{"entries":[{"name":"action.yml","mode":"0100644","size":752,"lastCommit":{"hash":"9789c4ecdc01d571331c14e5197514b53839de4b","oid":"9789c4ecdc01d571331c14e5197514b53839de4b","parents":["77e20cc1076f351f6ba37b545a20b21c6207cbe5"],"author":{"name":"Zijun Yu","email":"zijun.yu.joey@gmail.com","when":"2026-03-14T13:56:55+08:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-03-14T07:56:55+02:00"},"message":"ggml : add OpenVINO backend (#15307)\n\n* Update build doc\n\n* Add cgraph tensor output name to OV op name\n\n* Update openvino build instructions\n\n* Add initial NPU support\n\n* draft NPU support version 2: prefill + kvcache\n\n* NPU support version 2: prefill + kvcache\n\n* Change due to ggml cgraph changes, not correct yet\n\n* Change due to ggml cgraph changes, llama-3.2 CPU work\n\n* Add AMD64 to CMakeLists\n\n* Change due to ggml cgraph changes, all device work\n\n* Refactor: clean, fix warning\n\n* Update clang-format\n\n* Statful transformation for CPU GPU\n\n* Add SwiGLU\n\n* Fuse to SDPA\n\n* Replace Concat with Broadcast in MulMat for GQA\n\n* Pull out indices creation for kv cache update\n\n* Refactor: remove past_token_len from extra_inputs\n\n* Fix Phi3 SwiGLU and SoftMax\n\n* Pull out sin cos from rope\n\n* Reduce memory: free ov weights node after graph conversion\n\n* Fix CPY due to cgraph change\n\n* Added OpenVINO CI/CD. Updated docs\n\n* Fix llama-cli\n\n* Fix Phi3 ROPE; Add test-backend-ops\n\n* Fix NPU\n\n* Fix llama-bench; Clang-format\n\n* Fix llama-perplexity\n\n* temp. changes for mark decomp\n\n* matmul in fp32\n\n* mulmat input conversion fix\n\n* mulmat type conversion update\n\n* add mark decomp pass\n\n* Revert changes in fuse_to_sdpa\n\n* Update build.md\n\n* Fix test-backend-ops\n\n* Skip test-thread-safety; Run ctest only in ci/run.sh\n\n* Use CiD for NPU\n\n* Optimize tensor conversion, improve TTFT\n\n* Support op SET_ROWS\n\n* Fix NPU\n\n* Remove CPY\n\n* Fix test-backend-ops\n\n* Minor updates for raising PR\n\n* Perf: RMS fused to OV internal RMS op\n\n* Fix after rebasing\n\n- Layout of cache k and cache v are unified: [seq, n_head, head_size]\n- Add CPY and FLASH_ATTN_EXT, flash attn is not used yet\n- Skip test-backend-ops due to flash attn test crash\n- Add mutex around graph conversion to avoid test-thread-safety fali in the future\n- Update NPU config\n- Update GPU config to disable SDPA opt to make phi-3 run\n\n* Change openvino device_type to GPU; Enable flash_attn\n\n* Update supports_buft and supports_op for quantized models\n\n* Add quant weight conversion functions from genai gguf reader\n\n* Quant models run with accuracy issue\n\n* Fix accuracy: disable cpu_repack\n\n* Fix CI; Disable test-backend-ops\n\n* Fix Q4_1\n\n* Fix test-backend-ops: Treat quantized tensors as weights\n\n* Add NPU Q4_0 support\n\n* NPU perf: eliminate zp\n\n* Dequantize q4_1 q4_k q6_k for NPU\n\n* Add custom quant type: q8_1_c, q4_0_128\n\n* Set m_is_static=false as default in decoder\n\n* Simpilfy translation of get_rows\n\n* Fix after rebasing\n\n* Improve debug util; Eliminate nop ReshapeReshape\n\n* STYLE: make get_types_to_requant a function\n\n* Support BF16 model\n\n* Fix NPU compile\n\n* WA for npu 1st token acc issue\n\n* Apply EliminateZP only for npu\n\n* Add