Issues

export_hf Qwen3-ASR: the exported audio encoder attends within 13-token chunks instead of transformers' `n_window_infer` windows (104 tokens)

google-ai-edge/litert-torch#1296 · opened

on Oct 2, 2026

npu_export: Qwen3-0.6B and Gemma-4-E2B bundles compiled for SM8550 (Hexagon v73) produce garbage on the NPU, while Gemma-3-1B from the same pipeline is correct

google-ai-edge/litert-torch#1290 · opened

on Sep 30, 2026

4

litert-torch 0.9.x silently produces wrong output for non-contiguous constants (correct on 0.8.0)

google-ai-edge/litert-torch#1238 · opened

on Sep 16, 2026

3

Provide ability to export a model with an MTP drafter section

google-ai-edge/litert-torch#1214 · opened

on Sep 2, 2026

2

`DEPTHWISE_CONV_2D` with `depth_multiplier > 1` is ~860x slower than the equivalent op via ONNX→QNN on the same device

google-ai-edge/litert-torch#1211 · opened

on Sep 2, 2026

2

export_hf: `parse_chat_template` never renders the generation prompt — `model.prefix` is taken from an assistant history turn, so a thinking-only model ships with a scaffold its template never produces (Qwen3-4B-Thinking-2507: `<think>\n\n</think>\n\n` instead of `<think>\n`; the reasoning lands in the answer text)

google-ai-edge/litert-torch#1209 · opened

on Sep 1, 2026

2

export_hf: the automatic thought-channel declaration is skipped on the structured-template path and keyed on a literal `<think>` — a thinking model can ship with no `LlmMetadata.channels` and no warning

google-ai-edge/litert-torch#1206 · opened

on Aug 30, 2026

tokenizer_to_sentencepiece_lib: byte-level BPE vocabs lose their byte semantics in the SentencePiece proto — standalone Latin-1/Ext-A characters get the byte token's id, unencodable characters become the UNK piece, and with no unk_token that piece is pad/eos (SmolLM3: <|im_end|> unmatchable in prompts and emitted for emoji)

google-ai-edge/litert-torch#1205 · opened

on Aug 30, 2026

2

npu_export: the gemma3-270m preset's mask concat lands just over a size threshold where the HTP applies it to the wrong rows

google-ai-edge/litert-torch#1184 · opened

on Aug 23, 2026

3

Gemma 4 E2B zero-adapter PTQ exports emit malformed JSON under constrained decoding

google-ai-edge/litert-torch#1183 · opened

on Aug 22, 2026

6

npu_export Stage 2 calibrates non-Gemma models with Gemma's chat markers, so the quantized model degrades on chat-templated prompts

google-ai-edge/litert-torch#1181 · opened

on Aug 22, 2026

6

GPU backend (WebGPU/OpenCL via ml_drift) silently omits trailing generated content after a tag-like boundary; CPU backend produces correct output on identical model

google-ai-edge/litert-torch#1148 · opened

on Aug 1, 2026

3

No path to fused odml.* composite lowering for plain PyTorch models — attention always decomposes (devx)

google-ai-edge/litert-torch#1105 · opened

on Jul 10, 2026

3

export_hf int4: the named int4 recipe (dynamic_wi4_afp32) is channelwise and collapses small LLMs — expose/recommend a blockwise int4 recipe

google-ai-edge/litert-torch#1068 · opened

on Jun 22, 2026

12

lora conversion fails for gqa/mha models: `_from_tensor_generator` omits num_query_groups factor in q/k/v out-dims

google-ai-edge/litert-torch#1066 · opened

on Jun 21, 2026

3

nn.LSTM cannot convert with a dynamic sequence length (time axis specialized)

google-ai-edge/litert-torch#1062 · opened

on Jun 19, 2026

2

conv_transpose1d weight constants deduplicated by shape — 2nd op silently reuses 1st's weights

google-ai-edge/litert-torch#1061 · opened

on Jun 19, 2026

6

Conversion fails: torch.uint8 unsupported in MLIR lowering (blocks Qwen3.5 / GatedDeltaNet)

