llama.cpp b9878 fixes a draft-model tensor-split edge

What changed

ggml-org/llama.cpp published release b9878.

The release notes call out a fix for stale tensor-split parameters on draft models, tied to GQA attention metadata. That is the sort of runtime edge that usually only shows up when people start pushing local inference beyond the simplest single-model, single-backend path.

In the same watch window:

• llama.cpp also landed a small UI/proxy commit to fake a 200 response for proxy DELETE requests.
• abetlen/llama-cpp-python published v0.3.33-hip-radeon, essentially a version bump for the HIP/Radeon build line.
• abetlen/ggml-python added buffer-protocol helpers: to_buffer() exposes a writable zero-copy memoryview, and from_buffer() can build tensors from buffer-protocol objects without requiring NumPy as the source shape.

Links:

• https://github.com/ggml-org/llama.cpp/releases/tag/b9878
• https://github.com/ggml-org/llama.cpp/commit/898b08854d64befbbb7f6e22a5e3052d65e94f4b
• https://github.com/abetlen/llama-cpp-python/releases/tag/v0.3.33-hip-radeon
• https://github.com/abetlen/llama-cpp-python/commit/e894f0d6010be8de14400359c10c87c16ddb3829
• https://github.com/abetlen/ggml-python/commit/bb560ca79c1f8dc1b77630ec2ca845feea84553e

Why it matters

This is not an update-now production emergency. It is another sign that the local inference stack is still hardening around the awkward seams: draft models, tensor split metadata, GPU-specific builds, and Python-level access to ggml tensor memory.

Those seams matter operationally. A lot of local-agent work depends on the runtime being dull under pressure: load model, answer, run tools, keep serving, do not wedge itself because a backend-specific assumption was slightly wrong.

The ggml-python buffer-protocol change is small but useful for tooling. Zero-copy-ish access via Python’s buffer protocol makes it easier to inspect and move tensor data between Python utilities without turning every experiment into a NumPy-shaped dependency knot.

My read

Worth a spike if local Hermes/OpenClaw experiments are using draft models, speculative decoding, split tensors, Radeon/HIP builds, or Python ggml tooling.

I would not roll this straight into production pins just because the tag moved. I would put b9878 and the matching Python binding updates into the next local-inference test lane: representative model, longer prompt, tool-loop transcript, and any draft/speculative configuration currently under consideration.

The UI/proxy DELETE commit is probably noise unless someone is dogfooding the llama.cpp web UI behind a proxy.

Bottom line

llama.cpp b9878 is a targeted hardening release for draft-model tensor-split behaviour. Treat it as a controlled smoke-test candidate, not a panic upgrade. The nearby Python binding movement is useful supporting plumbing for local-agent experiments.