{"entries":[{"name":"CMakeLists.txt","mode":"0100644","size":2706,"lastCommit":{"hash":"b7feacf7f39d79518e6ca48a9fcf697c5244b585","oid":"b7feacf7f39d79518e6ca48a9fcf697c5244b585","parents":["6ad70c5a779f770a37b91c51d88daf30313fca08"],"author":{"name":"Kevin Pouget","email":"kpouget@redhat.com","when":"2026-01-28T10:49:40+01:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-01-28T17:49:40+08:00"},"message":"ggml: new backend for Virglrenderer API Remoting acceleration (v2) (#18718)\n\n","tree":"2fc4256fa14c6828c8176c20c1ccbf3a7ae6db6b"}},{"name":"apir_cs_ggml-rpc-front.cpp","mode":"0100644","size":3505,"lastCommit":{"hash":"015deb90485b9ebfff492e10fe7080a0439202b6","oid":"015deb90485b9ebfff492e10fe7080a0439202b6","parents":["2ceda3f6622661af839c19767705f33fb9f6cdd2"],"author":{"name":"Kevin Pouget","email":"kpouget@redhat.com","when":"2026-02-04T03:46:18+01:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-02-04T10:46:18+08:00"},"message":"ggml-virtgpu: make the code thread safe (#19204)\n\n* ggml-virtgpu: regenerate_remoting.py: add the ability to deprecate a function\n\n* ggml-virtgpu: deprecate buffer_type is_host remoting\n\nnot necessary\n\n* ggml-virtgpu: stop using static vars as cache\n\nThe static init isn't thread safe.\n\n* ggml-virtgpu: protect the use of the shared memory to transfer data\n\n* ggml-virtgpu: make the remote calls thread-safe\n\n* ggml-virtgpu: backend: don't continue if couldn't allocate the tensor memory\n\n* ggml-virtgpu: add a cleanup function for consistency\n\n* ggml-virtgpu: backend: don't crash if buft->iface.get_max_size is missing\n\n* fix style and ordering\n\n* Remove the static variable in apir_device_get_count\n\n* ggml-virtgpu: improve the logging\n\n* fix review minor formatting changes","tree":"44248faa120c04d95352f4a966efa82f6cbdfa7a"}},{"name":"backend","mode":"0040000","size":0,"lastCommit":{"hash":"153d324bcf86d220b235ca010eeb11213f32b5d1","oid":"153d324bcf86d220b235ca010eeb11213f32b5d1","parents":["b3df57286c336255fd02c2162b0b5fe3adc2965e"],"author":{"name":"Ruben Ortlam","email":"rortlam@redhat.com","when":"2026-08-11T08:20:46+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-08-11T09:20:46+03:00"},"message":"llama: add default load-mode auto, which avoids mmap on iGPUs (#26081)\n\n* llama: add new default load-mode auto which picks mmap unless a non-Metal iGPU is used\n\n* Update ggml/src/ggml-hexagon/ggml-hexagon.cpp\n\nCo-authored-by: Max Krasnyansky \n\n* set mmap_support to false on OpenCL backend\n\n* fix order of load modes\n\n* use -1 for auto\n\n* resolve load mode auto earlier to correctly pick gpu host or cpu memory\n\n* add load mode auto to llama-bench\n\n* bump virtgpu api version, regenerate docs\n\n---------\n\nCo-authored-by: Piotr Wilkin (ilintar) \nCo-authored-by: Max Krasnyansky \nCo-authored-by: Georgi Gerganov ","tree":"f34c7980beeeddf5776f0c96b55e78f620b02ff4"}},{"name":"ggml-backend-buffer-type.cpp","mode":"0100644","size":3519,"lastCommit":{"hash":"153d324bcf86d220b235ca010eeb11213f32b5d1","oid":"153d324bcf86d220b235ca010eeb11213f32b5d1","parents":["b3df57286c336255fd02c2162b0b5fe3adc2965e"],"author":{"name":"Ruben Ortlam","email":"rortlam@redhat.com","when":"2026-08-11T08:20:46+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-08-11T09:20:46+03:00"},"message":"llama: add default load-mode auto, which avoids mmap on iGPUs (#26081)\n\n* llama: add new default load-mode auto which picks mmap unless a non-Metal iGPU is used\n\n* Update ggml/src/ggml-hexagon/ggml-hexagon.cpp\n\nCo-authored-by: Max Krasnyansky \n\n* set mmap_support to false on OpenCL backend\n\n* fix order of load modes\n\n* use -1 for auto\n\n* resolve load mode auto earlier to correctly pick gpu host or cpu memory\n\n* add load mode auto to llama-bench\n\n* bump virtgpu api version, regenerate docs\n\n---------\n\nCo-authored-by: Piotr Wilkin (ilintar) \nCo-authored-by: Max Krasnyansky \nCo-authored-by: Georgi Gerganov ","tree":"f34c7980beeeddf5776f0c96b55e78f620b02ff4"}},{"name":"ggml-backend-buffer.cpp","mode":"0100644","size":5435,"lastCommit":{"hash":"d6f3030047f85a98b009189e76f441fe818ea44d","oid":"d6f3030047f85a98b009189e76f441fe818ea44d","parents":["009a1133268d040a7c574a7b9c95413b0be369a9"],"author":{"name":"Johannes Gäßler","email":"johannesg@5d6.de","when":"2026-04-09T16:42:19+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-04-09T16:42:19+02:00"},"message":"ggml: backend-agnostic tensor parallelism (experimental) (#19378)\n\n* ggml: backend-agnostic tensor parallelism\n\n* support for