bridge.recipes.diffusion_gemma.diffusion_gemma#

Training recipe for the public DiffusionGemma 26B-A4B checkpoint.

Module Contents#

Functions#

diffusion_gemma_26b_sft_config

Return the 8-GPU EP=8 block-diffusion SFT configuration.

Data#

_HF_PATH

API#

bridge.recipes.diffusion_gemma.diffusion_gemma._HF_PATH#

‘google/diffusiongemma-26B-A4B-it’

bridge.recipes.diffusion_gemma.diffusion_gemma.diffusion_gemma_26b_sft_config(
*,
train_path: str | None = None,
validation_path: str | None = None,
test_path: str | None = None,
hf_model_path: str = _HF_PATH,
) → megatron.bridge.training.config.ConfigContainer#

Return the 8-GPU EP=8 block-diffusion SFT configuration.

Pass train_path to train on JSONL split files. Without it, the recipe uses synthetic data so it can be constructed for smoke tests and recipe discovery.