All pull requests

Pull requests

• [QwixDotGeneral] Fold the scale into the unquantized operand when it varies along a contracting axis

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

  ·

• [QwixDotGeneral] Accumulate in f32 instead of f16 when scales are applied on output

  #409·copybara-service[bot] opened Oct 13, 2026Bot·

  ·

• Add support for Linen models in process_prequantized_params.

  #408·copybara-service[bot] opened Oct 12, 2026Bot·

  ·

• Shard Linen params in process_prequantized_params using their partitioning metadata.

  #407·copybara-service[bot] opened Oct 12, 2026Bot·

  ·

• Refactor template sharding resolution into a new _resolve_template_sharding helper.

  #406·copybara-service[bot] opened Oct 12, 2026Bot·

  ·

• Implement unbiased aligned FP8 + INT4 multi-pass residual quantization in QWIX.

  #405·copybara-service[bot] opened Oct 9, 2026Bot·

  ·

• Support hardware-compressed 1:4 structured sparse FP8/S4 lowering.

  #403·copybara-service[bot] opened Oct 5, 2026Bot·

  ·

• Enable output-scaled dot_general for symmetric 4-bit W4A16 when contracting tile sizes meet hardware MXU minimums.

  #401·copybara-service[bot] opened Oct 5, 2026Bot·

  ·

• Dequantize fine-grained 4-bit operands to FP8 on input in W4A8 dot_general.

  #400·copybara-service[bot] opened Oct 5, 2026Bot·

  ·

• Resolve Pyrefly and pytype type-checking suppressions across Qwix.

  #399·copybara-service[bot] opened Oct 1, 2026Bot·

  ·

• Internal

  #395·copybara-service[bot] opened Sep 23, 2026Bot·

  ·

• Internal

  #394·copybara-service[bot] opened Sep 23, 2026Bot·

  ·

• Internal

  #393·copybara-service[bot] opened Sep 23, 2026Bot·

  ·

• Internal

  #392·copybara-service[bot] opened Sep 23, 2026Bot·

  ·

• Internal

  #391·copybara-service[bot] opened Sep 23, 2026Bot·

  ·

• Internal

  #390·copybara-service[bot] opened Sep 23, 2026Bot·

  ·

• Internal

  #389·copybara-service[bot] opened Sep 23, 2026Bot·

  ·

• Leave uncalibrated activations in float during ODML conversion.

  #388·copybara-service[bot] opened Sep 23, 2026Bot·

  ·

• Add quantize_linen_class to quantize Linen classes directly.

  #387·copybara-service[bot] opened Sep 23, 2026Bot·

  ·

• Locate weights by the module that declares them, not the one that reads them.

  #386·copybara-service[bot] opened Sep 23, 2026Bot·

  ·

• Fix the degenerate-scale guard for float64 calibration arrays.

  #385·copybara-service[bot] opened Sep 23, 2026Bot·

  ·

• Treat minimum and maximum as same-scale ops.

  #384·copybara-service[bot] opened Sep 23, 2026Bot·

  ·

• Restart Qwix op ids on every module invocation.

  #383·copybara-service[bot] opened Sep 23, 2026Bot·

  ·

• Apply per-tensor activation scales at the rank of the tensor being quantized.

  #382·copybara-service[bot] opened Sep 23, 2026Bot·

  ·

• Add Gemma 4 E2B MLP quantization and multi-pass sweep configuration.

  #381·copybara-service[bot] opened Sep 22, 2026Bot·

  ·