feat: Support Nemotron-nano-v3 Omni AutoModel Path by yuekaizhang · Pull Request #2362 · NVIDIA-NeMo/RL

yuekaizhang · 2026-04-29T10:25:44Z

This PR follows the nano-v3-omni mbridge training branch to add AutoModel backend support for Nemotron-Nano-Omni.

Signed-off-by: Yuekai Zhang <yuekaiz@cw-dfw-cs-001-vscode-02.cm.cluster> Signed-off-by: Yuekai Zhang <zhangyuekai@foxmail.com>

Signed-off-by: Yuekai Zhang <zhangyuekai@foxmail.com>

Add .claude/settings.local.json, .codex, and .humanize/ to .gitignore as these are local tool configuration/cache files that should not be tracked. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> Signed-off-by: Yuekai Zhang <zhangyuekai@foxmail.com>

- Add MMPRTinyDataset class with HF download fallback and local cache support - Add format_mmpr_tiny_dataset for OpenAI-API message conversion - Port verl_geo3k reward function from old Megatron-Bridge implementation - Register mmpr-tiny in DATASET_REGISTRY and vlm_hf_data_processor - Register verl_geo3k reward in VLMVerifyWorker - Add mathruler dependency to pyproject.toml for answer grading - Create debug-friendly YAML config with step_400 checkpoint - Create launch script with uv sync pre-step - Create CoT prompt file with \boxed{} instruction Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> Signed-off-by: Yuekai Zhang <zhangyuekai@foxmail.com>

- Remove self.preprocessor from MMPRTinyDataset to prevent double formatting (vlm_hf_data_processor already handles format dispatch) - Add pylatexenc to pyproject.toml as transitive dependency of mathruler (mathruler does not declare it in its own metadata) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> Signed-off-by: Yuekai Zhang <zhangyuekai@foxmail.com>

- Fix prompt file brace escaping: \boxed{} -> \boxed{{}} so str.format() has exactly one replacement placeholder (fixes AC-2/AC-5) - Add explicit split parameter validation to MMPRTinyDataset with ValueError for unsupported splits (fixes AC-1 negative test) - Regenerate uv.lock with mathruler and pylatexenc entries (fixes AC-6) - Add unit tests for dataset formatting, split validation, prompt file format compatibility, and verl_geo3k_reward (14 tests, all pass) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> Signed-off-by: Yuekai Zhang <zhangyuekai@foxmail.com>

Signed-off-by: Yuekai Zhang <zhangyuekai@foxmail.com>

Add 3 processor smoke tests using a stub NemotronNanoVLV2Processor: - test_processor_produces_valid_datum_spec: verifies DatumSpec fields - test_prompted_text_contains_boxed_literal: verifies \boxed{} survives - test_placeholder_conversion_for_nemotron_processor: verifies <image> placeholder and question text in vllm_content Uses a tiny 1x1 PNG fixture for image resolution. All 17 MMPR tests pass (11 dataset + 6 reward). Signed-off-by: Yuekai Zhang <zhangyuekai@foxmail.com>

- Stub now captures the exact text arg passed to __call__ - Assert exact equality: vllm_content == "<image>\n" + prompted_question - Assert exactly one <image> token in output (no duplicates) - Negative assertion: raw dataset string "<image>\nQuestion" not in output - Assert captured __call__ text matches expected tokenizer input - All tests run with -p no:testmon --override-ini='addopts=' Signed-off-by: Yuekai Zhang <zhangyuekai@foxmail.com>

Clean up partial state (stale images_dir, parquet, temp dir) before re-downloading when the ready marker is absent. Prevents shutil.move from nesting images/images when images_dir already exists from an interrupted prior attempt. Signed-off-by: Yuekai Zhang <zhangyuekai@foxmail.com>

MMPR-Tiny has 4,192 rows with multiple images (up to 11). The previous code truncated to images[0], losing visual context for multi-image questions. - _load_mmpr_tiny_from_cache: keep all image paths instead of [imgs[0]] - format_mmpr_tiny_dataset: split question on <image> and <image_N> placeholders, interleave image content items with text segments - Add tests for multi-image and numbered-placeholder formatting Signed-off-by: Yuekai Zhang <zhangyuekai@foxmail.com>

Signed-off-by: root <zhangyuekai@foxmail.com> Signed-off-by: Yuekai Zhang <zhangyuekai@foxmail.com>

