google-ai-edge/LiteRT

View on GitHub

LiteRT, successor to TensorFlow Lite. is Google's On-device framework for high-performance ML & GenAI deployment on edge platforms, via efficient conversion, runtime, and optimization

Stars: 3.4k
Forks: 454
Open beginner issues: 0
Indexed issues: 89
Avg merge: 2d 5h
Merged PRs (30d): 283

Dominant language: C++
License: Apache-2.0
Last GitHub push: Sep 19, 2026
Latest indexed: Sep 20, 2026
Contributing guide: Contributing guide
Code of conduct: Code of conduct
Beginner labels: No beginner labels indexed

Beginner friendly Everything

Open All

89 open issues indexed

[Tensor SDK][AVF] Device assignment of the Tensor TPU mailbox context (Pixel 11): will app-launched pVMs get it, and is Tensor G5 in scope? Open

Difficulty 5/5 Over a week Newbie friendliness 20/100

google-ai-edge/LiteRT#10081 · Opened Sep 19, 2026
CompiledModel.from_file on a truncated .tflite ends the process with SIGSEGV (ai-edge-litert 2.2.0, macOS arm64); from_buffer and 2.1.6 raise Open

Difficulty 4/5 3-5 days Newbie friendliness 56/100

google-ai-edge/LiteRT#10080 · Opened Sep 19, 2026
Add `py.typed` and type hints to Python package Open

type:feature

google-ai-edge/LiteRT#9973 · 1 comment · 1 assignee · Opened Sep 15, 2026
WebGPU accelerator: `--enable_profiling` fails with "UNIMPLEMENTED: Profile is not implemented" (macOS, LiteRT-LM v0.16.0 and v0.17.0) Open

status:awaiting LiteRTer status:awaiting PR review type:accelerators type:bug type:gpu delegate

google-ai-edge/LiteRT#9967 · 1 comment · 3 assignees · Opened Sep 15, 2026
The memory data printed by Tflite's benchmark is different from the actual test results Open

status:awaiting user response status:stale type:bug

google-ai-edge/LiteRT#9899 · 2 comments · 1 assignee · Opened Sep 11, 2026
Add Signed Process Domain (SignedPD) support and fallback in HtpBackend::Init. Open

status:awaiting PR review type:feature

google-ai-edge/LiteRT#9754 · 2 comments · 1 assignee · Opened Sep 3, 2026
LiteRT.js / WASM: fp16-quantized PP-OCRv5 recognizer runs 21× slower than its fp32 parent on XNNPACK Open

status:awaiting LiteRTer type:bug

google-ai-edge/LiteRT#9663 · 2 comments · 3 assignees · Opened Aug 31, 2026
LiteRT.js / WebGPU: Matcha-TTS flow-matching decoder output diverges from WASM (mean abs 0.34, deterministic, Apple Metal) Open

status:awaiting LiteRTer type:bug

google-ai-edge/LiteRT#9662 · 1 comment · 3 assignees · Opened Aug 31, 2026
LiteRT.js / WebGPU: PP-OCRv5 recognizer deterministically misreads text vs WASM (fp32 weights too, Apple Metal) Open

status:awaiting LiteRTer type:bug

google-ai-edge/LiteRT#9661 · 2 comments · 2 assignees · Opened Aug 31, 2026
LiteRT.js / WebNN: Failed to execute 'conv2d' on 'MLGraphBuilder' for efficientvit Segmentation demo Open

status:awaiting user response type:bug

google-ai-edge/LiteRT#9548 · 4 comments · 2 assignees · Opened Aug 25, 2026
GPU delegate: explicit-dequantize int8 FULLY_CONNECTED fails kernel creation with a transposed BHWC shape mismatch (Metal / OpenCL / WebGPU) Open

status:awaiting LiteRTer type:bug

google-ai-edge/LiteRT#9523 · 2 comments · 3 assignees · Opened Aug 24, 2026
CompiledModel::CreateOutputBuffers creates 4 bytes size buffer for CPU accelerator Open

status:awaiting LiteRTer type:bug

google-ai-edge/LiteRT#9518 · 4 comments · 1 assignee · Opened Aug 23, 2026
MediaTek MT6878 NPU dispatch: SIGSEGV at first inference after "contradictory buffer requirements" warning for KV-cache index tensors Open

status:awaiting LiteRTer type:bug

google-ai-edge/LiteRT#9500 · 4 comments · 1 assignee · Opened Aug 21, 2026
MediaTek MT6878 AOT compiler: SIGSEGV in NeuronCompilation_getSupportedOperations for non-transposed KV-cache LLM graphs Open

status:awaiting LiteRTer type:bug

google-ai-edge/LiteRT#9499 · 1 assignee · Opened Aug 21, 2026
samsung_runtime/ (and mediatek_runtime/) missing from litert_npu_runtime_libraries*.zip in v2.2.0 Open

