Issues

• Failed build: MaxText Package Tests

  build failed
  AI-Hypercomputer/maxtext#5630 · Filter by author github-actions opened on Oct 9, 2026
  18

• MaxText GPU Docker Image Build Failure

  gpu-docker-image-build-failure
  AI-Hypercomputer/maxtext#5624 · Filter by author github-actions opened on Oct 9, 2026
  1

• MaxText TPU Docker Image Build Failure

  tpu-docker-image-build-failure
  AI-Hypercomputer/maxtext#5623 · Filter by author github-actions opened on Oct 9, 2026
  1

• Olmo 3: to_huggingface has no HF_SHAPE entry, and the exported config.json has the wrong eos_token_id, max_position_embeddings and RoPE format

  AI-Hypercomputer/maxtext#5587 · Filter by author abhishekraok opened on Oct 7, 2026

• Olmo 3: YaRN is applied to sliding-window layers; the trained model uses it on full-attention layers only (+0.105 nats at init)

  AI-Hypercomputer/maxtext#5544 · Filter by author abhishekraok opened on Oct 5, 2026
  1

• olmo3-7b-pt default HF id (main revision) uses YaRN, but the config uses default RoPE (KL 0.35 vs HF)

  AI-Hypercomputer/maxtext#5524 · Filter by author abhishekraok opened on Oct 3, 2026

• SFT: multi-turn conversations are templated per round, inserting extra system prompts and EOS between turns

  AI-Hypercomputer/maxtext#5522 · Filter by author abhishekraok opened on Oct 3, 2026
  2

• [Bug] _compat_top_k shim does not handle traced inputs → ShardingTypeError in tpu-inference gather_logprobs on JAX 0.11

  bug
  AI-Hypercomputer/maxtext#5440 · Filter by author piowag opened on Sep 29, 2026
  1

• bug: Numeric divergence between the mHC layer and its Pallas kernel

  bug
  AI-Hypercomputer/maxtext#5195 · Filter by author denis-mil opened on Sep 10, 2026
  3

• LoRA under scan_layers=True restacks the whole base model, once per adapter

  AI-Hypercomputer/maxtext#5129 · Filter by author hodaaaaaaaaaa opened on Sep 3, 2026

• checkpoint_conversion: save fails with Orbax NoEntryError when --lazy_load_tensors is combined with a single simulated device

  bug
  AI-Hypercomputer/maxtext#5106 · Filter by author janbernloehr opened on Sep 2, 2026

• Multi-host config validation assumes 1 device, so global_batch_size_to_train_on is derived wrong

  AI-Hypercomputer/maxtext#4933 · Filter by author WandLZhang opened on Aug 18, 2026

• get_context_partition_and_sub_seq reads context_autoregressive, so MoE subgroup splitting never engages under ici_context_parallelism

  AI-Hypercomputer/maxtext#4931 · Filter by author WandLZhang opened on Aug 18, 2026

• [Feature Proposal]: Transition MaxText Checkpoint Conversion to Autoscaling CPU Node Pool (n2-highmem-64)

  feature request
  AI-Hypercomputer/maxtext#4888 · Filter by author RexBearIU opened on Aug 14, 2026
  2

• Multimodal SFT: Conversion errors to HF Safetensors

  bug
  AI-Hypercomputer/maxtext#4712 · Filter by author dgouju opened on Aug 3, 2026
  1

  

• Rampup batch size settings are not validated in the Pydantic config

  AI-Hypercomputer/maxtext#4690 · Filter by author Atishyy27 opened on Jul 31, 2026
  

• partial_rotary_factor is ignored when use_mrope is true (affects qwen3.5-35b-a3b and qwen3.5-397b-a17b)

  AI-Hypercomputer/maxtext#4616 · Filter by author lettt325 opened on Jul 26, 2026

• [bug] Mid-Training Evaluation Loop Spawns Subprocess Claiming /dev/vfio/0 ('Device or resource busy')

  bug
  AI-Hypercomputer/maxtext#4612 · Filter by author karajendran opened on Jul 25, 2026
  1

• [bug] Dynamic TPU Metric Protobuf Import Causes Linkage Panic ('Check failed: GeneratedDatabase()->Add(...)')

  bug
  AI-Hypercomputer/maxtext#4611 · Filter by author karajendran opened on Jul 25, 2026
  1

• Qwen3-Next GDN hardcodes Precision.HIGHEST, ignoring config.matmul_precision — is this intentional?

  AI-Hypercomputer/maxtext#4510 · Filter by author Noman654 opened on Jul 16, 2026
  2

  

• cudnn_flash_jax silently attends across packed sequence boundaries (decoder_segment_ids ignored in training)

  bug
  AI-Hypercomputer/maxtext#4476 · Filter by author bzantium opened on Jul 15, 2026
  1

  

• Qwen3-Next: sequential GDN paths (lax.scan recurrence, solve_triangular) bound TPU step time

  AI-Hypercomputer/maxtext#4347 · Filter by author bzantium opened on Jul 3, 2026
  1

  

• Refactor: Decouple GMM kernel backend dispatch from quantization logic in MoE sparse_matmul

  feature request
  AI-Hypercomputer/maxtext#3646 · Filter by author pengchengneo opened on Apr 12, 2026
  1

  

• [Bug] Qwix Quantization Qwen3-14B: jax._src.dtypes.InvalidInputException

  bug
  AI-Hypercomputer/maxtext#3520 · Filter by author mrzzy opened on Mar 30, 2026
  2

  

• Conversion scripts fail on standard TPU VMs for large models

  bug
  AI-Hypercomputer/maxtext#3418 · Filter by author zzxslp opened on Mar 14, 2026
  3