Add recipes for Llama3.3 70B and Llama4 Scout #13
Conversation
nvpohanh
force-pushed
the
dev/nvpohanh/llama-v0.1
branch
from
838cee6 to
6bc5cea
Compare
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Summary of Changes
Hello @nvpohanh, I'm Gemini Code Assist1! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!
This pull request introduces two new quick start recipes, one for the Llama 3.3 70B Instruct model and another for the Llama 4 Scout Instruct model. These recipes provide detailed instructions for deploying and serving these large language models using vLLM, specifically optimized for NVIDIA Blackwell architecture GPUs with FP8 and NVFP4 quantization.
Highlights
- New Model Recipes: I've added comprehensive quick start recipes for deploying two new Llama Instruct models: Llama 3.3 70B and Llama 4 Scout.
- vLLM and Blackwell Optimization: These recipes detail setting up vLLM for efficient inference, including Docker image building, server launch commands with specific parameters for FP8/NVFP4 quantization, and optimizations tailored for NVIDIA Blackwell GPUs.
- Deployment and Validation Guidance: Each recipe provides step-by-step instructions covering model access, environment setup, server configuration, and includes methods for basic testing, accuracy verification using
lm_eval, and performance benchmarking withbenchmark_serving.py.
Using Gemini Code Assist
The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.
Invoking Gemini
You can request assistance from Gemini at any point in your pull request via creating an issue comment (i.e. comment on the pull request page) using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands.
| Feature | Command | Description |
|---|---|---|
| Code Review | /gemini review |
Performs a code review for the current pull request in its current state. |
| Pull Request Summary | /gemini summary |
Provides a summary of the current pull request in its current state. |
| Comment | @gemini-code-assist | Responds in comments when explicitly tagged, both in issue comments and review comments. |
| Help | /gemini help |
Displays a list of available commands. |
Customization
To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.
Limitations & Feedback
Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments or fill out our survey to provide feedback.
You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension.
Footnotes
-
Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩
There was a problem hiding this comment.
Code Review
This PR adds two new recipes for Llama 3.3 70B and Llama 4 Scout on NVIDIA Blackwell hardware. The recipes are well-structured and provide detailed steps for setup and validation.
I've found a critical issue in the run_accuracy.sh script in both recipe files, which will prevent it from running correctly. I've also pointed out a few minor typos and formatting inconsistencies in the documentation that would improve clarity and quality.
Overall, great additions! Just a few corrections needed to make these recipes ready for users.
nvpohanh
force-pushed
the
dev/nvpohanh/llama-v0.1
branch
from
6bc5cea to
7e967a2
Compare
This is just the first working version. We will continue to improve on the recipes. Signed-off-by: Po-Han Huang <pohanh@nvidia.com>
nvpohanh
force-pushed
the
dev/nvpohanh/llama-v0.1
branch
from
7e967a2 to
c8783b9
Compare
| - `--gpu-memory-utilization`: The fraction of GPU memory to be used for the model executor. We recommend setting it to `0.9` to use up to 90% of the GPU memory. | ||
| - `--compilation-config`: Configuration for vLLM compilation stage. We recommend setting it to `'{"pass_config": {"enable_fi_allreduce_fusion": true}, "custom_ops": ["+rms_norm"], "level": 3}'` to enable all the necessary fusions for the best performance. | ||
| - We are trying to enable these fusions by default so that this flag is no longer needed in the future. | ||
| - `--enable-chunked-prefill`: Enable chunked prefill stage. We recommend always adding this flag for best performance. |
There was a problem hiding this comment.
Chunked prefill is enabled by default, so this argument is redundant.
2 participants
This is just the first working version. We will continue to improve on the recipes.