ROCm: compress the embedded GPU code; upgrade llama.cpp b11247 → b11256 - #470

Summary

• ROCm GPU code compressed (--offload-compress). The Windows ROCm jllama.dll in llama-*-all-windows-x86-64-jar-with-dependencies.jar was ~1 GB. Each HIP translation unit embeds one code object per GPU target (the whole kernel set, once for each of the 23 architectures), and clang stores those bundles uncompressed by default.
  • The jar hid the size (234 MB zipped), but LlamaLoader extracts the library to the temp dir on every start. In the all-backends fat jar, ROCm is tried right after CUDA, so this happens on nearly every machine without an NVIDIA card.
  • ggml-hip is now compiled with --offload-compress, scoped to that target and its source language (HIP on Linux, CXX on Windows). The HIP runtime decompresses each bundle when the module loads.
  • Upstream llama.cpp does not do this.
• CI guard: both ROCm jobs run the new .github/verify-hip-offload-compressed.py after the build.
  • It prints the library size and the compressed/uncompressed bundle counts, also into the job summary.
  • It fails on any uncompressed bundle (__CLANG_OFFLOAD_BUNDLE__), or if there is no compressed one at all, so losing the flag reds the job instead of shipping the 1 GB library again.
• llama.cpp b11247 → b11256: 9 commits, 22 KiB diff, one step. No project source change is needed.
  • #29632 moves several upstream tools to llama_backend_init(), which jllama, TTS and the trainer already call.
  • fs_write_atomic() (#29642) is unused here; the scheduler graph_inputs pass (#29634) and faster GGUF duplicate checks (#29598) are internal.
  • server-schema.cpp and server-task.cpp are untouched, so the request/response contract is unchanged.
  • All nine patches apply unchanged, and the drop-checks show all are still needed.
  • No upstream release.yml change, so the CUDA/ROCm/OpenVINO pins stay.
  • History rows are appended to docs/history/llama-cpp-breaking-changes.md.

Test plan

• Fresh configure on b11256: all nine patches applied (stamp checked); CPU build of jllama + jllama_test succeeds
• ctest: 590/590 C++ tests pass
• NativeLibraryLoadSmokeTest 4/4 (linked build-info matches the b11256 pin)
• The check script was tested on synthetic files: pass, uncompressed bundle, no bundle, no library, and matches straddling chunk boundaries
• Not verifiable here (no ROCm toolchain in this environment): the ROCm build with --offload-compress and the resulting DLL/.so size. The two ROCm jobs' new "Verify the GPU code is compressed" step and its job-summary table show both.
• CI is green on this branch

Related issues / PRs

Refs #468 (b11237 → b11247)

Checklist

• My commits follow the repository's commit style
• No security-sensitive changes

🤖 Generated with Claude Code

https://claude.ai/code/session_01FtfgoazykGcQSCR3TYBmTQ

Generated by Claude Code