Repository navigation
Clarify NVIDIA email signing guidance - #5699
Merged
wujingyue merged 1 commit intoJul 16, 2026
Merged
Conversation
Signed-off-by: Jingyue Wu <jingyuew@nvidia.com>
|
Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually. Contributors can view more details about this message here. |
wujingyue
commented
Jul 7, 2026
| - Commit PR changes with both `-s` and `-S`: `-s` adds the required | ||
| `Signed-off-by` trailer, and `-S` signs the commit so copy-pr-bot and `/ok to | ||
| test` can verify the pushed commit without manually specifying the SHA. | ||
| Megatron Core engineers at NVIDIA should sign using their NVIDIA emails so they |
Contributor
Author
There was a problem hiding this comment.
Is this right? I got kicked out from certain user groups because I had been signing with my personal email. But I'm not sure about the machinery behind.
Phlip79
approved these changes
Jul 16, 2026
Contributor
|
🔄 Merge queue validation started! You can track the progress here: https://github.com/NVIDIA/Megatron-LM/actions/runs/29470436835 |
chochowski
pushed a commit
to chochowski/Megatron-LM
that referenced
this pull request
Jul 20, 2026
Signed-off-by: Jingyue Wu <jingyuew@nvidia.com> Signed-off-by: mchochowski <mchochowski@nvidia.com>
terminator123
pushed a commit
to 021ai/Megatron-LM
that referenced
this pull request
Aug 3, 2026
Signed-off-by: Jingyue Wu <jingyuew@nvidia.com>
svcnvidia-nemo-ci
pushed a commit
to dimapihtar/Megatron-LM
that referenced
this pull request
Aug 4, 2026
Signed-off-by: Jingyue Wu <jingyuew@nvidia.com> Signed-off-by: Dmytro Pykhtar <dpykhtar@nvidia.com>
BestJuly
added a commit
to BestJuly/Megatron-LM
that referenced
this pull request
Sep 23, 2026
* Short-circuit condition to avoid copying from GPU memory in `ChainedOptimizer` (NVIDIA#5623) Signed-off-by: Hristo Filaretov <hfilaretov@nvidia.com> Co-authored-by: Michał Marcinkiewicz <43240942+mmarcinkiewicz@users.noreply.github.com> * Increase Megatron-FSDP overlap test dim to 8192 for reliable overlap (NVIDIA#5770) Signed-off-by: Jingyue Wu <jingyuew@nvidia.com> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * Inference: Add profile endpoints to chat completions. (NVIDIA#5611) Signed-off-by: Siddharth Singh <sidsingh@nvidia.com> * Inference: Do not route pad/dummy tokens to any expert (NVIDIA#4922) Signed-off-by: Siddharth Singh <sidsingh@nvidia.com> * Inference: Add load aware routing to prefix caching. (NVIDIA#5607) Signed-off-by: Siddharth Singh <sidsingh@nvidia.com> * fix(clip_grads): handle empty grads_for_norm in inf-norm and p-norm paths (NVIDIA#5530) Signed-off-by: Min Htet Myet <mattralminn@email.com> Signed-off-by: Mattral <mattralminn@gmail.com> * test(gpt): AUT-830 mark tp1_pp4_vp1_resume_torch_decoupled_lr flaky on h100 (NVIDIA#5801) Signed-off-by: svcnemo-autobot <svcnemo-autobot@nvidia.com> * Delegate reasoning token retention to the chat template in multi-turn conversations (NVIDIA#5276) * Various ModelOpt fixes: QAD test for CICD, use model builder config instead of model provider, allow loading teacher ckpt independently of student (NVIDIA#4520) Signed-off-by: Asha Anoosheh <aanoosheh@nvidia.com> * build: Update Transformer Engine to 2.17 (NVIDIA#5680) Signed-off-by: Ajay <abalasa@nvidia.com> Signed-off-by: Ajay Balasa <abalasa@nvidia.com> Signed-off-by: [Your Name] <your.email@example.com> * Mamba prefix caching fixes (NVIDIA#5502) Signed-off-by: Keshav Santhanam <ksanthanam@nvidia.com> Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com> * Exercise nested MFSDP CUDA graph capture (NVIDIA#5796) Signed-off-by: Jingyue Wu <jingyuew@nvidia.com> * Implement Quantile Balancing in MoE (NVIDIA#5349) Signed-off-by: Haoran Zhang <haoranz@nvidia.com> * Unset NCCL overrides for MFSDP v2 tests (NVIDIA#5794) Signed-off-by: Jingyue Wu <jingyuew@nvidia.com> * Fix MegatronFSDP root module hook dispatch (NVIDIA#5808) Signed-off-by: Jingyue Wu <jingyuew@nvidia.com> * Pin cudnn-fe and cuTeDSL version (NVIDIA#5812) Signed-off-by: Ajay Balasa <abalasa@nvidia.com> * chore: rotate oncall schedule * Inference: Add the nemotron_v3 reasoning parser (NVIDIA#5634) Signed-off-by: Siddharth Singh <sidsingh@nvidia.com> * Missing moe_router_dtype causes unexpected downcast in ModelOpt example (NVIDIA#5810) Signed-off-by: Jinhang Choi <jinhangc@nvidia.com> Co-authored-by: Cory Ye <44509866+cspades@users.noreply.github.com> * Avoid FSDP unit terminology in MFSDP v2 (NVIDIA#5793) Signed-off-by: Jingyue Wu <jingyuew@nvidia.com> Signed-off-by: svcnvidia-nemo-ci <svcnvidia-nemo-ci@nvidia.com> Co-authored-by: svcnvidia-nemo-ci <svcnvidia-nemo-ci@nvidia.com> * Fix configured norm epsilon in MambaLayer (NVIDIA#5750) Signed-off-by: Shanhao <5197744+shanhaoli@users.noreply.github.com> * [refactor] Common combined-1F1B schedule-plan base (1/4 of NVIDIA#4798) (NVIDIA#4941) Signed-off-by: Yan Xu <yxu1@nvidia.com> Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * chore(tests): AUT-851 move NCCL defaults from run_ci_test.sh to conftest (NVIDIA#5826) Signed-off-by: svcnemo-autobot <svcnemo-autobot@nvidia.com> * Set num_splits to 0 for FA4 inference (NVIDIA#5804) Signed-off-by: Keshav Santhanam <ksanthanam@nvidia.com> * ci: Enhance nightly/mr/weekly error reporting (NVIDIA#5831) Signed-off-by: Ajay Balasa <abalasa@nvidia.com> * feat(docker): Add NCCL installation script and install NCCL 2.30.4 (NVIDIA#5815) Signed-off-by: Ajay Balasa <abalasa@nvidia.com> * IMA fix by making the copy of book keeping buffer to GPU blocking (NVIDIA#5715) Signed-off-by: Shanmugam Ramasamy <111910568+shanmugamr1992@users.noreply.github.com> * Add GPTModel to HybridModel migration guide (NVIDIA#5698) Signed-off-by: Philip Petrakian <ppetrakian@nvidia.com> * Pair frozen FSDP backward hooks (NVIDIA#5710) Signed-off-by: Jingyue Wu <jingyuew@nvidia.com> * Clarify NVIDIA email signing guidance (NVIDIA#5699) Signed-off-by: Jingyue Wu <jingyuew@nvidia.com> * Overlap FSDP communication with compute (NVIDIA#5719) Signed-off-by: Jingyue Wu <jingyuew@nvidia.com> * chore(skills): add Regent Open Plugin manifest (NVIDIA#5840) Signed-off-by: oliver könig <okoenig@nvidia.com> Co-authored-by: Claude Fable 5 <noreply@anthropic.com> * chore(skills): remove Open Plugin manifest (superseded) (NVIDIA#5842) Signed-off-by: oliver könig <okoenig@nvidia.com> * return prefix cache hits data from the chat completions api (NVIDIA#5609) Signed-off-by: Siddharth Singh <sidsingh@nvidia.com> * Refactor data parallel coordinator to enable modular handlers (NVIDIA#5550) Signed-off-by: Keshav