opendatahub-io / kubeflow-sdk / 29480045503

83%

push

github

feat: Add SpeculativeDecodingTrainer with TRAIN_ONLY and DATA_ONLY mode support (#117)

* feat: add SpeculativeDecodingTrainer for TRAIN_ONLY mode

Add SpeculativeDecodingTrainer supporting TRAIN_ONLY mode for training
speculative decoding draft models (Eagle3) using the speculators library.

- Training script injection via inspect.getsource()
- In-memory progression tracking with HTTP metrics endpoint
- PVC mount deduplication for shared storage
- Support for both pvc:// URIs and direct paths for output_dir
  and hidden_states_path
- Comprehensive validation and 45 unit tests

* fix: remove unused save_path from instrumentation functions

* feat: add DATA_ONLY mode support to SpeculativeDecodingTrainer

* remove num_nodes from SpeculativeDecodingTrainer

* fix: write termination message directly on phase completion for DATA_ONLY progression tracking

* fix: add target_layer_ids support, data_path validation, sidecar PVC check, and thread-safe termination

* feat: add AutoConfig layer detection, strict PVC validation, and HF_TOKEN forwarding to sidecar

* fix: disable trust_remote_code in AutoConfig.from_pretrained

* fix: pass HF_TOKEN to AutoConfig for gated model support

* fix: inject _REGEN_SCRIPT_B64 into rendered training script

950 of 1257 new or added lines in 6 files covered. (75.58%)

10140 of 12280 relevant lines covered (82.57%)

0.83 hits per line