Signed-off-by: Yuekai Zhang <zhangyuekai@foxmail.com>

…nting Signed-off-by: Yuekai Zhang <zhangyuekai@foxmail.com>

…sound_projection to resume training Signed-off-by: Yuekai Zhang <zhangyuekai@foxmail.com>

Signed-off-by: Yuekai Zhang <zhangyuekai@foxmail.com>

yuekaizhang · 2026-06-15T10:16:23Z

/ok to test 4177252

github-actions · 2026-06-15T10:17:11Z

❌ Submodule Fast-Forward Check Failed

Check based on commit: 4177252 (PR #2362 from nemotron)

❌ Submodules that need attention:

Automodel: ❌ Commits have DIVERGED from a common ancestor
TARGET (main branch): https://github.com/NVIDIA-NeMo/Automodel/commits/6de0c3612156c1d882b6e951af9b0f5d9af29d4a/
CURRENT (PR #2362 from nemotron): https://github.com/NVIDIA-NeMo/Automodel/commits/24b47e856263d313b942f0ed666c63fff83306b4/

Please ensure all submodule commits are fast-forwards of the main branch before merging.

yuekaizhang · 2026-06-15T13:32:04Z

@jinglinglingling Hi Linglin, could you please review the PR again? Thanks!

Signed-off-by: Yuekai Zhang <zhangyuekai@foxmail.com>

github-actions · 2026-06-16T05:29:33Z

❌ Submodule Fast-Forward Check Failed

Check based on commit: 2a49359 (PR #2362 from nemotron)

❌ Submodules that need attention:

Automodel: ❌ Commits have DIVERGED from a common ancestor
TARGET (main branch): https://github.com/NVIDIA-NeMo/Automodel/commits/6de0c3612156c1d882b6e951af9b0f5d9af29d4a/
CURRENT (PR #2362 from nemotron): https://github.com/NVIDIA-NeMo/Automodel/commits/24b47e856263d313b942f0ed666c63fff83306b4/

Please ensure all submodule commits are fast-forwards of the main branch before merging.

Resolve the unresolved merge conflict between the nemotron (1970h) and main (2000h) branches by setting the nightly compute budget to 2070 hours. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Signed-off-by: Yuekai Zhang <zhangyuekai@foxmail.com>

yuekaizhang · 2026-06-16T05:38:52Z

/ok to test 86c1c1f

github-actions · 2026-06-16T05:38:59Z

❌ Submodule Fast-Forward Check Failed

Check based on commit: 86c1c1f (PR #2362 from nemotron)

❌ Submodules that need attention:

Automodel: ❌ Commits have DIVERGED from a common ancestor
TARGET (main branch): https://github.com/NVIDIA-NeMo/Automodel/commits/6de0c3612156c1d882b6e951af9b0f5d9af29d4a/
CURRENT (PR #2362 from nemotron): https://github.com/NVIDIA-NeMo/Automodel/commits/24b47e856263d313b942f0ed666c63fff83306b4/

Please ensure all submodule commits are fast-forwards of the main branch before merging.

yuki-97

thanks @yuekaizhang , overall LGTM and left some minor comments.

Rename the reward function and its "verl_geo3k" config registration name to "geo3k", and note in the docstring that it has the same behavior as verl's geo3k reward function. Updates the function definition, the VLM environment dispatch/import, the unit test (file + call sites), the exemplar/recipe YAMLs, and the docs. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Signed-off-by: root <zhangyuekai@foxmail.com>

github-actions · 2026-06-16T09:28:33Z

❌ Submodule Fast-Forward Check Failed

Check based on commit: 15cc685 (PR #2362 from nemotron)

❌ Submodules that need attention:

Automodel: ❌ Commits have DIVERGED from a common ancestor
TARGET (main branch): https://github.com/NVIDIA-NeMo/Automodel/commits/6de0c3612156c1d882b6e951af9b0f5d9af29d4a/
CURRENT (PR #2362 from nemotron): https://github.com/NVIDIA-NeMo/Automodel/commits/24b47e856263d313b942f0ed666c63fff83306b4/

Please ensure all submodule commits are fast-forwards of the main branch before merging.

yuekaizhang · 2026-06-16T09:46:17Z

/ok to test 15cc685

yuekaizhang · 2026-06-16T10:27:25Z

/ok to test c4f6247

github-actions · 2026-06-16T10:27:39Z

❌ Submodule Fast-Forward Check Failed

Check based on commit: c4f6247 (PR #2362 from nemotron)