Santhanam <ksanthanam@nvidia.com> * Pass device IDs to cleanup barrier (NVIDIA#5702) Signed-off-by: Jingyue Wu <jingyuew@nvidia.com> * Fix averaging for MoE z-loss metric tracking (NVIDIA#3199) Signed-off-by: Markus Schnös <markus.schnoes@deepl.com> Co-authored-by: Markus Schnös <markus.schnoes@deepl.com> * test(mfsdp): AUT-881 mark test_overlaps_communication_and_compute flaky (NVIDIA#5848) Signed-off-by: svcnemo-autobot <svcnemo-autobot@nvidia.com> * Inference: Extend default cuda-graph coverage to 512 tokens (NVIDIA#5797) * Fix for sequence-level aux MoE loss being dependent on batch size (NVIDIA#5798) Signed-off-by: Oleg Sudakov <osudakov@nvidia.com> Signed-off-by: Oleg Sudakov <oleg.sudakov@outlook.com> Co-authored-by: Fei Wu <33940270+YangFei1990@users.noreply.github.com> * Allow parameterless FSDP root modules (NVIDIA#5711) Signed-off-by: Jingyue Wu <jingyuew@nvidia.com> * Fix issue where parameter groups with different min/max LRs get overridden at checkpoint load time (NVIDIA#4705) Signed-off-by: John St John <jstjohn@nvidia.com> Signed-off-by: John St. John <jstjohn@nvidia.com> * fix bug where Gemma4 is not working with recompute_granularity = "full" (NVIDIA#5324) Signed-off-by: Yury Parfenov <4665475+warpuv@users.noreply.github.com> Co-authored-by: Guihong Li <guihongl@nvidia.com> * Avoid extra MFSDP v2 model-weight sync memcpy (NVIDIA#5834) Signed-off-by: Jingyue Wu <jingyuew@nvidia.com> * [experimental] Add experimental/agent_compose placeholder with preview pointer (NVIDIA#5639) Signed-off-by: Yan Bai <bayan@nvidia.com> * Stabilize mfsdp_v2 overlap test by enlarging the model (NVIDIA#5846) Signed-off-by: Jingyue Wu <jingyuew@nvidia.com> * Print important dependencies (NVIDIA#5814) Signed-off-by: Gautham Kollu <gkollu@nvidia.com> * Test zero-CTA copy-engine all-gather (NVIDIA#5858) Signed-off-by: Jingyue Wu <jingyuew@nvidia.com> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * Route Lion through DistributedOptimizer and support single-moment checkpointing (NVIDIA#5742) Signed-off-by: Deepak Narayanan <dnarayanan@nvidia.com> Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com> * Fix broken remove_sharded_tensors public API and re-enable its unit test (NVIDIA#5759) Signed-off-by: Rui Zhu <rui.zhu.rz399@yale.edu> Co-authored-by: Claude Fable 5 <noreply@anthropic.com> Co-authored-by: Dmytro Pykhtar <37850217+dimapihtar@users.noreply.github.com> * Make the model larger and higher mb size to make reduce flakiness (NVIDIA#5816) Signed-off-by: Gautham Kollu <gkollu@nvidia.com> * chore(beep boop 🤖): Bump (main) (2026-07-20) * Update active oncall to Phlip79 this week (NVIDIA#5896) Signed-off-by: Philip Petrakian <ppetrakian@nvidia.com> * Overlap async scheduling phases (NVIDIA#5549) Signed-off-by: Lawrence McAfee <lmcafee@nvidia.com> * fix(deps): AUT-1323 use released dev dependency sources (NVIDIA#6273) Signed-off-by: NVIDIA NeMo Bot <nemo-bot@nvidia.com> Co-authored-by: NVIDIA NeMo Bot <nemo-bot@nvidia.com> * build: AUT-1363 verify downloaded yq binaries (NVIDIA#6344) Signed-off-by: svcnemo-autobot <svcnemo-autobot@nvidia.com> * cp: chore(ci): bump community workflow to v1.8.8 (NVIDIA#6345) Signed-off-by: oliver könig <okoenig@nvidia.com> Signed-off-by: svcnemo-autobot <svcnemo-autobot@nvidia.com> Co-authored-by: oliver könig <okoenig@nvidia.com> * fix(ci): AUT-1396 stabilize docker cache and external test queue (NVIDIA#6348) Signed-off-by: svcnemo-autobot <svcnemo-autobot@nvidia.com> * fix(optimizer): AUT-1363 avoid grad threshold check during capture (NVIDIA#6355) Signed-off-by: Hristo Filaretov <hfilaretov@nvidia.com> Signed-off-by: svcnemo-autobot <svcnemo-autobot@nvidia.com> Co-authored-by: Hristo Filaretov <hello@hgf.sh> Co-authored-by: Michał Marcinkiewicz <43240942+mmarcinkiewicz@users.noreply.github.com> * [dev] Refactor CSA structure: Move CSA implementation helpers into csa_utils dir (NVIDIA#6372) Signed-off-by: Hongxiao Bai <hongxiaob@nvidia.com> * [dev] Optimize MTP contiguous packed-CP rolls (NVIDIA#6246) Signed-off-by: Hongxiao Bai <hongxiaob@nvidia.com> * [Dev] Update moe recipes (NVIDIA#6335) Signed-off-by: Xin Yao <xiny@nvidia.com> * fix(ci): redact sensitive environment values from test output (NVIDIA#6408) Signed-off-by: oliver könig <okoenig@nvidia.com> * [dev] [Deepseek-V4] Fix fused CSA indexer loss normalization and compact attention indices (NVIDIA#6349) Signed-off-by: Hongxiao Bai <hongxiaob@nvidia.com> * ci(auth): treat svcnemo-autobot as internal (NVIDIA#6437) Signed-off-by: oliver könig <okoenig@nvidia.com> * [DEV] fix(multimodal_dev): make --record-memory-history work on the model_provider entry (NVIDIA#6277) Signed-off-by: xuwenc <xuwenc@nvidia.com> Co-authored-by: Claude Fable 5 <noreply@anthropic.com> Co-authored-by: Li Tao <lit@nvidia.com> * fix(ci): AUT-1427 skip Claude labeling for external authors (NVIDIA#6395) Signed-off-by: svcnemo-autobot <svcnemo-autobot@nvidia.com> * fix: AUT-1504 abort when SSO list download fails on dev (NVIDIA#6517) Signed-off-by: svcnemo-autobot <svcnemo-autobot@nvidia.com> * ci: AUT-1408 backport GitHub App authentication to dev (NVIDIA#6435) Signed-off-by: svcnemo-autobot <svcnemo-autobot@nvidia.com> * fix(transformer): AUT-1497 support dual RoPE full recompute (NVIDIA#6519) Signed-off-by: svcnemo-autobot <svcnemo-autobot@nvidia.com> * ci: Do not upload docker oci artifact attestations (NVIDIA#6543) Signed-off-by: Charlie Truong <chtruong@nvidia.com> * Fix non-fused mHC projection for empty sequences (NVIDIA#6371) Signed-off-by: bvolpato <brunocvcunha@gmail.com> * [dev] Refactor: extract and split common logic between GDN & GDN2 (cherry-pick NVIDIA#5843) (NVIDIA#6088) Signed-off-by: Yuzhong Wang <yuzhongw@nvidia.com> Co-authored-by: claude[bot] <209825114+claude[bot]@users.noreply.github.com> * [Dev] Allow FP8/FP4 with AbsorbedMLA up-projection recompute (NVIDIA#6178) Signed-off-by: Yunsheng Ni <yunshengn@nvidia.com> Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com> * [dev] Add chunked optimizer-state and master-weight offload (NVIDIA#6244) Signed-off-by: Hongxiao Bai <hongxiaob@nvidia.com> * [Dev] Support mHC selective recompute with CUDA graphs under EP a2a overlap (NVIDIA#5841) Signed-off-by: Jingqi Yang <jingqiny@nvidia.com> Signed-off-by: jingqiny-99 <jingqiny@nvidia.com> * [Dev] Recompute EP A2A overlap activations per layer instead of per model chunk (NVIDIA#6311) Signed-off-by: Pingtian Li <pingtianl@nvidia.com> Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com> * [dev] moe(perf): Refactor CP layout organization