[danielhanchen]

Based on cpu-speed (#36). One new ggml patch, no source change.

Problem

The CPU build compiles out ggml's llamafile tinyBLAS GEMM (no sgemm.cpp object in the pin's or #36's build logs), so every Q8_0 matmul runs one vec_dot per output element. On a Q8_0 Z-Image DiT that dot product is 84% of the denoise time (per-op profile, Zen 2), and ComfyUI --cpu is 1.69x faster than the pin on Z-Image Q8_0 at 512 px.

Fix

ggml-patches/0017 builds tinyBLAS on x86 and routes Q8_0 weights through it. Every other type keeps its current path. GGML_CPU_SGEMM_TYPES selects the types (default q8_0).

Evidence

Colab CPU VMs (Intel Broadwell, 4 cores / 8 threads, 50 GB), #36 vs #36 + this patch, same VM, interleaved, 3 repeats:

VM	Z-Image Q8_0, 512 px, 1 step	#36	this branch
vm1	sampling	162.2 s [153.1, 201.5]	100.9 s [99.2, 139.3] (1.60x)
vm1	generate	239.7 s	153.2 s
vm5	sampling	161.3 s [145.2, 189.8]	100.9 s [100.6, 103.6] (1.60x)
vm5	generate	216.1 s	153.6 s

• Pixels are identical to the same source built without the patch on the same VM.
• Q4_K_M rows are unchanged, as expected (they already have a fast path).

Not covered

AVX512, Zen 4/5, hybrid Intel, Windows and ARM CPUs were not timed; only 4-core cloud VMs.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-11-03T20:50:11.371036Z	ca0d2bc	Manual request

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[danielhanchen]

@codex review

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. What shall we delve into next?

Reviewed commit: ca0d2bc42c

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".