Repository navigation
Add packed INT4 GPTQ decode and page sizes 64 and 128 - #560
Open
krisclarkdev wants to merge 7 commits into
Open
krisclarkdev wants to merge 7 commits into
krisclarkdev wants to merge 7 commits into
Conversation
…e-1 decode graph path that reached the measured rate. Co-authored-by: Cursor <cursoragent@cursor.com>
Streaming chat completions accept return_token_ids so llama-benchy can time a generate, and the server stays on the page-1 decode path without speculative decoding. Co-authored-by: Cursor <cursoragent@cursor.com>
.specify/ and specs/ are local notes and do not belong in the GPTQ pull request. Co-authored-by: Cursor <cursoragent@cursor.com>
The no_buffer mamba cache rejects every page size except 1, so paged decode never started. extra_buffer is now selected for the Intel XPU linear backend, and the fused GDN kernel keeps the live state. Co-authored-by: Cursor <cursoragent@cursor.com>
Spec Kit plans stay untracked. The ignore file no longer lists editors this repo does not use. Co-authored-by: Cursor <cursoragent@cursor.com>
3 tasks done
Wide layers share four columns and long-K layers keep the two-column pipeline. Vocabulary rows stay in one pipelined pass, and Gated DeltaNet verify uses the SYCL ops the base wheel does not ship. Co-authored-by: Cursor <cursoragent@cursor.com>
3 of 4 tasks
Speed up M=4 packed GEMV and multi-row lm_head for NEXTN
This branch has not been deployed
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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
extra_bufferso the mamba cache no longer rejects every page size except 1, and the XPU paged attention kernel captures the decode graph..specify/andspecs/stay ignored.This branch contains the work in #554 plus the page-size change. Merging this covers that pull request.
llama-benchy 0.4.0 on node hal, model
Qwen3.8-27B-TURBO-DF735-882-NM-DAU-GPTQ-Int4-g128-tuned, one run, concurrency 1, speculative decoding off:The configured context length is 131072. The KV pool holds 81152 tokens, so one request cannot fill that window. The XPU paged decode kernel accepts 64 and 128 only.
Test plan
.specify/orspecs/filesMade with Cursor