GPT-OSS, Qwen 3 MoE\n\n* partial Vulkan fix\n\n* add support for 4/8 GPUs\n\n* unconditional peer access\n\n* re-use buffers + ggml contexts\n\n* fix output pattern\n\n* NCCL support\n\n* GGML: HIP: add RCCL support\n\n* Remove shfl and AllReduce from backend interface\n\n* move allocation workaround out of ggml-alloc.c\n\n* 2d tensor set/get support\n\n* Fix the seg fault without NCCL\n\n* Apply suggestion from JohannesGaessler\n\n* support for tensor dims % n_devs != 0\n\n* fix view_offs scaling\n\n* arbitrary num. of GPUs/tensor split\n\n* fix compilation\n\n* better granularity estimate\n\n* Support device-specific host buffer types if all underlying backends expose the same type. This allows using pinned memory instead of pageable memory for CUDA.\n\nFix compilation errors.\n\n* partial Qwen 3 Next support\n\n* Fix qwen3 30b (#8)\n\n* Fix crash with Qwen-30B-A3B Q4_0\n\nQwen-30B-A3B Q4_0 has an intermediate dimension of 768. Using a granularity of 256 forces an uneven split between GPUs, which is not supported by the current implementation.\n\n* Decide block size based on tensor quantization type\n\n* Fix crashes due to KV cache serialization (#9)\n\nKV cache serialization requires non-zero offsets on the tensor. Add support in the meta backend to set/get a tensor with a non-zero offset.\n\n* metal : fix build (#7)\n\n* static memory allocations, fix usage count\n\n* fix tensor granularity\n\n* more even memory distribution\n\n* use BF16 for allreduce\n\n* rebase fixup\n\n* better error message for unsupported architectures\n\n* Fix device mismatch during scatter of allReduce. (#11)\n\nThere is a mismatch between the dst buffer device and the backend device, causing the use of sync copies\n\n* Enable the previous allreduce implementation. It is better in both perf and stability (#12)\n\n* delay AllReduce for Moe for less I/O\n\n* build : clean-up compile warnings\n\n* backend : move most of the meta backend API to ggml-backend-impl.h\n\n* cont : hide unused public API in the implementation\n\n* llama : use llama_device + remove ggml_backend_dev_is_meta()\n\n* ggml-backend : remove unused alloc include\n\n* minor : remove regex include\n\n* ggml : introduce ggml-ext.h for staging new APIs\n\n* rebase fixup\n\n* fix tests\n\n* llama : more robust logic for determining Meta devices (#16)\n\n* llama : more robust logic for determining Meta devices\n\n* cont : fix devs size check\n\nCo-authored-by: Johannes Gäßler \n\n* cont : fix log type\n\nCo-authored-by: Johannes Gäßler \n\n---------\n\nCo-authored-by: Johannes Gäßler \n\n* disable roundtrip for meta backend\n\n* fix arch selection\n\n* Qwen 3.5 support\n\n* fix Gemma 4 MoE\n\n* fix OpenVino, SYCL\n\n* fix test-llama-archs for CPU-only builds\n\n* Fix Qwen 3.5 MoE\n\n* disable meta backend tests for WebGPU\n\n* tests : filter CPU-based devices from the Meta backend tests (#17)\n\n* meta : formatting, naming, indentation (#18)\n\n* formatting : llama-model.cpp\n\n* formatting : ggml-ext.h\n\n* formatting : ggml-backend-meta.cpp\n\n* meta : add TODO\n\n* add documentation\n\n* better error messages\n\n* fix GPT-OSS\n\n---------\n\nCo-authored-by: Carl Philipp Klemm \nCo-authored-by: Gaurav Garg \nCo-authored-by: Georgi Gerganov ","tree":"b5a5a9103d2576517dadc371794151a06ef39e66"}},{"name":"ggml-backend-device.cpp","mode":"0100644","size":6133,"lastCommit":{"hash":"153d324bcf86d220b235ca010eeb11213f32b5d1","oid":"153d324bcf86d220b235ca010eeb11213f32b5d1","parents":["b3df57286c336255fd02c2162b0b5fe3adc2965e"],"author":{"name":"Ruben Ortlam","email":"rortlam@redhat.com","when":"2026-08-11T08:20:46+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-08-11T09:20:46+03:00"},"message":"llama: add default load-mode auto, which avoids mmap on iGPUs (#26081)\n\n* llama: add new default load-mode auto which picks mmap unless a non-Metal iGPU is used\n\n* Update ggml/src/ggml-hexagon/ggml-hexagon.cpp\n\nCo-authored-by: Max Krasnyansky \n\n* set mmap_support to false on OpenCL backend\n\n* fix order of load modes\n\n* use -1 for auto\n\n* resolve load mode auto earlier to correctly pick gpu host or cpu memory\n\n* add load mode auto to llama-bench\n\n* bump virtgpu api version, regenerate docs\n\n---------\n\nCo-authored-by: Piotr Wilkin (ilintar) \nCo-authored-by: Max Krasnyansky \nCo-authored-by: Georgi Gerganov ","tree":"f34c7980beeeddf5776f0c96b55e78f620b02ff4"}},{"name":"ggml-backend-reg.cpp","mode":"0100644","size":7095,"lastCommit":{"hash":"ffaafde16ffebd2853467f0dd833625a726ce08e","oid":"ffaafde16ffebd2853467f0dd833625a726ce08e","parents":["efba35a860954e78e72177b0d6b0c88b239d8cb1"],"author":{"name":"Kevin Pouget","email":"kpouget@redhat.com","when":"2026-02-26T13:00:57+01:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-02-26T20:00:57+08:00"},"message":"ggml-virtgpu: improve the reliability of the code (#19846)\n\n* ggml-virtgpu-backend: validate the consistency of the received objects\n\nThis patch adds consistency checks in the\nggml-virtgpu-backend (running on the host side) to ensure that the\ndata received from the guest is consistent (valid pointers, valid\nsizes and offsets).