Repository navigation
feat: add tensor overrides and batch size controls for hybrid offloading - #283
Merged
Merged
Conversation
Phase 2 of hybrid GPU/CPU offloading support (ref #280). Adds three new InferenceServiceSpec fields for fine-grained inference tuning: - tensorOverrides: regex-based tensor placement (--override-tensor) - batchSize: prompt processing batch size (--batch-size) - uBatchSize: decoding micro-batch size (--ubatch-size) These controls are particularly useful for hybrid MoE workloads where batch size tuning amortizes PCIe overhead and tensor overrides enable expert-level placement decisions beyond the typed moeCPUOffload flag. This feature was implemented by Qwen3.6-35B-A3B running via llama.cpp with hybrid CPU/GPU offloading on dual RTX 5060 Ti GPUs, deployed by LLMKube itself with the Phase 1 offloading support from #281. Signed-off-by: Christopher Maher <chris@mahercode.io>
Merged
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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Summary
Phase 2 of hybrid GPU/CPU offloading (ref #280). Adds three new
InferenceServiceSpecfields for fine-grained inference tuning:tensorOverrides([]string): regex-based tensor placement via--override-tensor. Enables expert-level placement decisions beyond the typedmoeCPUOffloadflag (e.g.,[\"exps=CPU\", \"blk\\\\.(0|1)\\\\.ffn_.*=CUDA0\"])batchSize(*int32, 1-16384): prompt processing batch size via--batch-size. Tuning this amortizes PCIe overhead in hybrid MoE workloadsuBatchSize(*int32, min 1): decoding micro-batch size via--ubatch-sizeNotable implementation detail
This feature was implemented by Qwen3.6-35B-A3B running on dual RTX 5060 Ti GPUs with Phase 1 hybrid offloading from #281. The model handled the full implementation in 32 minutes of active agent time: 56 tool calls, 100% success rate, 9 unit tests plus integration test updates, Helm CRD sync, and all verification commands green.
Test plan
Ref #280 (Phase 2 of 3, Phase 3 still tracked in the issue)