microsoft / microsoft/onnxruntime
[Feature Request] Multi-Head Latent Attention(DeepSeek) support on CPU/NPU
Nobody has claimed this yet.
- Dominant language
- C++
- Stars
- 21.9k
- Forks
- 4.2k
- Avg merge
- 4d 11h
- Merged PRs (30d)
- 184
Description
### Describe the feature request
DeepSeek's models use Multi-head Latent Attention, the current ONNX model [https://huggingface.co/onnxruntime/DeepSeek-R1-Distill-ONNX](url) release leverages GroupQueryAttention.
Is MLA on roadmap for ONNXRT?
### Describe scenario use case
Lower KV cache footprint with Multi-Head Latent Attention improving mobile and edge inference
Contributor guide
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Research direction
Start by reviewing the linked DeepSeek-R1-Distill-ONNX release and its use of GroupQueryAttention, then compare that with the requested Multi-Head Latent Attention behavior. The issue names no ONNX Runtime entry point or test; completion criteria would need to define CPU/NPU support and validation of the lower KV-cache footprint.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- cpp, machine-learning
- Domain
- machine-learning, performance
- Issue type
- Feature
- Difficulty
- 5/5
- Estimated time
- Over a week
- Activity status
- Stale
- Clarity
- Needs clarification
- Newbie friendliness
- 20/100