\n\n* ggml-virtgpu-backend: add fallback/skips for optional ggml backend methods\n\n```\n 1. bck->iface.synchronize(bck)\n 2. buft->iface.get_alloc_size(buft, op)\n 3. buft->iface.get_max_size(buft)\n```\n\nthese three methods are optional in the GGML interface. `get_max_size`\nwas already properly defaulted, but `backend sychronize` and `butf\nget_max_size` would have segfaulted the backend if not implemented.\n\n* ggml-virtgpu-backend: fix log format missing argument\n\n* ggml-virtgpu-backend: improve the abort message\n\n* ggml-virtgpu-backend: more safety checks\n\n* ggml-virtgpu-backend: new error code\n\n* ggml-virtgpu-backend: initialize all the error codes\n\n* ggml-virtgpu: add a missing comment generated by the code generator\n\n* ggml-virtgpu: add the '[virtgpu]' prefix to the device/buffer names\n\n* ggml-virtgpu: apir_device_buffer_from_ptr: improve the error message\n\n* ggml-virtgpu: shared: make it match the latest api_remoting.h of Virglrenderer APIR\n\n(still unmerged)\n\n* ggml-virtgpu: update the code generator to have dispatch_command_name in a host/guest shared file\n\n* ggml-virtgpu: REMOTE_CALL: fail if the backend returns an error\n\n* docs/backend/VirtGPU.md: indicate that the RAM+VRAM size is limed to 64 GB with libkrun\n\n* ggml-virtgpu: turn off clang-format header ordering for some of the files\n\nCompilation breaks when ordered alphabetically.\n\n* ggml-virtgpu: clang-format\n\n* ggml-virtgpu/backend/shared/api_remoting: better comments for the APIR return codes","tree":"1108fa7c54dcd0d4e44379807b15037d9e891a42"}},{"name":"ggml-backend.cpp","mode":"0100644","size":2601,"lastCommit":{"hash":"660b1b4bdc6fedc18e8c3d87a945ffb51f91c547","oid":"660b1b4bdc6fedc18e8c3d87a945ffb51f91c547","parents":["c20c44514a06a3fd70e3d2e8b830812047360e5b"],"author":{"name":"Ruben Ortlam","email":"rortlam@redhat.com","when":"2026-04-30T17:37:13+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-04-30T17:37:13+02:00"},"message":"vulkan: add get/set tensor 2d functions (#22514)\n\n* vulkan: add get/set_tensor_2d functions\n\n* fix backend interface comments\n\n* Update ggml/src/ggml-metal/ggml-metal.cpp\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"8f83772fd44aacda2b55e35ad3521b9820add96e"}},{"name":"ggml-remoting.h","mode":"0100644","size":2598,"lastCommit":{"hash":"ffaafde16ffebd2853467f0dd833625a726ce08e","oid":"ffaafde16ffebd2853467f0dd833625a726ce08e","parents":["efba35a860954e78e72177b0d6b0c88b239d8cb1"],"author":{"name":"Kevin Pouget","email":"kpouget@redhat.com","when":"2026-02-26T13:00:57+01:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-02-26T20:00:57+08:00"},"message":"ggml-virtgpu: improve the reliability of the code (#19846)\n\n* ggml-virtgpu-backend: validate the consistency of the received objects\n\nThis patch adds consistency checks in the\nggml-virtgpu-backend (running on the host side) to ensure that the\ndata received from the guest is consistent (valid pointers, valid\nsizes and offsets).\n\n* ggml-virtgpu-backend: add fallback/skips for optional ggml backend methods\n\n```\n 1. bck->iface.synchronize(bck)\n 2. buft->iface.get_alloc_size(buft, op)\n 3. buft->iface.get_max_size(buft)\n```\n\nthese three methods are optional in the GGML interface. `get_max_size`\nwas already properly defaulted, but `backend sychronize` and `butf\nget_max_size` would have segfaulted the backend if not implemented.\n\n* ggml-virtgpu-backend: fix log format missing argument\n\n* ggml-virtgpu-backend: improve the abort message\n\n* ggml-virtgpu-backend: more safety checks\n\n* ggml-virtgpu-backend: new error code\n\n* ggml-virtgpu-backend: initialize all the error codes\n\n* ggml-virtgpu: add a missing comment generated by the code generator\n\n* ggml-virtgpu: add the '[virtgpu]' prefix to the device/buffer names\n\n* ggml-virtgpu: apir_device_buffer_from_ptr: improve the error message\n\n* ggml-virtgpu: shared: make it match the latest api_remoting.h of Virglrenderer APIR\n\n(still unmerged)\n\n* ggml-virtgpu: update the code generator to have dispatch_command_name in a host/guest shared file\n\n* ggml-virtgpu: REMOTE_CALL: fail if the backend returns an error\n\n* docs/backend/VirtGPU.md: indicate that the RAM+VRAM size is limed to 64 GB with libkrun\n\n* ggml-virtgpu: turn off clang-format header ordering for some of the files\n\nCompilation breaks when ordered alphabetically.