GeGLU\n\n* Fix Hunyuan\n\n* Support iSWA\n\n* Fix NPU accuracy\n\n* Fix ROPE accuracy when freq_scale != 1\n\n* Minor: not add attention_size_swa for non-swa model\n\n* Minor refactor\n\n* Add Q5_K to support phi-3-q4_k_m\n\n* Requantize Q6_K (gs16) to gs32 on GPU\n\n* Fix after rebasing\n\n* Always apply Eliminate_ZP to fix GPU compile issue on some platforms\n\n* kvcachefusion support\n\n* env variable GGML_OPENVINO_DISABLE_SDPA_OPTIMIZATION added\n\n* Fix for Phi3\n\n* Fix llama-cli (need to run with --no-warmup)\n\n* Fix add_sliced_mask; Revert mulmat, softmax; Remove input attention_size, iSWA model not working\n\n* fix after rebasing\n\n* Fix llama-3-8b and phi3-mini q4_0 NPU\n\n* Update to OV-2025.3 and CMakeLists.txt\n\n* Add OV CI cache\n\n* Apply CISC review and update CI to OV2025.3\n\n* Update CI to run OV dep install before build\n\n* Update OV dockerfile to use OV2025.3 and update build docs\n\n* Style: use switch in supports_ops\n\n* Style: middle ptr and ref align, omit optional struct keyword\n\n* NPU Unify PD (#14)\n\n* Stateless. Fix llama-cli llama-server\n\n* Simplify broadcast op in attention\n\n* Replace get_output_tensor+memcpy with set_output_tensor\n\n* NPU unify PD. Unify dynamic and static dims\n\n* Clean placeholders in ggml-openvino.cpp\n\n* NPU unify PD (handled internally)\n\n* change graph to 4d, support multi sequences\n\n* Fix llama-bench\n\n* Fix NPU\n\n* Update ggml-decoder.cpp\n\nHitting error while compiling on windows:\n\nerror C3861: 'unsetenv': identifier not found\n\nReason: unsetenv() is a POSIX function; it doesn’t exist on Windows. Visual Studio (MSVC) won’t recognize it.\n\nProposed fix: Use _putenv_s() (Windows equivalent)\nThis is supported by MSVC and achieves the same effect: it removes the environment variable from the process environment.\n\nThis keeps cross-platform compatibility.\n\n* Update ggml-decoder.cpp\n\n* Update ggml-decoder.cpp\n\n* Update ggml-decoder.cpp\n\n* Update ggml-decoder.cpp\n\n* Update ggml-decoder.cpp\n\n* Remove the second decoder for node. Moving the function into the model decoder\n\n* Fix error for naive\n\n* NPU prefill chunking\n\n* NPU fix llama-bench\n\n* fallback naive run with accuracy issue\n\n* NPU support llma-perplexity -b 512 --no-warmup\n\n* Refactor: split ov_graph_compute for dynamic and static\n\n* remove unused API GgmlOvDecoder::get_output_stride(const std::string & name)\n\n* minor update due to ov 2025.4\n\n* remove unused API GgmlOvDecoder::get_output_names()\n\n* remove unused API get_output_shape(const std::string & name)\n\n* Modified API GgmlOvDecoder::get_output_type(const std::string & name)\n\n* Removed API GgmlOvDecoder::get_output_op_params(const std::string & name)\n\n* Removed API get_output_ggml_tensor(const std::string & name)\n\n* Removed API m_outputs\n\n* Removed m_output_names\n\n* Removed API GgmlOvDecoder::get_input_names()\n\n* Removed API GgmlOvDecoder::get_input_stride(const std::string& name)\n\n* Removed API get_input_type\n\n* Removed API get_input_type\n\n* Removed API GgmlOvDecoder::get_input_shape(const std::string & name)\n\n* Removed API GgmlOvDecoder::get_input_op_params(const std::string & name)\n\n* Fix error for decoder cache\n\n* Reuse cached decoder\n\n* GPU