RepoJournal
Local LLMs Local LLMs
46 wires and counting

$ follow Local LLMs

Keep up with Local LLMs in about 3 minutes: what actually shipped — the commits, pull requests, releases, and security advisories that matter.

or

fair warning: these emails are deeply technical. diffs, version numbers, CVEs, benchmark deltas. if that's not your idea of a good read, this isn't your newsletter.

Folds into your digest — weekly by default, monthly if you prefer. Unsubscribe in one click.

$ status

wire 2026-08-30
stories 147
RSS Archive Topics

© 2026 RepoJournal Home Showcase Explore How it works Privacy

$ the-wire · showcase

llama.cpp Vulkan memory bloat fixed, vLLM deprecates PyAV

By RepoJournal · Filed · About Local LLMs

jeffbolznv jhen0409 nikwen Isotr0py connorcarpenter15 yudigege86 xiezhq-hermann
12 people shipped this
jeffbolznv @jeffbolznv 1 cited jhen0409 @jhen0409 1 cited nikwen @nikwen 1 cited Isotr0py @Isotr0py 1 cited connorcarpenter15 @connorcarpenter15 1 cited yudigege86 @yudigege86 1 cited xiezhq-hermann @xiezhq-hermann 1 cited ch-wan @ch-wan 1 cited nicois @nicois am17an @am17an TheEpicDolphin @TheEpicDolphin joeltg @joeltg

llama.cpp fixes a Vulkan memory bloat bug that caused extreme VRAM usage, while vLLM begins deprecating its PyAV video decoder backend over performance concerns.

llama.cpp's Vulkan backend now pads K instead of N in mul_mat_id [1]. The N padding was unnecessary for mul_mat_id and caused extreme memory bloat [1]. This is a breaking change for any fork patching around the old behavior. A separate tuning PR updates fa-vec tunings for M1 Max [2], while another adds remaining fa-vec tunings for M4 Pro [3]. vLLM deprecates the PyAV video decoder backend due to poor performance [4]. The Rust frontend now accepts audio and video media inputs over gRPC [5]. vLLM also fixes speculative decoding on ROCm by reaching FULL cudagraphs for AITER MLA [6]. sglang's HiCache buffer mode now decides staged-fetch fate against the live tree, preventing silent skips of pins when node ids go stale [7]. "When the lookup missed, it silently skipped the pin and fetched anyway" [7]. A stacked config series moves the resolution pipeline's dispatcher off the record [8].

Action items

References

  1. [1] vulkan: Change mul_mat_id to pad K rather than N ↗ ggml-org/llama.cpp
  2. [2] metal : add fa-vec tunings for M1 Max ↗ ggml-org/llama.cpp
  3. [3] metal : add remaining fa-vec tunings for M4 Pro ↗ ggml-org/llama.cpp
  4. [4] [Multimodal] Deprecate PyAV video decoder backend ↗ vllm-project/vllm
  5. [5] [Rust Frontend][gRPC] Add audio and video media inputs ↗ vllm-project/vllm
  6. [6] [ROCm][MLA] Reach FULL cudagraphs for AITER MLA speculative decoding ↗ vllm-project/vllm
  7. [7] [HiCache] buffer mode: decide staged-fetch fate against the live tree ↗ sgl-project/sglang
  8. [8] config: the resolution pipeline's dispatcher leaves the record ↗ sgl-project/sglang

$ ls local-llm/reviews/ # the step-back reads

the week in review Vulkan top_k radix select arrives for long rows the month in review Speculative decoding consolidates across the stack as DeepSeek V4 lands
Topics: Local LLMs AI / ML Full archive →

Quick answers

What shipped in Local LLMs on August 30, 2026?
llama.cpp fixes a Vulkan memory bloat bug that caused extreme VRAM usage, while vLLM begins deprecating its PyAV video decoder backend over performance concerns. In total, 69 commits, 69 pull requests, and 9 releases landed.
Who contributed to Local LLMs on August 30, 2026?
12 developers shipped this update, including jeffbolznv, jhen0409, nicois, nikwen, am17an, Isotr0py, connorcarpenter15, and yudigege86, and 4 more.
What were the notable Local LLMs updates?
vulkan: Change mul_mat_id to pad K rather than N, metal : add fa-vec tunings for M1 Max, and metal : add remaining fa-vec tunings for M4 Pro.
♥

$ log Local LLMs --recent

Vulkan adds TQ1_0, Spark2.5 lands in llama.cpp 180 shipped vLLM closes scale-out multimodal handoff hole 86 shipped vLLM Fast Start maps weights zero-copy; GLMGA sampling capped 256 shipped Ollama enables speculative decoding under structured output, vLLM adds LoRA and fusion 259 shipped

$ log --all # elsewhere on the wire

Meta React Native adds Stable API guards to renderer attributed string FastAPI & Pydantic Pydantic fixes Mypy plugin crash, tightens datetime constraints

$ ls showcases/

Rails Django Laravel Node.js Go Vercel Vue.js Kubernetes HashiCorp Arch Linux Linux Anthropic

$ init your-desk

RepoJournal can write one of these for any repo you care about. Sign in with GitHub, pick your repos and sources, and your desk gets a wire like this one every morning.