status:awaiting LiteRTer type:bug

google-ai-edge/LiteRT#9482 · 2 comments · 3 assignees · Opened Aug 21, 2026
[run_model][GPU] Reported inference time measures OpenCL enqueue instead of completed execution Open

status:awaiting LiteRTer status:awaiting PR review type:bug

google-ai-edge/LiteRT#9453 · 1 comment · 2 assignees · Opened Aug 20, 2026
[qwen3 ASR] Incompatible Neuron SDK major version: bytecode compiled with 8.2.30, but runtime is 9.2.0 Open

status:awaiting user response type:bug

google-ai-edge/LiteRT#9364 · 14 comments · 2 assignees · Opened Aug 17, 2026
ML Drift GPU (Mali): DRQ int8 FULLY_CONNECTED + ADD on rank-3 input fails kernel creation — "Unable to parse bc coord for BATCH axis" (2.1.3 & 2.1.5) Open

status:awaiting LiteRTer type:bug

google-ai-edge/LiteRT#9277 · 2 comments · 2 assignees · Opened Aug 13, 2026
GPU accelerator: rank-3 PAD on a non-innermost axis returns row-shifted output (correct on CPU) Open

status:awaiting LiteRTer type:bug

google-ai-edge/LiteRT#9272 · 3 comments · 2 assignees · Opened Aug 13, 2026
Conv2D with `same` padding + asymmetric stride returns all-zeros in the LiteRT interpreter Open

status:awaiting LiteRTer type:bug

google-ai-edge/LiteRT#9256 · 4 comments · 2 assignees · Opened Aug 12, 2026
Integer overflow to out-of-bounds in vendored safetensors.h (OOB read/write, ASAN-confirmed) Open

status:awaiting LiteRTer status:awaiting PR review type:bug

google-ai-edge/LiteRT#9255 · 3 comments · 2 assignees · Opened Aug 12, 2026
[JIT][Android 16][MediaTek][D8300] LiteRT-LM native Gemma 3 270M aborts during NPU compilation; CPU/GPU cannot create executor Open

status:awaiting LiteRTer type:bug

google-ai-edge/LiteRT#9250 · 3 comments · 2 assignees · Opened Aug 12, 2026
GPU delegate: SUM accumulates in fp16, so a reduction whose total exceeds 65504 returns NaN (EfficientNet SE blocks) Open

status:awaiting LiteRTer type:bug

google-ai-edge/LiteRT#9249 · 2 comments · 3 assignees · Opened Aug 12, 2026
GPU delegate: DIV with identical inputs and REDUCE_MAX with const inputs split a small perceiver-style block into 3 partitions (2.1.6) Open

status:awaiting LiteRTer type:bug type:gpu delegate

google-ai-edge/LiteRT#9192 · 2 assignees · Opened Aug 9, 2026
GPU delegate: BROADCAST_TO is rejected even at rank 4, so Tensor.expand (GQA repeat_kv) takes any transformer off the GPU Open

type:feature type:gpu delegate

google-ai-edge/LiteRT#9191 · 3 comments · 1 assignee · Opened Aug 9, 2026
Security: Integer Overflow in WebNN Div Op via INT32_MIN/-1 Open

status:awaiting LiteRTer type:bug

google-ai-edge/LiteRT#9189 · 12 comments · 3 assignees · Opened Aug 9, 2026
Conversion lowers reduction-style einsum to an invalid tf.BatchMatMulV2 Open

status:awaiting LiteRTer type:bug

google-ai-edge/LiteRT#9023 · 4 comments · 2 assignees · Opened Jul 31, 2026
Division by a constant zero returns FLT_MAX instead of infinity Open

status:awaiting LiteRTer type:bug

google-ai-edge/LiteRT#9021 · 4 comments · 2 assignees · Opened Jul 31, 2026
XNNPACK COS returns values outside [-1, 1] for large float32 inputs Open

status:awaiting LiteRTer type:bug

google-ai-edge/LiteRT#9020 · 4 comments · 2 assignees · Opened Jul 31, 2026
LiteRT ML Drift source-quantized INT8 FC produces row-sum-dependent error Open

status:awaiting LiteRTer type:bug

google-ai-edge/LiteRT#8979 · 3 comments · 2 assignees · Opened Jul 29, 2026
[Bug] no-delegate builtin interpreter corrupts FULLY_CONNECTED output via memory-planner buffer aliasing (XNNPACK/preserve_all_tensors correct) Open

status:awaiting LiteRTer type:bug

google-ai-edge/LiteRT#8928 · 2 comments · 2 assignees · Opened Jul 27, 2026
GPU delegate ignores `BATCH_MATMUL` `adj_x` and `adj_y` options Open