❌ Submodules that need attention:

Automodel: ❌ Commits have DIVERGED from a common ancestor
TARGET (main branch): https://github.com/NVIDIA-NeMo/Automodel/commits/6de0c3612156c1d882b6e951af9b0f5d9af29d4a/
CURRENT (PR #2362 from nemotron): https://github.com/NVIDIA-NeMo/Automodel/commits/24b47e856263d313b942f0ed666c63fff83306b4/

Please ensure all submodule commits are fast-forwards of the main branch before merging.

yuekaizhang · 2026-06-16T10:38:38Z

/ok to test 04c62f0

github-actions · 2026-06-16T10:39:17Z

❌ Submodule Fast-Forward Check Failed

Check based on commit: 04c62f0 (PR #2362 from nemotron)

❌ Submodules that need attention:

Automodel: ❌ Commits have DIVERGED from a common ancestor
TARGET (main branch): https://github.com/NVIDIA-NeMo/Automodel/commits/6de0c3612156c1d882b6e951af9b0f5d9af29d4a/
CURRENT (PR #2362 from nemotron): https://github.com/NVIDIA-NeMo/Automodel/commits/24b47e856263d313b942f0ed666c63fff83306b4/

Please ensure all submodule commits are fast-forwards of the main branch before merging.

…y recipes Move and rename the Nemotron-Omni VLM GRPO configs into examples/configs/recipes/vlm/ following the VLM recipe naming convention, and run both as functional nightly tests: - vlm_grpo-nemotron-omni-30ba3b-clevr-1n8g-automodel-ep8.v1 (CLEVR-CoGenT, 1 node) - vlm_grpo-nemotron-omni-30ba3b-mmpr-4n8g-automodel-ep8.v1 (MMPR-Tiny, 4 nodes) Both configs are minimized snapshots that inherit from vlm_grpo_3B.yaml; the MMPR config carries cluster.num_nodes=4. The previous thin wrapper recipe (same mmpr name) is replaced by the full config under that name, and the nightly now points at these recipes directly via matching driver scripts. Nightly GPU-hours total stays at 2041h, under the 2070h budget. Update the Nemotron-Omni guide to the new config paths. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Signed-off-by: root <zhangyuekai@foxmail.com>

github-actions · 2026-06-16T12:21:26Z

❌ Submodule Fast-Forward Check Failed

Check based on commit: bc85d5f (PR #2362 from nemotron)

❌ Submodules that need attention:

Automodel: ❌ Commits have DIVERGED from a common ancestor
TARGET (main branch): https://github.com/NVIDIA-NeMo/Automodel/commits/6de0c3612156c1d882b6e951af9b0f5d9af29d4a/
CURRENT (PR #2362 from nemotron): https://github.com/NVIDIA-NeMo/Automodel/commits/24b47e856263d313b942f0ed666c63fff83306b4/

Please ensure all submodule commits are fast-forwards of the main branch before merging.

yuekaizhang · 2026-06-16T12:21:33Z

/ok to test bc85d5f

terrykong

PR #2362 — feat: Support Nemotron-nano-v3 Omni AutoModel Path

Nice work — this adds Nemotron-Nano-Omni 30B-A3B AutoModel recipes (CLEVR-CoGenT + MMPR-Tiny), the MMPR-Tiny loader, a geo3k reward, docs, and nightly tests. Reviewed by a team of agents (RL/config, Automodel SME, tests, bug-finder) plus an adversarial pass.

No blocking bugs found. Every bug candidate was verified and refuted: the geo3k brace parser is correct (validated with nested/multi-last/no-close/empty cases), the MMPR download-cache lock is safe, the placeholder <image> token count matches len(images), and AutoConfig.from_pretrained in ModelFlag.matches() runs only at worker init (not a hot path). Convergence curves vs. Megatron-Bridge are present in the comments, and the nightly GPU-hour bump (2000→2070) is justified (actual total ≈2041h).

The inline comments are all suggestions/nits — the most substantive is migrating off the deprecated enable_deepep knob (Automodel upstream did the same for its own Nemotron-Omni recipes in this bump). Existing unit tests all pass locally (geo3k 6/6, mmpr_tiny 10/10).

