Repository navigation
docs(readme): the NPU layer kernel and lm-head decode from source - #325
Merged
Merged
Conversation
PR Reviewer Guide 🔍(Review updated until commit b129a88)Here are some key observations to aid the review process:
|
The README still said the NPU layer kernel "is not yet built from source", while docs/npu.md Step 3d on the same main documents the opposite: both the dense layer kernel and lm-head run from source, ELF to ELF, with no xclbin at any point, one compiled program covering every context position, verified on Qwen3-0.6B, 1.7B and 4B against an fp64 reference (cosine >= 0.9999, greedy decode exact). Point the README at that, and keep the two qualifiers that are still true: the kernel sources stay private per CONTRIBUTING, and the work is not yet merged here. Recovered from the unreported docs-hrx-status branch, which is otherwise a pre-RFC-#213-stage-3 snapshot (it would re-document the Vulkan, ROCm, ZINC, lean and ComfyUI routes this repository removed, and it names docs/zinc.md, which #306 deleted). This one commit was the only part still additive.
bong-water-water-bong
force-pushed
the
docs/npu-from-source
branch
from
October 5, 2026 21:25
7d753ee to
b129a88
Compare
|
Persistent review updated to latest commit b129a88 |
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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
One stale sentence in the README, corrected.
The README said the NPU engine's "layer kernel is not yet built from source".
docs/npu.mdStep 3d — on the samemain— documents the opposite: both the dense layer kernel and lm-head run from source, ELF to ELF, with no xclbin at any point, one compiled program covering every context position, verified on Qwen3-0.6B, 1.7B and 4B against an fp64 reference (cosine similarity ≥ 0.9999 at every tested position, greedy decode exact end to end).The README now says that, and keeps the two qualifiers that are still true: the kernel sources stay private per CONTRIBUTING, and the work is not yet merged here.
Where it came from
Recovered from the unreported
docs-hrx-statusbranch (12 commits). Only this one commit was still additive. The rest is a pre-RFC-#213-stage-3 snapshot: applied as-is it would re-document the Vulkan, ROCm, ZINC, ROCmFPXleanand ComfyUI routes this repository has removed, and it editsdocs/zinc.md, which #306 deleted.docs/serve.mdanddocs/hrx.mdwere the worst of it — its lead-sentence rewrite re-asserts "one llama.cpp build with two GPU backends, ggml-hrx and ggml-vulkan".So the branch is superseded; this single fix is not.
Verified
tools/copyright.py --checkpasses.docs/npu.mdStep 3d now agree.Not to be merged: the release is on hold indefinitely.