llama.cpp fork sync

Tracking state of the bong-water-water-bong/llama.cpp fork (the Zaya1/paged-KV development line). No submodule — this document is the sync record. Bump it whenever the fork moves.

Current sync point

Fork branch: master
Merge commit: 0807d70be21914a2ed616b500f7b10258d029f7a
Date: 2026-08-01
Source PR: bong-water-water-bong/llama.cpp#2 — feat: paged KV cache — fixed-pool GPU tensors with CPU backing store
Local checkout: synced to this commit 2026-08-02 (was at PR head 329cb3241), rebuilt -DGGML_VULKAN=ON -DBUILD_SHARED_LIBS=OFF (static libs required by zaya_server import — without BUILD_SHARED_LIBS=OFF ggml builds shared and the import check in CMakeLists.txt silently fails). Vulkan backend verified on Strix Halo: llama-bench sees Radeon 8060S (RADV, KHR_coopmat); unified_server relinked against the synced statics.

What shipped in this sync

Feature: paged KV cache (PR #2)

Fixed-pool GPU KV tensor pool with CPU backing store
llama_kv_cache_paged + llama_kv_paged_scorer (heuristic scorer factory, drafter-model hook reserved)
Zaya arch support: LLM_ARCH_ZAYA, CCA attention (even layers) + MoE (odd layers), Q4NX/1BP conversion (conversion/zaya.py, gguf-py), ROCm ops (hip_zaya_ops.hip)
ONNX runtime integration (src/llama-onnx.{h,cpp}, gated on onnxruntime_FOUND)

CI/build fixes landed on top

Commit	Fix
`a14293be`	missing `ggml_graph_dump_txt` prototype (`-Werror=missing-prototypes`)
`3d4fa9cf`	unused `n_gpu_layers` (`-Werror=unused-parameter`)
`06c3be13`…`b0aad5d1`	`llama-onnx.cpp` only compiled when onnxruntime found + `#ifdef LLAMA_ONNX` guard
`ab2baf23`	dead `n_written` in `repool_simple` (macOS `-Wunused-but-set-variable`)
`de538c0`	`PRId64` in ZAYA_DEBUG fprintf (macOS `-Werror=format`)
`a8576f5`	test-llama-archs: zaya skipped with FIXME (see #1357)

Known issues (tracked in 1bit-monster)

#1357 — zaya CCA conv graph produces invalid shapes (ggml_ssm_conv + conv_1d_grouped yield n-4 outputs vs n expected; 1-token/seq reserve graphs abort). Arch never ran in llama.cpp; test excluded pending a fix with the intended conv semantics.
#1358 — macOS test-thread-safety leaks 231MB under leaks -atExit (only remaining red check; needs a macOS box).

Status

Linux CI (cpu x64/arm64, full test-llama-archs): green (verified locally too: 114 arch/config combos OK, exit 0)
macOS tvOS/visionOS: green; macOS x64/arm64: red on #1358
Self-hosted GPU jobs (strixhalo runners): pending

Build note

Fork default branch is prism; the paged-KV work lives on master (PR #2 was merged there directly). Zaya1 branch retained for further development.