All pull requests

Pull requests

• Addresses deferred follow-up items from C-W1, FP-2, and C-W2 reviews:

  #5847·copybara-service[bot] opened Oct 19, 2026Bot·

  ·
  

• Relocates third_party/py/maxtext/src/maxtext/m3/models/glm5_3_flash/glm5_config.py to third_party/py/maxtext/src/maxtext/m3/configs/glm5_3_flash_config.py per the m3 layout convention (src/maxtext/m3/configs/):

  #5846·copybara-service[bot] opened Oct 19, 2026Bot·

  ·
  

• Implements Work Item C-W2 (W3 return_aux contract, F49 heterogeneous super-block slot alignment, removal of **kwargs swallow, and FP-1 Glm5ScanTest stream dtype alignment):

  #5845·copybara-service[bot] opened Oct 19, 2026Bot·

  ·
  

• Removes pre-existing silent-default patterns in modeling_glm5_3_flash.py and glm5_3_flash_mhc_test.py identified by crew_ledger.py scan-diff --jax-rules on:

  #5844·copybara-service[bot] opened Oct 19, 2026Bot·

  ·
  

• Fuse token replication and expert sorting gather in RoutedMoE.permute.

  #5843·copybara-service[bot] opened Oct 19, 2026Bot·

  ·

• Implements F51b (F53, F58, F59) in third_party/py/maxtext/tests/unit/m3/glm5_3_flash_ckpt_parity_test.py and third_party/py/maxtext/tests/unit/m3/BUILD:

  #5842·copybara-service[bot] opened Oct 19, 2026Bot·

  ·
  

• Implements F51a (F52) in third_party/py/maxtext/tests/unit/m3/glm5_3_flash_ckpt_parity_test.py:

  #5841·copybara-service[bot] opened Oct 19, 2026Bot·

  ·
  

• Hardens Glm5MoETest.test_experts_tp_sharding_no_weight_all_gather in third_party/py/maxtext/tests/unit/m3/glm5_3_flash_moe_test.py (W-moe-hlo / F57):

  #5840·copybara-service[bot] opened Oct 19, 2026Bot·

  ·
  

• Hardens third_party/py/maxtext/tests/unit/m3/m3_smoke_test.py (W-smoke / F55):

  #5839·copybara-service[bot] opened Oct 19, 2026Bot·

  ·
  

• Add distill_share_frozen_teacher_weights and lti_utils.share_frozen_params

  #5838·copybara-service[bot] opened Oct 19, 2026Bot·

  ·

• Add distill_save_trainable_params_only and distill_skip_optimizer_checkpoint

  #5837·copybara-service[bot] opened Oct 19, 2026Bot·

  ·

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

  #5836·copybara-service[bot] opened Oct 19, 2026Bot·

  ·

• Add remat_unscanned_layers_with_scan to bound peak activation memory when

  #5835·copybara-service[bot] opened Oct 19, 2026Bot·

  ·

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

  #5834·copybara-service[bot] opened Oct 19, 2026Bot·

  ·

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

  #5833·copybara-service[bot] opened Oct 19, 2026Bot·

  ·

• Improve MaxText distillation trainer plumbing for frozen student parameters,

  #5832·copybara-service[bot] opened Oct 19, 2026Bot·

  ·

• Accept single-element regex list overrides (for example

  #5831·copybara-service[bot] opened Oct 19, 2026Bot·

  ·

• Support weight-only N:M sparsity in Qwix QT rules (QtRule.rhs_sparsity_rule),

  #5830·copybara-service[bot] opened Oct 19, 2026Bot·

  ·
  1

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

  #5829·copybara-service[bot] opened Oct 19, 2026Bot·

  ·
  1

• Fix safetensors_dynamic load timeout on many hosts

  #5828·ecnal-cienet opened Oct 19, 2026Collaborator·4 tasks done·

  ·
  3

• Add a bounded-inflight parallel worker pool and thread-safe safetensors shard cache to the GLM-5.3-Flash streaming checkpoint converter.

  #5827·copybara-service[bot] opened Oct 19, 2026Bot·

  ·
  1

  

• feat(ci): support configurable and RC Docker images in E2E and promotion workflows

  #5824·chiajunglien opened Oct 19, 2026Collaborator·4 tasks done·

  ·
  5

• move splash attention out of experimental ops

  #5823·copybara-service[bot] opened Oct 18, 2026Bot·

  ·
  1

• Extract the platform-neutral review criteria from maxtext-review.toml into

  #5822·copybara-service[bot] opened Oct 17, 2026Bot·

  ·
  

• Add end-to-end TPU test scripts and logit verification for Nemotron-3

  gemini-review
  #5820·AusarYao28 opened Oct 17, 2026Collaborator·4 tasks done·

  ·
  7