\n\n* ggml-virtgpu: clang-format\n\n* ggml-virtgpu/backend/shared/api_remoting: better comments for the APIR return codes","tree":"1108fa7c54dcd0d4e44379807b15037d9e891a42"}},{"name":"ggmlremoting_functions.yaml","mode":"0100644","size":4393,"lastCommit":{"hash":"015deb90485b9ebfff492e10fe7080a0439202b6","oid":"015deb90485b9ebfff492e10fe7080a0439202b6","parents":["2ceda3f6622661af839c19767705f33fb9f6cdd2"],"author":{"name":"Kevin Pouget","email":"kpouget@redhat.com","when":"2026-02-04T03:46:18+01:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-02-04T10:46:18+08:00"},"message":"ggml-virtgpu: make the code thread safe (#19204)\n\n* ggml-virtgpu: regenerate_remoting.py: add the ability to deprecate a function\n\n* ggml-virtgpu: deprecate buffer_type is_host remoting\n\nnot necessary\n\n* ggml-virtgpu: stop using static vars as cache\n\nThe static init isn't thread safe.\n\n* ggml-virtgpu: protect the use of the shared memory to transfer data\n\n* ggml-virtgpu: make the remote calls thread-safe\n\n* ggml-virtgpu: backend: don't continue if couldn't allocate the tensor memory\n\n* ggml-virtgpu: add a cleanup function for consistency\n\n* ggml-virtgpu: backend: don't crash if buft->iface.get_max_size is missing\n\n* fix style and ordering\n\n* Remove the static variable in apir_device_get_count\n\n* ggml-virtgpu: improve the logging\n\n* fix review minor formatting changes","tree":"44248faa120c04d95352f4a966efa82f6cbdfa7a"}},{"name":"include","mode":"0040000","size":0,"lastCommit":{"hash":"ffaafde16ffebd2853467f0dd833625a726ce08e","oid":"ffaafde16ffebd2853467f0dd833625a726ce08e","parents":["efba35a860954e78e72177b0d6b0c88b239d8cb1"],"author":{"name":"Kevin Pouget","email":"kpouget@redhat.com","when":"2026-02-26T13:00:57+01:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-02-26T20:00:57+08:00"},"message":"ggml-virtgpu: improve the reliability of the code (#19846)\n\n* ggml-virtgpu-backend: validate the consistency of the received objects\n\nThis patch adds consistency checks in the\nggml-virtgpu-backend (running on the host side) to ensure that the\ndata received from the guest is consistent (valid pointers, valid\nsizes and offsets).\n\n* ggml-virtgpu-backend: add fallback/skips for optional ggml backend methods\n\n```\n 1. bck->iface.synchronize(bck)\n 2. buft->iface.get_alloc_size(buft, op)\n 3. buft->iface.get_max_size(buft)\n```\n\nthese three methods are optional in the GGML interface. `get_max_size`\nwas already properly defaulted, but `backend sychronize` and `butf\nget_max_size` would have segfaulted the backend if not implemented.\n\n* ggml-virtgpu-backend: fix log format missing argument\n\n* ggml-virtgpu-backend: improve the abort message\n\n* ggml-virtgpu-backend: more safety checks\n\n* ggml-virtgpu-backend: new error code\n\n* ggml-virtgpu-backend: initialize all the error codes\n\n* ggml-virtgpu: add a missing comment generated by the code generator\n\n* ggml-virtgpu: add the '[virtgpu]' prefix to the device/buffer names\n\n* ggml-virtgpu: apir_device_buffer_from_ptr: improve the error message\n\n* ggml-virtgpu: shared: make it match the latest api_remoting.h of Virglrenderer APIR\n\n(still unmerged)\n\n* ggml-virtgpu: update the code generator to have dispatch_command_name in a host/guest shared file\n\n* ggml-virtgpu: REMOTE_CALL: fail if the backend returns an error\n\n* docs/backend/VirtGPU.md: indicate that the RAM+VRAM size is limed to 64 GB with libkrun\n\n* ggml-virtgpu: turn off clang-format header ordering for some of the files\n\nCompilation breaks when ordered alphabetically.\n\n* ggml-virtgpu: clang-format\n\n* ggml-virtgpu/backend/shared/api_remoting: better comments for the APIR return codes","tree":"1108fa7c54dcd0d4e44379807b15037d9e891a42"}},{"name":"regenerate_remoting.py","mode":"0100755","size":13170,"lastCommit":{"hash":"ffaafde16ffebd2853467f0dd833625a726ce08e","oid":"ffaafde16ffebd2853467f0dd833625a726ce08e","parents":["efba35a860954e78e72177b0d6b0c88b239d8cb1"],"author":{"name":"Kevin Pouget","email":"kpouget@redhat.com","when":"2026-02-26T13:00:57+01:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-02-26T20:00:57+08:00"},"message":"ggml-virtgpu: improve the reliability of the code (#19846)\n\n* ggml-virtgpu-backend: validate the consistency of the received objects\n\nThis patch adds consistency checks in the\nggml-virtgpu-backend (running on the host side) to ensure that the\ndata received from the guest is consistent (valid pointers, valid\nsizes and offsets).