Public story · 2026-08-29 · source-backed

A separate llama.cpp fix: `conv_transpose_2d` computed only the first batch.

Issue

2026-08-29

Confidence

source-backed

Sources

Redaction

redacted

Story

Build b10676 fixes ggml_compute_forward_conv_transpose_2d_impl, which left every batch after the first as zeros; both the src1 permutation and the main loop now iterate the batch dimension, and ggml_graph_plan scales the work buffer by src1 batch count. The Metal kernel had the identical defect, decoding no batch from the grid, so multi-batch results diverged from the CPU reference. A multi-batch case was added to test-backend-ops, which is what should have caught it originally. Two silent-wrong-output bugs in consecutive builds is a bad week for anyone treating llama.cpp output as ground truth. (GitHub)

Source trail

ggml-org/llama.cpp More from github.com

Entities

Build CPU GitHub

Provenance

Canonical issue: Ramsay Research Agent — August 29, 2026
AI generated: no
Story unit: 2026-08-29-a-separate-llama-cpp-fix-conv-transpose-2d-computed-only-the-first-batch
Labels: source-backed, canonical briefing excerpt

Fetching from the wire…

A separate llama.cpp fix: `conv_transpose_2d` computed only the first batch.

Story

Related stories

cjpais/Handy — Free Offline Speech-to-Text (15.7K stars, +94/day)

llama.cpp fixed a Vulkan bug that silently produced wrong tokens on AMD and NVIDIA.

Stanford's BLAST ships local sandbox orchestration as a 7 MB MIT-licensed binary.

llama.cpp fixed a SYCL path moving 4.56 GB per 2048-token prefill chunk.

gemma4.c runs Gemma 4 E2B in 700 lines of pure C at real time on a Ryzen 7 7700.

llama.cpp merged `--n-cpu-ffn`, giving dense models the offload flag MoE models already had.

Kimodo.cpp ports NVIDIA's text-to-motion model to GGML

oMLX splits prefill across ANE, CPU and GPU for a measured 36% gain, and cuts compile memory 35.8 GB to 4.7 GB.

Source trail

Entities

Provenance

A separate llama.cpp fix: `conv_transpose_2d` computed only the first batch.

Story

Related stories

cjpais/Handy — Free Offline Speech-to-Text (15.7K stars, +94/day)

llama.cpp fixed a Vulkan bug that silently produced wrong tokens on AMD and NVIDIA.

Stanford's BLAST ships local sandbox orchestration as a 7 MB MIT-licensed binary.

llama.cpp fixed a SYCL path moving 4.56 GB per 2048-token prefill chunk.

gemma4.c runs Gemma 4 E2B in 700 lines of pure C at real time on a Ryzen 7 7700.

llama.cpp merged `--n-cpu-ffn`, giving dense models the offload flag MoE models already had.

Kimodo.cpp ports NVIDIA's text-to-motion model to GGML

oMLX splits prefill across ANE, CPU and GPU for a measured 36% gain, and cuts compile memory 35.8 GB to 4.7 GB.

Source trail

Entities

Provenance