All pull requests

Pull requests

• Parallelize f32->qd8/qdu8 convert across a row when batch < threads

  #11990·gadgilrajeev opened Nov 11, 2026·

  ·

• Fix RISC-V RVV qu8/qs8 dispatch and scalar fallbacks in dwconv-config

  #11989·copybara-service[bot] opened Nov 11, 2026Bot·

  ·

• Replace Python build identifier generator with POSIX shell sha256sum.

  #11987·copybara-service[bot] opened Nov 11, 2026Bot·

  ·

• Add YNNPACK runtime backend for text decoder models.

  #11986·copybara-service[bot] opened Nov 10, 2026Bot·

  ·

• Add YNNPACK conv, pool, select, and transformer composite lowerings.

  #11985·copybara-service[bot] opened Nov 10, 2026Bot·

  ·

• Support affine blockwise quantization and row-only gather on YNNPACK.

  #11983·copybara-service[bot] opened Nov 10, 2026Bot·

  ·

• Optimize Qwen3.5-27B CPU memory usage, compile latency, and throughput.

  #11984·copybara-service[bot] opened Nov 10, 2026Bot·

  ·

• Unpack GGUF sub-block, 5-bit, and IQ4 LUT weights losslessly.

  #11982·copybara-service[bot] opened Nov 10, 2026Bot·

  ·

• Keep YNNPACK worker threads awake between short parallel loops.

  #11981·copybara-service[bot] opened Nov 10, 2026Bot·

  ·

• Optimize Gemma 4 (E2B, E4B, 26B-A4B, 31B) prefill and decode performance across ML Drift GPU (Metal & OpenCL), XNNPACK CPU, YNNPACK CPU, and LiteRT-LM Tensor API executor:

  #11980·copybara-service[bot] opened Nov 10, 2026Bot·

  ·

• Speed up YNNPACK quantized LLMs and make graph compilation stack-safe.

  #11979·copybara-service[bot] opened Nov 10, 2026Bot·

  ·

• Bring up Qwen3.5-27B and dense Qwen3.5 family presets on Tensor API.

  #11978·copybara-service[bot] opened Nov 10, 2026Bot·

  ·

• Lower GGUF Q8_0, Q4_1, and Q4_K weights natively on ML Drift GPUs.

  #11977·copybara-service[bot] opened Nov 10, 2026Bot·

  ·

• Support Q5_K/Q6_K GGUF weights, lazy unpacking, and GPU weight cache.

  #11976·copybara-service[bot] opened Nov 10, 2026Bot·

  ·

• Add 9p minmax s8 and u8 maxpool microkernels for AVX2 (u32), AVX-512

  #11975·copybara-service[bot] opened Nov 10, 2026Bot·

  ·

• Disable ARM SME ACLE microkernels by default behind build option.

  #11974·copybara-service[bot] opened Nov 10, 2026Bot·

  ·

• When m <= step_m, there is only a single m tile, so B is not reused across

  #11973·copybara-service[bot] opened Nov 10, 2026Bot·

  ·

• Hoist b_scale and offset broadcasts out of the m loop in dequantize_dot

  #11971·copybara-service[bot] opened Nov 10, 2026Bot·

  ·

• Support YNN_UNKNOWN_RANK output rank inference, stricter parameter

  #11970·copybara-service[bot] opened Nov 10, 2026Bot·

  ·

• Implement dot_uint8_int4_int32_32x32x64_16x16x8_amxint8 and dot_int8_int4_int32_32x32x64_16x16x8_amxint8 microkernels using AVX-512 nibble unpacking into stack-allocated int8 AMX B-tile buffers.

  #11968·copybara-service[bot] opened Nov 10, 2026Bot·

  ·

• Fix KleidiAI ARM SME2 M=1 GEMV get_nr() overflow and n_step tiling

  #11967·copybara-service[bot] opened Nov 10, 2026Bot·

  ·

• Use unsigned 2-bit weights in x86 qd8-qc2w GEMM microkernels

  #11966·copybara-service[bot] opened Nov 10, 2026Bot·

  ·

• Let Tensor API blockwise quantization reference a named scales tensor.

  #11965·copybara-service[bot] opened Nov 10, 2026Bot·

  ·

• Add FP16 operator benchmarks to average_pooling_bench

  #11961·copybara-service[bot] opened Nov 10, 2026Bot·

  ·

• Fix Zen 5 f16 avgpool regression and add SIMD microkernels

  #11959·copybara-service[bot] opened Nov 9, 2026Bot·

  ·