Refactor rollout.py by file decompositions by fzyzcjy · Pull Request #899 · radixark/miles

fzyzcjy · 2026-04-06T02:14:02Z

Mechanical Move

Transform script: https://gist.github.com/fzyzcjy/f479621f729158542dc22db3d17f2929

One-click verification

python3 <(curl -sL https://gist.githubusercontent.com/fzyzcjy/f479621f729158542dc22db3d17f2929/raw/transform_rollout_split.py)

Requires mechanical move PRs to include a reproducible transform script as a gist. The script is fully self-contained: creates a worktree, replays the transformation with commits, diffs against the PR, and reports pass/fail. PR description includes a one-liner verification command.

…rm()

…tion

(cherry picked from commit b42aa4e)

Note: rollout_ft/0 was already merged into main as PR #897 and deleted from origin, so the cascade starts from main → rollout_ft/1. Conflict 1: miles/ray/rollout.py (modify/delete) - HEAD (rollout_ft/1) deleted the file as part of a mechanical split: commit d3fc26e "mechanically move" splits the original 1298-line miles/ray/rollout.py into the directory miles/ray/rollout/{addr_allocator, metrics,observability,rollout_manager,rollout_server,router_manager, server_group}.py. - origin/main modified the file (4 commits since merge-base): a772c33 zero_std all_zero_ratio/all_one_ratio metrics (#1034) 41615af weight staleness control for fully async rollout (#958) c198efa consistent hashing routing policy (#891) eaa36a2 heartbeat and id to session server (#866) Resolution: removed miles/ray/rollout.py (preserve mechanical split) and re-applied main's 7 hunks to the new directory files: - miles/ray/rollout/router_manager.py: + import uuid + router_args.policy = args.sglang_router_policy (in start_router) + args.session_server_instance_id = uuid.uuid4().hex (in start_session_server) - miles/ray/rollout/train_data_conversion.py: + train_data["weight_versions"] population (after multimodal block) + "weight_versions" added to per-DP-split key whitelist - miles/ray/rollout/metrics.py: + oldest_weight_version statistics + mixed_version_ratio in _compute_metrics_from_samples + zero_std/all_zero_percentage and all_one_percentage in _compute_zero_std_metrics Verified: all dependent symbols (Sample.weight_versions, Sample.oldest_weight_version, args.sglang_router_policy, args.session_server_instance_id) are already present in the merged tree from cleanly-merged peer files (miles/utils/types.py, miles/backends/sglang_utils/arguments.py, etc.). Verified: merge_diff_check.py shows the only "lost" deviations are PR #897 skill files (already in main) — no real loss.

fzyzcjy · 2026-05-05T14:01:44Z

Heads up — I merged origin/main into this branch (and cascaded the result down rollout_ft/2 … /32) so the chain stays current with main.

The conflict (only one in the entire 32-branch cascade)

miles/ray/rollout.py was a modify/delete:

this branch deleted it (the mechanical split into miles/ray/rollout/ introduced by d3fc26e62)
main still modified the original file via four PRs since the merge-base:
- Add heartbeat and id to session server #866 — heartbeat & id to session server (eaa36a246)
- Add consistent hashing routing policy for rollout #891 — consistent hashing routing policy (c198efa60)
- Add weight staleness control for fully async rollout #958 — weight staleness control for fully async rollout (41615af98)
- Add zero_std all_zero_ratio and all_one_ratio metrics #1034 — zero_std all_zero/all_one ratio metrics (a772c3302)

Resolution

Removed the orphaned miles/ray/rollout.py and re-applied each main-side hunk into the corresponding new module:

Hunk	Target
`import uuid`	`miles/ray/rollout/router_manager.py`
`router_args.policy = args.sglang_router_policy` (in `start_router`)	`miles/ray/rollout/router_manager.py`
`args.session_server_instance_id = uuid.uuid4().hex` (in `start_session_server`)	`miles/ray/rollout/router_manager.py`
`train_data["weight_versions"] = …`	`miles/ray/rollout/train_data_conversion.py`
`"weight_versions"` in per-DP key whitelist	`miles/ray/rollout/train_data_conversion.py`
`oldest_versions` / `mixed_version_ratio` block	`miles/ray/rollout/metrics.py`
`zero_std/all_zero_percentage` & `all_one_percentage` rewrite	`miles/ray/rollout/metrics.py`

Merge commit: 9a5e5b196. The full conflict + resolution detail is in the commit message body.