remove Q6_K requantization\n\n* NPU fix wrong model output shape\n\n* NPU fix q4 perf regression\n\n* Remove unused variable nodes\n\n* Fix decoder can_reuse for llama-bench\n\n* Update build.md for Windows\n\n* backend buffer: allocate on host\n\n* Use shared_buffer for GPU NPU; Refactor\n\n* Add ov_backend_host_buffer; Use cached remote context\n\n* Put kvcache on GPU\n\n* Use ggml_aligned_malloc\n\n* only use remote tensor for kvcache\n\n* only use remote tensor for kvcache for GPU\n\n* FIX: use remote tensor from singleton\n\n* Update build.md to include OpenCL\n\n* NPU always requant to q4_0_128\n\n* Optimize symmetric quant weight extraction: use single zp\n\n* Use Q8_0_C in token embd, lm_head, and for 5 and 6 bits quant\n\n* Update build.md\n\n* Support -ctk f32\n\n* Initial stateful graph support\n\n* Update ggml/src/ggml-openvino/ggml-decoder.cpp\n\nCo-authored-by: Yamini Nimmagadda \n\n* code cleanup\n\n* npu perf fix\n\n* requant to f16 for Q6 embed on NPU\n\n* Update ggml/src/ggml-openvino/ggml-decoder.cpp\n\n* Update ggml/src/ggml-openvino/ggml-openvino-extra.cpp\n\n* Create OPENVINO.md in llama.cpp backend docs\n\n* Update OPENVINO.md\n\n* Update OPENVINO.md\n\n* Update OPENVINO.md\n\n* Update build.md\n\n* Update OPENVINO.md\n\n* Update OPENVINO.md\n\n* Update OPENVINO.md\n\n* kq_mask naming fix\n\n* Syntax correction for workflows build file\n\n* Change ov backend buffer is_host to false\n\n* Fix llama-bench -p -n where p<=256\n\n* Fix --direct-io 0\n\n* Don't put kvcache on GPU in stateful mode\n\n* Remove hardcode names\n\n* Fix stateful shapes\n\n* Simplification for stateful and update output shape processing\n\n* Remove hardcode names\n\n* Avoid re-compilation in llama-bench\n\n* Extract zp directly instead of bias\n\n* Refactor weight tensor processing\n\n* create_weight_node accept non-ov backend buffer\n\n* remove changes in llama-graph.cpp\n\n* stateful masking fix (#38)\n\nFix for stateful accuracy issues and cl_out_of_resources error in stateful GPU with larger context sizes.\n\n* Fix test-backend-ops crash glu, get_rows, scale, rms_norm, add\n\n* hardcoded name handling for rope_freqs.weight\n\n* Suppress logging and add error handling to allow test-backend-ops to complete\n\n* Fix MUL_MAT with broadcast; Add unsupported MUL_MAT FLASH_ATTN cases\n\n* Use bias instead of zp in test-backend-ops\n\n* Update OV in CI, Add OV CI Tests in GH Actions\n\n* Temp fix for multithreading bug\n\n* Update OV CI, fix review suggestions.\n\n* fix editorconfig-checker, update docs\n\n* Fix tabs to spaces for editorconfig-checker\n\n* fix editorconfig-checker\n\n* Update docs\n\n* updated model link to be GGUF model links\n\n* Remove GGML_CPU_REPACK=OFF\n\n* Skip permuted ADD and MUL\n\n* Removed static variables from utils.cpp\n\n* Removed initializing non-existing variable\n\n* Remove unused structs\n\n* Fix test-backend-ops for OV GPU\n\n* unify api calling\n\n* Update utils.cpp\n\n* When the dim is dynamic, throw an error, need to is stastic forst\n\n* Add interface compute_model_outputs(), which get the model output through computing the node use count & status in the cgraph to avoid the flag using\n\n* No need to return\n\n* Fix test-backend-ops for OV GPU LNL\n\n* Fix