Upgrade llama.cpp from b9151 to b9172

[bernardladenthin]

Summary

Upgrades the pinned llama.cpp version from b9151 to b9172 and adds necessary build configuration to handle upstream changes in WebUI asset provisioning.

Changes Made

• CMakeLists.txt: Updated GIT_TAG from b9151 to b9172 and added set(LLAMA_BUILD_WEBUI OFF CACHE BOOL "" FORCE) to disable WebUI asset building (upstream now provisions assets at build time via HF bucket by default)
• README.md: Updated llama.cpp version badge and link from b9151 to b9172
• CLAUDE.md:
  • Updated pinned version reference from b9151 to b9172
  • Added changelog entries documenting upstream changes between b9151–b9172:
    • clip_has_whisper_encoder() removed from public API (no project impact)
    • WebUI assets now provisioned at build time; project disables asset download via LLAMA_BUILD_WEBUI=OFF
    • common_params::webui default made conditional on LLAMA_WEBUI_DEFAULT_ENABLED macro
    • Bug fix in common_reasoning_budget_clone (no API change)
    • AMD RDNA3 WMMA flash attention support in CUDA backend
    • Non-function Responses API tools now silently skipped in server-chat

Implementation Details

The LLAMA_BUILD_WEBUI=OFF cache variable is set before FetchContent_MakeAvailable() to prevent upstream from downloading WebUI assets during the build. This aligns with the project's build strategy and avoids unnecessary network I/O.

All upstream changes are either internal (CUDA backend, server implementation) or non-breaking for the Java bindings. No JNI code modifications required.

https://claude.ai/code/session_01DVizDEtXBVDaXciEoo9a8v

[claude]

Review Summary

✅ Strengths

1. Consistent version updates: All references to llama.cpp version are correctly updated from b9151 to b9172 across CMakeLists.txt, README.md, and CLAUDE.md
2. Proper CMake configuration order: The LLAMA_BUILD_WEBUI=OFF cache variable is correctly placed before FetchContent_MakeAvailable(), ensuring upstream respects the setting
3. Well-documented changelog: The CLAUDE.md additions clearly document the upstream changes and their project impact
4. Minimal, focused PR: The changes are scoped appropriately for a version upgrade with no unnecessary modifications

⚠️ Question for Clarification

WebUI asset provisioning: The CLAUDE.md documentation indicates two upstream mechanisms for WebUI provisioning:

• LLAMA_USE_PREBUILT_WEBUI=ON (default) — downloads prebuilt assets from HF bucket
• LLAMA_BUILD_WEBUI — builds assets from source

The PR sets LLAMA_BUILD_WEBUI=OFF (disabling source build), but since LLAMA_USE_PREBUILT_WEBUI=ON is the default, will the build still attempt to download prebuilt WebUI assets from the HF bucket?

Clarification needed: Should the configuration also include set(LLAMA_USE_PREBUILT_WEBUI OFF CACHE BOOL "" FORCE) to fully prevent WebUI asset downloads and achieve the stated goal of avoiding unnecessary network I/O?

If the current setting is confirmed to work as intended (i.e., setting only LLAMA_BUILD_WEBUI=OFF prevents all WebUI provisioning), please verify this is tested in the CI pipeline.

Summary

The PR is technically sound and well-documented. The clarification above is the only item that should be addressed before merging.

[claude]

Consider whether LLAMA_USE_PREBUILT_WEBUI also needs to be disabled. According to CLAUDE.md, upstream defaults to LLAMA_USE_PREBUILT_WEBUI=ON which downloads prebuilt WebUI assets. Setting LLAMA_BUILD_WEBUI=OFF prevents building from source, but the prebuilt download might still occur. If the goal is to fully avoid network I/O, you may also need:

set(LLAMA_USE_PREBUILT_WEBUI OFF CACHE BOOL "" FORCE)

Or confirm that LLAMA_BUILD_WEBUI=OFF alone is sufficient to prevent all WebUI asset provisioning.