Add Super VL and Mamba DisAgg compatibility to vLLM 0.28 - #44
TomerBN-Nvidia merged 7 commits into
Conversation
…llm-project#52929) Signed-off-by: Nave Assaf <nassaf@nvidia.com>
Signed-off-by: Daniel Serebrenik <daserebrenik@nvidia.com>
Port only the PR vllm-project#38 areas not covered by upstream vllm-project#52929 and vllm-project#53121: flat image-processor configuration, native projector and RADIO weight names, and the image-processor fallback. Backport of the residual semantic delta from TomerBN-Nvidia#38 (7c4ca09). Co-authored-by: OpenAI Codex <codex@openai.com> Signed-off-by: tbarnatan <tbarnatan@nvidia.com>
Route scheduler-tagged prompt-tail padding through the rollback-capable speculative decode state layout while preserving the uniform K+1 graph shape. Assisted-by: OpenAI Codex Signed-off-by: Nathan Sala <natouda@gmail.com>
Assisted-by: OpenAI Codex Signed-off-by: Nathan Sala <natouda@gmail.com>
Co-authored-by: Nick Hill <nickhill123@gmail.com> Signed-off-by: Nathan Sala <33957625+natsala13@users.noreply.github.com>
|
👋 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 You ask your reviewers to trigger select CI tests on top of 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 If you have any questions, please reach out to us on Slack at https://slack.vllm.ai. 🚀 |
13d3385
into
TomerBN-Nvidia:super_vl_evals_v0.28.0
5 participants
Summary
Lineage
TomerBN-Nvidia/vllm:super_vl_evals_v0.28.02cf0a6915ce544dc493a0990f2ea38d81601128a(v0.28.0)natsala13/vllm:version/v028-eval-pr55178e6ee5b84d677f3af92d3c47a33f570a7b99dd5b2The source branch is a clean seven-commit linear descendant of the target branch.
Important: the first four commits in this PR already exist, unchanged, on Tomer's separate
codex/super-vl-evals-v0280-compat-async-minimalbranch at0c4dd4e2b. They appear here because this PR intentionally targets the cleansuper_vl_evals_v0.28.0branch. The final three commits preserve the imported PR vllm-project#55178 history; the last commit is formatting-only.Validation
super_vl_evals_v0.28.0is clean and reports up to dategit diff --checkpasses