test-thread-safety\n\n* use the shape from infer request of output tensor create to avoid issue\n\n* fix dynamic output shape issue\n\n* fix issue for the unused node in tests\n\n* Remove unused lock\n\n* Add comment\n\n* Update openvino docs\n\n* update to OV release version 2026.0\n\n* add ci ov-gpu self hosted runner\n\n* fix editorconfig\n\n* Fix perplexity\n\n* Rewrite the model inputs finding mechanism (#54)\n\n* Rewrite the model inputs finding logistic\n\n* Put stateful shape handle in get input shape\n\n* Put the iteration logistic in func\n\n* Added ggml-ci-intel-openvino-gpu and doc update\n\n* .hpp files converted to .h\n\n* fix ggml-ci-x64-intel-openvino-gpu\n\n* Fix for stateful execution bug in llama-bench\n\n* Minor updates after stateful llama-bench fix\n\n* Update ggml/src/ggml-openvino/utils.cpp\n\nCo-authored-by: Yamini Nimmagadda \n\n* Remove multiple get_shape calls\n\n* Bring back mutex into compute\n\n* Fix VIEW op, which slice the input node\n\n* Added token_len_per_seq existence check before slicing masks and moved node retrieval inside guarded block to prevent missing-key access\n\n* Temp. fix for test requant errors\n\n* Update to OV ggml-ci to low-perf\n\n* ci : temporary disable \"test-llama-archs\"\n\n* ci : cache v4 -> v5, checkout v4 -> v6, fix runner tag\n\n* docs : update url\n\n* Fix OV link in docker and Update docs\n\n---------\n\nCo-authored-by: Ravi Panchumarthy \nCo-authored-by: Cavus Mustafa \nCo-authored-by: Arshath \nCo-authored-by: XuejunZhai \nCo-authored-by: Yamini Nimmagadda \nCo-authored-by: Xuejun Zhai \nCo-authored-by: Georgi Gerganov ","tree":"72527f32be2bb161ca1878b0481b08ee76ab1f28"}}],"lastCommit":{"hash":"9789c4ecdc01d571331c14e5197514b53839de4b","oid":"9789c4ecdc01d571331c14e5197514b53839de4b","parents":["77e20cc1076f351f6ba37b545a20b21c6207cbe5"],"author":{"name":"Zijun Yu","email":"zijun.yu.joey@gmail.com","when":"2026-03-14T13:56:55+08:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-03-14T07:56:55+02:00"},"message":"ggml : add OpenVINO backend (#15307)\n\n* Update build doc\n\n* Add cgraph tensor output name to OV op name\n\n* Update openvino build instructions\n\n* Add initial NPU support\n\n* draft NPU support version 2: prefill + kvcache\n\n* NPU support version 2: prefill + kvcache\n\n* Change due to ggml cgraph changes, not correct yet\n\n* Change due to ggml cgraph changes, llama-3.2 CPU work\n\n* Add AMD64 to CMakeLists\n\n* Change due to ggml cgraph changes, all device work\n\n* Refactor: clean, fix warning\n\n* Update clang-format\n\n* Statful transformation for CPU GPU\n\n* Add SwiGLU\n\n* Fuse to SDPA\n\n* Replace Concat with Broadcast in MulMat for GQA\n\n* Pull out indices creation for kv cache update\n\n* Refactor: remove past_token_len from extra_inputs\n\n* Fix Phi3 SwiGLU and SoftMax\n\n* Pull out sin cos from rope\n\n* Reduce memory: free ov weights node after graph conversion\n\n* Fix CPY due to cgraph change\n\n* Added OpenVINO CI/CD. Updated docs\n\n* Fix llama-cli\n\n* Fix Phi3 ROPE; Add test-backend-ops\n\n* Fix NPU\n\n* Fix llama-bench; Clang-format\n\n* Fix llama-perplexity\n\n* temp. changes for mark decomp\n\n* matmul in fp32\n\n* mulmat input conversion fix\n\n* mulmat type conversion update\n\n* add mark decomp