Adam/h3 rvm posttraining - #1806

Open

Abecid wants to merge 141 commits into

hao-ai-lab:mainfrom

Abecid:adam/h3-rvm-posttraining

Commits on Aug 9, 2026

[feat] VSA for MiniMax H3: packed mixed-modality sparse attention (FA4 CuTe sm10x), exempt/compete mask policies, gate-compress branch, per-request sparsity/schedule knobs, attention-mass probe
SolitaryThinker
committed
281d07d
[misc] VSA-H3: /simplify pass — dedupe with the Wan VSA backend, drop dead state
SolitaryThinker
committed
7dece04
[train] MiniMax H3 DMD2 distillation via packed dual-modality adapter
SolitaryThinker
committed
dc70667
Merge remote-tracking branch 'origin/vsa-h3' into feat/h3-dmd2-vsa
SolitaryThinker
committed
7f81df6
[train] VSA-H3 DMD2 overfit: per-role backends, VSA metadata wiring, wandb guard
SolitaryThinker
committed
d81f1f1
docs: on-tray handoff for the VSA-H3 DMD2 overfit experiment
SolitaryThinker
committed
eb2cd4e

Commits on Aug 11, 2026

[bugfix] Triton block-sparse backward: scale QK^T after the dot, keep K raw

Will Lin
and
claude
committed
95f4f54
[feat] per-rank sharded base-weight cache for fast training relaunches

Will Lin
and
claude
committed
10dee7b
[train] H3 VSA/DMD2 training fixes: bf16 boundaries, validation offload, buffer-free AdamW

Will Lin
and
claude
committed
69fb5af
[train] MiniMax-H3 config suite: SFT sparsity sweep, data-free DMD2, 32-GPU Slurm launch, README

Will Lin
and
claude
committed
a20b69a
docs: Slurm handoff for the 32x GB200 data-free DMD2 run

Will Lin
and
claude
committed
b2a72b3

Commits on Aug 12, 2026

[train] DMD2: per-modality loss normalization for packed multi-modality models

Will Lin
and
claude
committed
19fb3ba
[train] DMD2: skip the unconditional teacher forward at guidance scale 1.0

Will Lin
and
claude
committed
8e06166
[train] DMD2: reuse the generator rollout for the critic loss on student steps

Will Lin
and
claude
committed
6c5bc85
[train] DMD2: intermediate-latent visualization callback (legacy training/ parity)

Will Lin
and
claude
committed
038384e

Commits on Aug 14, 2026

[train] MiniMax-H3: text-only schema for data-free DMD2 preprocessing

Will Lin
and
claude
committed
987b539
[train] run ops: checkpoint start-step gating, LR reset on resume, W&B fallback, weights-only DCP export

Will Lin
and
claude
committed
b65b37f
[train] enforce fp32 master weights in every training entry point

Will Lin
and
claude
committed
33e104e
[train] DMD2: sigma-space score sampling + shift-aware H3 schedule (v5 config)

Will Lin
and
claude
committed
1aeb90a
[train] H3 v5 config: requeue-safe resume_from_checkpoint=latest

Will Lin
and
claude
committed
b45233b
[train] H3 v5 config: keep-4 checkpoint retention (fp32 saves are ~528 GiB)

Will Lin
and
claude
committed
dd9276d
[train] dcp_to_diffusers: --link-base space-lean export + stale shard-index cleanup

Will Lin
and
claude
committed
deaf60e
[misc] lessons: preflight ruleset distilled from H3 DMD2 v1-v5 launches

Will Lin
and
claude
committed
80adb0f

Commits on Aug 15, 2026

[bugfix] resume latest: skip incomplete checkpoints (require dcp/.metadata)

Will Lin
and
claude
committed
0246a62
[misc] correct fp32 checkpoint size to measured 741 GiB (was ~528 estimate)