for Qwen3.5-style hybrid attention model (NVIDIA#6387) Signed-off-by: Yuzhong Wang <yuzhongw@nvidia.com> Co-authored-by: Haoran Zhang <haoranz@nvidia.com> * [lite] Add MagiAttention as an optional core-attention backend for qwen3_moe (NVIDIA#6478) Signed-off-by: Tailai Ma <58548582+xiaoyao0115@users.noreply.github.com> Co-authored-by: Claude Fable 5 <noreply@anthropic.com> * [dev] Overlap CP communication for indexer and compressed KV (NVIDIA#5691) Signed-off-by: xiaoyao0115 <1804647152@qq.com> * [dev] Replace DP balance all-to-all rerouting with all-gather (NVIDIA#6378) Signed-off-by: xiaoyao0115 <1804647152@qq.com> * [dev] moe(perf): Minor improvements of pre-GDR fused path. (NVIDIA#6658) Signed-off-by: Yuzhong Wang <yuzhongw@nvidia.com> Co-authored-by: CarlosGomes98 <cgomes@nvidia.com> * [Dev] Generalize FSDP persistent communication buffers (NVIDIA#6316) Signed-off-by: hongbinl <hongbinl@nvidia.com> * [Dev] Add Ling-V3 Tiny hybrid model training support (NVIDIA#6556) Signed-off-by: fkuner <784819644@qq.com> Signed-off-by: Yuzhong Wang <yuzhongw@nvidia.com> Signed-off-by: Fan Kun <784819644@qq.com> Co-authored-by: Yuzhong Wang <yuzhongw@nvidia.com> Co-authored-by: claude[bot] <209825114+claude[bot]@users.noreply.github.com> * ci: centralize container builds and add cache population workflow (NVIDIA#6646) Signed-off-by: Ajay Balasa <abalasa@nvidia.com> Signed-off-by: Charlie Truong <chtruong@nvidia.com> * ci: prioritize baseline cache for merge groups (NVIDIA#6721) Signed-off-by: Charlie Truong <chtruong@nvidia.com> * ci: use base PR cache as fallback (NVIDIA#6774) Signed-off-by: Charlie Truong <chtruong@nvidia.com> * Set provenance: false during docker build Signed-off-by: Charlie Truong <chtruong@nvidia.com> * [dev](Fix): Make the mHC attention CUDA-graph split opt-in (NVIDIA#6661) Signed-off-by: Jingqi Yang <jingqiny@nvidia.com> Signed-off-by: jingqiny-99 <jingqiny@nvidia.com> * feat: support unified model bagel with generation and understanding (NVIDIA#3635) Signed-off-by: sophiayyya <yangjingyi0730@163.com> Signed-off-by: root <root@eos0294.eos.clusters.nvidia.com> Signed-off-by: Sophia Yang <sopyang@cw-dfw-cs-001-dc-02.cm.cluster> Signed-off-by: Deyu Fu <deyuf@nvidia.com> Signed-off-by: Sophia Yang <sopyang@cw-dfw-cs-001-vscode-01.cm.cluster> Co-authored-by: root <root@eos0294.eos.clusters.nvidia.com> Co-authored-by: Sophia Yang <sopyang@cw-dfw-cs-001-dc-02.cm.cluster> Co-authored-by: Deyu Fu <deyuf@nvidia.com> Co-authored-by: Sophia Yang <sopyang@cw-dfw-cs-001-vscode-01.cm.cluster> * [Dev] Fuse standard RoPE packing for MLA and DSA (NVIDIA#6343) Signed-off-by: Robin Zhang <robinz@nvidia.com> * [dev] add DSv4 Flash THD 4K and 64K recipe (NVIDIA#6911) Signed-off-by: Hongxiao Bai <hongxiaob@nvidia.com> * [Dev] Add configurable attention latent normalization epsilon (NVIDIA#6204) Signed-off-by: janEbert <janpabloe@nvidia.com> Signed-off-by: Robin Zhang <robinz@nvidia.com> Co-authored-by: janEbert <janpabloe@nvidia.com> * [Dev] Add TileLang DSA and indexer weights projection precision controls (NVIDIA#6279) Signed-off-by: ℍ𝕠𝕝𝕝𝕠𝕨 𝕄𝕒𝕟 <hollowman@opensuse.org> Signed-off-by: Robin Zhang <robinz@nvidia.com> Co-authored-by: ℍ𝕠𝕝𝕝𝕠𝕨 𝕄𝕒𝕟 <hollowman@opensuse.org> * fix(deps): AUT-2172 pin compatible TileLang runtime (NVIDIA#6946) Signed-off-by: svcnemo-autobot <svcnemo-autobot@nvidia.com> * Fix hybrid MTP partial CUDA graph capture (NVIDIA#6583) Signed-off-by: guihong-nv <guihongl@nvidia.com> * [dev] Fix mock data, hybrid recompute, CUDA graph reuse, and Hash MoE selection (NVIDIA#6704) Signed-off-by: Hongxiao Bai <hongxiaob@nvidia.com> Signed-off-by: Deyu Fu <deyuf@nvidia.com> Co-authored-by: Deyu Fu <deyuf@nvidia.com> * [Dev] Support paged stash with TE whole-MoE CUDA graphs (NVIDIA#6022) Signed-off-by: Robin Zhang <robinz@nvidia.com> Co-authored-by: Charlie Truong <chtruong@nvidia.com> * [Dev] Support device-initiated grouped linear without TE op fuser (NVIDIA#6847) Signed-off-by: Zhongbo Zhu <42691305+zhongbozhu@users.noreply.github.com> Signed-off-by: hongbinl <hongbinl@nvidia.com> Co-authored-by: Zhongbo Zhu <42691305+zhongbozhu@users.noreply.github.com> * [dev] lite: Sync Megatron Lite experimental updates for deepseek v4 (NVIDIA#6397) Signed-off-by: Yan Bai <bayan@nvidia.com> * [Dev] fix: account for DSA top-k sparsity and indexer cost in FLOPs (NVIDIA#6753) Signed-off-by: Yunsheng Ni <yunshengn@nvidia.com> Co-authored-by: Claude Fable 5 <noreply@anthropic.com> * build: define ssm and te extras consumed by Dockerfile.ci.lts (NVIDIA#7004) Signed-off-by: oliver könig <okoenig@nvidia.com> Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com> * [Dev] Add global-batch quantile balancing (NVIDIA#6637) Signed-off-by: Harry Zhou <hhanyu@nvidia.com> Signed-off-by: Deyu Fu <deyuf@nvidia.com> * chore: AUT-2153 sync main through 3005e9c into dev (NVIDIA#6907) Signed-off-by: Hristo Filaretov <hfilaretov@nvidia.com> Signed-off-by: Jingyue Wu <jingyuew@nvidia.com> Signed-off-by: svcnemo-autobot <svcnemo-autobot@nvidia.com> Co-authored-by: Hristo Filaretov <hello@hgf.sh> Co-authored-by: Michał Marcinkiewicz <43240942+mmarcinkiewicz@users.noreply.github.com> Co-authored-by: Jingyue Wu <jingyuew@nvidia.com> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * [dev] fix: avoid duplicate offload of recomputed activations (NVIDIA#7044) Signed-off-by: Hongxiao Bai <hongxiaob@nvidia.com> * [Dev] ci: guard examples/multimodal_dev with a unit-test bucket (NVIDIA#6083) Signed-off-by: Li Tao <lit@nvidia.com> Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com> * [Dev] Add RMSNorm latent up-projection (NVIDIA#6804) Signed-off-by: Harry Zhou <hhanyu@nvidia.com> * cp: AUT-2161 route external jobs through ephemeral v2 (NVIDIA#6909) (NVIDIA#7065) Signed-off-by: svcnemo-autobot <svcnemo-autobot@nvidia.com> Signed-off-by: Charlie Truong <chtruong@nvidia.com> Co-authored-by: svcnemo-autobot <svcnemo-autobot@nvidia.com> * fix(sync): preserve reviewable argument history Signed-off-by: svcnemo-autobot <svcnemo-autobot@nvidia.com> (cherry picked from commit 562bada) * fix(sync): defer unauditable triage integration Signed-off-by: svcnemo-autobot <svcnemo-autobot@nvidia.com> (cherry picked from commit 796e9a2) * fix(sync): defer coupled triage consumers Signed-off-by: svcnemo-autobot <svcnemo-autobot@nvidia.com> (cherry picked from commit 69084d2) * style(sync): format inherited Python changes Signed-off-by: svcnemo-autobot <svcnemo-autobot@nvidia.com> (cherry picked from commit cdf6af0) * fix: restore omitted sync dependencies