Generated by Claude Code

- Migrate off deprecated enable_deepep knob to dispatcher=deepep + experts=torch_mm in both CLEVR and MMPR recipes (avoids grouped_gemm dep) - Remove leftover NRL_DEBUG_PICKLE debug env var from MMPR recipe - Fix docs max_total_sequence_length 10000 -> 8192 to match MMPR recipe - Update mmpr_tiny.py copyright year to 2026 - Add geo3k tests for last-wins multi-boxed and nested/unbalanced braces - Declare experts/dispatcher and freeze_config in Automodel TypedDicts Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Signed-off-by: root <zhangyuekai@foxmail.com>

# Conflicts: # tests/unit/test_recipes_and_test_suites.py

github-actions · 2026-06-17T14:15:31Z

❌ Submodule Fast-Forward Check Failed

Check based on commit: a3ca738 (PR #2362 from nemotron)

❌ Submodules that need attention:

Automodel: ❌ Commits have DIVERGED from a common ancestor
TARGET (main branch): https://github.com/NVIDIA-NeMo/Automodel/commits/6de0c3612156c1d882b6e951af9b0f5d9af29d4a/
CURRENT (PR #2362 from nemotron): https://github.com/NVIDIA-NeMo/Automodel/commits/24b47e856263d313b942f0ed666c63fff83306b4/

Please ensure all submodule commits are fast-forwards of the main branch before merging.

yuekaizhang · 2026-06-17T14:16:46Z

/ok to test a3ca738

qiaochuz-nv · 2026-06-30T05:50:52Z

Doc cross-check for docs/guides/nemotron-3-nano-omni.md: I verified the stated recipe/config details against the resolved PR recipes, including inheritance from examples/configs/vlm_grpo_3B.yaml. The doc looks technically accurate.

Two non-blocking doc nits from the POR run:

MMPR-Tiny data prep: the doc says MMPR-Tiny is auto-downloaded from HF / no manual prep needed. That is true with network access, but the standard air-gapped/offline test environment sets HF_HUB_OFFLINE=1, so OpenGVLab/MMPR-Tiny (notably images.zip) must be pre-staged in the HF dataset cache. Consider adding a note for offline clusters.
MMPR launch snippet: the snippet uses a placeholder CONTAINER and injects PYTHONPATH=...:/path/to/automodel-omni, which reads like the Omni code may not yet be part of the base container. If that is temporary/intended, no issue; otherwise it may be worth clarifying the expected container/code path.

Neither is a blocker for the PR.

qiaochuz-nv · 2026-06-30T16:59:26Z

POR Test Summary — Nemotron-3-Nano-Omni-30B VLM GRPO (PR #2362)

PR under test: NVIDIA-NeMo/RL #2362 — Nemotron-Nano-v3 Omni AutoModel VLM GRPO support
MR: https://gitlab-master.nvidia.com/dl/DLQASH/nmfw_tests/-/merge_requests/270
Branch: qiaochuz/por_2362_nano_omni_vlm → main
Commit: e786d6b (3 files, 152 insertions — test logic only; local infra reverted)
Status: ✅ Both testcases REAL GREEN

Model & Backend Under Test

Item	Value
Model	`nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16`
Backend	AutoModel
Parallelism	Expert Parallel = 8 / vLLM Tensor Parallel = 8 (triton MoE backend)
Towers	Vision + audio frozen; LM trained
Max sequence length	8192

What the Tests Do

Both drivers are a 1:1 mirror of the PR's own oracle scripts
(tests/test_suites/vlm/vlm_grpo-nemotron-omni-30ba3b-*.v1.sh):

Entrypoint: examples/run_vlm_grpo.py --config <recipe>
Step shrink: grpo.max_num_steps=10 (CI-sized; full run is longer)
Logging: wandb off, tensorboard on, GPU monitor on, checkpointing on
Assertion (gated): dump TB logs → metrics.json; only if train/loss reached
step 10, run check_metrics.py 'max(data["train/reward"]) > 0.5'

Function Tests

1. CLEVR — function tier (`testcases/pipelines/`)

test_vlm_grpo_nemotron_omni_30ba3b_clevr_1n8g_automodel_ep8.sh

Field	Value
Task	CLEVR-CoGenT
Scale	1 node × 8 GPU
Execution	Interactive
Timeout	7200 s
Recipe	`vlm_grpo-nemotron-omni-30ba3b-clevr-1n8g-automodel-ep8.v1.yaml`

