[dzannotti]

Reverts #155 (upstream llama.cpp sync of 2026-10-08). After it merged, agents stopped working on master.

What's reverted

• Merge upstream llama.cpp (2026-10-08) and the final halo-box/llama.cpp master #155: the upstream sync itself.
• quantize : choose quant types to hit a bpw or file size target (upstream #15550) #156 (quantize to a bpw or file-size target): it builds directly on the imatrix code that Merge upstream llama.cpp (2026-10-08) and the final halo-box/llama.cpp master #155 brought in (common/imatrix-loader.cpp, src/llama-quant.cpp), so Merge upstream llama.cpp (2026-10-08) and the final halo-box/llama.cpp master #155 can't be reverted without it. It will need to land again after the sync is redone.

#118 (Vulkan QSA scratch) is kept. It has no dependency on #155.

Checks

• Clean reverts: git revert -m 1 of 918108a (quantize : choose quant types to hit a bpw or file size target (upstream #15550) #156), then of 172a530 (Merge upstream llama.cpp (2026-10-08) and the final halo-box/llama.cpp master #155), with no conflicts.
• Resulting code: identical to master before Merge upstream llama.cpp (2026-10-08) and the final halo-box/llama.cpp master #155 (6aa81b8) plus vulkan : grow the context-sized scratch in 64 MiB steps #118's 21-line change to ggml-vulkan.cpp, and nothing else.
• HIP build (gfx1151): llama-server builds OK. The Vulkan build is left to CI.
• Tool-calling smoke test (Qwen3-1.7B Q8_0, --jinja): a get_weather request returns finish_reason: tool_calls with a valid call.

🤖 Generated with Claude Code

https://claude.ai/code/session_01G5NBdiBqRmGomhSRXBoucx