All pull requests

Pull requests

• 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

• Add Nemotron-3-Nano-Omni configuration and checkpoint conversion

  gemini-review
  #5819·AusarYao28 opened Oct 17, 2026Collaborator·4 tasks done·

  ·
  8

• Add Nemotron-3 hybrid Mamba-2 / Attention / MoE model architecture

  gemini-review
  #5818·AusarYao28 opened Oct 17, 2026Collaborator·4 tasks done·

  ·
  5

• Support non-gated MoE experts, squared ReLU, and hybrid Mamba FLOPs

  gemini-review
  #5817·AusarYao28 opened Oct 17, 2026Collaborator·4 tasks done·

  ·
  7

• Baseline TPU v5p profiles (per_device_batch_size=1, max_target_length=4096) show training step time is dominated by Gated DeltaNet triangular solves (44%) and QSA indexer mask gathers (28%), followed by MoE 512-expert collectives (9%) and MoE ragged-dot compute (3%), while dense QSA attention takes 0.2% (but drives peak HBM).

  #5816·copybara-service[bot] opened Oct 17, 2026Bot·

  ·
  1

  

• Adds calculate_qwen3_8_flash_next_tflops_training_per_device in src/maxtext/utils/maxtext_utils.py and wires it into calculate_tflops_training_per_device and get_shared_expert_mlp_dim to accurately compute training TFLOPs for the Qwen3.8-Flash-Next hybrid architecture:

  #5815·copybara-service[bot] opened Oct 17, 2026Bot·

  ·
  1

  

• DeepSeek-V4.1-Flash E2E training fixes, _scatter_bool 9.7x speedup, and peak-roofline TFLOPs

  #5812·Rohan-Bierneni opened Oct 17, 2026Collaborator·

  ·
  2

• Tag RC Docker images during build and run CI, E2E, and performance tests in parallel

  #5811·SurbhiJainUSC opened Oct 17, 2026Collaborator·4 tasks done·

  ·
  2

• Add GLM-5.3-Flash conversion and HF logits integration

  #5809·hengtaoguo opened Oct 16, 2026Collaborator·2/2·

  ·
  4

• Add GLM-5.3-Flash model layers and configuration

  #5808·hengtaoguo opened Oct 16, 2026Collaborator·1/2·

  ·
  4

• Add minimal m3 training loop for profiling

  gemini-reviewpull ready
  #5807·bvandermoon opened Oct 16, 2026Collaborator·4 tasks done·

  ·
  14

• HF streaming (7/7): Padded row reads, E8M0 scales, bounded writes, checkpoint quantization_config

  #5806·snehalv2002 opened Oct 16, 2026Collaborator·3 of 4 tasks·

  ·
  4

• [2/2] Add generic audio generation with Qwen3-Omni support

  #5804·hengtaoguo opened Oct 16, 2026Collaborator·2/2·3 of 4 tasks·

  ·
  7

• [1/2] Add Qwen3-Omni speech components and checkpoint support

  #5803·hengtaoguo opened Oct 16, 2026Collaborator·1/2·3 of 4 tasks·

  ·
  5

• Add DeepSeek-V4.1-Flash (text) with on-the-fly FP8/FP4 checkpoint loading

  #5802·snehalv2002 opened Oct 16, 2026Collaborator·3 of 4 tasks·

  ·
  6

• Add talker for Qwen3-Omni to generate audio outputs

  #5801·hengtaoguo opened Oct 16, 2026Collaborator·4 tasks·

  ·
  5

• Qwen3.5-397B RL trainer on 128 v7x: fp8 fake-quant hoist, fused SwiGLU backward, attention backward savings, the wi forward GMM's weight pipeline depth and the GDN context-parallel m scan as config keys

  #5799·parambole opened Oct 16, 2026Collaborator·

  ·
  3

• Fall back to the bf16 combine reduce-scatter in the DSv3 eval step.

  #5797·copybara-service[bot] opened Oct 16, 2026Bot·

  ·

• Introduce a custom scheduling pass that can help improve the scheduling

  #5795·Tixxx opened Oct 16, 2026·

  ·
  4

• Fix and speed up CPU checkpoint conversion and forward_pass_logit_checker

  pull ready
  #5794·chiajunglien opened Oct 16, 2026Collaborator·4 tasks done·

  ·
  2

• Enable 4 virtual CPU devices in cpu-integration and move sharding/smoke tests to CPU

  #5790·copybara-service[bot] opened Oct 16, 2026Bot·

  ·
  1

  

• Right-size GPU CI worker cap and prune stale deselected test keys from per_test_baseline.json

  #5789·copybara-service[bot] opened Oct 16, 2026Bot·

  ·
  1

  

• Gate full TPU Pathways test suites to scheduled runs and run smoke train tests on PRs

  #5788·copybara-service[bot] opened Oct 16, 2026Bot·

  ·
  1

  

• Reduce redundant kernel shape sweeps and model dimensions in unit tests.

  #5787·copybara-service[bot] opened Oct 16, 2026Bot·

  ·
  1