Skip to content
Permalink

Comparing changes

Choose two branches to see what’s changed or to start a new pull request. If you need to, you can also or learn more about diff comparisons.

Open a pull request

Create a new pull request by comparing changes across two branches. If you need to, you can also . Learn more about diff comparisons here.
base repository: abetlen/llama-cpp-python
Failed to load repositories. Confirm that selected base ref is valid, then try again.
Loading
base: main
Choose a base ref
...
head repository: ByCzech/llama-cpp-python
Failed to load repositories. Confirm that selected head ref is valid, then try again.
Loading
compare: main
Choose a head ref
Checking mergeability… Don’t worry, you can still create the pull request.
  • 19 commits
  • 17 files changed
  • 2 contributors

Commits on Aug 27, 2026

  1. feat: DEB packages - Add modular CPU, Vulkan, CUDA, ROCm and SYCL bac…

    …kend packages
    
    * split llama-cpp-python into core and dynamically loadable backend packages
    * add CPU backend with multiple optimized CPU variants and OpenBLAS
    * add Vulkan backend
    * add co-installable CUDA 12.4, 12.9 and 13.3 backends using update-alternatives
    * add ROCm 7.14 HIP backend with support for current AMD GPU architectures
    * add Intel oneAPI SYCL backend with oneMKL, oneDNN and Level Zero support
    * add transitional packages for legacy package names
    * add backend alternatives handling and private shared-library dependency resolution
    * add CUDA 12.9 compatibility handling for Debian Trixie/glibc 2.41
    * isolate external CUDA, ROCm and oneAPI toolchains from Debian system libraries
    * update package dependencies, shlibs handling, cleanup rules and runtime RPATHs
    Jaroslav Mikulík
    Jaroslav Mikulík committed Aug 27, 2026
    Configuration menu
    Copy the full SHA
    bb4a154 View commit details
    Browse the repository at this point in the history
  2. fix: filled missing maintainer

    Jaroslav Mikulík
    Jaroslav Mikulík committed Aug 27, 2026
    Configuration menu
    Copy the full SHA
    e62d620 View commit details
    Browse the repository at this point in the history
  3. feat: DEB packages - Add modular CPU, Vulkan, CUDA, ROCm and SYCL bac…

    …kend packages
    
    * split llama-cpp-python into core and dynamically loadable backend packages
    * add CPU backend with multiple optimized CPU variants and OpenBLAS
    * add Vulkan backend
    * add co-installable CUDA 12.4, 12.9 and 13.3 backends using update-alternatives
    * add ROCm 7.14 HIP backend with support for current AMD GPU architectures
    * add Intel oneAPI SYCL backend with oneMKL, oneDNN and Level Zero support
    * add transitional packages for legacy package names
    * add backend alternatives handling and private shared-library dependency resolution
    * add CUDA 12.9 compatibility handling for Debian Trixie/glibc 2.41
    * isolate external CUDA, ROCm and oneAPI toolchains from Debian system libraries
    * update package dependencies, shlibs handling, cleanup rules and runtime RPATHs
    
    FIxed dependencies for CUDA 12.9
    Jaroslav Mikulík
    Jaroslav Mikulík committed Aug 27, 2026
    Configuration menu
    Copy the full SHA
    c6079e0 View commit details
    Browse the repository at this point in the history
  4. Merge branch 'deb-pkg'

    #Conflicts:
    #	README.md
    Jaroslav Mikulík
    Jaroslav Mikulík committed Aug 27, 2026
    Configuration menu
    Copy the full SHA
    46244d0 View commit details
    Browse the repository at this point in the history

Commits on Oct 10, 2026

  1. update debian package

    Jaroslav Mikulík
    Jaroslav Mikulík committed Oct 10, 2026
    Configuration menu
    Copy the full SHA
    c1bf8a4 View commit details
    Browse the repository at this point in the history
  2. Merge tag 'v0.3.36'