pass\n\n* Revert changes in fuse_to_sdpa\n\n* Update build.md\n\n* Fix test-backend-ops\n\n* Skip test-thread-safety; Run ctest only in ci/run.sh\n\n* Use CiD for NPU\n\n* Optimize tensor conversion, improve TTFT\n\n* Support op SET_ROWS\n\n* Fix NPU\n\n* Remove CPY\n\n* Fix test-backend-ops\n\n* Minor updates for raising PR\n\n* Perf: RMS fused to OV internal RMS op\n\n* Fix after rebasing\n\n- Layout of cache k and cache v are unified: [seq, n_head, head_size]\n- Add CPY and FLASH_ATTN_EXT, flash attn is not used yet\n- Skip test-backend-ops due to flash attn test crash\n- Add mutex around graph conversion to avoid test-thread-safety fali in the future\n- Update NPU config\n- Update GPU config to disable SDPA opt to make phi-3 run\n\n* Change openvino device_type to GPU; Enable flash_attn\n\n* Update supports_buft and supports_op for quantized models\n\n* Add quant weight conversion functions from genai gguf reader\n\n* Quant models run with accuracy issue\n\n* Fix accuracy: disable cpu_repack\n\n* Fix CI; Disable test-backend-ops\n\n* Fix Q4_1\n\n* Fix test-backend-ops: Treat quantized tensors as weights\n\n* Add NPU Q4_0 support\n\n* NPU perf: eliminate zp\n\n* Dequantize q4_1 q4_k q6_k for NPU\n\n* Add custom quant type: q8_1_c, q4_0_128\n\n* Set m_is_static=false as default in decoder\n\n* Simpilfy translation of get_rows\n\n* Fix after rebasing\n\n* Improve debug util; Eliminate nop ReshapeReshape\n\n* STYLE: make get_types_to_requant a function\n\n* Support BF16 model\n\n* Fix NPU compile\n\n* WA for npu 1st token acc issue\n\n* Apply EliminateZP only for npu\n\n* Add GeGLU\n\n* Fix Hunyuan\n\n* Support iSWA\n\n* Fix NPU accuracy\n\n* Fix ROPE accuracy when freq_scale != 1\n\n* Minor: not add attention_size_swa for non-swa model\n\n* Minor refactor\n\n* Add Q5_K to support phi-3-q4_k_m\n\n* Requantize Q6_K (gs16) to gs32 on GPU\n\n* Fix after rebasing\n\n* Always apply Eliminate_ZP to fix GPU compile issue on some platforms\n\n* kvcachefusion support\n\n* env variable GGML_OPENVINO_DISABLE_SDPA_OPTIMIZATION added\n\n* Fix for Phi3\n\n* Fix llama-cli (need to run with --no-warmup)\n\n* Fix add_sliced_mask; Revert mulmat, softmax; Remove input attention_size, iSWA model not working\n\n* fix after rebasing\n\n* Fix llama-3-8b and phi3-mini q4_0 NPU\n\n* Update to OV-2025.3 and CMakeLists.txt\n\n* Add OV CI cache\n\n* Apply CISC review and update CI to OV2025.3\n\n* Update CI to run OV dep install before build\n\n* Update OV dockerfile to use OV2025.3 and update build docs\n\n* Style: use switch in supports_ops\n\n* Style: middle ptr and ref align, omit optional struct keyword\n\n* NPU Unify PD (#14)\n\n* Stateless. Fix llama-cli llama-server\n\n* Simplify broadcast op in attention\n\n* Replace get_output_tensor+memcpy with set_output_tensor\n\n* NPU unify PD. Unify dynamic and static dims\n\n* Clean placeholders in ggml-openvino.cpp\n\n* NPU unify PD (handled internally)\n\n* change graph to 4d, support multi sequences\n\n* Fix llama-bench\n\n* Fix NPU\n\n* Update ggml-decoder.cpp\n\nHitting error while compiling on windows:\n\nerror C3861: 'unsetenv': identifier not found\n\nReason: unsetenv() is a POSIX function; it doesn’t exist on Windows. Visual Studio (MSVC) won’t recognize it.