\n\n* ggml-virtgpu-backend: add fallback/skips for optional ggml backend methods\n\n```\n 1. bck->iface.synchronize(bck)\n 2. buft->iface.get_alloc_size(buft, op)\n 3. buft->iface.get_max_size(buft)\n```\n\nthese three methods are optional in the GGML interface. `get_max_size`\nwas already properly defaulted, but `backend sychronize` and `butf\nget_max_size` would have segfaulted the backend if not implemented.\n\n* ggml-virtgpu-backend: fix log format missing argument\n\n* ggml-virtgpu-backend: improve the abort message\n\n* ggml-virtgpu-backend: more safety checks\n\n* ggml-virtgpu-backend: new error code\n\n* ggml-virtgpu-backend: initialize all the error codes\n\n* ggml-virtgpu: add a missing comment generated by the code generator\n\n* ggml-virtgpu: add the '[virtgpu]' prefix to the device/buffer names\n\n* ggml-virtgpu: apir_device_buffer_from_ptr: improve the error message\n\n* ggml-virtgpu: shared: make it match the latest api_remoting.h of Virglrenderer APIR\n\n(still unmerged)\n\n* ggml-virtgpu: update the code generator to have dispatch_command_name in a host/guest shared file\n\n* ggml-virtgpu: REMOTE_CALL: fail if the backend returns an error\n\n* docs/backend/VirtGPU.md: indicate that the RAM+VRAM size is limed to 64 GB with libkrun\n\n* ggml-virtgpu: turn off clang-format header ordering for some of the files\n\nCompilation breaks when ordered alphabetically.\n\n* ggml-virtgpu: clang-format\n\n* ggml-virtgpu/backend/shared/api_remoting: better comments for the APIR return codes","tree":"1108fa7c54dcd0d4e44379807b15037d9e891a42"}},{"name":"virtgpu-apir.h","mode":"0100644","size":366,"lastCommit":{"hash":"b7feacf7f39d79518e6ca48a9fcf697c5244b585","oid":"b7feacf7f39d79518e6ca48a9fcf697c5244b585","parents":["6ad70c5a779f770a37b91c51d88daf30313fca08"],"author":{"name":"Kevin Pouget","email":"kpouget@redhat.com","when":"2026-01-28T10:49:40+01:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-01-28T17:49:40+08:00"},"message":"ggml: new backend for Virglrenderer API Remoting acceleration (v2) (#18718)\n\n","tree":"2fc4256fa14c6828c8176c20c1ccbf3a7ae6db6b"}},{"name":"virtgpu-forward-backend.cpp","mode":"0100644","size":2003,"lastCommit":{"hash":"ffaafde16ffebd2853467f0dd833625a726ce08e","oid":"ffaafde16ffebd2853467f0dd833625a726ce08e","parents":["efba35a860954e78e72177b0d6b0c88b239d8cb1"],"author":{"name":"Kevin Pouget","email":"kpouget@redhat.com","when":"2026-02-26T13:00:57+01:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-02-26T20:00:57+08:00"},"message":"ggml-virtgpu: improve the reliability of the code (#19846)\n\n* ggml-virtgpu-backend: validate the consistency of the received objects\n\nThis patch adds consistency checks in the\nggml-virtgpu-backend (running on the host side) to ensure that the\ndata received from the guest is consistent (valid pointers, valid\nsizes and offsets).\n\n* ggml-virtgpu-backend: add fallback/skips for optional ggml backend methods\n\n```\n 1. bck->iface.synchronize(bck)\n 2. buft->iface.get_alloc_size(buft, op)\n 3. buft->iface.get_max_size(buft)\n```\n\nthese three methods are optional in the GGML interface. `get_max_size`\nwas already properly defaulted, but `backend sychronize` and `butf\nget_max_size` would have segfaulted the backend if not implemented.\n\n* ggml-virtgpu-backend: fix log format missing argument\n\n* ggml-virtgpu-backend: improve the abort message\n\n* ggml-virtgpu-backend: more safety checks\n\n* ggml-virtgpu-backend: new error code\n\n* ggml-virtgpu-backend: initialize all the error codes\n\n* ggml-virtgpu: add a missing comment generated by the code generator\n\n* ggml-virtgpu: add the '[virtgpu]' prefix to the device/buffer names\n\n* ggml-virtgpu: apir_device_buffer_from_ptr: improve the error message\n\n* ggml-virtgpu: shared: make it match the latest api_remoting.h of Virglrenderer APIR\n\n(still unmerged)\n\n* ggml-virtgpu: update the code generator to have dispatch_command_name in a host/guest shared file\n\n* ggml-virtgpu: REMOTE_CALL: fail if the backend returns an error\n\n* docs/backend/VirtGPU.md: indicate that the RAM+VRAM size is limed to 64 GB with libkrun\n\n* ggml-virtgpu: turn off clang-format header ordering for some of the files\n\nCompilation breaks when ordered alphabetically.