Conversation
Collaborator
Author
|
/sweep full-sweep --single-node --runner-type h200 --model-prefix dsr1 --seq-lens 1k1k --max-conc 1 |
Contributor
|
@cquil11 Kicking off a sweep. Run: https://github.com/InferenceMAX/InferenceMAX/actions/runs/20312209087 |
YukioZzz
added a commit
that referenced
this pull request
Aug 15, 2026
Build the SGLang-pinned ROCm/mori #341 source into a derived node-local image so write-mode runs use native wait_all and fail fast instead of silently falling back to polling.
YukioZzz
added a commit
that referenced
this pull request
Aug 15, 2026
Adds the Kimi-K3 MXFP4 multi-node agentic-coding arm (kimik3-fp4-mi355x-vllm-disagg-agentic-mtp): 1P(TP8) x 1D(TP8) with DSpark speculative decoding, MoRIIO KV transfer and the vLLM SimpleCPUOffloadConnector DRAM tier, at concurrency 1 (no offload) and 8/12/16 (DRAM offload). The engine side needs vLLM PR #51052 (Kimi-K3 hybrid attention/KDA transfer over MoRIIO). It ships as a validated full-file overlay rather than a unified diff, because the diff loses hunks against each new nightly image. MoRIIO WRITE mode additionally needs IOEngine.wait_all from ROCm/mori #341, which no published wheel carries, so job.slurm builds that pinned commit into a node-local derived image before serving. Prefill and decode must agree on the transfer direction, so the read/write decision is made once for the whole job and pushed to both containers: if the derived image cannot be produced on every node, the job serves READ mode on the base image instead of failing or serving a mismatched pair. Set MORI_WAITALL_REQUIRED=1 to keep a hard failure instead. Multi-node failures were previously only visible through the Slurm stdout stream, and decode-node logs were never retained at all. Per-node engine, router and container logs plus the mori build transcript are now gathered from every allocated node, on both the normal and the interrupted path, and uploaded with the Slurm stdout/stderr in the server-log artifact.
YukioZzz
added a commit
to YukioZzz/InferenceX
that referenced
this pull request
Aug 19, 2026
…AI#2602, #51052) Add the formal kimik3-fp4-mi355x-vllm-disagg-agentic-mtp recipe (1P1D TP8, DSpark speculative decoding, MoRIIO KV transfer, optional vLLM SimpleCPU offload) and the minimum harness needed to run it on cluster:mi355x-amds. Engine-side K3 hybrid attention/KDA over MoRIIO ships as k3_moriio_51052.patch (vLLM #51052). MoRIIO write mode builds ROCm/mori SemiAnalysisAI#341 (IOEngine.wait_all) into a per-node derived image before serve, with READ fallback when the build does not reach every node.
4 tasks
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
1 participant
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.