Signed-off-by: svcnemo-autobot <svcnemo-autobot@nvidia.com> (cherry picked from commit 4d9b765) * fix: close remaining sync test failures Signed-off-by: svcnemo-autobot <svcnemo-autobot@nvidia.com> (cherry picked from commit 9f3a66c) * fix: repair remaining sync unit failures Signed-off-by: svcnemo-autobot <svcnemo-autobot@nvidia.com> (cherry picked from commit e07e13b) * style: format dynamic engine assignment Signed-off-by: svcnemo-autobot <svcnemo-autobot@nvidia.com> (cherry picked from commit b25258e) * fix: restore async scheduling token commit Signed-off-by: svcnemo-autobot <svcnemo-autobot@nvidia.com> (cherry picked from commit 56834c1) * fix(inference): restore synchronized async and cache contracts Signed-off-by: svcnemo-autobot <svcnemo-autobot@nvidia.com> (cherry picked from commit ed1a2db) * fix(modelopt): avoid unsupported recipe API Signed-off-by: svcnemo-autobot <svcnemo-autobot@nvidia.com> (cherry picked from commit e884bf3) * fix(modelopt): normalize legacy quantization configs Signed-off-by: svcnemo-autobot <svcnemo-autobot@nvidia.com> (cherry picked from commit a659231) * fix(modelopt): support legacy quantization configs Signed-off-by: svcnemo-autobot <svcnemo-autobot@nvidia.com> (cherry picked from commit e4ba000) * fix(modelopt): support legacy prefill signatures Signed-off-by: svcnemo-autobot <svcnemo-autobot@nvidia.com> (cherry picked from commit 3192094) * fix(modelopt): restore migrated model builders Signed-off-by: svcnemo-autobot <svcnemo-autobot@nvidia.com> (cherry picked from commit 316a437) * fix(config): resolve model config defaults at call time Signed-off-by: svcnemo-autobot <svcnemo-autobot@nvidia.com> (cherry picked from commit ab30e83) * fix(gpt): repair hyperconnection and MTP specs Signed-off-by: svcnemo-autobot <svcnemo-autobot@nvidia.com> (cherry picked from commit 95da38c) * test(hybrid): refresh QAD total-loss golden values Signed-off-by: svcnemo-autobot <svcnemo-autobot@nvidia.com> (cherry picked from commit d8f1e40) * test: align quantile balancing coverage with dev Signed-off-by: svcnemo-autobot <svcnemo-autobot@nvidia.com> * fix(moe): preserve sequence aux loss batch scaling Signed-off-by: svcnemo-autobot <svcnemo-autobot@nvidia.com> * [Dev] Add SiTU-GLU through Transformer Engine (NVIDIA#6673) Signed-off-by: Harry Zhou <hhanyu@nvidia.com> Signed-off-by: Deyu Fu <deyuf@nvidia.com> Co-authored-by: Deyu Fu <deyuf@nvidia.com> * [dev] Fix premature handler advancement in fine-grained activation offloading (NVIDIA#6783) Signed-off-by: HaochenYuan <haocheny@nvidia.com> Signed-off-by: hongbinl <hongbinl@nvidia.com> Co-authored-by: hongbinl <hongbinl@nvidia.com> * [DEV] fix(multimodal_dev): derive mock samples from the sample index (NVIDIA#6287) Signed-off-by: xuwenc <xuwenc@nvidia.com> Co-authored-by: Xin Yao <xiny@nvidia.com> * [DEV] fix(multimodal_dev): align the Qwen3.5-VL fp32 RoPE wrappers with apply_rotary_pos_emb (NVIDIA#6611) Signed-off-by: xuwenc <xuwenc@nvidia.com> Co-authored-by: Xin Yao <xiny@nvidia.com> * [dev] [fix] Skip auxiliary loss for hash MoE layers (NVIDIA#7079) Signed-off-by: Hongxiao Bai <hongxiaob@nvidia.com> Co-authored-by: Xin Yao <xiny@nvidia.com> * [dev] [DeepSeek-V4] Add compact BF16 and MXFP8 DSA indexer support (NVIDIA#5992) Signed-off-by: Hongxiao Bai <hongxiaob@nvidia.com> * feat(moe): add a megakernel backend with MOK integration (NVIDIA#6572) Signed-off-by: qizhang <qizhang@nvidia.com> Signed-off-by: Xin Yao <xiny@nvidia.com> Co-authored-by: Xin Yao <xiny@nvidia.com> * [MoE] Preserve HybridEP capacity across CUDA graph replay (NVIDIA#6999) Signed-off-by: Gao <gdeng@nvidia.com> * [dev] DeepSeek-V4 CSA optimization: fuse some THD-related ops (NVIDIA#7064) Signed-off-by: Hongxiao Bai <hongxiaob@nvidia.com> * [Dev] Enable dense routing maps for Flex dispatch (NVIDIA#6614) Signed-off-by: Harry Zhou <hhanyu@nvidia.com> Signed-off-by: Deyu Fu <deyuf@nvidia.com> Co-authored-by: Deyu Fu <deyuf@nvidia.com> Co-authored-by: Xin Yao <xiny@nvidia.com> * [Dev] feat(multimodal): migrate Qwen3.5-VL to HybridModel (NVIDIA#6315) Signed-off-by: Li Tao <lit@nvidia.com> * [Dev] fix: Preserve offload events in Hybrid mHC CUDA graphs (NVIDIA#7014) Signed-off-by: jingqiny-99 <jingqiny@nvidia.com> * [dev] agent_compose: adopt Core fusions and main_grad accumulation on the DSv4 path (NVIDIA#7127) Signed-off-by: Yan Bai <bayan@nvidia.com> Signed-off-by: leikaixiang.shyoshyo <leikaixiang.shyoshyo@bytedance.com> Co-authored-by: leikaixiang.shyoshyo <leikaixiang.shyoshyo@bytedance.com> * [Dev] Remove unused legacy hybrid CP implementation (NVIDIA#7203) Signed-off-by: Tailai Ma <tailaim@nvidia.com> * chore: AUT-2153 sync exact 100 main commits into dev (NVIDIA#7094) Signed-off-by: Ajay Balasa <abalasa@nvidia.com> Signed-off-by: Evgenii Zheltonozhskii <zheltonozhskiy@gmail.com> Signed-off-by: Yongqiang Wang <yongqiang.seagull@gmail.com> Signed-off-by: Achyuthan Sivasankar <achyuthan.sivasankar@gmail.com> Signed-off-by: svcnvidia-nemo-ci <svcnvidia-nemo-ci@nvidia.com> Signed-off-by: Jingyue Wu <jingyuew@nvidia.com> Signed-off-by: Philip Petrakian <ppetrakian@nvidia.com> Signed-off-by: Teodor-Dumitru Ene <teodord.ene@gmail.com> Signed-off-by: Ali Arda Eker <aeker@nvidia.com> Signed-off-by: Charlie Truong <chtruong@nvidia.com> Signed-off-by: ykarnati <ykarnati@nvidia.com> Signed-off-by: William Dykas <wdykas@oci-hsg-cs-001-vscode-03.cm.cluster> Signed-off-by: William Dykas <wdykas@oci-nrt-cs-001-vscode-01.cm.cluster> Signed-off-by: zhongboz <zhongboz@nvidia.com> Signed-off-by: Zhongbo Zhu <zhongboz@nvidia.com> Signed-off-by: Siddharth Singh <sidsingh@nvidia.com> Signed-off-by: shanmugamr1992 <shanmugamr@nvidia.com> Signed-off-by: Keshav Santhanam <ksanthanam@nvidia.com> Signed-off-by: Shijie Wang <jaywan@nvidia.com> Signed-off-by: [Your Name] <your.email@example.com> Signed-off-by: svcnemo-autobot <svcnemo-autobot@nvidia.com> Signed-off-by: Jon Barker <jbarker@aws-dfw-cs-001-login-01.cm.cluster> Signed-off-by: wdykas <wdykas@nvidia.com> Signed-off-by: Chen Cui <chcui@nvidia.com> Signed-off-by: janEbert <janpabloe@nvidia.com> Signed-off-by: oliver könig <okoenig@nvidia.com> Signed-off-by: HaochenYuan <haocheny@nvidia.com> Signed-off-by: Hollow Man <hollowman@opensuse.org> Signed-off-by: William Dykas <wdykas@nvidia.com> Signed-off-by: Cory Ye <cye@nvidia.com> Signed-off-by: Laura Dang <laurad@nvidia.com> Signed-off-by: Shiqing Fan <shiqingf@nvidia.com> Signed-off-by: Deepak Narayanan <deepakn94@gmail.com> Signed-off-by: Jiangfei Duan <jiangfeid@nvidia.com> Signed-off-by: Philip