 Zhiyao Jiang
|
031fe3d727
|
ci: pin the DSv4 bshd/thd parity test to the miles impl (#3162)
|
2026-09-09 15:59:02 -07:00 |
|
|
Shi-DongandZhichenzzz
|
f01f9a3166
|
fix: use host weights when syncing an offloaded actor (#3128)
Signed-off-by: Shi Dong <shi.dong@radixark.ai>
Co-authored-by: Zhichenzzz <zczeng@uw.edu>
|
2026-09-09 15:29:59 -07:00 |
|
|
Yueming Yuan
|
fa7939f899
|
ci(lark): report every failing test in a job (#3168)
|
2026-09-09 15:17:32 -07:00 |
|
|
Yueming Yuan
|
b6dc4301b5
|
fix(ci): keep the Nemotron-H plugin importable without megatron.bridge (#3169)
|
2026-09-09 15:17:16 -07:00 |
|
|
Shi-DongandZhichenzzz
|
f1eac9bad4
|
Fix missing Nemotron-H Mamba convolution mappings (#3163)
Co-authored-by: Zhichenzzz <zczeng@uw.edu>
|
2026-09-09 14:43:08 -07:00 |
|
|
Shi-Dong
|
021d311e89
|
Fix Nemotron-H forward compatibility when MTP is disabled (#3164)
|
2026-09-09 14:42:53 -07:00 |
|
|
Shi-Dong
|
d8dd0dec8d
|
Scan full responses for repetition metrics (#3115)
|
2026-09-09 12:04:50 -07:00 |
|
|
Yueming YuanandAlan Livshin
|
327eefa0fb
|
ci(lark): explain each failed job on the CI status card (#3152)
Co-authored-by: Alan Livshin <alan.livshin@alivshin-mbp.tail134ba0.ts.net>
|
2026-09-09 11:31:52 -07:00 |
|
|
Yueming Yuan
|
f207d590b0
|
ci: repeat a failed test's last output in the suite summary (#3149)
|
2026-09-09 11:28:40 -07:00 |
|
|
Yueming Yuan
|
aa46eada5b
|
[CI] disable the two FT determinism tests #3148 missed (#3160)
|
2026-09-09 10:57:44 -07:00 |
|
|
Xinyu JiangandZhiyao Jiang
|
8d9826eacc
|
[AMD] docker: retire the rocm700 variants, add a rocm10-mi35x variant (#2854)
Co-authored-by: Zhiyao Jiang <jessicajiang324@gmail.com>
|
2026-09-08 22:37:39 -07:00 |
|
|
|
7bf05390d7
|
[feat] Enable THD packed-sequence training for DeepSeek-V4-Flash, with context parallelism (#2038)
Co-authored-by: Xinyu Jiang <xinyuj2@andrew.cmu.edu>
Co-authored-by: guapisolo <guapisolo@gmail.com>
Co-authored-by: Yueming Yuan <yym022502@gmail.com>
|
2026-09-08 22:04:54 -07:00 |
|
|
Tao LinandClaude Fable 5
|
4f8e0b86f0
|
examples: launcher for the in-process Harbor path (#2836)
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
|
2026-09-08 21:52:41 -04:00 |
|
|
Yueming Yuan
|
bf482051a8
|
Bound the trainer's GPU memory during the LoRA weight sync (#3147)
|
2026-09-08 18:11:40 -07:00 |
|
|
Tao LinandClaude Fable 5
|
4501b2655d
|
smoke: manual golden episodes on real sandbox APIs (#2868)
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
|
2026-09-08 20:00:13 -04:00 |
|
|
Yueming Yuan
|
7d31a5ff1c
|
[CI] temporarily disable failing FT tests (#3148)
|
2026-09-08 16:30:30 -07:00 |
|
|
Yueming Yuan
|
aaab50cde1
|
Support --colocate-memory-peak-device gpu under LoRA (#3100)
|
2026-09-08 16:11:06 -07:00 |
|
|
Tao LinandClaude Fable 5
|
4dff6f2625
|
examples: in-process Harbor agent function (#2806)
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
|
2026-09-08 18:14:35 -04:00 |
|
|
Yueming Yuan
|
3de96596f1
|
Dump debug train data once per DP/CP shard instead of on every rank (#3141)
|
2026-09-07 23:10:32 -07:00 |
|
|
Zhichen Zeng
|
db3d6b56c8
|
Bump sglang to v0.5.19 (#3124)
|
2026-09-07 22:04:14 -07:00 |
|
|
|
91e40b5423
|
Decide the pipeline stage from the parallel state in compute_advantages_and_returns (#3125)
Co-authored-by: Qijia Shen <qijia.shen@eigent.ai>
Co-authored-by: yueming-yuan <yym022502@gmail.com>
|
2026-09-07 21:22:21 -07:00 |
|
|
Jiajun Li
|
78747f6912
|
refactor(docs): nest CI pages under developer (#3134)
|
2026-09-06 23:25:38 -07:00 |
|
|
Jiajun LiandDavid
|
4a58e4e101
|
fix(rollout): drop groups with missing rewards (#2814)
