feat: add vllm router - #272
Merged
Merged
Conversation
![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 the installation of vllm-router to both the CANN and CUDA Dockerfiles, and resolves a shared library loading issue for Mooncake on CANN by updating the RUNPATH of libtransfer_engine.so using patchelf. It also includes a patch to fix DecodeBenchConnector for backends with tuple-based KV caches, such as Ascend MLA. The review feedback points out that patchelf may not be pre-installed in the CANN base image, which could cause build failures, and suggests conditionally installing it during the build process.
Collaborator
|
Rebase onto the latest main branch, then Since the CANN vLLM fork hasn't fixed the 002-related issue, stick with 003. |
Every one of the six vLLM Dockerfiles carried none of it, so a disaggregated group had no router to run: the Ascend recipe fell back to an engine example script with no metrics and no circuit breaker, and the CUDA one could not be deployed at all. CANN and CUDA first because those are the two recipes GPUStack ships a PD mode for. The wheel is prebuilt for both aarch64 and x86_64, so this is an install rather than a build. The build fails if the binary is missing afterwards. Without that check the failure surfaces at deploy time as a container that exits immediately, one layer away from anything that explains why.
`_fill_blocks` assumed every layer's KV cache is a single block-indexed tensor
and dereferenced `kv_cache.device` directly. vllm-ascend's MLA backend registers
it as a tuple (kv_c_cache, k_pe_cache), so the first request that reaches
`start_load_kv` raises
AttributeError: 'tuple' object has no attribute 'device'
which propagates out of the worker and kills EngineCore -- the server then 500s
every request and the container exits. Measured on 910B2 with GLM-4.7-Flash
(MLA) on vLLM 0.23.0: `capD-1xtp2` and `capD-1xtp4` both died ~40 s after the
benchmark started.
Both tuple members are block-indexed with the same num_blocks, so the existing
single-tensor path applies to each unchanged. The fill body moves into
`_fill_block_tensor` and is called once per member.
Scope: block-indexed layouts only. Non-block-indexed state (Mamba / linear
attention, where upstream fills the whole buffer) is left alone; upstream main
handles that separately via `_fill_state_tensor` and `CircularBufferSpec`.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.