Result: REAL GREEN — run 5532636, 10 steps, reward 0.83 (>0.5), val acc 0.713

2. MMPR-tiny — e2e tier (`testcases/e2e/`)

test_vlm_grpo_nemotron_omni_30ba3b_mmpr_4n8g_automodel_ep8.sh

Field	Value
Task	MMPR-tiny (geo3k / MMPR reward, `vlm_environment` path)
Scale	4 nodes × 8 GPU = 32 GPU
Execution	sbatch via `ray.sub` (batched-job model)
Recipe	`vlm_grpo-nemotron-omni-30ba3b-mmpr-4n8g-automodel-ep8.v1.yaml`

Result: REAL GREEN (verified ×2) — run 5533255, SLURM COMPLETED 0:0,
runner failures=0, reward 0.799 (>0.5), step 10/10, val acc 0.664

Coverage

Covers both VLM GRPO recipes that PR #2362 introduces, across two scales and two execution models:

Single-node interactive (CLEVR) + multi-node batched ray.sub (MMPR) — exercises the batched-job submission path the PR adds.
New mmpr-tiny dataset + geo3k / MMPR reward function + vlm_environment path.
AutoModel VLM with frozen vision/audio towers, MoE EP=8 / TP=8 triton backend.

Tags: codecoverage, por_v06, current_por, rc_p0

Doc Check — `docs/guides/nemotron-3-nano-omni.md`

The recipes run green exactly as the PR ships them — no functional/run-breaking doc issue.
5 minor cosmetic/consistency doc nits were noted during the doc-vs-recipe re-check; no PR
comment or patch filed (per decision to leave doc fix out of scope).

Artifacts

Test	Run ID	Cluster
CLEVR	`5532636`	EOS
MMPR	`5533255`	EOS, nodes eos[0548,0558-0559,0563]

Remote logs base: /lustre/fsw/coreai_dlalgo_ci/qiaochuz/nemo_llm_nightly/test_suite/rl/logs/
Container: nemo-rl-nightly-20260629.sqsh

yuekaizhang and others added 28 commits April 29, 2026 02:55

bump to vllm 0.19 latest

9d4a9cc

Signed-off-by: Yuekai Zhang <yuekaiz@cw-dfw-cs-001-vscode-02.cm.cluster> Signed-off-by: Yuekai Zhang <zhangyuekai@foxmail.com>

close optimizer save

9c97187

Signed-off-by: Yuekai Zhang <zhangyuekai@foxmail.com>

update config multi node

5cdd202

Signed-off-by: Yuekai Zhang <zhangyuekai@foxmail.com>

add 3rd party to ignore

2571a86

Signed-off-by: Yuekai Zhang <zhangyuekai@foxmail.com>

chore: add testmondata to .gitignore

6512967

Signed-off-by: Yuekai Zhang <zhangyuekai@foxmail.com>

filter multi image examples

6e6890b

Signed-off-by: root <zhangyuekai@foxmail.com> Signed-off-by: Yuekai Zhang <zhangyuekai@foxmail.com>

update ignore

08a1a44

Signed-off-by: root <zhangyuekai@foxmail.com> Signed-off-by: Yuekai Zhang <zhangyuekai@foxmail.com>

replaced mathruler

fd107b8

Signed-off-by: Yuekai Zhang <zhangyuekai@foxmail.com>

add 32 node config

690d898

Signed-off-by: Yuekai Zhang <zhangyuekai@foxmail.com>

vllm gpu mem 0.2, illegal instruction

2a7f8a5

Signed-off-by: Yuekai Zhang <zhangyuekai@foxmail.com>

enable activation checkpointing

cc802e1

Signed-off-by: Yuekai Zhang <zhangyuekai@foxmail.com>

automodel image_flag support, dynamic resolution, activation checkpoi…

ee75c34

…nting Signed-off-by: Yuekai Zhang <zhangyuekai@foxmail.com>

Post-hoc fused_attn=True on RADIO ViT blocks, Freeze sound_encoder / …

a89262a

…sound_projection to resume training Signed-off-by: Yuekai Zhang <zhangyuekai@foxmail.com>

change 4node yaml to small batch

c1b22bd

Signed-off-by: Yuekai Zhang <zhangyuekai@foxmail.com>

update 1 node debug yamml

1d9705c

Signed-off-by: Yuekai Zhang <zhangyuekai@foxmail.com>

update vllm 0.20.0

a81b5a6

Signed-off-by: Yuekai Zhang <zhangyuekai@foxmail.com>

update config

9e08086

Signed-off-by: Yuekai Zhang <zhangyuekai@foxmail.com>

remove useless scripts

eeb93f7

Signed-off-by: Yuekai Zhang <zhangyuekai@foxmail.com>

chore: apply pre-commit auto-fixes (ruff, ruff-format, taplo)

