Google Git Google Git
Sign in
chromium/external/github.com/google-ai-edge/LiteRT/HEAD
  1. 135bab0 Merge pull request #11292 from graham0824:dev/jiunkaiy/kanana2.0 by Copybara-Service · 56 minutes ago upstream/main
  2. b5f1030 Derive the NPUW weights-bank name from the shared weight pool. by Tommy Chiang · 6 hours ago
  3. b5a379b Internal changes only. by Andrew Zhang · 7 hours ago
  4. 98f0617 Tune compiler plugin targets. by Andrew Zhang · 7 hours ago
  5. 625f6c0 Update pinned Eigen commit for XLA and TensorFlow. by Antonio Sanchez · 12 hours ago
  6. 14a78f3 Preserve the temporary 4-level 2-bit mid-riser codebook (`int2_in_int4`) during Hadamard MSE scale refinement (`scale_refinement="mse_grid"`). by Google AI Edge · 12 hours ago
  7. aa9e40c Stage spatial tensor bindings while background command buffer preparation runs. by Fengwu Yao · 15 hours ago
  8. 1487c17 Support blockwise quantized (e.g. centered int2) EMBEDDING_LOOKUP tables. by Majid Dadashi · 16 hours ago
  9. 7e4a566 Discover `.dylib` vendor dispatch and compiler plugin libraries on macOS in `Environment.create()`. by Andrew Zhang · 16 hours ago
  10. 3011fa6 Merge pull request #11341 from Cbaoj:intel-openvino-npu-hal-run-gated by Copybara-Service · 16 hours ago
  11. 8e37c21 Merge pull request #11658 from JeevakaPrabu:wcl_npu_support by Copybara-Service · 16 hours ago
  12. dd75860 Add LiteRT vision evaluation backend execution engine and CLI runner. by Changming Sun · 16 hours ago
  13. 5062b2a Merge pull request #11657 from JeevakaPrabu:log_platform_info by Copybara-Service · 16 hours ago
  14. 61506ee Upgrade TFLite Metal GPU delegate language selection up to Metal 3.2. by Google AI Edge · 17 hours ago
  15. 95fa3bb Support NPU AOT on-device execution and multi-SoC targeting in ATS. by Gerardo Carranza · 18 hours ago
  16. dae6964 Make LiteRT-Swift the Swift package for apps and the LiteRT manifest dev-only. by Jun Jiang · 18 hours ago
  17. 4fc19a4 Use the MSVC overflow intrinsics in `CheckedInt`. by Quentin Khan · 20 hours ago
  18. bfcc35c Emit named variables before composite kernel output writes for post-op linking. by Google AI Edge · 20 hours ago
  19. 2fb44c5 Fix `CheckedInt` operations for operands of different types. by Quentin Khan · 21 hours ago
  20. b312775 Remove 2-bit fixed-point QNN cases from dump_graph by Google AI Edge · 24 hours ago
  21. c30db29 Count the NVIDIA answer check's CPU and NPU parts only; read a run directory once. by Google AI Edge · 34 hours ago
  22. 6318508 Add the NVIDIA benchmark's Windows build script and the mobile-QAT llama.cpp comparison. by Google AI Edge · 35 hours ago
  23. a75f382 Add clock_check.tool, the host clock check of the Linux NVIDIA benchmarks. by Google AI Edge · 35 hours ago
  24. add8e67 Document the NVIDIA measurements against llama.cpp and the tables. by Google AI Edge · 36 hours ago
  25. a3d1479 Add the LiveCodeBench and GPQA evaluations of the NVIDIA benchmark and their tables' processing. by Google AI Edge · 36 hours ago
  26. 2a197bb Add the llama.cpp comparison runners of the NVIDIA benchmark. by Google AI Edge · 36 hours ago
  27. 5d6d53c Add the native Windows runners of the NVIDIA benchmark. by Google AI Edge · 36 hours ago
  28. 4c681c3 Add the NVIDIA benchmark's model, platform and workload descriptors and the table processing, with its test. by Google AI Edge · 36 hours ago
  29. 227b953 Add the dense GEMV dependent launch, off by default and capture-safe when on. by Google AI Edge · 36 hours ago
  30. 0198711 Run Qwen 3.8 27B on TensorRT-RTX: gated DeltaNet plugins, partial RoPE, SwiGLU, and the qwen3_5 partition policy. by Google AI Edge · 36 hours ago
  31. 3eda073 Add a block-scaled INT8 tensor-core GEMM with its copies spread over the slices. by Google AI Edge · 36 hours ago
  32. 03f26ff Keep a copy-on-write model view read-only on Windows until it is written. by Google AI Edge · 36 hours ago
