knowledge(coreml): ANE CPU-scheduled matmul research lead - #76
Merged
Conversation
Records an unverified third-party claim (reviewer comment on the Surgical Inference draft, from the ds4-ssd repo author): a private ANE matmul op schedulable directly from the CPU with no Core ML recompile, int8 weight transfers for bandwidth-bound stages, and the prefill/decode distinction that makes 'LLMs don't benefit from the ANE' a decode-only claim. Framed as a lead to investigate, not confirmed guidance; includes a verification checklist and links to existing neural-engine notes.
Adds an 'If proven right' section making the payoff explicit: the prefill/decode split would let LLM prefill (compute-bound) run on the ANE off the GPU while decode stays bandwidth-bound elsewhere, reopening on-device LLM inference and a potential future 'llm' model class. Kept conditional on the existing verification checklist.
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
What
Adds
knowledge/coreml/ane-cpu-scheduled-matmul.mdcapturing a research lead relevant to mobius's ANE-targeting mission, plus an index bullet inknowledge/coreml/AGENTS.md.Source
A reviewer comment on the Surgical Inference draft (M. Mireles), left by the author of the
ds4-ssdrepo, responding to the paper's "reverse engineering the ANE" section. The claims:.mlmodelc→predictcycle ("ANE as a BLAS backend").Why it matters here
Sharpens two positions in our existing ANE notes, and — if it holds — reopens a direction we currently write off:
llmmodel class. Kept conditional on verification.Status
Explicitly recorded as unverified — a lead to investigate, not confirmed guidance, and not a committed scope change. The note includes a verification checklist (locate the API surface via
ds4-ssd, reproduce the M4 prefill number withcoreml-cli, measure int8 vs FP16 transfer, confirm iOS availability/OS fragility) and links to the vendoredneural-engine/reverse-engineering docs.No code or model changes.