Monk <pmonk@nvidia.com> Signed-off-by: Jimmy Zhang <jiemingz@nvidia.com> Signed-off-by: Xuanteng Huang <xuantengh@nvidia.com> Signed-off-by: guihong-nv <guihongl@nvidia.com> Signed-off-by: asolergi-nv <asolergibert@nvidia.com> Signed-off-by: Antoni-Joan Solergibert <asolergibert@nvidia.com> Signed-off-by: ilml <tolong@nvidia.com> Signed-off-by: Tom Long <tolong@nvidia.com> Signed-off-by: Deepak Narayanan <dnarayanan@nvidia.com> Signed-off-by: YangFei1990 <feiw@nvidia.com> Signed-off-by: Siddhartha Raman Sundara Raman <sraman@nvidia.com> Signed-off-by: kunlunl <kunlunl@nvidia.com> Signed-off-by: Lawrence McAfee <lmcafee@nvidia.com> Signed-off-by: Björn Buschkämper <bjoern.buschkaemper@gmail.com> Signed-off-by: Chaitra Sathe <csathe@nvidia.com> Signed-off-by: Will Dykas <wdykas@nvidia.com> Signed-off-by: shanmugamr1992 <shanmugamr1992@gmail.com> Signed-off-by: Asha Anoosheh <aanoosheh@nvidia.com> Signed-off-by: Yan Bai <bayan@nvidia.com> Co-authored-by: Ajay <abalasa@nvidia.com> Co-authored-by: Evgenii Zheltonozhskii <zheltonozhskiy@gmail.com> Co-authored-by: Maanu Grover <maanug@nvidia.com> Co-authored-by: claude[bot] <209825114+claude[bot]@users.noreply.github.com> Co-authored-by: Yongqiang Wang <yongqiang.seagull@gmail.com> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Co-authored-by: achyuthan.s <113010327+Achyuthan-S@users.noreply.github.com> Co-authored-by: svcnvidia-nemo-ci <svcnvidia-nemo-ci@nvidia.com> Co-authored-by: Jingyue Wu <jingyuew@nvidia.com> Co-authored-by: Philip Petrakian <ppetrakian@nvidia.com> Co-authored-by: Teodor-Dumitru Ene <34819528+tdene@users.noreply.github.com> Co-authored-by: Ali Arda Eker <aeker1@binghamton.edu> Co-authored-by: Charlie Truong <chtruong@nvidia.com> Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com> Co-authored-by: Yashaswi Karnati <144376261+yashaswikarnati@users.noreply.github.com> Co-authored-by: wdykas <73254672+wdykas@users.noreply.github.com> Co-authored-by: William Dykas <wdykas@oci-hsg-cs-001-vscode-03.cm.cluster> Co-authored-by: William Dykas <wdykas@oci-nrt-cs-001-vscode-01.cm.cluster> Co-authored-by: Zhongbo Zhu <42691305+zhongbozhu@users.noreply.github.com> Co-authored-by: Siddharth Singh <sidsingh@nvidia.com> Co-authored-by: Shanmugam Ramasamy <111910568+shanmugamr1992@users.noreply.github.com> Co-authored-by: Keshav Santhanam <ksanthanam@nvidia.com> Co-authored-by: Shijie <jaywan@nvidia.com> Co-authored-by: [Co-author Name] <coauthor.email@example.com> Co-authored-by: Jon Barker <jbarker@nvidia.com> Co-authored-by: Jon Barker <jbarker@aws-dfw-cs-001-login-01.cm.cluster> Co-authored-by: Chen Cui <chcui@nvidia.com> Co-authored-by: janEbert <janpabloe@nvidia.com> Co-authored-by: oliver könig <okoenig@nvidia.com> Co-authored-by: HaochenYuan <106647990+HaochenYuan@users.noreply.github.com> Co-authored-by: ℍ𝕠𝕝𝕝𝕠𝕨 𝕄𝕒𝕟 <hollowman@opensuse.org> Co-authored-by: Cory Ye <44509866+cspades@users.noreply.github.com> Co-authored-by: Laura Dang <laurad@nvidia.com> Co-authored-by: Shiqing Fan <shiqingf@nvidia.com> Co-authored-by: Jieming Zhang <jiemingz@nvidia.com> Co-authored-by: Deepak Narayanan <deepakn94@gmail.com> Co-authored-by: Jiangfei Duan <jiangfeid@nvidia.com> Co-authored-by: Philip Monk <phil@pcmonk.me> Co-authored-by: Jimmy Zhang <133159885+jiemingz@users.noreply.github.com> Co-authored-by: Xuanteng Huang <44627253+xuantengh@users.noreply.github.com> Co-authored-by: Guihong Li <guihongl@nvidia.com> Co-authored-by: Antoni-Joan Solergibert <asolergibert@nvidia.com> Co-authored-by: Tom Long <tolong@nvidia.com> Co-authored-by: Deepak Narayanan <dnarayanan@nvidia.com> Co-authored-by: Fei Wu <33940270+YangFei1990@users.noreply.github.com> Co-authored-by: Xin Yao <xiny@nvidia.com> Co-authored-by: Siddhartha Raman Sundara Raman <sraman@nvidia.com> Co-authored-by: kunlunl <kunlunl@nvidia.com> Co-authored-by: Lawrence McAfee <85179052+lmcafee-nvidia@users.noreply.github.com> Co-authored-by: Jorge Albericio <jalbericiola@nvidia.com> Co-authored-by: Teodor-Dumitru Ene <teodord.ene@gmail.com> Co-authored-by: Björn Buschkämper <bjoern.buschkaemper@gmail.com> Co-authored-by: Dmytro Pykhtar <37850217+dimapihtar@users.noreply.github.com> Co-authored-by: nvcsathe <csathe@nvidia.com> Co-authored-by: Will Dykas <wdykas@nvidia.com> Co-authored-by: Philip Monk <pmonk@nvidia.com> Co-authored-by: shanmugamr1992 <shanmugamr1992@gmail.com> Co-authored-by: Asha Anoosheh <aanoosheh@nvidia.com> Co-authored-by: Yan Bai <bayan@nvidia.com> * revert(sync): revert AUT-2153 sync exact 100 main commits into dev (NVIDIA#7094) (NVIDIA#7283) Signed-off-by: oliver könig <okoenig@nvidia.com> * [dev] preserve mcore/TE parameter attributes on MOK-aliased expert weights across Module._apply (NVIDIA#7289) Signed-off-by: HaochenYuan <haocheny@nvidia.com> * [dev] Load-balanced context-parallel DSA indexer (NVIDIA#6058) Signed-off-by: jingqiny-99 <jingqiny@nvidia.com> Co-authored-by: Claude Fable 5 <noreply@anthropic.com> * Recompute RoPE during bwd pass for DSv4 (NVIDIA#7036) Signed-off-by: Ravi Ghadia <rghadia@nvidia.com> * [dev] Make fused CSA Top-K deterministic across no-grad and grad forwards (NVIDIA#7169) Signed-off-by: Bo Li <22713281+bobboli@users.noreply.github.com> Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com> * [dev](perf): DSA pre-attention: plan-driven scorer dispatch, host-built CP layout metadata (NVIDIA#6206) Signed-off-by: jingqiny-99 <jingqiny@nvidia.com> * [dev] Enable mHC attention CUDA graph split for HybridStack and packed THD (NVIDIA#7385) Signed-off-by: Hongxiao Bai <hongxiaob@nvidia.com> * [PP] Skip shape exchange for fixed packed sequences (NVIDIA#7000) Signed-off-by: Gao <gdeng@nvidia.com> Co-authored-by: Claude Opus 5 <noreply@anthropic.com> * [Dev] Add pipeline-parallel support to multimodal training in multimodal_dev code path (NVIDIA#6081) Signed-off-by: Li Tao <lit@nvidia.com> * [Dev] Skip attention-mask static inputs for mask-less attention modules (NVIDIA#6797) Signed-off-by: Li Tao <lit@nvidia.com> * revert: drop the early multimodal_dev test/CI cherry-picks superseded upstream Upstream NVIDIA#6083 (a1fa3e5, "ci: guard examples/multimodal_dev with a unit-test bucket") is the final, reviewed version of four early commits carried on dev_mdp by the same author. Keeping both would leave an early and a final version of the same work in one tree. Reverted: 816ccf3 Fix mRoPE parity test stub missing new config fields 62e10ca test(examples/multimodal_dev): repair stale rope_utils config stub 0194190 test(examples/multimodal_dev): make the standalone