chore: ⬆️ Update ggml-org/llama.cpp to 0021a77de0a8966059dc94548fb3b96654e0bb12

[localai-bot]

Changes: https://github.com/ggml-org/llama.cpp/compare/84e908c625fb60992b4cdef8180fb12fa9b4c4bf..0021a77de0a8966059dc94548fb3b96654e0bb12

[mudler]

@localai-org-maint-bot pick this up, fix CI

[localai-org-maint-bot]

Pushed the patch-context refresh on current master. The full llama.cpp patch series now applies cleanly to the new pin, and all 14 standalone C++ unit-test files pass. The new bot-assisted commit intentionally has no Signed-off-by; a human must satisfy DCO per the repository policy.

[mudler]

@localai-org-maint-bot CI fails

[localai-org-maint-bot]

Pushed 972310a53 to fix the remaining patch failure. The previous refresh kept pre-series context around the TTS insertion; after patches 0001 and 0002 changed the same file, GNU patch rejected 0003. The narrowed hunk now matches the post-series insertion point, and the complete four-patch stack applies cleanly against llama.cpp 0d9ceae1. DCO still needs human action; this bot cannot add a human attestation.

[mudler]

@localai-org-maint-bot another one

[localai-org-maint-bot]

Pushed f684a352f. The new pin changed the slot-reset, prompt-batch, and pre_decode() contexts, so GNU patch was fuzz-applying stale hunks before the L4T compile failed. Both server patches now apply exactly against 4df29be4, git diff --check passes, and all 14 standalone C++ unit-test files pass. The full CMake build was unavailable on this host; the new CI run will validate that path. DCO needs human action because this bot cannot add a human attestation.

[mudler]

@localai-org-maint-bot again

[localai-org-maint-bot]

Pushed a5e7ed1fc. The remaining gRPC failures came from the TTS patch still targeting pre-score-patch line offsets, so GNU patch rejected its pre_decode() hunk. I refreshed the TTS hunk locations against llama.cpp 4df29be4; the complete score + TTS patch stack applies cleanly with git apply, git diff --check passes, and all 14 standalone C++ unit-test files pass. The Go toolchain cache restore error is runner corruption, not a code failure. DCO still needs human action because this bot cannot add a human attestation.

[localai-org-maint-bot]

Pushed 1420ad057 after rebasing onto current master. The new llama.cpp pin creates the batch threadpool during common_init_from_params(), but LocalAI left cpuparams_batch.n_threads at the -1 sentinel; that became the CI allocation of exactly 2^64 bytes in both gRPC jobs. The server now resolves the sentinel from the requested inference thread count before model initialization, while preserving an explicit threads_batch override. The focused regression and all 15 standalone C++ unit-test files pass, both patch files apply cleanly to 4df29be4, and git diff --check passes. The pushed run will validate the full container path. DCO still needs human action because this bot cannot add a human attestation.

[localai-org-maint-bot]

Restored the full fix stack after automation regenerated the branch at f4463f294 and dropped it. The branch now points to 4f78c8c2e on current master with llama.cpp 0021a77de. The complete patch series applies cleanly to that revision, git diff --check passes, and all 15 standalone C++ unit-test files pass. The full backend matrix is running on the pushed head. The bot-created commits intentionally contain no human DCO attestation.