fix(customizer): grpo sft full serialization fix

[soluwalana]

Summary

A successful one-step GRPO all_weights run with policy_backend=automodel already writes HuggingFace safetensors (policy/weights/model/shard-*.safetensors). Publication always sent that tree to NeMo-RL's DCP converter, which requires a .metadata file and failed with FileNotFoundError from convert_dcp_to_hf. Users never got a registered model or fileset from a completed full-weight GRPO job.

After this change, Automodel full-weight checkpoints are copied (preferring model/consolidated/) and only DTensor V1 DCP trees still go through convert_dcp_to_hf.

Related Issue

NVBug 6740834. QA cross-check: DevTest T6259695 / job rl-14f91222c14d (also tracked against 6716627).

Changes

• Detect an already-HuggingFace full-weight tree (find_hf_full_weight_root) and copy it for publication: flatten .hf_metadata (config.json, tokenizer), attach policy/tokenizer when needed, strip the FSDP2 architecture prefix, and leave optimizer state behind.
• Rename a single-rank, single-file Automodel shard (shard-00001-model-00001-of-00001.safetensors) to model.safetensors so HuggingFace/from_pretrained and fileset checks accept it. Multi-rank or multi-file shard layouts are left as-is with a warning — those need Automodel consolidation, not a rename.
• Keep DCP conversion only when .metadata is present (DTensor V1). Fail with an explicit path if neither an HF tree nor DCP metadata exists.
• For Automodel all_weights GRPO, emit checkpointing.model_save_format=safetensors and save_consolidated=every so training writes model/consolidated/ when the pinned Automodel honors it. The string "every" is used instead of True/"final": NeMo-RL setattrs this past Automodel's bool→enum normalizer, and it never passes is_final_checkpoint. Omit both keys for DTensor V1 (V1 rejects model_save_format) and for LoRA.

Type of Change

• Code change (feature, bug fix, or refactor)
• Code change with documentation updates
• Documentation only
• Contributor tooling or automation
• CI, build, or test infrastructure

Quality Gates

• Tests added or updated for changed behavior
• Existing tests cover changed behavior — justification:
• Tests not applicable — justification:
• Documentation updated for user-visible behavior
• Documentation not applicable — justification: Publication bugfix only; no API, CLI, or user-facing docs change.

Verification

• Pull request title follows the repository's Conventional Commit format
• Every commit includes an appropriate Signed-off-by: trailer
• uv run pre-commit run -a passes, or any blocked checks are identified below
• Targeted tests pass, or tests are marked not applicable above
• No secrets, API keys, or credentials are included

Targeted validation:

.venv/bin/pytest services/rl/tests/test_checkpoints.py services/rl/tests/test_grpo_config.py -q

95 passed.

uv run pre-commit run -a was not run on this branch.

Verified that one step run fixed the HF export.

2026-09-08 23:22:25,692 - nmp.rl.tasks.training.backends.nemo_rl.backend - INFO - Processing created checkpoint
2026-09-08 23:22:25,693 - nmp.rl.tasks.training.backends.nemo_rl.backend - INFO - Copying HuggingFace full-weight checkpoint from /var/run/scratch/job/training/checkpoints/step_1/policy/weights/model/consolidated to /var/run/scratch/job/output_model
2026-09-08 23:22:27,326 - nmp.rl.tasks.training.backends.nemo_rl.checkpoints - INFO - Copying tokenizer from /var/run/scratch/job/training/checkpoints/step_1/policy/tokenizer to /var/run/scratch/job/output_model
2026-09-08 23:22:27,338 - nmp.rl.tasks.training.chat_templates - INFO - Applied chat template to output checkpoint
2026-09-08 23:22:27,694 - httpx - INFO - HTTP Request: PUT http://nemo-platform-api:8080/apis/jobs/v2/workspaces/default/jobs/rl-a453b2521a17/steps/grpo-training/tasks/task-d58fc6d0-a61c-4775-88b9-6dcf46ae5c9d "HTTP/1.1 200 OK"
2026-09-08 23:22:27,695 - nmp.rl.tasks.training.runner - INFO - Result written to: /var/run/scratch/job/training/rl_training_result.json

[github-actions]

Suite	Lines Covered	Line Rate	Branch Rate
Unit Tests	35562/45376	78.4%	62.9%
Integration Tests	21162/43151	49.0%	22.3%

[github-actions]

🌿 Preview your docs: https://nvidia-preview-pr-1868-nvbug-6740834-solu.docs.buildwithfern.com/nemo-platform