Repository navigation
Per-island task lists and per-island placement - #255
Merged
Merged
Conversation
michaellchung
force-pushed
the
s23-pr-13-remove-unreachable
branch
from
October 14, 2026 11:32
702c83f to
29d87ca
Compare
michaellchung
changed the base branch from
s23-pr-13-remove-unreachable
to
main
October 14, 2026 11:32
`yeto launch --rl-island-tasks 'l0=web1-10,pwn1-10;l1=web11-20,pwn11-20'` (or a JSON file) gives every training island its own subset of the tasks in --data. The launcher filters the --data JSONL by each row's metadata.task_id on the launching machine, writes one prompt file per island under the run directory, mounts it in place of --data and sends the island its SHA256 (--data-sha256). The island checks the file before training and records task ids and SHA256 in rl_driver_start; the run manifest and dry-run plan carry the same record. Before any cloud call the launcher refuses: --data that is not one local JSONL, external learners, island ids out of range, islands without a list, unknown or repeated task ids, rows without a task id, and training-island eval (P3-Q3: eval runs on the eval island). Unequal task counts only warn (P3-Q2). A JSON spec is rewritten inline so the head's second hop rebuilds the same bytes from its staged --data. The full-parameter training contract moves to v6: evaluation settings and the Miles eval flags are island-local, so islands with different (or no) eval settings are not refused by the syncer. All full-parameter contract values change (old test runs cannot resume, as in P0-D5). Parsing and filtering live in the backend-neutral yeto/rl/island_tasks.py; --rl-island-tasks is a neutral CLI flag, --data-sha256 sits next to --data in the Miles island flags. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Review fixes for rl-per-island-tasks: - _check_island_tasks computes the plan once and caches it on args; island tasks, the run manifest and the dry run read the cache. The cache is not replayed to the head, which plans again from its staged --data; a test checks both hops give the same per-island SHA256. - Islands are built only for --gpu entries (elastic adds none); a learner id outside the plan is refused with the planned ids listed. - The SFT learner and diffusion sample paths refuse --rl-island-tasks. - island_tasks.py takes the flag name from its caller for error messages. - The help says a JSON spec file path must not contain '='. - Drop a duplicated stale comment in island_entry.py. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Two islands with different placements (P4 mixed island placement) must hash the same training contract. The trainer shape, GPU split, rollout engine size, weight transfer path and the colocate/offload switches are island-local now, like the data and port flags. Cherry-picked onto the per-island tasks work (P3). Conflicts resolved: - island_entry.py: the placement flags and switches join P3's island-local eval flags and switches in one set each; the contract keeps using P3's _dense_shared_miles_argv helper; the docstring names both changes. - test_rl_launcher.py: kept both P3's eval check and P4's placement check. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
--rl-island-placement ISLAND:KEY=VALUE (keys placement, standby-gpus, use-gpus-per-node) gives one island its own placement values; the other islands keep the run-wide flags. yeto.rl.engine.island_placement parses and checks the overrides and builds each island's args copy; the launcher derives every island-side value (shape checks, minimum nodes, learner command, capability check, ports rejections) from that copy. Refused before any cloud work: mixed placement under the legacy (strict) schedule, --rl-elastic or --rl-overlap-eval in a mixed run (the error names the colocated islands), and per-island standby GPUs on a colocated island. A colocated island of a mixed run drops the inherited rollout and standby GPU counts. The run manifest and dry-run rows record every island's placement and AlgorithmSpec hash. The island driver emits rl_round_wall (placement, round and rollout seconds) every round. A static test keeps placement reads in the launcher on per-island args. Cherry-picked onto the per-island tasks work (P3) and the 1.7b.4b algorithm move. Conflict resolved in launcher.py _prepare_rl_args: kept the per-island require_ports_supported loop and dropped the _prepare_ports_algorithm call here, since 1.7b.4b already calls it later. