Skip to content

Update fork#34

Open
AleHD wants to merge 3476 commits into
swiss-ai:mainfrom
NVIDIA:main
Open

Update fork#34
AleHD wants to merge 3476 commits into
swiss-ai:mainfrom
NVIDIA:main

Conversation

@AleHD

@AleHD AleHD commented Feb 14, 2025

Copy link
Copy Markdown
Collaborator

Opening this PR to track changes done upstream.

@TJ-Solergibert TJ-Solergibert self-assigned this Mar 18, 2025
@AleHD AleHD mentioned this pull request Apr 10, 2025
Closed
chtruong814 and others added 28 commits May 22, 2026 11:37
@chtruong814
fix: Fix multi-node functional test phase sync (#4924)
4db6fa4 
Signed-off-by: Charlie Truong <chtruong@nvidia.com>
@shanmugamr1992 @claude
Perf tests (#4917)
686aa8c 
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
@balasaajay
fix(cuda_graphs): handle TE 2.15 removal of FP8GlobalStateManager.set…
0beaa53 
…_skip_fp8_weight_update_tensor (#4874)
@Phlip79
Fix paged stashing test submodules lookup (#4925)
6c1bd6e
@sraman-rgb @yaox12 @gautham-kollu
Add TEFusedDenseMLP for Dense+Grouped GEMM fusion on SM100+ (#4318) (#…
fa7a23b 
…4786)

Co-authored-by: Siddhartha Raman S <sraman@login-lyris02.lyris.clusters.nvidia.com>
Co-authored-by: Xin Yao <xiny@nvidia.com>
Co-authored-by: gautham-kollu <gkollu@nvidia.com>
Co-authored-by: Siddhartha Raman S <sraman@login-lyris01.lyris.clusters.nvidia.com>
@WanZzzzzz @Phlip79
Fix mxfp8 param gather numerical issue when DP overlap is off (#4800)
5f79118 
Signed-off-by: qiyuw <qiyuw@nvidia.com>
Co-authored-by: Philip Petrakian <ppetrakian@nvidia.com>
@WanZzzzzz
[MXFP8/FP4-param-gather] Post processing after forced param AG in eval (
08bad7a 
#4562)

Signed-off-by: qiyuw <qiyuw@nvidia.com>
@balasaajay
ci: Update training script paths in BERT and T5 (#4939)
3fb34c6
@maanug-nv
Various training utils (#4872)
34560c4 
Signed-off-by: Maanu Grover <maanug@nvidia.com>
@chtruong814
ci: restore perf test torchrun logs (#4951)
f7f584d 
Signed-off-by: Charlie Truong <chtruong@nvidia.com>
@deepakn94 @claude
Fix get_batch return order to ignore BlendedDataset provenance fiel…
4bd8bb3 
…ds (#4952)

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
@ko3n1g
test(release): add release goldens for deepseekv3/nemotron3 and set t…
be2b2cd 
…p2pp2 exit-interval (#4932)

Signed-off-by: oliver könig <okoenig@nvidia.com>
@github-actions
chore(beep boop 🤖): Bump (main) (2026-05-25)
3b2b6e7
@ko3n1g
test: enable NVTE_CUTEDSL_FUSED_GROUPED_MLP via pytest fixture (#4931)
2f754f4 
Signed-off-by: oliver könig <okoenig@nvidia.com>
@chtruong814
Avoid offsetting functional test master port (#4973)
2bd9fd5 
Signed-off-by: Charlie Truong <chtruong@nvidia.com>
@Devil1716
Fix elastification unwrap_model import (#4972)
4415119
@ko3n1g
test: re-enable paged stashing MoE tests (#4978)
432d76b 
Signed-off-by: oliver könig <okoenig@nvidia.com>
@ko3n1g @claude
test(ci): re-enable 8experts2parallel_multi_dist_optimizer_instances_…
ff64743 
…1node (#4984)

Signed-off-by: oliver könig <okoenig@nvidia.com>
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
@balasaajay
ci: Add support for MBridge job gating based on PR labels (#4926)
0e5cd0e
@ko3n1g @claude
test: re-enable test_pp2_create_cudagraphs_first_stage on TE 2.15+ (#…
08c368a 
…4985)

Signed-off-by: oliver könig <okoenig@nvidia.com>
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
@ko3n1g
fix(tests): initialize num_microbatches calculator in vision cudagrap…
6ce6fac 
…h tests (#4986)

Signed-off-by: oliver könig <okoenig@nvidia.com>
@balasaajay
ci: Add allow_failure flag to gpt and moe recipes that are failing in…
859b719 
… nightlies (#4905)
@deepakn94 @claude
Drain predecessor reduce-scatter at dispatch time (#4940)
88e7ab0 
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
@balasaajay
nightly(ci): Update golden values for functional t5 tests (#4995)
e6b2bd8
@github-actions
chore: rotate oncall schedule
7521ecb
@yanring @buptzyb
@Victarry @Phlip79 @shifangx
[main] Refactor and Improve MoE Logginginit commit (#3431)
873678a 
Co-authored-by: Robin Zhang <robinz@nvidia.com>
Co-authored-by: Dennis Liu <denliu@nvidia.com>
Co-authored-by: Philip Petrakian <ppetrakian@nvidia.com>
Co-authored-by: Shifang Xu <shifangx@nvidia.com>
@ko3n1g
ci: validate release branch-rules (#4929)
4e52a9e 
Signed-off-by: oliver könig <okoenig@nvidia.com>
@cspades
[Megatron-FSDP] Add conditional param.grad dereferencing logic to sup…
67b2f38 
…port full-iteration (FWD-BWD) CUDA graphability. (#4663)