  33. 5caf179 Take unwritten mapped pages from the working set without reading them on Windows, and calibrate the NVIDIA weight holder from aliased CUDA memory. by Google AI Edge · 36 hours ago
  34. f8e92a0 Say when nvguard's paired performance runs were measured hours apart. by Google AI Edge · 36 hours ago
  35. 222d876 Document the 26B dependent-launch case and the dumps of a dispatch before the directory fix in the nvguard README. by Google AI Edge · 2 days ago
  36. 067326c Let nvguard go on past a failed probe, read non-finite dumps, and note a device path the baseline lacks too. by Google AI Edge · 2 days ago
  37. fb4e143 Run nvguard on Windows ARM64 from Git Bash. by Google AI Edge · 2 days ago
  38. 22d39ed Enable the Windows ARM64 NVIDIA SDK toolchains in the OSS build. by Google AI Edge · 2 days ago
  39. bf49eb6 Fix GatedDeltaNet GPU shader linking with downstream elementwise ops. by Fengwu Yao · 4 days ago
  40. 8f57652 Support dynamic range FULLY_CONNECTED with N-D blockwise quantized filters. by Majid Dadashi · 4 days ago
  41. b691f00 Merge pull request #11004 from graham0824:dev/hungjuiw/fix-conv2d-shared-filter-quant-axis by Copybara-Service · 4 days ago
  42. 1a8c5b2 Merge pull request #11293 from graham0824:dev/chuntl/enhance_gcc11.2_cmake by Copybara-Service · 4 days ago
  43. 7d1e7ce Add LiteRT vision evaluation schema, preprocessing, data I/O, and metrics. by Changming Sun · 4 days ago
  44. 7a14da7 Fix the LiteRT Swift Metal buffer test and surface test failures. by Jun Jiang · 4 days ago
  45. 454c96b Use simd_sum for the qkv_norm_rope RMS reduction on Metal. by Fengwu Yao · 4 days ago
  46. fa1a040 Explain compilation failures caused by a requested accelerator that is not registered. by Tenghui Zhu · 4 days ago
  47. e022cd1 Resolve LiteRT system runtime ABI version via environment options by Terry Heo · 4 days ago
  48. 9095bd7 Fix flaky accelerator_memory_apple_test and re-enable it. by Marissa Ikonomidis · 4 days ago
  49. 1e66025 Skip all-zero zero-point tensors for symmetric blockwise int4 weights. by Fengwu Yao · 4 days ago
  50. 49fc72e litert: Expose create_tensor_buffer_requirements_with_alignment in LiteRtRuntimeContext. by Google AI Edge · 4 days ago
  51. a903981 Remove unused `_reshape_data_for_blockwise` from `common_quantize`. by Maria Lyubimtseva · 4 days ago
  52. 2c0d67a Reject non-shared blockwise-quantized EMBEDDING_LOOKUP on GPU (closes #11400). by Aravind Murali · 4 days ago
  53. fb46fcf Merge pull request #11613 from graham0824:dev/chuntl/set_perf_mode_for_loading by Copybara-Service · 4 days ago
  54. c6e165a Refactor Kotlin Environment options JNI plumbing. by Chunlei Niu · 4 days ago
  55. c727190 Add Wildcat Lake (WCL, NPU5020) SoC support to OpenVINO plugin by Jeevaka Prabu Badrappan · 4 days ago
  56. 30e76b3 Log the effective NPU platform in the OpenVINO version info by Jeevaka Prabu Badrappan · 4 days ago
  57. 3060a1f Reject filter zero points in the hybrid FULLY_CONNECTED kernels. by Quentin Khan · 4 days ago
  58. 315a68d Qualcomm AI Engine Direct - Support GQA transformation. by jiunkaiy · 4 days ago
  59. 2201237 Qualcomm AI Engine Direct - Add HTP context load vote option. by chuntl-qti · 4 days ago
  60. 156ea9a Load Wasm loader scripts in module workers in @litertjs/wasm-utils. by Tenghui Zhu · 4 days ago
  61. c220946 Remove unused hooks dependency from Google Tensor dispatch_api BUILD target to fix OSS build of dispatch_api_so. by Google AI Edge · 4 days ago
  62. 229d1b5 Support float16 in TFL GREATER and BROADCAST_TO. by Majid Dadashi · 4 days ago
  63. 648fec7 Support float16 in TFL GREATER and BROADCAST_TO. by Majid Dadashi · 4 days ago
  64. ea5116f Run the odml.gated_delta_block composite as one GPU kernel per Qwen3.5 GatedDeltaNet layer. by Fengwu Yao · 4 days ago
  65. 151856d Avoid unconditional /proc/self/smaps_rollup walks in Linux GetMemoryUsage. by Volodymyr Kysenko · 4 days ago
