How google-deepmind/gemma works
Gemma is a JAX library for running and fine-tuning open-weight Gemma models. The primary workflow loads checkpoint parameters, constructs a model and sampler, then tokenizes a prompt (optionally with images) and generates a response; the README documents multi-turn, multimodal chat. Separate capabilities include model-family implementations, data and fine-tuning utilities, and a diffusion-based training/sampling stack. The diffusion adapter’s Sudoku and PubMedQA data and evaluation stages are represented from the tree and index; their internal wiring is not inferred where excerpts do not show it.
Main entry points: _sampler.py Text sampler · sft_model.py Diffusion fine-tuning · __init__.py Gemma API · _checkpoint.py Checkpoint loading
Components 25
Public API and inference
- Gemma API
gemma/gm/__init__.py - Checkpoint loading
gemma/gm/ckpts/_checkpoint.py - Text tokenizer
gemma/gm/text/_tokenizer.py - Text sampler
gemma/gm/text/_sampler.py - Generation loop
gemma/gm/text/_sampler_loop.py
Models and modalities
- Gemma model families
gemma/gm/nn/gemma4/_gemma4.py - Transformer modules
gemma/gm/nn/gemma4/_modules.py - Positional embeddings
gemma/gm/math/_positional_embeddings.py - Vision processing
gemma/gm/nn/gemma4/vision/_preprocessing.py - Audio model
gemma/gm/nn/gemma4/audio/_model.py
Data and fine-tuning
- Task data
gemma/gm/data/_tasks.py - Data transforms
gemma/gm/data/_functional.py - Preference losses
gemma/gm/losses/_dpo.py - LoRA adaptation
gemma/peft/_lora.py - Quantization
gemma/peft/_quantization.py
Diffusion workflows
- Diffusion sampler
gemma/diffusion/_sampler.py - Diffusion transformer
gemma/diffusion/_transformer.py - Adapter datasets
gemma/diffusion/hackable_diffusion_adapter/data/data.py - Gemma diffusion network
gemma/diffusion/hackable_diffusion_adapter/hd/hd_gemma_network.py - Autoregressive state
gemma/diffusion/hackable_diffusion_adapter/hd/hd_gemma_ar_state_handler.py - Diffusion fine-tuning
gemma/diffusion/hackable_diffusion_adapter/hd/sft_model.py - Task evaluation
gemma/diffusion/hackable_diffusion_adapter/eval/ar_eval.py
Tools and research
- Tool orchestration
gemma/gm/tools/_manager.py - T5 Gemma research
gemma/research/t5gemma/t5gemma.py
Other components
- Application user
Connections 22
- Application user → Gemma API: uses
- Gemma API → Checkpoint loading: exposes
gemma/gm/__init__.py - Gemma API → Text sampler: exposes
gemma/gm/__init__.py - Application user → Checkpoint loading: loads parameters
- Application user → Text sampler: submits prompt
- Checkpoint loading → Text sampler: supplies parameters
- Gemma model families → Text sampler: configures model
- Text sampler → Text tokenizer: tokenizes input
gemma/gm/text/_sampler.py - Text sampler → Generation loop: runs generation
gemma/gm/text/_sampler.py - Text sampler → Vision processing: accepts images
- Transformer modules → Positional embeddings: uses embeddings
gemma/gm/nn/gemma4/_modules.py - Task data → Data transforms: uses transforms
gemma/gm/data/_tasks.py - Task data → Text tokenizer: uses tokenizer
gemma/gm/data/_tasks.py - Application user → LoRA adaptation: fine-tunes with
- Application user → Quantization: uses quantization
- Gemma diffusion network → Diffusion transformer: uses model
gemma/diffusion/hackable_diffusion_adapter/hd/hd_gemma_network.py - Autoregressive state → Gemma diffusion network: calls encoder
gemma/diffusion/hackable_diffusion_adapter/hd/hd_gemma_ar_state_handler.py - Diffusion fine-tuning → Autoregressive state: uses state handler
gemma/diffusion/hackable_diffusion_adapter/hd/sft_model.py - Diffusion fine-tuning → Gemma diffusion network: uses network
gemma/diffusion/hackable_diffusion_adapter/hd/sft_model.py - Diffusion fine-tuning → Task evaluation: uses evaluator
gemma/diffusion/hackable_diffusion_adapter/hd/sft_model.py - T5 Gemma research → Transformer modules: uses modules
gemma/research/t5gemma/t5gemma.py - Tool orchestration → Application user: serves tool calls