Verification

Reverse-diff check: every + line in git diff badc784f2 9a5e5b196 -- 'miles/ray/rollout/' corresponds 1:1 to a + line in main's pre-merge rollout.py patch — no missing or extra additions.
mechanical-refactor-verify one-click ran against the unchanged BASE_COMMIT=1fb7e93c / TARGET_COMMIT=badc784f — PASS (this PR's mechanical split itself is unaffected; the verification compares against the pre-merge target which is still reachable).
Auxiliary files from the four main PRs (e.g. Sample.weight_versions / oldest_weight_version in miles/utils/types.py, --sglang-router-policy in miles/backends/sglang_utils/arguments.py, session_server_instance_id plumbing in miles/rollout/session/sessions.py and openai_endpoint_utils.py) all auto-merged cleanly and are present in the merged tree.
/2 → /32 in the cascade were all auto-merged with zero conflicts.

miles/ray/rollout.py: modify/delete conflict. rollout_ft/1 split this file into the miles/ray/rollout/ package; main (#1164 Bump sglang to v0.5.12) added env var SGLANG_OPT_USE_CUSTOM_ALL_REDUCE_V2 to ServerGroup. Re-applied main's new env key to its moved location miles/ray/rollout/server_group.py and kept the file deletion.

fzyzcjy · 2026-05-31T01:06:31Z

Verified the mechanical refactor via the one-click transform script:

python3 <(curl -sL https://gist.githubusercontent.com/fzyzcjy/f479621f729158542dc22db3d17f2929/raw/transform_rollout_split.py)

Result: PASS: transform reproduces the commit exactly.

The script checks out BASE_COMMIT = 1fb7e93c, runs the transform (line-range split of miles/ray/rollout.py into the miles/ray/rollout/ package), runs pre-commit run --all-files, and diffs against TARGET_COMMIT = badc784f. The diff is empty, confirming this is a pure mechanical move with no semantic changes. ✅

fzyzcjy added 30 commits April 6, 2026 09:19

empty

8659b57

abs import

e23d43d

mechanically move

d3fc26e

import

981710b

fmt

3943b37

mv and rename

ade062d

extract

6302d20

more

fb84c33

fmt

6a06221

more

f0bdf5a

extract

5e03bbe

more

e14676e

fmt

8ee59cd

keep worktree after verification for inspection

6352440

extract scaffold into utils.py, transform script only defines transfo…

6313761

…rm()

rename utils.py to mechanical_refactor_verify_utils.py

a07fe93

fix py3.10 compat: type alias and Optional syntax

3ea792a

simplify docstring: point to SKILL.md for usage

55d8d03

rename _run to exec_command

3666739

replace MechanicalVerifier class with verify_mechanical_refactor func…

167d5f3

…tion

remove RunFn, transform calls exec_command directly

704d0fe

add git_add_and_commit helper

b0406a0

move ruff format step into verify_mechanical_refactor

f581104

rename root to dir_root, remove run param from transform

85ba9a6

add gist update command, one gist per PR

40d6eef

clarify PR scope: mechanical only, semantic changes go to separate PR

597749e

remove relationship to existing refactor workflows section

2010083

cp

f5bf690

Merge branch 'rollout_ft/0' into rollout_ft/1

118423b

fzyzcjy added 14 commits April 6, 2026 10:44

add dedent() util, fix shlex.quote in git_add_and_commit

f9c13ac

remove __pycache__ from tracking

ec850e5

fix py3.10 compat: Optional[str] and add missing import

f8a9af5

more

df42bc4

cp

8d545e1

Merge branch 'rollout_ft/0' into rollout_ft/1

a3fe2f6

use pre-commit instead of ruff format in verify scaffold

21e7532

remove __pycache__

a352b0c

fmt

3f32106

cp

ee35f3a

Merge branch 'rollout_ft/0' into rollout_ft/1

3d594ee

skip pre-commit commit when no files changed

b42aa4e

skip pre-commit commit when no files changed

1fb7e93

(cherry picked from commit b42aa4e)

Merge branch 'rollout_ft/0' into rollout_ft/1

badc784

Base automatically changed from rollout_ft/0 to main May 5, 2026 13:22

Merge main into rollout_ft/1

8128e97

fzyzcjy requested review from Zhichenzzz and guapisolo as code owners May 6, 2026 07:37

fzyzcjy force-pushed the rollout_ft/1 branch from 94aea49 to 8128e97 Compare May 6, 2026 07:47

Shi-Dong approved these changes May 6, 2026

View reviewed changes

fzyzcjy added 4 commits May 7, 2026 22:58

Merge origin/main into rollout_ft/1

1ea7392

Apply pre-commit fixes (ruff, black) to scripts/run_qwen3_4b_npu.py

17b38a8

Merge main into rollout_ft/1

05444e8

fzyzcjy merged commit 68a8f11 into main May 31, 2026
24 of 25 checks passed

fzyzcjy deleted the rollout_ft/1 branch May 31, 2026 01:16

Frankstein73 pushed a commit to sii-avalanche/miles that referenced this pull request Jun 1, 2026

Refactor rollout.py by file decompositions (radixark#899)

e1b6692

Navigation Menu

Uh oh!

Refactor rollout.py by file decompositions - #899

Refactor rollout.py by file decompositions#899
fzyzcjy merged 54 commits into
mainfrom
rollout_ft/1

fzyzcjy commented Apr 6, 2026 •

edited

Loading

Uh oh!

fzyzcjy commented May 5, 2026

Uh oh!

fzyzcjy commented May 31, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Conversation

fzyzcjy commented Apr 6, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Mechanical Move

One-click verification

Uh oh!

fzyzcjy commented May 5, 2026

The conflict (only one in the entire 32-branch cascade)

Resolution

Verification

Uh oh!

fzyzcjy commented May 31, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

fzyzcjy commented Apr 6, 2026 •

edited

Loading