\n\n* ggml-virtgpu: clang-format\n\n* ggml-virtgpu/backend/shared/api_remoting: better comments for the APIR return codes","tree":"1108fa7c54dcd0d4e44379807b15037d9e891a42"}},{"name":"virtgpu-forward-buffer-type.cpp","mode":"0100644","size":3515,"lastCommit":{"hash":"ffaafde16ffebd2853467f0dd833625a726ce08e","oid":"ffaafde16ffebd2853467f0dd833625a726ce08e","parents":["efba35a860954e78e72177b0d6b0c88b239d8cb1"],"author":{"name":"Kevin Pouget","email":"kpouget@redhat.com","when":"2026-02-26T13:00:57+01:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-02-26T20:00:57+08:00"},"message":"ggml-virtgpu: improve the reliability of the code (#19846)\n\n* ggml-virtgpu-backend: validate the consistency of the received objects\n\nThis patch adds consistency checks in the\nggml-virtgpu-backend (running on the host side) to ensure that the\ndata received from the guest is consistent (valid pointers, valid\nsizes and offsets).\n\n* ggml-virtgpu-backend: add fallback/skips for optional ggml backend methods\n\n```\n 1. bck->iface.synchronize(bck)\n 2. buft->iface.get_alloc_size(buft, op)\n 3. buft->iface.get_max_size(buft)\n```\n\nthese three methods are optional in the GGML interface. `get_max_size`\nwas already properly defaulted, but `backend sychronize` and `butf\nget_max_size` would have segfaulted the backend if not implemented.\n\n* ggml-virtgpu-backend: fix log format missing argument\n\n* ggml-virtgpu-backend: improve the abort message\n\n* ggml-virtgpu-backend: more safety checks\n\n* ggml-virtgpu-backend: new error code\n\n* ggml-virtgpu-backend: initialize all the error codes\n\n* ggml-virtgpu: add a missing comment generated by the code generator\n\n* ggml-virtgpu: add the '[virtgpu]' prefix to the device/buffer names\n\n* ggml-virtgpu: apir_device_buffer_from_ptr: improve the error message\n\n* ggml-virtgpu: shared: make it match the latest api_remoting.h of Virglrenderer APIR\n\n(still unmerged)\n\n* ggml-virtgpu: update the code generator to have dispatch_command_name in a host/guest shared file\n\n* ggml-virtgpu: REMOTE_CALL: fail if the backend returns an error\n\n* docs/backend/VirtGPU.md: indicate that the RAM+VRAM size is limed to 64 GB with libkrun\n\n* ggml-virtgpu: turn off clang-format header ordering for some of the files\n\nCompilation breaks when ordered alphabetically.\n\n* ggml-virtgpu: clang-format\n\n* ggml-virtgpu/backend/shared/api_remoting: better comments for the APIR return codes","tree":"1108fa7c54dcd0d4e44379807b15037d9e891a42"}},{"name":"virtgpu-forward-buffer.cpp","mode":"0100644","size":5894,"lastCommit":{"hash":"ffaafde16ffebd2853467f0dd833625a726ce08e","oid":"ffaafde16ffebd2853467f0dd833625a726ce08e","parents":["efba35a860954e78e72177b0d6b0c88b239d8cb1"],"author":{"name":"Kevin Pouget","email":"kpouget@redhat.com","when":"2026-02-26T13:00:57+01:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-02-26T20:00:57+08:00"},"message":"ggml-virtgpu: improve the reliability of the code (#19846)\n\n* ggml-virtgpu-backend: validate the consistency of the received objects\n\nThis patch adds consistency checks in the\nggml-virtgpu-backend (running on the host side) to ensure that the\ndata received from the guest is consistent (valid pointers, valid\nsizes and offsets).\n\n* ggml-virtgpu-backend: add fallback/skips for optional ggml backend methods\n\n```\n 1. bck->iface.synchronize(bck)\n 2. buft->iface.get_alloc_size(buft, op)\n 3. buft->iface.get_max_size(buft)\n```\n\nthese three methods are optional in the GGML interface. `get_max_size`\nwas already properly defaulted, but `backend sychronize` and `butf\nget_max_size` would have segfaulted the backend if not implemented.\n\n* ggml-virtgpu-backend: fix log format missing argument\n\n* ggml-virtgpu-backend: improve the abort message\n\n* ggml-virtgpu-backend: more safety checks\n\n* ggml-virtgpu-backend: new error code\n\n* ggml-virtgpu-backend: initialize all the error codes\n\n* ggml-virtgpu: add a missing comment generated by the code generator\n\n* ggml-virtgpu: add the '[virtgpu]' prefix to the device/buffer names\n\n* ggml-virtgpu: apir_device_buffer_from_ptr: improve the error message\n\n* ggml-virtgpu: shared: make it match the latest api_remoting.h of Virglrenderer APIR\n\n(still unmerged)\n\n* ggml-virtgpu: update the code generator to have dispatch_command_name in a host/guest shared file\n\n* ggml-virtgpu: REMOTE_CALL: fail if the backend returns an error\n\n* docs/backend/VirtGPU.md: indicate that the RAM+VRAM size is limed to 64 GB with libkrun\n\n* ggml-virtgpu: turn off clang-format header ordering for some of the files\n\nCompilation breaks when ordered alphabetically.