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The fake-engine tape golden gains the rl_round_wall events (volatile seconds scrubbed); the elastic wire comparison ignores round_seconds; the provenance source check follows the per-island ports rejection call. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ement Review follow-up for P4: - The launcher sends --rl-per-island-placement to every island of a --rl-island-placement launch; only then do the placement flags leave the full-parameter training contract (recorded as per_island_placement). A uniform launch keeps them, so islands with different GPU splits are still refused. --rollout-num-gpus-per-engine always stays in: it sets the engine TP and has no per-island value. - The launcher static check also covers rollout_num_gpus, and a new check keeps island task factory calls inside _island_task. - island_placement parses and resolves once per call; the dry run builds the placement rows once; the dedup key names every placement field. Cherry-picked onto the per-island tasks work (P3). Conflicts resolved (P4-Q6: one v6 contract carries both changes): - island_entry.py: P3's _dense_shared_miles_argv takes a per_island_placement keyword and adds the placement flags and switches only then; the eval switches stay always island-local. The schema stays yeto-dense-full-training-contract-v6, now with per_island_placement. - launcher.py dry_run_plan: builds both the task plan and the placement rows once. - test_rl_launcher.py: kept P3's eval check and P4's conditional placement checks. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A launch with both --rl-island-tasks and --rl-island-placement gives each island its own task file and its own placement in the dry run, while both islands carry the same algorithm hash and the per-island placement marker. Each feature's refusals (eval on training islands, mixed placement under legacy scheduling) still apply when both are given. A unit check pins the v6 dense contract: per-island data and placement leave it, the engine size stays in. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
michaellchung
force-pushed
the
s23-pr-14-per-island
branch
from
October 14, 2026 11:32
9422e10 to
e277bd3
Compare
michaellchung
marked this pull request as ready for review
October 14, 2026 11:32
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
2 participants
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.
Summary
Implements rl-per-island-tasks (
--rl-island-tasks) and rl-mixed-island-placement (--rl-island-placement ISLAND:KEY=VALUE), with one full-parameter training contract v6, and syncs the English OpenSpec changes.Motivation
P3-Q1..Q6 and P4-Q1..Q6.
What changed
yeto/rl/island_tasks.py: launcher filters--dataJSONL by task id on the launching machine, writes one prompt file per island, mounts it and sends--data-sha256; island checks it and records task ids inrl_driver_start; plan computed once and recomputed identically on the head.yeto/rl/engine/island_placement.py: per-island placement / standby GPUs / GPUs per node under inter-island elastic scheduling; launcher derives every island-side value from the per-island args copy; driver emitsrl_round_wall.Behavior changes / compatibility
yeto-dense-full-training-contract-v6, one bump for two changes: eval settings are island-local; with--rl-island-placementthe placement flags leave the contract (recorded asper_island_placement).--rollout-num-gpus-per-enginealways stays in. Old test runs cannot resume (P0-D5).--data, external learners (P3-Q6), training-island eval (P3-Q3), mixed placement under strict scheduling,--rl-elastic/--rl-overlap-evalin a mixed run. Unequal task counts only warn (P3-Q2).rl_round_wallevents.Tests
Local CPU tests: every test file this PR adds or changes: 288 passed in 10.55s. The tree of this branch equals s23-final; the full suite with cargo on s23-final: 11 failed (expected set below).
Known expected failures until the ports image is rebuilt for the new Miles pin (P0-1.11 GPU run):
test_rl_ports_image(2),test_rl_backend_identity::test_golden_samples_keep_both_old_hashes_and_carry_the_identity, and 8test_decoupling_goldensamples (golden is regenerated after the rebuild).Depends on
#254. Tree of this branch equals
s23-final(a9dce1e7). #256 (image pins) follows and removes the expected failures.Not verified on GPU
None of this PR has run on GPU yet. The regression GPU run for the S23 stack is P0-1.11 GPU run (in progress at the time of writing); results will be linked here before merge.
🤖 Generated with Claude Code