Skip to content

docs(upstream-sync): record today's 4 sync PRs #9

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
panwudi merged 1 commit into from
May 27, 2026
Merged

docs(upstream-sync): record today's 4 sync PRs #9

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
56 changes: 56 additions & 0 deletions docs/upstream-sync.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -101,6 +101,23 @@ commit)。
settings env var 问题,1 个 `test_boundary_snapshot_store` 是 full-suite
ordering flake(isolated 通过, 与本批改动无关).

- **2026-05-27 memory_guard_tier 收尾** — 三段 PR 把 C 组 `c645c9f` 重写
落地 + 5 个 follow-up 一起带回, 加 10 个独立 upstream 修复:
- **PR #5(backend)**:settings/process_memory_enforcer/engine_pool/server/cli
全切到 `memory_guard_tier`. 老字段 deprecated alias 一个 release.
`ModelTooLargeError.max_memory` -> `.ceiling`. 4511 pass.
- **PR #6(admin UI)**:两个 slider -> tier dropdown, 修了 PR-5 之后
admin POST 隐性 500. 4510 pass.
- **PR #7(5 个 follow-up)**:acd0533/4cfbc8b/3ef7b94/b129a19/64bd2a2.
引入 Metal wired-limit clamp + watermark-tier shrink +
tier-aware active-memory reclaim + `custom` tier. 4536 pass.
- **PR #8(10 个独立修复)**:boundary-store race 修(消掉
`test_cleanup_all_drains_queue` flake), per-engine MLX 线程/流,
VLM lazy state, profiles 重构. 4567 pass.
- **最终 baseline**:4567 pass / 3 known env-override fails / 36 skip
(boundary_snapshot flake 由 #1423 修掉, 不再算).
- 全部 sync/* PR 已 self-merge 进 main.

## 已引入(cherry-picked)

| 上游 commit | flyto commit | 内容 | 引入日期 |
Expand Down Expand Up @@ -153,6 +170,45 @@ cherry-pick 一律带 `-x`,commit message 里保留 "cherry picked from commit
| `ef1e842`(=7d640c1 #1417) | `ef1e842` | vlm: per-image lookup + whole-request fallback | 干净 |
| `5e394cf`(=1010fd3) | `5e394cf` | admin: 运行时 propagate `model_dirs` 到 OQManager + HFUploader | 干净 |

### 2026-05-27 memory_guard_tier 三段(PR-1 backend / PR-3 admin UI / 5 个 follow-up)

C 组 `c645c9f` 重写终于动手, 用 3 个 PR 分阶段落地(spike doc § 3.4):

| 阶段 | flyto PR | flyto commits | 内容 |
|---|---|---|---|
| spike doc | #4 | `0d2ec29` | 设计稿: `docs/memory-guard-tier-migration-spike.md` |
| PR-1 backend | #5 | `53ed139` `b9fa4a0` `07e46a6` | settings.py / process_memory_enforcer.py / engine_pool.py / server.py / cli.py: 把 `max_*_memory` 换成 `memory_guard_tier`. 老字段 / 老环境变量 / 老 CLI flag 保留 deprecated alias 一个 release. `ModelTooLargeError.max_memory` -> `.ceiling` |
| PR-3 admin UI | #6 | `f1c3d43` `80d7066` | 两个 slider 换成 tier dropdown; admin POST handler 修复(PR-1 backend 落地后, `routes.py` 写不存在的 `global_settings.model.max_model_memory` 会 500). i18n en/zh/zh-TW 补翻译; 其余 5 个语言先用英文占位 |

### 2026-05-27 C 组 5 个 follow-up(分支 `sync/memory-guard-tier-followups`,PR #7)

| 上游 commit | flyto commit | 内容 | 冲突处理 |
|---|---|---|---|
| `acd0533` | `bd9a159` | scheduler: adaptive prefill throttle + (legacy) user-explicit hard cap | settings.py / server.py / process_memory_enforcer.py 取 HEAD —— `user_explicit_max` / `max_process_memory_is_explicit` 已被 c645c9f 删, scheduler.py + 新 helper(`prefill_transient_tracker.py`)是本 commit 的实用部分; test 里 2 个 `user_explicit_max` 测试删 |
| `4cfbc8b` | `5b3fe20` | scheduler: 切到 watermark-tier shrink | 干净 |
| `3ef7b94` | `109ac76` | memory: clamp 到 effective Metal cap + sysctl 警告 | enforcer.py 主体 auto-merge; admin UI(routes/i18n/css/js)取 HEAD —— flyto 的 admin UI 是 PR-3 自己的形状, 上游 UI 改不直接适用; 测试取上游(`TestMetalWiredLimit`) |
| `b129a19`(#1425) | `7033c3b` | test: catch up renames from c645c9f + 沉默 enforcer 警告 | 干净 |
| `64bd2a2`(#1431) | `bdda9d2` | memory: tier-aware active-memory reclaim + Custom ceiling | settings.py `memory_guard_custom_ceiling_gb` 字段加; `MemoryGuardTier` Literal 加 `"custom"`; validate() 检查 custom > 0 ceiling; admin UI 取 HEAD(Custom 选项暴露留作后续) |

### 2026-05-28 独立修复批(分支 `sync/upstream-2026-05-28`,PR #8)

10 个跟 memory tier 无关的上游修复, 主要是 boundary-store race + 每引擎 stream + VLM 修复.

| 上游 commit | flyto commit | 内容 | 冲突处理 |
|---|---|---|---|
| `4f3a9b9`(#1423) | `7b2e849` | boundary-store: serialize cleanup_all + cleanup_request with writer thread | 干净 —— 消掉一直拖着的 `test_cleanup_all_drains_queue` flake |
| `bc1c427` | `0a65ddc` | boundary-store: drop unreachable shutdown(cleanup=) path | 干净 |
| `2916ab4`(#1422) | `89f3b99` | cache: 删 dead TieredCacheManager | 干净 |
| `56860b3`(#1304) | `fc26ab3` | engine: 每引擎线程 + mx.Stream, 消除 cross-engine 流污染 | scheduler.py / batch_generator.py 多处冲突 —— 取上游(`self._stream` 替模块级 `generation_stream`, 三段 phase timer 重构) |
| `a62f953` | `b7cb489` | engine: 删 redundant `_ensure_wired_limit` guard | 干净(依赖 56860b3) |
| `e6d8a3f`(#1445) | `c50d64e` | test(mtp): drop monkeypatch of removed `_get_generation_stream` | 干净 |
| `2e698ff`(#1437) | `f554f19` | scheduler: wait on generation_stream in store-cache worker | scheduler.py 主体 conflict —— 56860b3 已用更通用的 `_safe_sync_stream(self._stream)` 替, 取 HEAD; paged_ssd_cache 的非冲突部分留 |
| `9d5bed8` | `ebf2c21` | engine: VLM model lazy state 在 loader 线程实例化 | 干净 |
| `ff7522b` | `414b843` | load: checkpoint 无 mtp.* 权重时跳过 VLM MTPModule attach | 干净 |
| `0c881f5`(#1399) | `59d9e7e` | profiles: three-scope template contract + drop is_builtin emission | 干净 |

收尾 baseline: 4567 pass / 3 fail / 36 skip(3 fail 都是已知 OMLX_API_KEY env override flake; boundary_snapshot flake 由 #1423 修掉了).

## 确认已在 flyto(评估时已存在,勿重复引入)

- `11e6ea7` (#1224) chunked prefill 基座 —— flyto 早已有(换形状引入,
Expand Down