All pull requests

Pull requests

• Add exclude_module_path, step_context, and scan_index_context gating to QtRule and QtProvider

  #436·parsley9877 opened Oct 29, 2026·5 tasks done·

  ·

• Add parametric MultipassConfig, DtypeSpec, and multi-pass ragged_dot_qt support

  #435·parsley9877 opened Oct 29, 2026·5 tasks done·

  ·

• internal changes.

  #434·copybara-service[bot] opened Oct 28, 2026Bot·

  ·

• conv_general_qt: support grouped convolutions in the backward pass

  #433·AshishKumar4 opened Oct 27, 2026·

  ·

• Add an end-to-end training step test for the qwix.kernels fused FP8 matmul

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

  ·

• Add the Pallas dRHS kernel to the qwix.kernels fused FP8 matmul

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

  ·

• Make quantized NNX models picklable by reference.

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

  ·

• Add the Pallas dLHS kernel to the qwix.kernels fused FP8 matmul

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

  ·

• [QwixQt] Configure backward residual quantization and tiling in QtProvider.

  #426·copybara-service[bot] opened Oct 21, 2026Bot·

  ·

• [QwixODML] Replay weight layout transforms when baking per-channel scales in conversion.

  #425·copybara-service[bot] opened Oct 20, 2026Bot·

  ·

• Add opt_type=adam_pax_low_memory, a low-memory bf16 Adam variant, with

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

  ·

• Add distill_save_trainable_params_only and distill_skip_optimizer_checkpoint

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

  ·

• Add distill_share_frozen_teacher_weights and lti_utils.share_frozen_params

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

  ·

• Add remat_unscanned_layers_with_scan to bound peak activation memory when

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

  ·

• Add opt-in fused chunked output-head projection (distill_fused_lm_head and

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

  ·

• Reduce distillation loss memory usage by passing integer labels with an

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

  ·

• Add weight-only N:M sparsity (quantization="" with weight_sparsity_n/m) via

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

  ·

• Accept single-element regex list overrides (for example

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

  ·

• Improve MaxText distillation trainer plumbing for frozen student parameters,

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

  ·

• [QwixMxfpNumerics] Re-encode MXFP4 weights to float8_e5m2 for FP8 matmuls

  #415·copybara-service[bot] opened Oct 17, 2026Bot·

  ·

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

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

  ·