status:awaiting LiteRTer status:awaiting PR review type:bug

google-ai-edge/LiteRT#8913 · 3 comments · 2 assignees · Opened Jul 27, 2026
Importing tensor from `AHardwareBuffer` fails on Adreno 660 Open

type:feature

google-ai-edge/LiteRT#8903 · 2 comments · 1 assignee · Opened Jul 25, 2026
xdsl dependancy requirement (ai-edge-litert[model-utils]) Open

type:feature

google-ai-edge/LiteRT#8894 · 2 comments · 1 assignee · Opened Jul 24, 2026
Builtin no-delegate interpreter corrupts reused buffer under default memory planner (RESHAPE→RMSNorm region) Open

status:awaiting LiteRTer type:bug

google-ai-edge/LiteRT#8889 · 3 comments · 2 assignees · Opened Jul 24, 2026
[Google Tensor SDK Beta][Tensor G5] AOT compiler returns INTERNAL for a fully partitioned 3.48 GB multimodal TFLite model Open

status:awaiting LiteRTer type:bug type:tensor sdk

google-ai-edge/LiteRT#8884 · 3 comments · 2 assignees · Opened Jul 24, 2026
[Qualcomm NPU] Compiler plugin does not repartition after QNN VTCM allocation failure (Gemma 4 E2B, SM8845) Open

status:awaiting user response status:stale type:bug type:feature

google-ai-edge/LiteRT#8869 · 3 comments · 1 assignee · Opened Jul 23, 2026
Git-LFS problem - Release v2.1.6 artifacts contain fake binary content of 133 bytes in size Open

status:awaiting LiteRTer type:bug

google-ai-edge/LiteRT#8835 · 4 comments · 3 assignees · Opened Jul 21, 2026
LiteRtMetalAccelerator cannot be registered on iphone Open

status:awaiting LiteRTer type:bug

google-ai-edge/LiteRT#8787 · 6 comments · 2 assignees · Opened Jul 18, 2026
Missing OHOS platform support to LiteRT Open

status:awaiting PR review type:feature

google-ai-edge/LiteRT#8691 · 3 comments · 1 assignee · Opened Jul 14, 2026
GPU delegate: odml.scaled_dot_product_attention composites on Hiera window-attention shapes hard-crash the device (driver hang → watchdog reboot, Pixel 8a) Open

status:awaiting LiteRTer type:bug type:gpu delegate

google-ai-edge/LiteRT#8620 · 3 comments · 2 assignees · Opened Jul 10, 2026
GPU delegate silently mis-computes a fully-delegated graph (SAM 2.1 mask decoder, corr 0.265 vs CPU); trigger is attention rank in graph context — not fp16, not LayerNorm Open

status:awaiting LiteRTer type:bug type:gpu delegate

google-ai-edge/LiteRT#8619 · 5 comments · 2 assignees · Opened Jul 10, 2026
LSTM / UNIDIRECTIONAL_SEQUENCE_LSTM: heap OOB read of weights rank in Prepare Open

status:awaiting LiteRTer status:awaiting PR review type:bug

google-ai-edge/LiteRT#8610 · 2 comments · 2 assignees · Opened Jul 10, 2026
Interpreter: heap OOB read on an out-of-range node input tensor index Open

status:awaiting LiteRTer status:awaiting PR review type:bug

google-ai-edge/LiteRT#8609 · 3 comments · 2 assignees · Opened Jul 10, 2026
BATCH_MATMUL: heap OOB read on a rank<2 operand Open

status:awaiting LiteRTer status:awaiting PR review type:bug

google-ai-edge/LiteRT#8608 · 2 comments · 2 assignees · Opened Jul 10, 2026
GPU delegate: an ADD that is both a graph output and consumed returns one of its operands, not the sum Open

status:awaiting LiteRTer type:bug type:gpu delegate

google-ai-edge/LiteRT#8599 · 2 comments · 2 assignees · Opened Jul 10, 2026
GPU delegate regression: RELU_0_TO_1 was supported in 2.1.3, rejected in 2.1.5 and 2.1.6 Open

status:awaiting LiteRTer type:bug

google-ai-edge/LiteRT#8598 · 3 comments · 2 assignees · Opened Jul 10, 2026
[feature windows]: Better D3D interop Open

type:feature

google-ai-edge/LiteRT#8586 · 4 comments · 1 assignee · Opened Jul 9, 2026
[TFLite] FULLY_CONNECTED reference kernel should perform requantization using integer kernels instead of float Open

status:awaiting LiteRTer type:support

google-ai-edge/LiteRT#8567 · 4 comments · 2 assignees · Opened Jul 9, 2026
SPM checkout fails: LFS smudge error Open

status:awaiting LiteRTer type:bug

google-ai-edge/LiteRT#8521 · 12 comments · 6 reactions · 3 assignees · Opened Jul 7, 2026

google-ai-edge/LiteRT

Get new issues in your inbox