[WS1][CUDA][MiniMax-H3] runtime modes, ablation overlay, Diffusers adapter, training provider, LoRA gate (RFC #420 stack 6/7) by fusheng-ji · Pull Request #574 · RL-Align/RL-Kernel

feat(h3): add projection rows (audio in/out, text context, QKV) on a fixed-order Triton GEMM

fusheng-ji

committed

7e9475d

feat(h3): projection rows Hopper evidence, charts and lint

fusheng-ji

committed

a6533fa

test(h3): make the single-unit audio_output shard case an explicit fail-closed test instead of a skip

fusheng-ji

committed

e109e16

fix(h3): bitwise row-permutation invariant d_weight/d_bias via logical keys

fusheng-ji

and

walker-ai

committed

c8a9bae

feat(h3): block graph passes logical row keys; skip the gather for already ordered rows

fusheng-ji

and

walker-ai

committed

5f7536c

[WS1][CUDA][MiniMax-H3] ws1_one_h3_block + full-model qualification: block 0 node-by-node and 50-block/generation gates

fusheng-ji

committed

17d0e9e

docs(h3): do not link to the fingerprint page, it lives in a sibling pull request

fusheng-ji

committed

3cf51bf

feat(h3): add architecture fingerprint (h3_arch_fingerprint)

fusheng-ji

committed

6dd9572

feat(h3): add operator trace ledger (h3_operator_trace)

fusheng-ji

committed

c74472b

feat(h3): add sealed trajectory artifact (h3_trajectory_artifact)

fusheng-ji

committed

efbac11

style(h3): black/isort/flake8 for the foundation rows

fusheng-ji

committed

bf06181

fix(h3): type the trajectory step body so mypy accepts StepRecord(**body)

fusheng-ji

committed

ea1687f

feat(h3): runtime modes, ablation overlay, Diffusers adapter, training provider, LoRA gate and CPS transition

fusheng-ji

committed

19a1e2a

feat(h3): token_refiner_block row - refiner graph, trace stage, shard weight loader, tests, doc

fusheng-ji

committed

3b1bd2b

feat(h3): ws1_full_h3_transformer and ws1_dual_trajectory harnesses with H100 evidence

fusheng-ji

committed

ae184d9

test(h3): measurement report for the runtime integration rows

fusheng-ji

committed

4d8f72e

feat(h3): token refiner Hopper evidence, charts and lint

fusheng-ji

committed

7ccd4a0

feat(h3): full transformer and dual trajectory Hopper evidence charts and lint

fusheng-ji

committed

d7c1952

test(h3): runtime report covers 925 and 4253 tokens and tolerates OOM

fusheng-ji

committed

51ad1db

style(h3): lint refiner report tool

fusheng-ji

committed

cd1b0fd

style(h3): black for the full-transformer chart tool

fusheng-ji

committed

8acada2

test(h3): runtime report prints OOM tracebacks and can skip ablations

fusheng-ji

committed

f63494e

fix(h3): satisfy mypy in the refiner report and the operator-trace stage import

fusheng-ji

committed

6de320d

style(h3): docstring length in the full-transformer chart tool

fusheng-ji

committed

eb7840b

test(h3): load the verl-omni scheduler by path instead of editing sys.path

fusheng-ji

committed

af3dcae

Navigation Menu

Repository navigation

[WS1][CUDA][MiniMax-H3] runtime modes, ablation overlay, Diffusers adapter, training provider, LoRA gate (RFC #420 stack 6/7) - #574

[WS1][CUDA][MiniMax-H3] runtime modes, ablation overlay, Diffusers adapter, training provider, LoRA gate (RFC #420 stack 6/7)#574
fusheng-ji wants to merge 30 commits into
RL-Align:test-h3from
fusheng-ji:feat/h3-runtime-integration

Commits on Oct 22, 2026

feat(h3): add projection rows (audio in/out, text context, QKV) on a fixed-order Triton GEMM

feat(h3): projection rows Hopper evidence, charts and lint

test(h3): make the single-unit audio_output shard case an explicit fail-closed test instead of a skip

fix(h3): bitwise row-permutation invariant d_weight/d_bias via logical keys

feat(h3): block graph passes logical row keys; skip the gather for already ordered rows

[WS1][CUDA][MiniMax-H3] ws1_one_h3_block + full-model qualification: block 0 node-by-node and 50-block/generation gates

docs(h3): do not link to the fingerprint page, it lives in a sibling pull request

feat(h3): add architecture fingerprint (h3_arch_fingerprint)

feat(h3): add operator trace ledger (h3_operator_trace)

feat(h3): add sealed trajectory artifact (h3_trajectory_artifact)

style(h3): black/isort/flake8 for the foundation rows

fix(h3): type the trajectory step body so mypy accepts StepRecord(**body)

feat(h3): runtime modes, ablation overlay, Diffusers adapter, training provider, LoRA gate and CPS transition

feat(h3): token_refiner_block row - refiner graph, trace stage, shard weight loader, tests, doc

feat(h3): ws1_full_h3_transformer and ws1_dual_trajectory harnesses with H100 evidence

test(h3): measurement report for the runtime integration rows

feat(h3): token refiner Hopper evidence, charts and lint

feat(h3): full transformer and dual trajectory Hopper evidence charts and lint

test(h3): runtime report covers 925 and 4253 tokens and tolerates OOM

style(h3): lint refiner report tool

style(h3): black for the full-transformer chart tool

test(h3): runtime report prints OOM tracebacks and can skip ablations

fix(h3): satisfy mypy in the refiner report and the operator-trace stage import

style(h3): docstring length in the full-transformer chart tool

test(h3): load the verl-omni scheduler by path instead of editing sys.path

Commits on Oct 25, 2026

Merge test-h3 into feat/h3-one-block

Merge local/h3merge/h3-one-block into feat/h3-projection-rows

Merge local/h3merge/h3-projection-rows into feat/h3-token-refiner

Merge local/h3merge/h3-token-refiner into feat/h3-full-transformer

Merge local/h3merge/h3-full-transformer into feat/h3-runtime-integration