harnesses pytest tests e0df736 ci: run the examples/multimodal_dev suite as a unit-test bucket Evidence that upstream fully absorbed them: test_thd_correctness.py is byte-identical between dev_mdp and origin/dev, and the upstream conftest.py is a strict superset (+57 lines) of ours, additionally carrying a cleanup hook for the NCCL process group and restricting the exit-code-5 remap to the --experimental pass. The upstream versions arrive with the subsequent merge of origin/dev. Signed-off-by: Li Tao <lit@nvidia.com> * revert: drop the early multimodal_dev PP support superseded by upstream NVIDIA#6081 Upstream c82b949 ("[Dev] Add pipeline-parallel support to multimodal training in multimodal_dev code path (NVIDIA#6081)", 2026-09-18) is the final, reviewed version of dev_mdp's 3f29946, by the same author. It landed upstream after the sync research was written, which is why PR_MAPPING.md still records 3f29946 as having no upstream counterpart. Upstream's version is both larger and already expressed on top of the NVIDIA#6315 HybridModel refactor: test_pp_support.py 553 lines vs our 401, plus a new test_run_script_fsdp_pp.py, conftest hooks and a data/mock.py fix. Carrying ours as well would leave two PP implementations in one tree. Two hunks could not be reverted mechanically, because later MDP commits build on them; both are resolved to keep the MDP work and drop only the PP-specific parts: - models/base.py: the MDP stack refactored the reverted region into _prepare_decoder_inputs(). Ours is kept as-is here and reconciled against upstream in the merge commit instead. - models/qwen35_vl/model.py: build_vision_encoder (MDP's replicated encoder domain) is kept; pre_process/post_process/vp_stage are dropped. The PP>1 rejection in pretrain_multimodal.py is restored by this revert and removed again by the merge, where upstream's PP support arrives. Signed-off-by: Li Tao <lit@nvidia.com> * fix(multimodal_dev): bind emit_pixels ahead of the THD/BSHD split The merge unified the two independent reasons to skip pixel materialization -- upstream's non-first pipeline stage (include_pixel_values) and MDP's owner-sharded capture window (suppress_pixels) -- into a single emit_pixels predicate, but computed it inside the packed/THD branch of pack_or_pad_batch() while also consulting it from the BSHD branch. Any BSHD collation therefore raised UnboundLocalError: cannot access local variable 'emit_pixels' where it is not associated with a value caught by examples/multimodal_dev/tests/test_thd_e2e.py:: TestPackOrPadBatchPadded::test_equal_lengths. Hoist both bindings above the layout split; both layouts consult them and neither depends on packing. Signed-off-by: Li Tao <lit@nvidia.com> * fix(mdp): give MockSFTLowLevelDataset a vocab_size for length-only draws Upstream now rejects MockSFTLowLevelDataset in "file"/"distribution" mode unless vocab_size >= 2, because that class can also synthesize token ids: ValueError: Generated mock data requires vocab_size >= 2. If tokenizer.vocab_size is unavailable, set vocab_size in the mock dataset config JSON; got None. MDP builds it for one purpose only -- to read `.sequence_lengths` and draw per-sample total token lengths -- and never materializes a token, so it had no reason to supply a vocabulary and did not. Supply a nominal vocab_size when the caller's --mdp-mock-dataset-config-json payload does not already carry one. vocab_size is not an input to generate_lognormal_samples(), so the drawn lengths, and therefore the scenario pool and every rank's view of it, are bit-identical to before. Without this the whole mdp_mock data path -- not just the tests -- fails at startup after the merge. Caught by examples/multimodal_dev/tests/test_thd_static_packing.py:: TestMockLengthDistribution (5 tests). Signed-off-by: Li Tao <lit@nvidia.com> * fix(mdp): drive the MDP proxy runner from --hybrid-layer-pattern Since the Qwen3.5-VL HybridModel migration, the model factory rejects a run that does not pass --hybrid-layer-pattern: ValueError: Qwen3.5-VL uses HybridModel and requires --hybrid-layer-pattern. Use GEGEGE*E per four MoE blocks (or G-G-G-*- for dense blocks), and append /*E or /*- for each MTP depth. run_mdp_experiments.sh is fork-only, so upstream could not update it; it still described the GDN placement with --linear-attention-freq 4, which HybridModel ignores. Every MDP proxy run therefore failed at startup after the merge. Port the expansion upstream added to run_qwen35_vl.sh: ATTN_CADENCE is one period of the attention layout ('G' GatedDeltaNet, '*' full attention), repeated and truncated to NUM_LAYERS, each block followed by its MLP symbol ('E' MoE / '-' dense), then one "/<last-attn><mlp>" per MTP depth. The default cadence "GGG*" is exactly what --linear-attention-freq 4 expressed, so the model is unchanged: the study shape (NUM_LAYERS=8, NUM_EXPERTS=8, MTP_NUM_LAYERS=1) yields GEGEGE*EGEGEGE*E/*E. GDN=0 degenerates to all-full-attention. --linear-attention-freq is dropped rather than kept alongside, since TransformerConfig skips its assertion for hybrid models and keeping it would create a second source of truth. Signed-off-by: Li Tao <lit@nvidia.com> * fix(mdp): pass --position-embedding-type mrope and --mrope-section in the proxy runner The Qwen3.5-VL factory now validates both explicitly: ValueError: Qwen3.5-VL requires --mrope-section 11 11 10, but got None. The section split is tied to kv_channels and ROTARY_PERCENT; a mismatch silently misplaces the T/H/W channel boundaries. Before the HybridModel migration, MultimodalModel hard-coded position_embedding_type="mrope" when constructing the decoder, so the proxy runner could pass --position-embedding-type rope and still get an MRoPE model. The migrated model threads the CLI value through and rejects anything but mrope, and takes the section split from --mrope-section rather than an internal constant. Pass both. 11 11 10 matches MROPE_SECTION in models/qwen35_vl/configuration.py and sums to half the rotary dimension (kv_channels * rotary_percent / 2 = 256 * 0.25 / 2 = 32), i.e. the same split the model used implicitly before, so the trained model is unchanged. Signed-off-by: Li Tao <lit@nvidia.com> * fix(mdp): let the capture window, not pipeline rank, decide who keeps pixels Upstream NVIDIA#6081 drops pixel_values off every non-first pipeline stage, on the premise that only the first stage holds the vision encoder. That premise does not hold under MDP: the encoder is a separate replicated domain, and which rank must materialize a