\n\nProposed fix: Use _putenv_s() (Windows equivalent)\nThis is supported by MSVC and achieves the same effect: it removes the environment variable from the process environment.\n\nThis keeps cross-platform compatibility.\n\n* Update ggml-decoder.cpp\n\n* Update ggml-decoder.cpp\n\n* Update ggml-decoder.cpp\n\n* Update ggml-decoder.cpp\n\n* Update ggml-decoder.cpp\n\n* Remove the second decoder for node. Moving the function into the model decoder\n\n* Fix error for naive\n\n* NPU prefill chunking\n\n* NPU fix llama-bench\n\n* fallback naive run with accuracy issue\n\n* NPU support llma-perplexity -b 512 --no-warmup\n\n* Refactor: split ov_graph_compute for dynamic and static\n\n* remove unused API GgmlOvDecoder::get_output_stride(const std::string & name)\n\n* minor update due to ov 2025.4\n\n* remove unused API GgmlOvDecoder::get_output_names()\n\n* remove unused API get_output_shape(const std::string & name)\n\n* Modified API GgmlOvDecoder::get_output_type(const std::string & name)\n\n* Removed API GgmlOvDecoder::get_output_op_params(const std::string & name)\n\n* Removed API get_output_ggml_tensor(const std::string & name)\n\n* Removed API m_outputs\n\n* Removed m_output_names\n\n* Removed API GgmlOvDecoder::get_input_names()\n\n* Removed API GgmlOvDecoder::get_input_stride(const std::string& name)\n\n* Removed API get_input_type\n\n* Removed API get_input_type\n\n* Removed API GgmlOvDecoder::get_input_shape(const std::string & name)\n\n* Removed API GgmlOvDecoder::get_input_op_params(const std::string & name)\n\n* Fix error for decoder cache\n\n* Reuse cached decoder\n\n* GPU remove Q6_K requantization\n\n* NPU fix wrong model output shape\n\n* NPU fix q4 perf regression\n\n* Remove unused variable nodes\n\n* Fix decoder can_reuse for llama-bench\n\n* Update build.md for Windows\n\n* backend buffer: allocate on host\n\n* Use shared_buffer for GPU NPU; Refactor\n\n* Add ov_backend_host_buffer; Use cached remote context\n\n* Put kvcache on GPU\n\n* Use ggml_aligned_malloc\n\n* only use remote tensor for kvcache\n\n* only use remote tensor for kvcache for GPU\n\n* FIX: use remote tensor from singleton\n\n* Update build.md to include OpenCL\n\n* NPU always requant to q4_0_128\n\n* Optimize symmetric quant weight extraction: use single zp\n\n* Use Q8_0_C in token embd, lm_head, and for 5 and 6 bits quant\n\n* Update build.md\n\n* Support -ctk f32\n\n* Initial stateful graph support\n\n* Update ggml/src/ggml-openvino/ggml-decoder.cpp\n\nCo-authored-by: Yamini Nimmagadda \n\n* code cleanup\n\n* npu perf fix\n\n* requant to f16 for Q6 embed on NPU\n\n* Update ggml/src/ggml-openvino/ggml-decoder.cpp\n\n* Update ggml/src/ggml-openvino/ggml-openvino-extra.cpp\n\n* Create OPENVINO.md in llama.cpp backend docs\n\n* Update OPENVINO.md\n\n* Update OPENVINO.md\n\n* Update OPENVINO.md\n\n* Update build.md\n\n* Update OPENVINO.md\n\n* Update OPENVINO.md\n\n* Update OPENVINO.md\n\n* kq_mask naming fix\n\n* Syntax correction for workflows build file\n\n* Change ov backend buffer is_host to false\n\n* Fix llama-bench -p -n where p<=256\n\n* Fix --direct-io 0\n\n* Don't put kvcache on GPU in stateful mode\n\n* Remove hardcode names\n\n* Fix stateful shapes\n\n* Simplification for stateful and update output shape processing\n\n* Remove hardcode names\n\n* Avoid re-compilation in llama-bench\n\n* Extract zp directly instead of bias\n\n* Refactor weight tensor processing\n\n* create_weight_node accept non-ov backend buffer\n\n* remove changes in llama-graph.cpp\n\n* stateful masking fix (#38)\n\nFix for stateful accuracy issues and cl_out_of_resources error in stateful GPU with larger context sizes.