  66. 6d2fc2c Scope target and CPU reference executors locally in AtsInferenceTest. by Gerardo Carranza · 5 days ago
  67. 0151809 This is an internal change by Google AI Edge · 5 days ago
  68. 00b60cf Migrate C++ LiteRT option encoders to OptionsEncoder. by Chunlei Niu · 5 days ago
  69. 4d08010 Refactor Kotlin CompiledModel options JNI plumbing. by Chunlei Niu · 5 days ago
  70. 9aa8be5 Pass expected dispatch functions from selected signatures to the Dispatch API by Andrew Zhang · 5 days ago
  71. 3660c21 Automated Code Change by Google AI Edge · 5 days ago
  72. 6b99af6 Build Windows ARM64 LiteRT wheels and prebuilts with Bazel and MSVC. by Andrew Zhang · 5 days ago
  73. 646d778 Fix stale graph edges left by composite inlining. by Andrew Zhang · 5 days ago
  74. a8345b4 Exclude models that overflow the 1 MB Windows stack in YNNPACK ATS. by Gerardo Carranza · 5 days ago
  75. c3f98f7 Package Qualcomm NPU plugins in win_arm64 wheels and SDK. by Andrew Zhang · 5 days ago
  76. d5d65f9 Run Python compiler plugins in-process from the LiteRT AOT ApplyPlugin component. by Andrew Zhang · 5 days ago
  77. c15df94 intel_openvino: add a build flag to leave out the NPU HAL integration by Cathy Bao · 5 days ago
  78. 94903d7 intel_openvino: work around C++ runtime mismatch in ov::Exception catch by Cathy Bao · 5 days ago
  79. 1dde2fc intel_openvino: run NPU HAL inference through callbacks by Cathy Bao · 5 days ago
  80. 5803cf7 intel_openvino: use NPU HAL-resolved model priority by Vijeetkumar Benni · 5 days ago
  81. 2bcdaff Copy misaligned host inputs in CustomOpDispatcher. by Fengwu Yao · 5 days ago
  82. 0711a61 Fix out-of-bounds read in NVIDIA bytecode step-feed tests. by Google AI Edge · 5 days ago
  83. d6f217f Link Apple dylibs with -headerpad_max_install_names so that consumers can rewrite their install names and rpaths with install_name_tool. by Google AI Edge · 5 days ago
  84. bd6969d Compare file identity, not path strings, in the NVIDIA Windows mmap test. by Google AI Edge · 5 days ago
  85. c0a549e Move the memory measurement library from LiteRT-LM to LiteRT. by Marissa Ikonomidis · 5 days ago
  86. 5fb2186 Return errors from LiteRT.js wasm bindings instead of throwing them. by Google AI Edge · 5 days ago
  87. 09c618a Give the NVIDIA step feed bytecode tests a 4-byte weight store run. by Google AI Edge · 5 days ago
  88. 8c1609b Add `litert_model`: a Python LiteRT IR backed by the C++ `LiteRtModelT`. by Andrew Zhang · 5 days ago
  89. b3d4b2f Improve NeuroPilot USDK adapter library loading on Android. by Andrew Zhang · 5 days ago
  90. b192b82 Hoist LayerKvCache into shared transformer block header. by Misha Gutman · 5 days ago
  91. 9761f56 Exclude failing models in builtin_cpu_models_ats and ynnpack_fallback_cpu_models_ats. by Gerardo Carranza · 5 days ago
  92. e397e90 Decouple Gemma 3 model configuration from Safetensor loading by splitting Safetensor shape inference into safetensor_model_config. by Ping Yu · 5 days ago
  93. 60d1625 Serialize blockwise-quantized weights directly without Dequantize. by Ping Yu · 5 days ago
  94. f2f74b3 Qualcomm AI Engine Direct - Make OE Linux toolchain ABI selection explicit. by chuntl-qti · 5 days ago
  95. 9e20ee6 Further simplify unit tests by hungjuiw · 5 days ago
  96. 6bec501 Automated Code Change by Chris Kennelly · 6 days ago
  97. 0e73b9f Make channels-last the only layout of GPU short_conv_step. by Fengwu Yao · 6 days ago
  98. 41e9f19 Fix uniform_quantize for large (>32 MiB) >2D tensors. by Maria Lyubimtseva · 6 days ago
  99. 03923b0 Add Kronecker product cross-block orthogonal rotations (`Q_1 (x) H_d`), deterministic Rademacher sign flips (`diag(s_H) H_d`), and post-OCTAV MSE scale refinement to decomposed Hadamard quantization in `litert_quantizer`. by Google AI Edge · 6 days ago
  100. 477c0aa Reject unimplemented quantizations in the builtin DEQUANTIZE kernel. by Quentin Khan · 6 days ago