Skip to content

docs: improve lora guide and demo notebook #4277

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
copybara-service merged 1 commit into from
Jun 29, 2026
+9 −15
Merged

docs: improve lora guide and demo notebook #4277

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
8 changes: 4 additions & 4 deletions docs/guides/lora_model_bringup.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -30,12 +30,12 @@ To enable LoRA support for a new model, follow these two simple steps:

The target model architecture must already be implemented and supported as a base model in MaxText.

- The JAX/NNX model definition should be located under `src/maxtext/models/` (e.g., \[gemma3.py\](../../src/maxtext/models/gemma3.py)).
- The JAX/NNX model definition should be located under `src/maxtext/models/` (e.g., [gemma3.py](https://github.com/AI-Hypercomputer/maxtext/blob/main/src/maxtext/models/gemma3.py)).
- The model configurations must be registered and runnable for baseline pre-training or full fine-tuning.

### Step 1.2: Define Trainable LoRA Target Modules

Add a recommended target pattern for your model architecture prefix in \[src/maxtext/configs/post_train/lora_module_path.yml\](../../src/maxtext/configs/post_train/lora_module_path.yml):
Add a recommended target pattern for your model architecture prefix in [src/maxtext/configs/post_train/lora_module_path.yml](https://github.com/AI-Hypercomputer/maxtext/blob/main/src/maxtext/configs/post_train/lora_module_path.yml):

```yaml
your_model_prefix: "decoder/layers/.*(self_attention/(query|key|value|out)|mlp/(wi_0|wi_1|wo))"
Expand Down Expand Up @@ -69,7 +69,7 @@ If you want to perform decoding or run high-performance serving on your adapted
To add weight mapping for vLLM decode:

1. **Create a Weight Mapping Config**:
Create a new file in \[src/maxtext/integration/tunix/weight_mapping/\](../../src/maxtext/integration/tunix/weight_mapping/) (e.g., `your_model.py`) defining a mapping dataclass. You can refer to \[gemma3.py\](../../src/maxtext/integration/tunix/weight_mapping/gemma3.py) or \[llama3.py\](../../src/maxtext/integration/tunix/weight_mapping/llama3.py) as templates.
Create a new file in [src/maxtext/integration/tunix/weight_mapping/](https://github.com/AI-Hypercomputer/maxtext/blob/main/src/maxtext/integration/tunix/weight_mapping/) (e.g., `your_model.py`) defining a mapping dataclass. You can refer to [llama3.py](https://github.com/AI-Hypercomputer/maxtext/blob/main/src/maxtext/integration/tunix/weight_mapping/llama3.py) as a template.

Your class should specify:

Expand All @@ -78,7 +78,7 @@ To add weight mapping for vLLM decode:
- `lora_to_hf_mappings()`: Custom mapping for LoRA weights if they require different handling.

2. **Register the Mapping**:
Register your new class in \[src/maxtext/integration/tunix/weight_mapping/__init__.py\](../../src/maxtext/integration/tunix/weight_mapping/__init__.py) inside the `StandaloneVllmWeightMapping` class:
Register your new class in [src/maxtext/integration/tunix/weight_mapping/__init__.py](https://github.com/AI-Hypercomputer/maxtext/blob/main/src/maxtext/integration/tunix/weight_mapping/__init__.py) inside the `StandaloneVllmWeightMapping` class:

```python
# Inside StandaloneVllmWeightMapping
Expand Down
16 changes: 5 additions & 11 deletions src/maxtext/examples/lora_llama3_demo.ipynb
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -79,6 +79,10 @@
" # Install uv, a fast Python package installer\n",
" !pip install uv\n",
" \n",
" # Set the torch backend to CPU for uv\n",
" import os\n",
" os.environ[\"UV_TORCH_BACKEND\"] = \"cpu\"\n",
" \n",
" # Install MaxText and post-training dependencies\n",
" !uv pip install -e .[tpu-post-train] --resolution=lowest\n",
" !install_tpu_post_train_extra_deps"
Expand Down Expand Up @@ -215,16 +219,6 @@
"outputs": [],
"source": [
"if not epath.Path(MODEL_CHECKPOINT_PATH).exists():\n",
" # Install torch for the conversion script\n",
" print(\"Installing torch...\")\n",
" subprocess.run(\n",
" [\n",
" sys.executable, \"-m\", \"pip\", \"install\",\n",
" \"torch\", \"--index-url\", \"https://download.pytorch.org/whl/cpu\"\n",
" ],\n",
" check=True\n",
" )\n",
"\n",
" print(\"Converting checkpoint from HuggingFace...\")\n",
" env = os.environ.copy()\n",
" env[\"JAX_PLATFORMS\"] = \"cpu\"\n",
Expand Down Expand Up @@ -607,4 +601,4 @@
},
"nbformat": 4,
"nbformat_minor": 5
}
}
Loading