Will Lin
and
claude
committed
08e18d3

Commits on Aug 30, 2026

docs(rvm): add FastH3 RVM task specification
Abecid
committed
f2640e8
feat(rvm): add signed velocity-matching utilities
Abecid
committed
5f8b7df
test(rvm): cover reward normalization and signed surrogate
Abecid
committed
e01d42a
feat(rvm): add frozen-reference LoRA context
Abecid
committed
386895e
feat(rvm): add LoRA-enabled FastH3 model adapter
Abecid
committed
b396dd7
feat(rvm): export MiniMax H3 RVM model
Abecid
committed
86d103a
feat(rvm): add exact four-step FastH3 sampler
Abecid
committed
104d48a
feat(rvm): export FastH3 RVM primitives
Abecid
committed
f8a292f
feat(rvm): add video tensor conversion for reward models
Abecid
committed
96bed9b
fix(rvm): keep LoRA optimizer masters in fp32
Abecid
committed
78f1145
fix(rvm): validate the public FastH3 schedule explicitly
Abecid
committed
184f850
feat(rvm): add batched VideoAlign rewards
Abecid
committed
f04375f
feat(rvm): add HPSv3 video reward adapters
Abecid
committed
c397344
feat(rvm): add paper-faithful RAFT dynamic-tracking reward
Abecid
committed
1316d94
feat(rvm): register configurable RVM reward mixture
Abecid
committed
52bf652
fix(rvm): preserve uint8 validation and reward videos
Abecid
committed
ba7c8d6
perf(rvm): keep decoded reward videos in uint8
Abecid
committed
aa75413
fix(rvm): normalize uint8 inputs before RAFT
Abecid
committed
1b6a8df
fix(rvm): normalize uint8 smoke rewards
Abecid
committed
234ed0a
fix(rvm): log uint8 validation videos without rescaling
Abecid
committed
b41cc9d
feat(rvm): add managed FastH3 reward velocity matching loop
Abecid
committed
2aa7f09
feat(rvm): expose RVM training method
Abecid
committed
ec86e76
config(rvm): add one-GPU FastH3 correctness smoke
Abecid
committed
02f8f97
config(rvm): add eight-GPU exact RVM baseline
Abecid
committed
b9c8226
config(rvm): add audio-preserving eight-GPU run
Abecid
committed
6382b4e
config(rvm): add conservative full-anchor ablation
Abecid
committed
6e7bd9b
config(rvm): add full 180-update FastH3 run
Abecid
committed
87b81d9
feat(rvm): export sharded checkpoints as inference LoRA
Abecid
committed
5fbb324
feat(rvm): prepare deterministic VidProM H3 prompt splits
Abecid
committed
851f959
test(rvm): add production reward-stack preflight
Abecid
committed
2388d04
feat(rvm): add FastH3 LoRA inference entrypoint
Abecid
committed
8070a9e
feat(rvm): index local checkpoints and validation artifacts
Abecid
committed
426b74c
chore(rvm): add shared workflow environment
Abecid
committed
bf3aca2
chore(rvm): add reproducible conda environment setup
Abecid
committed
ccc3274
chore(rvm): download FastH3 and reward checkpoints
Abecid
committed
6f8bd57
chore(rvm): add deterministic prompt encoding workflow
Abecid
committed
85ec02b
test(rvm): add strict FastH3 inference smoke
Abecid
committed
7ec319b
test(rvm): add one-GPU preflight gate
Abecid
committed
5a91bae
feat(rvm): add one-GPU training smoke launcher
Abecid
committed
c069e4c
feat(rvm): add eight-GPU learning-rate sweep
Abecid
committed
cc224cc
feat(rvm): add RVM anchor sweep
Abecid
committed
1ec7330
feat(rvm): add full FastH3 RVM launcher
Abecid
committed
3e507ee
feat(rvm): add exact-resume launcher
Abecid
committed
b32063b
feat(rvm): add checkpoint-to-LoRA export script
Abecid
committed
f430126
feat(rvm): add post-training LoRA inference script
Abecid
committed
333bff8
docs(rvm): add end-to-end FastH3 experiment guide
Abecid
committed
fce2931
docs(rvm): add GPU-agent methodology and recovery runbook
Abecid
committed
8629f4c
test(rvm): enforce FastH3 experiment invariants
Abecid
committed
a2fb390
fix(rvm): preserve reward import exception context
Abecid
committed
183654b

Commits on Aug 31, 2026

feat(rvm): persist pilot validation reward metrics
Abecid
committed
9ff5c53
feat(rvm): expose local-metrics pilot method
Abecid
committed
d42390f
feat(rvm): add one-GPU Modal correctness config
Abecid
committed
6ec99f3
feat(rvm): add four-GPU Modal pilot config
Abecid
committed
ad3e23e
feat(rvm): allow container-managed Python environments
Abecid
committed
7a0102c
feat(rvm): persist reward preflight scores as JSON
Abecid
committed
ad30c49
feat(rvm): add one- and four-GPU Modal pilot launcher
Abecid
committed
74c0d0d

Commits on Sep 1, 2026

[bugfix]: stabilize H3 RVM Modal training
Abecid
committed
99b2ed7
[docs]: explain H3 RVM reward trends
Abecid
committed
6e05e4b
fix(rvm): match batch-global reward scaling and continuous time
Abecid
committed
9907bc4
feat(rvm): expose motion saturation diagnostics
Abecid
committed
3645dbd
feat(rvm): wire faithful scale-up configs and gates
Abecid
committed
ce5a3c0
fix(rvm): keep Modal preflight compatible with faithful configs
Abecid
committed
b738146
docs(rvm): align launch guidance with faithful scale-up path
Abecid
committed
47e9a8e
refactor(rvm): move smoke setup into portable scripts
Abecid
committed
593e8b1
refactor(rvm): reduce Modal to a thin test wrapper
Abecid
committed
32cf037
docs(rvm): separate Modal testing from custom-node training
Abecid
committed
c27d33f
fix(rvm): pass Modal smoke options as one payload
Abecid
committed
292ba58