\n\n* ggml-virtgpu: clang-format\n\n* ggml-virtgpu/backend/shared/api_remoting: better comments for the APIR return codes","tree":"1108fa7c54dcd0d4e44379807b15037d9e891a42"}},{"name":"virtgpu-forward-device.cpp","mode":"0100644","size":5755,"lastCommit":{"hash":"153d324bcf86d220b235ca010eeb11213f32b5d1","oid":"153d324bcf86d220b235ca010eeb11213f32b5d1","parents":["b3df57286c336255fd02c2162b0b5fe3adc2965e"],"author":{"name":"Ruben Ortlam","email":"rortlam@redhat.com","when":"2026-08-11T08:20:46+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-08-11T09:20:46+03:00"},"message":"llama: add default load-mode auto, which avoids mmap on iGPUs (#26081)\n\n* llama: add new default load-mode auto which picks mmap unless a non-Metal iGPU is used\n\n* Update ggml/src/ggml-hexagon/ggml-hexagon.cpp\n\nCo-authored-by: Max Krasnyansky \n\n* set mmap_support to false on OpenCL backend\n\n* fix order of load modes\n\n* use -1 for auto\n\n* resolve load mode auto earlier to correctly pick gpu host or cpu memory\n\n* add load mode auto to llama-bench\n\n* bump virtgpu api version, regenerate docs\n\n---------\n\nCo-authored-by: Piotr Wilkin (ilintar) \nCo-authored-by: Max Krasnyansky \nCo-authored-by: Georgi Gerganov ","tree":"f34c7980beeeddf5776f0c96b55e78f620b02ff4"}},{"name":"virtgpu-forward-impl.h","mode":"0100644","size":2866,"lastCommit":{"hash":"ffaafde16ffebd2853467f0dd833625a726ce08e","oid":"ffaafde16ffebd2853467f0dd833625a726ce08e","parents":["efba35a860954e78e72177b0d6b0c88b239d8cb1"],"author":{"name":"Kevin Pouget","email":"kpouget@redhat.com","when":"2026-02-26T13:00:57+01:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-02-26T20:00:57+08:00"},"message":"ggml-virtgpu: improve the reliability of the code (#19846)\n\n* ggml-virtgpu-backend: validate the consistency of the received objects\n\nThis patch adds consistency checks in the\nggml-virtgpu-backend (running on the host side) to ensure that the\ndata received from the guest is consistent (valid pointers, valid\nsizes and offsets).\n\n* ggml-virtgpu-backend: add fallback/skips for optional ggml backend methods\n\n```\n 1. bck->iface.synchronize(bck)\n 2. buft->iface.get_alloc_size(buft, op)\n 3. buft->iface.get_max_size(buft)\n```\n\nthese three methods are optional in the GGML interface. `get_max_size`\nwas already properly defaulted, but `backend sychronize` and `butf\nget_max_size` would have segfaulted the backend if not implemented.\n\n* ggml-virtgpu-backend: fix log format missing argument\n\n* ggml-virtgpu-backend: improve the abort message\n\n* ggml-virtgpu-backend: more safety checks\n\n* ggml-virtgpu-backend: new error code\n\n* ggml-virtgpu-backend: initialize all the error codes\n\n* ggml-virtgpu: add a missing comment generated by the code generator\n\n* ggml-virtgpu: add the '[virtgpu]' prefix to the device/buffer names\n\n* ggml-virtgpu: apir_device_buffer_from_ptr: improve the error message\n\n* ggml-virtgpu: shared: make it match the latest api_remoting.h of Virglrenderer APIR\n\n(still unmerged)\n\n* ggml-virtgpu: update the code generator to have dispatch_command_name in a host/guest shared file\n\n* ggml-virtgpu: REMOTE_CALL: fail if the backend returns an error\n\n* docs/backend/VirtGPU.md: indicate that the RAM+VRAM size is limed to 64 GB with libkrun\n\n* ggml-virtgpu: turn off clang-format header ordering for some of the files\n\nCompilation breaks when ordered alphabetically.\n\n* ggml-virtgpu: clang-format\n\n* ggml-virtgpu/backend/shared/api_remoting: better comments for the APIR return codes","tree":"1108fa7c54dcd0d4e44379807b15037d9e891a42"}},{"name":"virtgpu-forward.gen.h","mode":"0100644","size":3687,"lastCommit":{"hash":"153d324bcf86d220b235ca010eeb11213f32b5d1","oid":"153d324bcf86d220b235ca010eeb11213f32b5d1","parents":["b3df57286c336255fd02c2162b0b5fe3adc2965e"],"author":{"name":"Ruben Ortlam","email":"rortlam@redhat.com","when":"2026-08-11T08:20:46+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-08-11T09:20:46+03:00"},"message":"llama: add default load-mode auto, which avoids mmap on iGPUs (#26081)\n\n* llama: add new default load-mode auto which picks mmap unless a non-Metal iGPU is used\n\n* Update ggml/src/ggml-hexagon/ggml-hexagon.cpp\n\nCo-authored-by: Max Krasnyansky \n\n* set mmap_support to false on OpenCL backend\n\n* fix order of load