\n\n* Fix test-backend-ops crash glu, get_rows, scale, rms_norm, add\n\n* hardcoded name handling for rope_freqs.weight\n\n* Suppress logging and add error handling to allow test-backend-ops to complete\n\n* Fix MUL_MAT with broadcast; Add unsupported MUL_MAT FLASH_ATTN cases\n\n* Use bias instead of zp in test-backend-ops\n\n* Update OV in CI, Add OV CI Tests in GH Actions\n\n* Temp fix for multithreading bug\n\n* Update OV CI, fix review suggestions.\n\n* fix editorconfig-checker, update docs\n\n* Fix tabs to spaces for editorconfig-checker\n\n* fix editorconfig-checker\n\n* Update docs\n\n* updated model link to be GGUF model links\n\n* Remove GGML_CPU_REPACK=OFF\n\n* Skip permuted ADD and MUL\n\n* Removed static variables from utils.cpp\n\n* Removed initializing non-existing variable\n\n* Remove unused structs\n\n* Fix test-backend-ops for OV GPU\n\n* unify api calling\n\n* Update utils.cpp\n\n* When the dim is dynamic, throw an error, need to is stastic forst\n\n* Add interface compute_model_outputs(), which get the model output through computing the node use count & status in the cgraph to avoid the flag using\n\n* No need to return\n\n* Fix test-backend-ops for OV GPU LNL\n\n* Fix test-thread-safety\n\n* use the shape from infer request of output tensor create to avoid issue\n\n* fix dynamic output shape issue\n\n* fix issue for the unused node in tests\n\n* Remove unused lock\n\n* Add comment\n\n* Update openvino docs\n\n* update to OV release version 2026.0\n\n* add ci ov-gpu self hosted runner\n\n* fix editorconfig\n\n* Fix perplexity\n\n* Rewrite the model inputs finding mechanism (#54)\n\n* Rewrite the model inputs finding logistic\n\n* Put stateful shape handle in get input shape\n\n* Put the iteration logistic in func\n\n* Added ggml-ci-intel-openvino-gpu and doc update\n\n* .hpp files converted to .h\n\n* fix ggml-ci-x64-intel-openvino-gpu\n\n* Fix for stateful execution bug in llama-bench\n\n* Minor updates after stateful llama-bench fix\n\n* Update ggml/src/ggml-openvino/utils.cpp\n\nCo-authored-by: Yamini Nimmagadda \n\n* Remove multiple get_shape calls\n\n* Bring back mutex into compute\n\n* Fix VIEW op, which slice the input node\n\n* Added token_len_per_seq existence check before slicing masks and moved node retrieval inside guarded block to prevent missing-key access\n\n* Temp. fix for test requant errors\n\n* Update to OV ggml-ci to low-perf\n\n* ci : temporary disable \"test-llama-archs\"\n\n* ci : cache v4 -> v5, checkout v4 -> v6, fix runner tag\n\n* docs : update url\n\n* Fix OV link in docker and Update docs\n\n---------\n\nCo-authored-by: Ravi Panchumarthy \nCo-authored-by: Cavus Mustafa \nCo-authored-by: Arshath \nCo-authored-by: XuejunZhai \nCo-authored-by: Yamini Nimmagadda \nCo-authored-by: Xuejun Zhai \nCo-authored-by: Georgi Gerganov ","tree":"72527f32be2bb161ca1878b0481b08ee76ab1f28"}}