Signed-off-by: Cory Ye <cye@nvidia.com>
github-actions Bot and others added 30 commits June 24, 2026 10:06
@github-actions
chore: rotate oncall schedule
cc0c960
@shanmugamr1992 @claude
Add inference functions to support MCore-/MBridge- training refactor …
4d44e37 
…and remove legacy modelbuilder functions (#5169)

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
@ko3n1g @claude
ci: launch GB200 unit tests via launch_on_gb200 marker (#5477)
5a256f3 
Signed-off-by: oliver könig <okoenig@nvidia.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
@ko3n1g @claude
build: install flash_mla from source in the CI image (#5481)
82de1b8 
Signed-off-by: oliver könig <okoenig@nvidia.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
@HollowMan6
[split 2/4] Scale DSA indexer loss in pipeline schedules (#5244)
0b0d985 
Signed-off-by: Hollow Man <hollowman@opensuse.org>
@ko3n1g
ci: check megatron.training imports in installation test (#5458)
0938eb7 
Signed-off-by: oliver könig <okoenig@nvidia.com>
@muyihao
Fix merges_file kwarg name in HuggingFaceTokenizer (#5406)
239959b 
Signed-off-by: yanghao.666 <yanghao.666@jd.com>
@Phlip79
Automated community request assignment (#5147)
3330d12 
Signed-off-by: Philip Petrakian <ppetrakian@nvidia.com>
@ilml
Clean up training.py module header (dedupe + reorganize imports/globa…
311416f 
…ls) (#5469)

Signed-off-by: ilml <tolong@nvidia.com>
@yashaswikarnati
Thread process groups through training checkpoint paths (#5486)
9038381 
Signed-off-by: ykarnati <ykarnati@nvidia.com>
@Phlip79
Narrow oncall responsibilities (#5490)
5863721 
Signed-off-by: Philip Petrakian <ppetrakian@nvidia.com>
@yashaswikarnati
Add MIMO forward step and per-token loss for hetero training (#5376)
1c1d6b5 
Signed-off-by: ykarnati <ykarnati@nvidia.com>
@yashaswikarnati @claude
Add Nemotron6-MoE VLM model provider for MIMO example (#5374)
ea967a7 
Signed-off-by: ykarnati <ykarnati@nvidia.com>
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
@ko3n1g @claude
ci: auto-retry test-data download in container-build job (#5498)
7168714 
Signed-off-by: oliver könig <okoenig@nvidia.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
@tdene
Force RL inference to CP=1 (#5423)
3bfd87b 
Signed-off-by: Teodor-Dumitru Ene <teodord.ene@gmail.com>
@deepakn94 @claude
Merge cu_seqlens across micro-batch for THD attention (#5454)
2a43e0d 
Signed-off-by: Deepak Narayanan <dnarayanan@nvidia.com>
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
@HollowMan6
[split 4/4] Enable DSA CP and THD hooks (#5246)
da482cf 
Signed-off-by: Hollow Man <hollowman@opensuse.org>
@sraman-rgb
Fix fused MLA down projection with tensor parallelism (#5383)
e1b8454 
Signed-off-by: Siddhartha Raman <sraman@nvidia.com>
@adityasingh2400
Fix NameError in is_flashinfer_min_version when check_equality=False (#…
8bafe7c 
…4961)

Signed-off-by: Aditya Singh <adisin650@gmail.com>
@Phlip79
Add hybrid FSDP unit module support (#4329)
da42015 
Signed-off-by: Philip Petrakian <ppetrakian@nvidia.com>
@ko3n1g @claude
fix: set DATA_PATH for moe-dynamic-inference recipe (#5506)
476228d 
Signed-off-by: oliver könig <okoenig@nvidia.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
@yeyu-nvidia @claude
Add --qad-train-target {base|mtp|both} for QAD / MTP QAT (frozen-base…
c0d7848 
…, frozen-MTP, or co-train) (#4785)

Signed-off-by: Ye Yu <yeyu@nvidia.com>
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
@zhongbozhu
[Main] Generalized fix for mxfp8 param gather (#5236)
0552f29 
Signed-off-by: Zhongbo Zhu <zhongboz@nvidia.com>
@wujingyue
Add CUDA graph training iteration test (#5417)
5949478 
Signed-off-by: Jingyue Wu <wujingyue@gmail.com>
@lauradang
test: restore G/G + lag=19 for gpt_grpo_tp4_pp1_dp2_8b throughput tes…
847de23 
…ts (#5514)

Signed-off-by: Laura Dang <laurad@nvidia.com>
@github-actions
Update copy-pr-bot.yaml [skip ci]
990ced9
@ko3n1g @claude
ci: cache-from a single coherent buildcache donor (#5509)
cbaa6eb 
Signed-off-by: oliver könig <okoenig@nvidia.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
@chtruong814
ci: Use GB300 for Github CI tests (#5520)
ed59a0a 
Signed-off-by: Charlie Truong <chtruong@nvidia.com>
@ko3n1g @claude
ci: pin HF_HUB_CACHE to bind-mounted cache for gpt-oss-20b inference …
0ff7226 
…test (#5512)

Signed-off-by: oliver könig <okoenig@nvidia.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
@deepakn94 @claude
Add inter-document attention masking to GPTDataset (#5298)
f88b85f 
Signed-off-by: Deepak Narayanan <dnarayanan@nvidia.com>
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

@TJ-Solergibert TJ-Solergibert

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.