Repository navigation
feat(runtime): make mixed prefill selectable at load time - #290
Draft
shadownetX wants to merge 1 commit into
Draft
shadownetX wants to merge 1 commit into
shadownetX wants to merge 1 commit into
Conversation
Mixed-enabled cores capture both separate and mixed execution programs, but serving cannot exclude the mixed programs without replacing the export. Load-time selection can isolate their scheduling impact using the same core. Add paired native CLI flags and propagate the optional selection to every rank. The loader validates the captured catalog, excludes mixed graphs when disabled, and rejects explicit enable when no eligible mixed graph remains. Omitting both flags retains existing selection. The existing graph and kernel pruning path removes unused resources without altering captured storage. This permits controlled mixed-on/off serving comparisons without re-export or changes to scheduler policy, MTP selection, or cache behavior. Add CPU regressions for graph selection, retained separate resources, startup tuning dependencies, invalid catalogs, CLI forwarding, and scheduler progress without mixed graphs. Targeted CPU regressions pass in CI; native CUDA loading, performance, and multi-rank replay remain unqualified. Co-authored-by: ChatGPT <support@openai.com> Design-Signed-Off-By: ChatGPT (GPT-6); design-review=ed22
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Motivation
Mixed-prefill exports already contain both separate and mixed execution graphs. However, switching between these execution strategies currently requires a new export, which can be particularly expensive when using the
audittuning profile (4h 18m on a single DGX Spark).Allowing runtime selection makes it possible to adapt serving to different workloads and compare both strategies using the exact same exported core.
Changes
--mixed-prefilland--no-mixed-prefilltoflare run.No changes to the core format, scheduler policy, speculative decoding, or prefix-cache behavior.
Validation
The GPU comparison will cover TTFT, generation throughput, memory usage, and graph launches.
This PR will remain in draft until GPU validation is complete.