobs scrambled across requests). Per-handle FIFOasyncio.Lockrestores the client's submission-order contract;delete_modelholds it too (delete-during-optim_step crash class).2.
3f72f75— engage CP only above 8K max_seq_len. The cookbook passesmax_seq_len=8192at create, which flipped auto-parallelism to CP=2; at CP>1 the seam returns per-CP-rank logprob chunks (each exactly pad-to-CP-multiple/2 long) because there is no full-sequence reassembly yet — clients see half-length per-sample logprobs. 8K fits without CP for supported model sizes; threshold raised until the seam stitches CP chunks (tracked as an open gap).3.
6b3e1d9— durable gate harnesses + image default. G1/G2/G3 harnesses land inscripts/gates/(previously pod-local /tmp, lost on every pod recreate); miles profile default image bumped tomiles_tinker_seam:20260730.Companion miles-side fix (upstream
get_rollout_datatuple return): GavinZhu-GMI/miles#4.