microbatch's pixels is decided by the capture window, so a rank that is not PP stage 0 can still be the owner. The merge applied both rules at once, which broke every MDP run with PP > 1: the collate still emits image_grid_thw on every stage, so a stage that dropped the pixels but kept the grids tripped MDP's payload check MdpConfigurationError: MDP: microbatch 1 violates: pixel data and grid metadata either both exist or both are absent (items without pixels). Keep the pipeline rule for the native path and defer to MDP's own pixel_capture_suppressed() when MDP is enabled. Signed-off-by: Li Tao <lit@nvidia.com> * fix(mdp): express VPP through the hybrid layer pattern, not a CLI flag Since the Qwen3.5-VL HybridModel migration, virtual pipeline stages are derived from the '|' segments of --hybrid-layer-pattern, and passing the old flag is rejected outright: AssertionError: --num-virtual-stages-per-pipeline-rank should not be used with --hybrid-layer-pattern. Virtual pipeline stages are derived from the number of | segments in the pattern. run_mdp_experiments.sh still emitted --num-virtual-stages-per-pipeline-rank, so every VPP>1 configuration failed at argument validation -- in the proxy matrix that is the ep-overlap cell, which needs VPP>1 because EP overlap requires it when PP>1. Drop the flag and split the pattern into PP x VPP segments instead. The study shape (NUM_LAYERS=8, PP=2, VPP=2) now yields GEGE|GE*E|GEGE|GE*E/*E. VPP=1 deliberately emits no '|' at all, leaving the segmentation implicit: Megatron splits evenly, which is the same layout, so the pattern for every other cell is byte-identical to before and their results stay comparable. Signed-off-by: Li Tao <lit@nvidia.com> --------- Signed-off-by: Hristo Filaretov <hfilaretov@nvidia.com> Signed-off-by: Jingyue Wu <jingyuew@nvidia.com> Signed-off-by: Siddharth Singh <sidsingh@nvidia.com> Signed-off-by: Min Htet Myet <mattralminn@email.com> Signed-off-by: Mattral <mattralminn@gmail.com> Signed-off-by: svcnemo-autobot <svcnemo-autobot@nvidia.com> Signed-off-by: Asha Anoosheh <aanoosheh@nvidia.com> Signed-off-by: Ajay <abalasa@nvidia.com> Signed-off-by: Ajay Balasa <abalasa@nvidia.com> Signed-off-by: [Your Name] <your.email@example.com> Signed-off-by: Keshav Santhanam <ksanthanam@nvidia.com> Signed-off-by: Haoran Zhang <haoranz@nvidia.com> Signed-off-by: Jinhang Choi <jinhangc@nvidia.com> Signed-off-by: svcnvidia-nemo-ci <svcnvidia-nemo-ci@nvidia.com> Signed-off-by: Shanhao <5197744+shanhaoli@users.noreply.github.com> Signed-off-by: Yan Xu <yxu1@nvidia.com> Signed-off-by: Shanmugam Ramasamy <111910568+shanmugamr1992@users.noreply.github.com> Signed-off-by: Philip Petrakian <ppetrakian@nvidia.com> Signed-off-by: oliver könig <okoenig@nvidia.com> Signed-off-by: Markus Schnös <markus.schnoes@deepl.com> Signed-off-by: Oleg Sudakov <osudakov@nvidia.com> Signed-off-by: Oleg Sudakov <oleg.sudakov@outlook.com> Signed-off-by: John St John <jstjohn@nvidia.com> Signed-off-by: John St. John <jstjohn@nvidia.com> Signed-off-by: Yury Parfenov <4665475+warpuv@users.noreply.github.com> Signed-off-by: Yan Bai <bayan@nvidia.com> Signed-off-by: Gautham Kollu <gkollu@nvidia.com> Signed-off-by: Deepak Narayanan <dnarayanan@nvidia.com> Signed-off-by: Rui Zhu <rui.zhu.rz399@yale.edu> Signed-off-by: Lawrence McAfee <lmcafee@nvidia.com> Signed-off-by: NVIDIA NeMo Bot <nemo-bot@nvidia.com> Signed-off-by: Hongxiao Bai <hongxiaob@nvidia.com> Signed-off-by: Xin Yao <xiny@nvidia.com> Signed-off-by: xuwenc <xuwenc@nvidia.com> Signed-off-by: Charlie Truong <chtruong@nvidia.com> Signed-off-by: bvolpato <brunocvcunha@gmail.com> Signed-off-by: Yuzhong Wang <yuzhongw@nvidia.com> Signed-off-by: Yunsheng Ni <yunshengn@nvidia.com> Signed-off-by: Jingqi Yang <jingqiny@nvidia.com> Signed-off-by: jingqiny-99 <jingqiny@nvidia.com> Signed-off-by: Pingtian Li <pingtianl@nvidia.com> Signed-off-by: Tailai Ma <58548582+xiaoyao0115@users.noreply.github.com> Signed-off-by: xiaoyao0115 <1804647152@qq.com> Signed-off-by: hongbinl <hongbinl@nvidia.com> Signed-off-by: fkuner <784819644@qq.com> Signed-off-by: Fan Kun <784819644@qq.com> Signed-off-by: sophiayyya <yangjingyi0730@163.com> Signed-off-by: root <root@eos0294.eos.clusters.nvidia.com> Signed-off-by: Sophia Yang <sopyang@cw-dfw-cs-001-dc-02.cm.cluster> Signed-off-by: Deyu Fu <deyuf@nvidia.com> Signed-off-by: Sophia Yang <sopyang@cw-dfw-cs-001-vscode-01.cm.cluster> Signed-off-by: Robin Zhang <robinz@nvidia.com> Signed-off-by: janEbert <janpabloe@nvidia.com> Signed-off-by: ℍ𝕠𝕝𝕝𝕠𝕨 𝕄𝕒𝕟 <hollowman@opensuse.org> Signed-off-by: guihong-nv <guihongl@nvidia.com> Signed-off-by: Zhongbo Zhu <42691305+zhongbozhu@users.noreply.github.com> Signed-off-by: Harry Zhou <hhanyu@nvidia.com> Signed-off-by: Li Tao <lit@nvidia.com> Signed-off-by: HaochenYuan <haocheny@nvidia.com> Signed-off-by: qizhang <qizhang@nvidia.com> Signed-off-by: Gao <gdeng@nvidia.com> Signed-off-by: leikaixiang.shyoshyo <leikaixiang.shyoshyo@bytedance.com> Signed-off-by: Tailai Ma <tailaim@nvidia.com> Signed-off-by: Evgenii Zheltonozhskii <zheltonozhskiy@gmail.com> Signed-off-by: Yongqiang Wang <yongqiang.seagull@gmail.com> Signed-off-by: Achyuthan Sivasankar <achyuthan.sivasankar@gmail.com> Signed-off-by: Teodor-Dumitru Ene <teodord.ene@gmail.com> Signed-off-by: Ali Arda Eker <aeker@nvidia.com> Signed-off-by: ykarnati <ykarnati@nvidia.com> Signed-off-by: William Dykas <wdykas@oci-hsg-cs-001-vscode-03.cm.cluster> Signed-off-by: William Dykas <wdykas@oci-nrt-cs-001-vscode-01.cm.cluster> Signed-off-by: zhongboz <zhongboz@nvidia.com> Signed-off-by: Zhongbo Zhu <zhongboz@nvidia.com> Signed-off-by: shanmugamr1992 <shanmugamr@nvidia.com> Signed-off-by: Shijie Wang <jaywan@nvidia.com> Signed-off-by: Jon Barker <jbarker@aws-dfw-cs-001-login-01.cm.cluster> Signed-off-by: wdykas <wdykas@nvidia.com> Signed-off-by: Chen Cui <chcui@nvidia.com> Signed-off-by: Hollow Man <hollowman@opensuse.org> Signed-off-by: William Dykas <wdykas@nvidia.com> Signed-off-by: Cory Ye <cye@nvidia.com> Signed-off-by: Laura Dang <laurad@nvidia.com> Signed-off-by: Shiqing Fan <shiqingf@nvidia.com> Signed-off-by: Deepak Narayanan <deepakn94@gmail.com> Signed-off-by: Jiangfei Duan <jiangfeid@nvidia.com> Signed-off-by: Philip Monk <pmonk@nvidia.com> Signed-off-by: Jimmy Zhang <jiemingz@nvidia.com> Signed-off-by: Xuanteng Huang <xuantengh@nvidia.com> Signed-off-by: asolergi-nv <asolergibert@nvidia.com> Signed-off-by: Antoni-Joan Solergibert <asolergibert@nvidia.com> Signed-off-by: ilml <tolong@nvidia.com> Signed-off-by: Tom Long <tolong@nvidia.com> Signed-off-by: YangFei1990 <feiw@nvidia.com> Signed-off-by: Siddhartha Raman Sundara Raman <sraman@nvidia.com> Signed-off-by: kunlunl <kunlunl@nvidia.com> Signed-off-by: Björn Buschkämper <bjoern.buschkaemper@gmail.com> Signed-off-by: Chaitra Sathe <csathe@nvidia.com> Signed-off-by: Will Dykas <wdykas@nvidia.com> Signed-off-by: shanmugamr1992 <shanmugamr1992@gmail.com> Signed-off-by: Ravi Ghadia <rghadia@nvidia.com> Signed-off-by: Bo Li <22713281+bobboli@users.noreply.github.com> Co-authored-by: Hristo Filaretov <hello@hgf.sh> Co-authored-by: Michał Marcinkiewicz <43240942+mmarcinkiewicz@users.noreply.github.com> Co-authored-by: Jingyue Wu <jingyuew@nvidia.com> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Co-authored-by: Siddharth Singh <sidsingh@nvidia.com> Co-authored-by: Min Htet Myet <88831350+Mattral@users.noreply.github.com> Co-authored-by: svcnemo-autobot <svcnemo-autobot@nvidia.com> Co-authored-by: Asha Anoosheh <aanoosheh@nvidia.com> Co-authored-by: Ajay <abalasa@nvidia.com> Co-authored-by: Keshav Santhanam <ksanthanam@nvidia.com> Co-authored-by: Haoran Zhang <haoranz@nvidia.com> Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com> Co-authored-by: Jinhang Choi <cepiross@gmail.com> Co-authored-by: Cory Ye <44509866+cspades@users.noreply.github.com> Co-authored-by: svcnvidia-nemo-ci <svcnvidia-nemo-ci@nvidia.com> Co-authored-by: Shanhao <suith27@gmail.com> Co-authored-by: Yan Xu <45385219+Connor-XY@users.noreply.github.com> Co-authored-by: Shanmugam Ramasamy <111910568+shanmugamr1992@users.noreply.github.com> Co-authored-by: Philip Petrakian <ppetrakian@nvidia.com> Co-authored-by: oliver könig <okoenig@nvidia.com> Co-authored-by: Marks101 <46690260+Marks101@users.noreply.github.com> Co-authored-by: Markus Schnös <markus.schnoes@deepl.com> Co-authored-by: OlegSudakov <oleg.sudakov@outlook.com> Co-authored-by: Fei Wu <33940270+YangFei1990@users.noreply.github.com> Co-authored-by: John St. John <jstjohn@users.noreply.github.com> Co-authored-by: Yury Parfenov <4665475+warpuv@users.noreply.github.com> Co-authored-by: Guihong Li <guihongl@nvidia.com> Co-authored-by: Yan Bai <bayan@nvidia.com> Co-authored-by: gautham-kollu <gkollu@nvidia.com> Co-authored-by: Deepak Narayanan <dnarayanan@nvidia.com> Co-authored-by: Rui Zhu <huthvincent@gmail.com> Co-authored-by: Dmytro Pykhtar <37850217+dimapihtar@users.noreply.github.com> Co-authored-by: Lawrence McAfee <85179052+lmcafee-nvidia@users.noreply.github.com> Co-authored-by: NVIDIA NeMo Bot <nemo-bot@nvidia.com> Co-authored-by: Hongxiao Bai <hongxiaob@nvidia.com> Co-authored-by: Xin Yao <xiny@nvidia.com> Co-authored-by: xuwchen <xuwenc@nvidia.com> Co-authored-by: Charlie Truong <chtruong@nvidia.com> Co-authored-by: Bruno Volpato <gh@brunovolpato.com> Co-authored-by: Yuzhong Wang <yuzhongw@nvidia.com> Co-authored-by: claude[bot] <209825114+claude[bot]@users.noreply.github.com> Co-authored-by: Yunsheng Ni <ni-ys13@tsinghua.org.cn> Co-authored-by: Jingqin Yang <jingqiny@nvidia.com> Co-authored-by: Pingtian Li <158665726+Wohox@users.noreply.github.com> Co-authored-by: Tailai Ma <58548582+xiaoyao0115@users.noreply.github.com> Co-authored-by: CarlosGomes98 <cgomes@nvidia.com> Co-authored-by: Hongbin Liu <lhb8125@users.noreply.github.com> Co-authored-by: Fan Kun <784819644@qq.com> Co-authored-by: sophiayyya <yangjingyi0730@163.com> Co-authored-by: root <root@eos0294.eos.clusters.nvidia.com> Co-authored-by: Sophia Yang <sopyang@cw-dfw-cs-001-dc-02.cm.cluster> Co-authored-by: Deyu Fu <deyuf@nvidia.com> Co-authored-by: Sophia Yang <sopyang@cw-dfw-cs-001-vscode-01.cm.cluster> Co-authored-by: Robin Zhang <robinz@nvidia.com> Co-authored-by: janEbert <janpabloe@nvidia.com> Co-authored-by: ℍ𝕠𝕝𝕝𝕠𝕨 𝕄𝕒𝕟 <hollowman@opensuse.org> Co-authored-by: Zhongbo Zhu <42691305+zhongbozhu@users.noreply.github.com> Co-authored-by: Yunsheng Ni <yunshengn@gmail.com> Co-authored-by: Harry Zhou <67385896+harryzhou2000@users.noreply.github.com> Co-authored-by: HaochenYuan <106647990+HaochenYuan@users.noreply.github.com> Co-authored-by: hongbinl <hongbinl@nvidia.com> Co-authored-by: QiZhangNV <163500791+QiZhangNV@users.noreply.github.com> Co-authored-by: Gao Deng <160076886+gdengk@users.noreply.github.com> Co-authored-by: leikaixiang.shyoshyo <leikaixiang.shyoshyo@bytedance.com> Co-authored-by: Evgenii Zheltonozhskii <zheltonozhskiy@gmail.com> Co-authored-by: Maanu Grover <maanug@nvidia.com> Co-authored-by: Yongqiang Wang <yongqiang.seagull@gmail.com> Co-authored-by: achyuthan.s <113010327+Achyuthan-S@users.noreply.github.com> Co-authored-by: Teodor-Dumitru Ene <34819528+tdene@users.noreply.github.com> Co-authored-by: Ali Arda Eker <aeker1@binghamton.edu> Co-authored-by: Yashaswi Karnati <144376261+yashaswikarnati@users.noreply.github.com> Co-authored-by: wdykas <73254672+wdykas@users.noreply.github.com> Co-authored-by: William Dykas <wdykas@oci-hsg-cs-001-vscode-03.cm.cluster> Co-authored-by: William Dykas <wdykas@oci-nrt-cs-001-vscode-01.cm.cluster> Co-authored-by: Shijie <jaywan@nvidia.com> Co-authored-by: [Co-author Name] <coauthor.email@example.com> Co-authored-by: Jon Barker <jbarker@nvidia.com> Co-authored-by: Jon Barker <jbarker@aws-dfw-cs-001-login-01.cm.cluster> Co-authored-by: Chen Cui <chcui@nvidia.com> Co-authored-by: Laura Dang <laurad@nvidia.com> Co-authored-by: Shiqing Fan <shiqingf@nvidia.com> Co-authored-by: Jieming Zhang <jiemingz@nvidia.com> Co-authored-by: Deepak Narayanan <deepakn94@gmail.com> Co-authored-by: Jiangfei Duan <jiangfeid@nvidia.com> Co-authored-by: Philip Monk <phil@pcmonk.me> Co-authored-by: Jimmy Zhang <133159885+jiemingz@users.noreply.github.com> Co-authored-by: Xuanteng Huang <44627253+xuantengh@users.noreply.github.com> Co-authored-by: Antoni-Joan Solergibert <asolergibert@nvidia.com> Co-authored-by: Tom Long <tolong@nvidia.com> Co-authored-by: Siddhartha Raman Sundara Raman <sraman@nvidia.com> Co-authored-by: kunlunl <kunlunl@nvidia.com> Co-authored-by: Jorge Albericio <jalbericiola@nvidia.com> Co-authored-by: Teodor-Dumitru Ene <teodord.ene@gmail.com> Co-authored-by: Björn Buschkämper <bjoern.buschkaemper@gmail.com> Co-authored-by: nvcsathe <csathe@nvidia.com> Co-authored-by: Will Dykas <wdykas@nvidia.com> Co-authored-by: Philip Monk <pmonk@nvidia.com> Co-authored-by: shanmugamr1992 <shanmugamr1992@gmail.com> Co-authored-by: Ravi Ghadia <40660742+ghadiaravi13@users.noreply.github.com> Co-authored-by: Bo Li <22713281+bobboli@users.noreply.github.com>
This branch was previously deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Validation