google-ai-edge/litert-torch#1060 · opened

on Jun 19, 2026

5

Failed to legalize operation 'tfl.transpose' when trying to convert model to Float16

google-ai-edge/litert-torch#1053 · opened

on Jun 15, 2026

3

Contribution offer: sparse-MoE building blocks for the Generative API + a phone-class MoE example (LFM2.5-8B-A1B / Qwen3-MoE)

google-ai-edge/litert-torch#1052 · opened

on Jun 12, 2026

3

GitHub Bug Report: macOS Dynamic Linker Symbol Collision / Silent SEGSEGV on Import

google-ai-edge/litert-torch#1049 · opened

on Jun 11, 2026

3

Clarify or fail fast for Gemma 4 QAT mobile Transformers checkpoints (quant_method=gemma)

google-ai-edge/litert-torch#1044 · opened

on Jun 9, 2026

2

Support audio encoder export for gemma-4-E2B-it

google-ai-edge/litert-torch#1039 · opened

on Jun 2, 2026

3

Gemma 4 E2B fine-tuned export loads in LiteRT-LM but tool-call fidelity regresses vs MLX baseline

google-ai-edge/litert-torch#1013 · opened

on May 1, 2026

6

litert_lm_builder missing 'gemma4' case in model type handler

google-ai-edge/litert-torch#1005 · opened

on Apr 26, 2026

8

Navigation Menu

Repository navigation

Issues

export_hf Qwen3-ASR: the exported audio encoder attends within 13-token chunks instead of transformers' `n_window_infer` windows (104 tokens)

npu_export: Qwen3-0.6B and Gemma-4-E2B bundles compiled for SM8550 (Hexagon v73) produce garbage on the NPU, while Gemma-3-1B from the same pipeline is correct

litert-torch 0.9.x silently produces wrong output for non-contiguous constants (correct on 0.8.0)

Provide ability to export a model with an MTP drafter section

`DEPTHWISE_CONV_2D` with `depth_multiplier > 1` is ~860x slower than the equivalent op via ONNX→QNN on the same device

export_hf: the automatic thought-channel declaration is skipped on the structured-template path and keyed on a literal `<think>` — a thinking model can ship with no `LlmMetadata.channels` and no warning

npu_export: the gemma3-270m preset's mask concat lands just over a size threshold where the HTP applies it to the wrong rows

Gemma 4 E2B zero-adapter PTQ exports emit malformed JSON under constrained decoding

npu_export Stage 2 calibrates non-Gemma models with Gemma's chat markers, so the quantized model degrades on chat-templated prompts

GPU backend (WebGPU/OpenCL via ml_drift) silently omits trailing generated content after a tag-like boundary; CPU backend produces correct output on identical model

No path to fused odml.* composite lowering for plain PyTorch models — attention always decomposes (devx)

export_hf int4: the named int4 recipe (dynamic_wi4_afp32) is channelwise and collapses small LLMs — expose/recommend a blockwise int4 recipe

lora conversion fails for gqa/mha models: `_from_tensor_generator` omits num_query_groups factor in q/k/v out-dims

nn.LSTM cannot convert with a dynamic sequence length (time axis specialized)

conv_transpose1d weight constants deduplicated by shape — 2nd op silently reuses 1st's weights

Conversion fails: torch.uint8 unsupported in MLIR lowering (blocks Qwen3.5 / GatedDeltaNet)

Failed to legalize operation 'tfl.transpose' when trying to convert model to Float16

Contribution offer: sparse-MoE building blocks for the Generative API + a phone-class MoE example (LFM2.5-8B-A1B / Qwen3-MoE)

GitHub Bug Report: macOS Dynamic Linker Symbol Collision / Silent SEGSEGV on Import

Clarify or fail fast for Gemma 4 QAT mobile Transformers checkpoints (quant_method=gemma)

Support audio encoder export for gemma-4-E2B-it

Gemma 4 E2B fine-tuned export loads in LiteRT-LM but tool-call fidelity regresses vs MLX baseline

litert_lm_builder missing 'gemma4' case in model type handler