d6adc41

Signed-off-by: Yuekai Zhang <zhangyuekai@foxmail.com>

revert: restore .gitignore to upstream version

23c4a0e

Signed-off-by: Yuekai Zhang <zhangyuekai@foxmail.com>

yuekaizhang requested review from a team as code owners April 29, 2026 10:25

Merge branch 'main' into nemotron

2a49359

Signed-off-by: Yuekai Zhang <zhangyuekai@foxmail.com>

jinglinglingling previously approved these changes Jun 16, 2026

View reviewed changes

yuki-97 reviewed Jun 16, 2026

View reviewed changes

Comment thread 3rdparty/Automodel-workspace/Automodel

Comment thread nemo_rl/environments/rewards.py Outdated

Comment thread tests/unit/rewards/test_geo3k_reward.py

yuki-97 previously approved these changes Jun 16, 2026

View reviewed changes

yuki-97 reviewed Jun 16, 2026

View reviewed changes

Comment thread examples/configs/recipes/vlm/vlm_grpo-nemotron-omni-30ba3b-mmpr-4n8g-automodel-ep8.v1.yaml

terrykong reviewed Jun 16, 2026

View reviewed changes

yuekaizhang and others added 2 commits June 17, 2026 07:08

Merge remote-tracking branch 'upstream/main' into fix_multi_image

a3ca738

# Conflicts: # tests/unit/test_recipes_and_test_suites.py

yuki-97 approved these changes Jun 17, 2026

View reviewed changes

Uh oh!

Conversation

yuekaizhang commented Apr 29, 2026

Uh oh!

yuekaizhang commented Jun 15, 2026

Uh oh!

github-actions Bot commented Jun 15, 2026

❌ Submodule Fast-Forward Check Failed

❌ Submodules that need attention:

Uh oh!

yuekaizhang commented Jun 15, 2026

Uh oh!

github-actions Bot commented Jun 16, 2026

❌ Submodule Fast-Forward Check Failed

❌ Submodules that need attention:

Uh oh!

yuekaizhang commented Jun 16, 2026

Uh oh!

github-actions Bot commented Jun 16, 2026

❌ Submodule Fast-Forward Check Failed

❌ Submodules that need attention:

Uh oh!

yuki-97 left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

github-actions Bot commented Jun 16, 2026

❌ Submodule Fast-Forward Check Failed

❌ Submodules that need attention:

Uh oh!

yuekaizhang commented Jun 16, 2026

Uh oh!

Uh oh!

yuekaizhang commented Jun 16, 2026

Uh oh!

github-actions Bot commented Jun 16, 2026

❌ Submodule Fast-Forward Check Failed

❌ Submodules that need attention:

Uh oh!

yuekaizhang commented Jun 16, 2026

Uh oh!

github-actions Bot commented Jun 16, 2026

❌ Submodule Fast-Forward Check Failed

❌ Submodules that need attention:

Uh oh!

github-actions Bot commented Jun 16, 2026

❌ Submodule Fast-Forward Check Failed

❌ Submodules that need attention:

Uh oh!

yuekaizhang commented Jun 16, 2026

Uh oh!

terrykong left a comment

Choose a reason for hiding this comment

PR #2362 — feat: Support Nemotron-nano-v3 Omni AutoModel Path

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

github-actions Bot commented Jun 17, 2026

❌ Submodule Fast-Forward Check Failed

❌ Submodules that need attention:

Uh oh!

yuekaizhang commented Jun 17, 2026

Uh oh!

qiaochuz-nv commented Jun 30, 2026

Uh oh!

qiaochuz-nv commented Jun 30, 2026

POR Test Summary — Nemotron-3-Nano-Omni-30B VLM GRPO (PR #2362)

Model & Backend Under Test

What the Tests Do

Function Tests

1. CLEVR — function tier (testcases/pipelines/)

2. MMPR-tiny — e2e tier (testcases/e2e/)

Coverage

1. CLEVR — function tier (`testcases/pipelines/`)

2. MMPR-tiny — e2e tier (`testcases/e2e/`)

Doc Check — `docs/guides/nemotron-3-nano-omni.md`