-
-
Notifications
You must be signed in to change notification settings - Fork 10.8k
Implement chunked prefill for Triton MLA attention backend #12800
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.
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
Conversation
|
👋 Hi! Thank you for contributing to the vLLM project. 💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels. Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging. To run CI, PR reviewers can either: Add 🚀 |
|
This pull request has merge conflicts that must be resolved before it can be |
12b15f3 to
f428ded
Compare
|
This pull request has merge conflicts that must be resolved before it can be |
2d969e0 to
d601380
Compare
34d9cd8 to
858eb26
Compare
858eb26 to
c37d92e
Compare
Signed-off-by: Lucas Wilkinson <lwilkinson@neuralmagic.com>
Fix alignment and modulo error in remainder for _gather_kv_cache Only use block_tables on chunked prefill Allow enabling chunked prefill Signed-off-by: Patrick Reiter Horn <patrick.horn@gmail.com> Signed-off-by: Yessen Kanapin <yessen@deepinfra.com>
c37d92e to
65571ba
Compare
|
This pull request has merge conflicts that must be resolved before it can be |
|
superseded by: #12639 |
Finishes the rest of @LucasWilkinson's PR #12639
Implement chunked prefill for Triton MLA backend (DeepSeek v2/v3)
I have been testing with DeepSeek-Coder-v2-Lite-Instruct, both short and long-context responses seem correct.
My prompt is:
{"model": "deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct", "max_tokens": 100, "seed": 6, "messages": [{"role": "user", "content": "Yes or no, will the following code compile:\n\n.... some long python script ..."}I think I fixed the issue with one block not being copied.