feat: Enable DPO training with HuggingFace inline provider (#2825)
Some checks failed
Integration Tests / discover-tests (push) Has been skipped
Integration Auth Tests / test-matrix (oauth2_token) (push) Failing after 7s
Integration Tests / record-tests (push) Has been skipped
Integration Tests / run-tests (push) Has been skipped
Vector IO Integration Tests / test-matrix (3.12, inline::milvus) (push) Failing after 22s
Python Package Build Test / build (3.13) (push) Failing after 16s
Test Llama Stack Build / generate-matrix (push) Successful in 19s
Vector IO Integration Tests / test-matrix (3.13, inline::milvus) (push) Failing after 21s
Vector IO Integration Tests / test-matrix (3.12, inline::sqlite-vec) (push) Failing after 31s
Test Llama Stack Build / build-custom-container-distribution (push) Failing after 32s
Test External API and Providers / test-external (venv) (push) Failing after 32s
Vector IO Integration Tests / test-matrix (3.13, remote::chromadb) (push) Failing after 36s
Vector IO Integration Tests / test-matrix (3.12, remote::chromadb) (push) Failing after 39s
Update ReadTheDocs / update-readthedocs (push) Failing after 31s
SqlStore Integration Tests / test-postgres (3.13) (push) Failing after 42s
Test Llama Stack Build / build-single-provider (push) Failing after 37s
Test External Providers Installed via Module / test-external-providers-from-module (venv) (push) Failing after 35s
Test Llama Stack Build / build-ubi9-container-distribution (push) Failing after 37s
Vector IO Integration Tests / test-matrix (3.13, remote::pgvector) (push) Failing after 40s
Vector IO Integration Tests / test-matrix (3.12, remote::pgvector) (push) Failing after 42s
Unit Tests / unit-tests (3.12) (push) Failing after 36s
Vector IO Integration Tests / test-matrix (3.13, inline::sqlite-vec) (push) Failing after 40s
SqlStore Integration Tests / test-postgres (3.12) (push) Failing after 45s
Test Llama Stack Build / build (push) Failing after 6s
Python Package Build Test / build (3.12) (push) Failing after 1m1s
Unit Tests / unit-tests (3.13) (push) Failing after 1m0s
Vector IO Integration Tests / test-matrix (3.13, inline::faiss) (push) Failing after 1m6s
Vector IO Integration Tests / test-matrix (3.12, inline::faiss) (push) Failing after 1m8s
Pre-commit / pre-commit (push) Successful in 1m50s

What does this PR do?

This PR adds support for Direct Preference Optimization (DPO) training
via the existing HuggingFace inline provider. It introduces a new DPO
training recipe, config schema updates, dataset integration, and
end-to-end testing to support preference-based fine-tuning with TRL.

Test Plan

Added integration test:

tests/integration/post_training/test_post_training.py::TestPostTraining::test_preference_optimize

Ran tests on both CPU and CUDA environments

---------

Co-authored-by: Ubuntu <ubuntu@ip-172-31-43-83.ec2.internal>
Co-authored-by: Ashwin Bharambe <ashwin.bharambe@gmail.com>
This commit is contained in:
Nehanth Narendrula 2025-07-31 02:33:36 -04:00 committed by GitHub
parent 2665f00102
commit cf73146132
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
7 changed files with 913 additions and 215 deletions

View file

@ -24,6 +24,10 @@ HuggingFace-based post-training provider for fine-tuning models using the Huggin
| `weight_decay` | `<class 'float'>` | No | 0.01 | |
| `dataloader_num_workers` | `<class 'int'>` | No | 4 | |
| `dataloader_pin_memory` | `<class 'bool'>` | No | True | |
| `dpo_beta` | `<class 'float'>` | No | 0.1 | |
| `use_reference_model` | `<class 'bool'>` | No | True | |
| `dpo_loss_type` | `Literal['sigmoid', 'hinge', 'ipo', 'kto_pair'` | No | sigmoid | |
| `dpo_output_dir` | `<class 'str'>` | No | ./checkpoints/dpo | |
## Sample Configuration