All pull requests

Pull requests

• Switch from ai-edge-quantizer to litert-quantizer.

  #1390·copybara-service[bot] opened Oct 26, 2026Bot·

  ·

• Pass a16w8 and use_float_input_output_normalizer from NpuPipelineConfig to quantize()

  #1389·dkozlov opened Oct 26, 2026·

  ·
  1

  1

• Apply final_logit_softcapping to Gemma 4 and Gemma 3n decode logits

  #1388·dkozlov opened Oct 26, 2026·

  ·
  1

• calib: select rms_norm decomposition ops and residual ADDs structurally for use_v2 exports

  #1387·dkozlov opened Oct 26, 2026·

  ·
  1

• export_hf: support Falcon-H1 and Zamba2 hybrid models

  #1380·copybara-service[bot] opened Oct 25, 2026Bot·

  ·

• export_hf: support Mamba2 hybrid models (Granite-4.0-H, Nemotron-H)

  #1379·copybara-service[bot] opened Oct 25, 2026Bot·

  ·

• Test that ReduceViewRankPass leaves a permute with a second user unfolded

  #1378·copybara-service[bot] opened Oct 25, 2026Bot·

  ·

• Fix constant aliasing in inline_consts caused by the data_ptr() cache key

  #1377·copybara-service[bot] opened Oct 25, 2026Bot·

  ·

• Point export_hf int4 docs at the blockwise dynamic_wi4b32_afp32 recipe

  #1376·copybara-service[bot] opened Oct 25, 2026Bot·

  ·

• Probe the chat template with a sentinel instead of searching the rendered output for the prompt

  #1375·copybara-service[bot] opened Oct 25, 2026Bot·

  ·

• tokenizer_to_sentencepiece: preserve byte-level BPE byte semantics

  #1374·copybara-service[bot] opened Oct 25, 2026Bot·

  ·

• Derive prompt_templates.model.prefix from the generation prompt

  #1373·copybara-service[bot] opened Oct 25, 2026Bot·

  ·

• npu_export: document the Qualcomm HTP prefill-mask size limit and warn when cache_length crosses it

  #1372·copybara-service[bot] opened Oct 25, 2026Bot·

  ·
  1

• Match FunctionGemma's fc format flag to the bundled chat template

  #1371·copybara-service[bot] opened Oct 25, 2026Bot·

  ·

• lfm2_vl: fold the SigLIP2 positional-embedding resize so the exported vision encoder compiles on GPU

  #1370·copybara-service[bot] opened Oct 25, 2026Bot·

  ·

• export_hf qwen3_asr: attend within the n_window_infer windows instead of 13-token chunks

  #1369·copybara-service[bot] opened Oct 25, 2026Bot·

  ·

• Add RecursiveSI/flint model extension and LiteRT-LM export support.

  #1367·copybara-service[bot] opened Oct 23, 2026Bot·

  ·

• Support MoE SiLU activation, attention sinks, and stacked per-expert weight loading.

  #1366·copybara-service[bot] opened Oct 23, 2026Bot·

  ·

• Just internal only change

  #1363·copybara-service[bot] opened Oct 22, 2026Bot·

  ·

• Fold slice and elementwise ops on DenseResourceElementsAttr constants.

  #1359·copybara-service[bot] opened Oct 22, 2026Bot·

  ·

• [Qwen3.5] Enable full GPU acceleration for Metal and OpenCL

  #1350·freedomtan opened Oct 20, 2026Contributor·

  ·

• Reject non-GELU-tanh experts for litert_moe and read num_experts_per_tok.

  #1349·john-rocky opened Oct 20, 2026Contributor·

  ·

• Reduce peak RAM usage during Gemma 4 E4B multimodal_lm export and support bfloat16 constants in inline_consts.

  #1347·copybara-service[bot] opened Oct 19, 2026Bot·

  ·

• Migrate production Python callers off the legacy TfLiteModelType enum.

  #1342·copybara-service[bot] opened Oct 16, 2026Bot·

  ·

• [export_hf] Resolve LFM2 fused gate_up_proj from w1/w3 in HF weights loader.

  #1337·copybara-service[bot] opened Oct 14, 2026Bot·

  ·