Commits on Sep 2, 2026

docs(rvm): plan selectable MJ-VIDEO reward profile
Abecid
committed
55d60ae
feat(rvm): add fixed reward calibration profiles
Abecid
committed
9848405
docs(rvm): record reward calibration implementation
Abecid
committed
450a873
feat(rvm): add source-aligned MJ-VIDEO aspect rewards
Abecid
committed
c3445e3
docs(rvm): record MJ-VIDEO adapter implementation
Abecid
committed
a205254
feat(rvm): add MJ-VIDEO assets and reward calibration workflow
Abecid
committed
1d2a307
docs(rvm): record MJ-VIDEO calibration workflow
Abecid
committed
59f4db7
feat(rvm): add selectable Physion MJ reward profile
Abecid
committed
95fe910
docs(rvm): record selectable Physion reward profile
Abecid
committed
9a2cb67
docs(rvm): add Physion profile guide and compatibility preflight
Abecid
committed
a9ea0b2
infra(rvm): support 8 and 16 H100 custom-node campaigns
Abecid
committed
c22a503
ci(rvm): add temporary MJ reward implementation audit
Abecid
committed
92f50be
ci(rvm): install FastVideo import dependencies for focused audit
Abecid
committed
6650100
ci(rvm): include remaining FastVideo utility dependency
Abecid
committed
b4961c6
test(rvm): distinguish compact one-H100 smoke geometry
Abecid
committed
c79aeef
fix(rvm): handle missing calibration bank directory
Abecid
committed
bab0ab2
fix(rvm): make MJ asset revisions immutable
Abecid
committed
809552c
ci(rvm): audit pinned MJ-VIDEO source import
Abecid
committed
7bb6dc9
docs(rvm): finalize MJ reward implementation progress report
Abecid
committed
3e62724
fix(rvm): calibrate rewards on training prompts not eval prompts
Abecid
committed
0c1e472
fix(rvm): keep reward calibration separate from held-out eval
Abecid
committed
6ae0e70
test(rvm): enforce calibration and eval split separation
Abecid
committed
a7e7dc8
docs(rvm): keep MJ calibration split separate from evaluation
Abecid
committed
5401621
docs(rvm): finalize MJ reward implementation audit
Abecid
committed
1322bb9
ci(rvm): remove temporary MJ reward audit workflow
Abecid
committed
0e2d591
docs(h3-rest): select scored-trajectory teacher distillation
Abecid
committed
afdc5f1
feat(h3-rest): add AMD and H3 trajectory utilities
Abecid
committed
5bdbfbd
test(h3-rest): cover reward-modulated trajectory math
Abecid
committed
2b81991
fix(h3-rest): harden REST math contracts
Abecid
committed
f11c75e
feat(h3-rest): add immutable scored-trajectory cache loader
Abecid
committed
fe9f817
test(h3-rest): cover immutable cache integrity
Abecid
committed
bdef944
feat(h3-rest): add LoRA-shard EMA stabilization
Abecid
committed
a9a0f0f
feat(h3-rest): add cache-backed FastH3 model adapter
Abecid
committed
3c5d98d
feat(h3-rest): export cache-backed H3 adapter
Abecid
committed
99a1418
feat(h3-rest): add scored-trajectory AMD trainer
Abecid
committed
dffd53b
feat(h3-rest): export REST distillation method
Abecid
committed
c5a1338
test(h3-rest): cover AMD loss signs and EMA state
Abecid
committed
a44263e
feat(h3-rest): add frozen full-H3 cache-builder adapter
Abecid
committed
aa013ec
feat(h3-rest): export full-H3 REST teacher adapter
Abecid
committed
68329f4
feat(h3-rest): add dense H3 student-aligned sampler
Abecid
committed
ea9b1ea
feat(h3-rest): add deterministic scored-trajectory cache builder
Abecid
committed
678cfa9
feat(h3-rest): add fail-closed cache verifier
Abecid
committed
6c3a57b
feat(h3-rest): add pinned teacher and cache environment
Abecid
committed
e114cce
feat(h3-rest): add selective pinned H3 teacher download
Abecid
committed
ff6cf0c
feat(h3-rest): add compact 4-GPU cache config
Abecid
committed
0a9245d
feat(h3-rest): add production 4-GPU cache config
Abecid
committed
6ed71c9
feat(h3-rest): add one-H100 optimizer smoke config
Abecid
committed
e63f63a
feat(h3-rest): add SP4xDP2 quality pilot config
Abecid
committed
a444bbc
feat(h3-rest): add one-pass 8-H100 full config
Abecid
committed
74907dd