modes\n\n* use -1 for auto\n\n* resolve load mode auto earlier to correctly pick gpu host or cpu memory\n\n* add load mode auto to llama-bench\n\n* bump virtgpu api version, regenerate docs\n\n---------\n\nCo-authored-by: Piotr Wilkin (ilintar) \nCo-authored-by: Max Krasnyansky \nCo-authored-by: Georgi Gerganov ","tree":"f34c7980beeeddf5776f0c96b55e78f620b02ff4"}},{"name":"virtgpu-shm.cpp","mode":"0100644","size":2741,"lastCommit":{"hash":"09294365a9d7a2b786584d59525b034622c3ed81","oid":"09294365a9d7a2b786584d59525b034622c3ed81","parents":["63d93d17336e41e4cc73a64451e5b1d2477abdb1"],"author":{"name":"JusteLeo","email":"leonard.adamo66@gmail.com","when":"2026-05-02T15:28:50+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-05-02T21:28:50+08:00"},"message":"ggml-virtgpu: fix circular dependency in headers (#22557)","tree":"578e0054fb25d8b802e07047cc2c1204a3478e2f"}},{"name":"virtgpu-shm.h","mode":"0100644","size":423,"lastCommit":{"hash":"b7feacf7f39d79518e6ca48a9fcf697c5244b585","oid":"b7feacf7f39d79518e6ca48a9fcf697c5244b585","parents":["6ad70c5a779f770a37b91c51d88daf30313fca08"],"author":{"name":"Kevin Pouget","email":"kpouget@redhat.com","when":"2026-01-28T10:49:40+01:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-01-28T17:49:40+08:00"},"message":"ggml: new backend for Virglrenderer API Remoting acceleration (v2) (#18718)\n\n","tree":"2fc4256fa14c6828c8176c20c1ccbf3a7ae6db6b"}},{"name":"virtgpu-utils.cpp","mode":"0100644","size":5955,"lastCommit":{"hash":"b7feacf7f39d79518e6ca48a9fcf697c5244b585","oid":"b7feacf7f39d79518e6ca48a9fcf697c5244b585","parents":["6ad70c5a779f770a37b91c51d88daf30313fca08"],"author":{"name":"Kevin Pouget","email":"kpouget@redhat.com","when":"2026-01-28T10:49:40+01:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-01-28T17:49:40+08:00"},"message":"ggml: new backend for Virglrenderer API Remoting acceleration (v2) (#18718)\n\n","tree":"2fc4256fa14c6828c8176c20c1ccbf3a7ae6db6b"}},{"name":"virtgpu-utils.h","mode":"0100644","size":2186,"lastCommit":{"hash":"b7feacf7f39d79518e6ca48a9fcf697c5244b585","oid":"b7feacf7f39d79518e6ca48a9fcf697c5244b585","parents":["6ad70c5a779f770a37b91c51d88daf30313fca08"],"author":{"name":"Kevin Pouget","email":"kpouget@redhat.com","when":"2026-01-28T10:49:40+01:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-01-28T17:49:40+08:00"},"message":"ggml: new backend for Virglrenderer API Remoting acceleration (v2) (#18718)\n\n","tree":"2fc4256fa14c6828c8176c20c1ccbf3a7ae6db6b"}},{"name":"virtgpu.cpp","mode":"0100644","size":19330,"lastCommit":{"hash":"09294365a9d7a2b786584d59525b034622c3ed81","oid":"09294365a9d7a2b786584d59525b034622c3ed81","parents":["63d93d17336e41e4cc73a64451e5b1d2477abdb1"],"author":{"name":"JusteLeo","email":"leonard.adamo66@gmail.com","when":"2026-05-02T15:28:50+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-05-02T21:28:50+08:00"},"message":"ggml-virtgpu: fix circular dependency in headers (#22557)","tree":"578e0054fb25d8b802e07047cc2c1204a3478e2f"}},{"name":"virtgpu.h","mode":"0100644","size":3189,"lastCommit":{"hash":"09294365a9d7a2b786584d59525b034622c3ed81","oid":"09294365a9d7a2b786584d59525b034622c3ed81","parents":["63d93d17336e41e4cc73a64451e5b1d2477abdb1"],"author":{"name":"JusteLeo","email":"leonard.adamo66@gmail.com","when":"2026-05-02T15:28:50+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-05-02T21:28:50+08:00"},"message":"ggml-virtgpu: fix circular dependency in headers (#22557)","tree":"578e0054fb25d8b802e07047cc2c1204a3478e2f"}}],"lastCommit":{"hash":"153d324bcf86d220b235ca010eeb11213f32b5d1","oid":"153d324bcf86d220b235ca010eeb11213f32b5d1","parents":["b3df57286c336255fd02c2162b0b5fe3adc2965e"],"author":{"name":"Ruben Ortlam","email":"rortlam@redhat.com","when":"2026-08-11T08:20:46+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-08-11T09:20:46+03:00"},"message":"llama: add default load-mode auto, which avoids mmap on iGPUs (#26081)\n\n* llama: add new default load-mode auto which picks mmap unless a non-Metal iGPU is used\n\n* Update ggml/src/ggml-hexagon/ggml-hexagon.cpp\n\nCo-authored-by: Max Krasnyansky \n\n* set mmap_support to false on OpenCL backend\n\n* fix order of load modes\n\n* use -1 for auto\n\n* resolve load mode auto earlier to correctly pick gpu host or cpu memory\n\n* add load mode auto to llama-bench\n\n* bump virtgpu api version, regenerate docs\n\n---------\n\nCo-authored-by: Piotr Wilkin (ilintar) \nCo-authored-by: Max Krasnyansky \nCo-authored-by: Georgi Gerganov ","tree":"f34c7980beeeddf5776f0c96b55e78f620b02ff4"}}