Co-authored-by: David <12414531+DavidBellamy@users.noreply.github.com>
|
2026-09-06 22:56:07 -07:00 |
|
|
Jiajun Li
|
5185f133e2
|
refactor(rollout): centralize common group filters (#3094)
|
2026-09-06 22:56:06 -07:00 |
|
|
Jiajun Li
|
51853a165a
|
refactor(rollout): rename shared filter output (#3095)
|
2026-09-06 22:56:06 -07:00 |
|
|
Shi-Dong
|
d2fc97ce58
|
Tolerate missing optional SGLang Anthropic helpers (#3114)
|
2026-09-04 14:48:36 -07:00 |
|
|
Zhichen ZengandVedant Gupta
|
7741150a29
|
feat(megatron): reset optimizer state by optimizer class (#3091)
Co-authored-by: Vedant Gupta <vedant.gupta@harmonic.fun>
|
2026-09-04 13:32:36 -07:00 |
|
|
fzyzcjy
|
dc2726907b
|
Fail CI runs on crashed async eval points or abnormal eval results (#2823)
|
2026-09-04 21:33:40 +08:00 |
|
|
fzyzcjy
|
a6d743f8bf
|
Await dynamic port probes instead of blocking the worker manager loop (#2898)
|
2026-09-04 15:03:33 +08:00 |
|
|
fzyzcjy
|
3408099cfc
|
Make the FT api server bind host a flag defaulting to loopback (#2897)
|
2026-09-04 15:03:07 +08:00 |
|
|
fzyzcjy
|
704009f41f
|
Reacquire the context lock safely when a released section is cancelled (#2896)
|
2026-09-04 15:02:39 +08:00 |
|
|
fzyzcjy
|
2495e4f817
|
Decide implicit healing after the config file overrides the args (#2895)
|
2026-09-04 15:02:12 +08:00 |
|
|
fzyzcjy
|
bed46c206f
|
Rename the trainer group and cell to TrainerController and TrainerCell (#2176)
|
2026-09-04 15:01:44 +08:00 |
|
|
fzyzcjy
|
db698299c5
|
Serialize cell suspension and fault injection against weight updates (#2894)
|
2026-09-04 15:01:16 +08:00 |
|
|
fzyzcjy
|
4db21c40ac
|
Make the train actor threaded only under fault tolerance (#2893)
|
2026-09-04 15:00:48 +08:00 |
|
|
fzyzcjy
|
d60cee6dd7
|
Temporarily remove vanilla fault tolerance for the eval fleet (#2497)
|
2026-09-04 15:00:21 +08:00 |
|
|
fzyzcjy
|
fb910cf3f4
|
Let workers deregister from the miles router (#2475)
|
2026-09-04 14:59:52 +08:00 |
|
|
fzyzcjy
|
1211351f1c
|
Report a rollout cell stuck initializing as unhealthy (#2474)
|
2026-09-04 14:59:23 +08:00 |
|
|
fzyzcjy
|
36b5cc6c90
|
Bound and wrap the free-port scan (#2473)
|
2026-09-04 14:58:56 +08:00 |
|
|
fzyzcjy
|
307c9c75f4
|
Run the weight equality check before leaving StateInitializing (#2472)
|
2026-09-04 14:58:29 +08:00 |
|
|
fzyzcjy
|
4f6be0661a
|
Fail the run when the FT api server cannot start (#2471)
|
2026-09-04 14:58:01 +08:00 |
|
|
fzyzcjy
|
3f74a8d8eb
|
Assert that the rollout executor keeps receiving weight versions (#2470)
|
2026-09-04 14:57:33 +08:00 |
|
|
fzyzcjy
|
84b41252cb
|
Publish the new weight version to the rollout executor (#2502)
|
2026-09-04 14:57:07 +08:00 |
|
|
fzyzcjy
|
38943d491c
|
Stop the fast suite from starting its own ray clusters and leaking accelerators (#2311)
|
2026-09-04 14:56:41 +08:00 |
|
|
fzyzcjy
|
7108fa998a
|
Route all trainer worker calls through the worker handle instead of Ray handles (#2175)
|
2026-09-04 14:56:11 +08:00 |
|
|
fzyzcjy
|
c4ee942164
|
Add a ray worker handle with death confirmation (#2174)
|
2026-09-04 14:55:42 +08:00 |
|
|
fzyzcjy
|
d034c9256e
|
Kill failed trainer workers through their own RPC channel instead of the manager (#2173)
|
2026-09-04 14:55:17 +08:00 |
|
|
fzyzcjy
|
e9cbd07400
|
Merge the now duplicated actor and rollout cell handlers (#2172)
|
2026-09-04 14:54:49 +08:00 |
|
|
fzyzcjy
|
dfdb532b9a
|
Remove the pending state in trainer cells (#2171)
|
2026-09-04 14:54:21 +08:00 |
|
|
fzyzcjy
|
4c324937d1
|
Remove the now-unused stopped trainer cell state (#2170)
|
2026-09-04 14:53:55 +08:00 |
|