    Jaroslav Mikulík
    Jaroslav Mikulík committed Oct 10, 2026
    Configuration menu
    Copy the full SHA
    9b63be4 View commit details
    Browse the repository at this point in the history
  3. Update llama.cpp

    Jaroslav Mikulík
    Jaroslav Mikulík committed Oct 10, 2026
    Configuration menu
    Copy the full SHA
    3e46029 View commit details
    Browse the repository at this point in the history
  4. update debpackage

    Jaroslav Mikulík
    Jaroslav Mikulík committed Oct 10, 2026
    Configuration menu
    Copy the full SHA
    d8d1836 View commit details
    Browse the repository at this point in the history
  5. Update llama.cpp

    Jaroslav Mikulík
    Jaroslav Mikulík committed Oct 10, 2026
    Configuration menu
    Copy the full SHA
    64ac075 View commit details
    Browse the repository at this point in the history
  6. Configuration menu
    Copy the full SHA
    4951955 View commit details
    Browse the repository at this point in the history
  7. Merge branch 'debian-pkg'

    Jaroslav Mikulík
    Jaroslav Mikulík committed Oct 10, 2026
    Configuration menu
    Copy the full SHA
    cea8514 View commit details
    Browse the repository at this point in the history
  8. Merge remote-tracking branch 'refs/remotes/origin/main'

    # Conflicts:
    #	debian/changelog
    #	debian/control
    #	debian/patches/0001-debian-dynamic-backend-path.patch
    #	debian/rules
    Jaroslav Mikulík
    Jaroslav Mikulík committed Oct 10, 2026
    Configuration menu
    Copy the full SHA
    acaa60a View commit details
    Browse the repository at this point in the history
  9. fix: escape ROCm version checks in Debian rules

    Make interprets an unescaped dollar followed by a closing parenthesis as
    a one-character variable reference, removing both characters before the
    recipe reaches the shell. This leaves an unterminated regex group and
    aborts the ROCm build before its compiler probe.
    
    Double both end-anchor dollars so grep receives the intended SDK 10.1
    and HIP 7.16 expressions. Bump the Debian revision to 0:0.3.36-3.
    
    Validated actual make -n expansion and grep execution with matching
    versions and negative version-stream cases.
    ByCzech
    ByCzech committed Oct 10, 2026
    Configuration menu
    Copy the full SHA
    46d384e View commit details
    Browse the repository at this point in the history

Commits on Oct 11, 2026

  1. fix: exclude unsupported gfx1250 kernels from the ROCm backend

    ROCm Core SDK 10.1 ships gfx1250 device libraries, but the vendored GGML
    classifies all __GFX12__ targets as RDNA4. Its flash-attention kernels
    then use v_dot2_f32_f16 inline assembly and gfx12 WMMA builtins that the
    compiler rejects for gfx1250.
    
    Remove only gfx1250 from ROCM101_TARGETS, retaining the other 24 targets
    and gfx908/MI100 support. Keep flash attention enabled for the remaining
    targets and retain the pinned upstream sources. Bump the Debian revision
    to 0:0.3.36-4.
    
    Validated actual Make expansion, unchanged remaining target order,
    changelog parsing and mbox application. GPU compilation requires the
    external ROCm SDK and has not been run in this workspace.
    ByCzech
    ByCzech committed Oct 11, 2026
    Configuration menu
    Copy the full SHA
    aaf5ede View commit details
    Browse the repository at this point in the history
  2. fix: restore gfx1250 with compatible HIP kernel dispatch

    Separate gfx1250 from RDNA4 in both HIP device macros and host feature
    selection. Keep native integer dot4 and wave32 tile kernels, use FP32
    products for the FP16 dot2 fallback, and retain hipBLAS matrix hardware
    eligibility without selecting incompatible custom GGML WMMA kernels.
    
    Restore gfx1250 in the ROCm 10.1 target list. Carry the source changes as
    a Debian quilt patch so the upstream orig tarball stays unmodified.
    This follows the previous packaging patch that excluded gfx1250.
    
