Add AMD GPU support for DeepSeek-V3, DeepSeek-R1 (MI300X/MI325X/MI355X + AITER)#278
Conversation
…X + AITER) - Add Step 1: uv-based vLLM ROCm installation - Add Step 2: vLLM server startup with AITER and AITER_MOE enabled - Add Step 3: benchmark script Tested and verified on 8x AMD MI300X and 8x AMD MI355X GPUs with AITER. Signed-off-by: Yuan Yue <yueyuan@amd.com>
Summary of ChangesHello, 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 significantly expands the hardware compatibility for DeepSeek-V3 and DeepSeek-R1 models by integrating official support for AMD MI300X, MI325X, and MI355X GPUs. It provides users with a complete workflow, from environment setup and vLLM server deployment to performance benchmarking, enabling broader accessibility and deployment options for these advanced AI models on AMD infrastructure. Highlights
🧠 New Feature in Public Preview: You can now enable Memory to help Gemini Code Assist learn from your team's feedback. This makes future code reviews more consistent and personalized to your project's style. Click here to enable Memory in your admin console. Changelog
Activity
Using Gemini Code AssistThe 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 by creating a comment using either
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 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. If you're interested in giving your feedback about your experience with Gemini Code Assist for GitHub and other Google products, sign up here. 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
|
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Code Review
This pull request adds documentation for running DeepSeek-V3 and DeepSeek-R1 models on AMD MI300X series GPUs. The changes are well-structured and provide clear installation and server launch instructions. However, I've identified a few areas for improvement, mainly concerning the benchmark instructions which are currently confusing and incomplete. I've also suggested minor formatting changes to improve the document's readability.
| ### Step 3: Run Benchmark | ||
|
|
||
| Open a new terminal and run the following command to execute the benchmark script. | ||
|
|
||
| ```bash | ||
| vllm bench serve \ | ||
| --model deepseek-ai/DeepSeek-V3 \ | ||
| --dataset-name random \ | ||
| --random-input-len 8000 \ | ||
| --random-output-len 1000 \ | ||
| --request-rate 10000 \ | ||
| --num-prompts 16 \ | ||
| --ignore-eos | ||
| ``` |
There was a problem hiding this comment.
The instructions for running the benchmark are confusing and the command is incomplete. The vllm bench serve command starts its own server, so the instruction 'Open a new terminal' might confuse users into thinking they need to connect to the server from Step 2. The benchmark command is also missing the necessary environment variables and parallelization arguments from Step 2 to run correctly on AMD GPUs.
I recommend rewriting this section for clarity and correctness. Here is a suggestion:
### Step 3: Run Benchmark
You can benchmark the model using the `vllm bench serve` command. This command will start a server, run the benchmark, and then exit. Note that the environment variables and server arguments from Step 2 must be included for the benchmark to run with the correct configuration.
```bash
SAFETENSORS_FAST_GPU=1 \
VLLM_USE_TRITON_FLASH_ATTN=0 \
VLLM_ROCM_USE_AITER=1 \
VLLM_ROCM_USE_AITER_MOE=1 \
vllm bench serve \
--model deepseek-ai/DeepSeek-V3 \
--tensor-parallel-size 8 \
--enable-expert-parallel \
--trust-remote-code \
--dataset-name random \
--random-input-len 8000 \
--random-output-len 1000 \
--request-rate 10000 \
--num-prompts 16 \
--ignore-eos|
|
||
|
|
There was a problem hiding this comment.
These extra blank lines are unnecessary. One blank line is sufficient to separate sections in Markdown.
|
|
||
| Recommended approaches by hardware type are: | ||
|
|
||
| MI300X/MI325X/MI355X |
There was a problem hiding this comment.
For better document structure and readability, consider making this line a sub-heading. This will help users quickly identify the hardware section.
| MI300X/MI325X/MI355X | |
| #### MI300X/MI325X/MI355X |
1 participant
Summary
This PR adds AMD GPU support for DeepSeek-V3 and DeepSeek-R1 on MI300X/MI325X/MI355X GPUs.
Changes
Hardware Tested
Related
Closes the AMD GPU support gap originally started in #144.