v1.1.2: air-gapped NeMo-RL v0.7.0 stack, multimodal SDG, validated finance tool-calling recipe - #6
Merged
Merged
Conversation
…nance tool-calling recipe Air-gapped stack on NeMo-RL v0.7.0. The trainer image is now built locally with one Gym virtual environment baked per component, so GRPO and SFT training needs no Gym source mount and resolves no packages at job time. Only sglang is still pulled as-is, and every image now ships for both amd64 and arm64. nvflow-client launcher image. A new container that submits and monitors workflows from outside the cluster, with air-gapped install scripts and tunnel-aware launching for collect_rollouts, training, and verify. Finance tool-calling recipe. The multi-turn SEC-search recipe is validated end to end on the new stack: SFT, GRPO, and evaluation all reproduce the prior baseline with no regression. Finance RL / GRPO. Multi-environment GRPO with per-environment judges and per-environment eval output directories. Gym-only stages, collect_rollouts and compute_rewards, now run on a CPU-only container instead of holding GPUs. Adds Nemotron Nano 30B and Qwen3-30B configs, and carries train and validation histograms into the consolidated W&B dashboard. Document-grounded SDG. Generalized domain interface so a new domain plugs in without forking the recipe, plus follow-up fixes. HopChain multimodal SDG recipe for open-source release. Local judge support for evaluation, so the demo recipes no longer need a hosted judge endpoint. Tooling and tests. A dependency-free rollout-trace viewer, and the unit-test suite grows from 2 modules to 16 covering rollout, verify, dataset, and JSONL utilities. Documentation accuracy pass. All markdown reconciled against the shipped behaviour, and the third-party license notice synced to the lockfile. Dependency and CVE floors raised across the images. Signed-off-by: Pritam Gundecha <pgundecha@nvidia.com>
psgundecha-nv
force-pushed
the
release/v1.1.2
branch
from
August 5, 2026 00:26
049e6df to
075830b
Compare
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
2 participants
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.
Air-gapped stack on NeMo-RL v0.7.0. The trainer image is now built locally with
one Gym virtual environment baked per component, so GRPO and SFT training needs
no Gym source mount and resolves no packages at job time. Only sglang is still
pulled as-is, and every image now ships for both amd64 and arm64.
nvflow-client launcher image. A new container that submits and monitors
workflows from outside the cluster, with air-gapped install scripts and
tunnel-aware launching for collect_rollouts, training, and verify.
Finance tool-calling recipe. The multi-turn SEC-search recipe is validated end
to end on the new stack: SFT, GRPO, and evaluation all reproduce the prior
baseline with no regression.
Finance RL / GRPO. Multi-environment GRPO with per-environment judges and
per-environment eval output directories. Gym-only stages, collect_rollouts and
compute_rewards, now run on a CPU-only container instead of holding GPUs. Adds
Nemotron Nano 30B and Qwen3-30B configs, and carries train and validation
histograms into the consolidated W&B dashboard.
Document-grounded SDG. Generalized domain interface so a new domain plugs in
without forking the recipe, plus follow-up fixes.
HopChain multimodal SDG recipe for open-source release.
Local judge support for evaluation, so the demo recipes no longer need a hosted
judge endpoint.
Tooling and tests. A dependency-free rollout-trace viewer, and the unit-test
suite grows from 2 modules to 16 covering rollout, verify, dataset, and JSONL
utilities.
Documentation accuracy pass. All markdown reconciled against the shipped
behaviour, and the third-party license notice synced to the lockfile.
Dependency and CVE floors raised across the images.