    Validation with the actual ROCm 10.1 clang compiler:
    - reproduce both original gfx1250 dot2 and WMMA compilation failures;
    - compile and link the complete ggml-hip backend for gfx1250;
    - compile the affected tile and MMA kernels for gfx1201 and gfx908;
    - compile a gfx1250 arithmetic/dispatch probe and run its host checks;
    - inspect generated gfx1250 ISA for native dot4 and FP32 products;
    - apply the Debian quilt series without fuzz and check the mbox with git am.
    
    No gfx1250 hardware inference or numerical validation was available.
    Native custom gfx1250 WMMA optimization remains separate work.
    ByCzech
    ByCzech committed Oct 11, 2026
    Configuration menu
    Copy the full SHA
    4a22fe2 View commit details
    Browse the repository at this point in the history
  3. fix: apply gfx1250 quilt patch across vendored revisions

    The RDNA4 hunk included GCN APU predicates introduced after the user's
    vendored llama.cpp revision 4df29be. dpkg-source rejected that hunk even
    though the RDNA4 predicate being replaced is identical in both revisions.
    
    Match the unique RDNA4 predicate without unrelated GCN/CDNA context.
    Preserve the original compatibility implementation and the upstream GCN
    classification on each revision. Keep the original source tarball intact.
    
    Validation:
    - dpkg-source --before-build applies both quilt patches without fuzz on
      pristine headers from 4df29be and 0c1e570;
    - generated 0c1e570 headers are byte-identical to the previous tested fix;
    - ROCm 10.1 clang compiles the failing gfx1250 flash tile and MMA kernels
      on 4df29be, plus MMVQ and a Q4_0 MMQ kernel;
    - a host/device dispatch probe compiles and its host assertions pass;
    - the mbox applies with git am and reproduces the same tree.
    
    GPU execution and numerical validation remain untested.
    ByCzech
    ByCzech committed Oct 11, 2026
    Configuration menu
    Copy the full SHA
    3e4b865 View commit details
    Browse the repository at this point in the history
  4. fix: synchronize vendored llama.cpp with the 0.3.36 bindings

    The Python 0.3.36 module binds llama_adapter_lora_init_from_file_ptr at
    import time. The retained 4df29be submodule does not export that API;
    upstream 0.3.36 pins 0c1e570 instead. Update the gitlink to that exact
    revision so the Python binding, core runtime, and backend plugins can
    be rebuilt from one matching source tree.
    
    Reject source headers lacking the required LoRA FILE API before any
    compilation, including stale source tarballs. Import the staged package
    before the separate accelerator builds; leave the source directory and
    clear library overrides so the check uses the packaged runtime.
    Declare the Python runtime dependencies required by that import check.
    
    Validation:
    - verify the gitlink pinned by the upstream 0.3.36 tag;
    - confirm the missing C API is absent in 4df29be and present in 0c1e570;
    - run the actual expanded Make header check against both revisions;
    - run the actual expanded staged-import recipe against the previously
      built 0c1e570 core and the 0.3.36 Python modules;
    - check the mbox with git am, including the resulting gitlink and tree.
    
    Unapply quilt patches before updating the submodule working directory,
    then reapply them. Rebuild the core and all plugins together. An orig
    source archive must also contain the matching upstream submodule.
    ByCzech
    ByCzech committed Oct 11, 2026
    Configuration menu
    Copy the full SHA
    b9ce865 View commit details
    Browse the repository at this point in the history
  5. fix: deleted obsolete file, update maintainer

    Jaroslav Mikulík
    Jaroslav Mikulík committed Oct 11, 2026
    Configuration menu
    Copy the full SHA
    043f9d6 View commit details
    Browse the repository at this point in the history
  6. fix: keep upstream sources clean for quilt packaging

    Jaroslav Mikulík
    Jaroslav Mikulík committed Oct 11, 2026
    Configuration menu
    Copy the full SHA
    4ca033b View commit details
    Browse the repository at this point in the history
Loading
⌂ Proxy