AI-Hypercomputer/maxtext

View on GitHub

A simple, performant, and scalable Jax LLM!

Stars: 2.4k
Forks: 607
Open beginner issues: 0
Indexed issues: 67
Avg merge: 2d 19h
Merged PRs (30d): 158

Dominant language: Python
License: Apache-2.0
Last GitHub push: Sep 12, 2026
Latest indexed: Sep 13, 2026
Contributing guide: Contributing guide
Code of conduct: No code of conduct
Beginner labels: good first issue

Open All

67 issues indexed so far

`attend_dtype` not used Open

bug

AI-Hypercomputer/maxtext#531 · 1 comment · 0 reactions · 1 assignee · Indexed Sep 13, 2026
Create a user friendly inference demo Open

inference

AI-Hypercomputer/maxtext#532 · 0 comments · 0 reactions · 1 assignee · Indexed Sep 13, 2026
Supported features Open

feature request

AI-Hypercomputer/maxtext#571 · 21 comments · 2 reactions · 2 assignees · Indexed Sep 13, 2026
Support beam search Open

feature request inference

AI-Hypercomputer/maxtext#594 · 0 comments · 0 reactions · 1 assignee · Indexed Sep 13, 2026
Support LoRA training Open

feature request

AI-Hypercomputer/maxtext#609 · 5 comments · 3 reactions · 1 assignee · Indexed Sep 13, 2026
DEFAULT_MASK_VALUE causes gradient explosion and nan loss on deep models Open

bug

AI-Hypercomputer/maxtext#614 · 2 comments · 2 reactions · 1 assignee · Indexed Sep 13, 2026
Multihost training collapses from time to time when loading the next batch Open

bug

AI-Hypercomputer/maxtext#786 · 3 comments · 0 reactions · 1 assignee · Indexed Sep 13, 2026
Support nsys profiler upload in all cases Open

bug good first issue

AI-Hypercomputer/maxtext#911 · 1 comment · 0 reactions · 1 assignee · Indexed Sep 13, 2026
PGLE doesn't work for Tensor Parallelism Open

AI-Hypercomputer/maxtext#1005 · 3 comments · 0 reactions · 2 assignees · Indexed Sep 13, 2026
Why logit checker has such a high tolerance? Open

AI-Hypercomputer/maxtext#1021 · 0 comments · 4 reactions · 1 assignee · Indexed Sep 13, 2026
llama GPU model with dcn fsdp + ici tp + cudnn flash attention broken Open

AI-Hypercomputer/maxtext#1093 · 2 comments · 0 reactions · 1 assignee · Indexed Sep 13, 2026
Apparent bug in the megablocks implementation Open

AI-Hypercomputer/maxtext#1183 · 14 comments · 2 reactions · 1 assignee · Indexed Sep 13, 2026
docker build issue Open

AI-Hypercomputer/maxtext#1212 · 0 comments · 0 reactions · 1 assignee · Indexed Sep 13, 2026
MFU drops significantly when using megablox with more experts Open

AI-Hypercomputer/maxtext#1256 · 8 comments · 1 reaction · 1 assignee · Indexed Sep 13, 2026
dependency error because of orbax-checkpoint version Open

AI-Hypercomputer/maxtext#1273 · 2 comments · 0 reactions · 1 assignee · Indexed Sep 13, 2026
a little error Open

AI-Hypercomputer/maxtext#1321 · 1 comment · 0 reactions · 1 assignee · Indexed Sep 13, 2026
The default setting of `param_scan_axis=1` hurts performance and memory consumption on GPUs Open

AI-Hypercomputer/maxtext#1382 · 1 comment · 0 reactions · 1 assignee · Indexed Sep 13, 2026
When using dcn-DP and dcn-FSDP together got error when saving checkpoint. Open

AI-Hypercomputer/maxtext#1434 · 1 comment · 0 reactions · 1 assignee · Indexed Sep 13, 2026
Please create direct conversion scripts from huggingface for Gemma3 models Open

AI-Hypercomputer/maxtext#1528 · 1 comment · 0 reactions · 1 assignee · Indexed Sep 13, 2026
Make maxtext_xpk_runner support other trainers Open

feature request

AI-Hypercomputer/maxtext#1552 · 0 comments · 0 reactions · 2 assignees · Indexed Sep 13, 2026
Fix "~/xpk" references Open

bug

AI-Hypercomputer/maxtext#1561 · 7 comments · 0 reactions · 1 assignee · Indexed Sep 13, 2026
Package hierarchy: should benchmarks, end_to_end, pedagogical_examples be under MaxText? Open

AI-Hypercomputer/maxtext#1635 · 3 comments · 0 reactions · 1 assignee · Indexed Sep 13, 2026
MaxText/prefix_cache.py vs jetstream/core/prefix_cache.py Open

AI-Hypercomputer/maxtext#1643 · 0 comments · 0 reactions · 2 assignees · Indexed Sep 13, 2026
execute bash setup.sh take too long due to dependency version not locked. Open

AI-Hypercomputer/maxtext#1744 · 2 comments · 0 reactions · 1 assignee · Indexed Sep 13, 2026
BUG (w FIX): Llama3 conversion to HF does not work Open

AI-Hypercomputer/maxtext#1774 · 3 comments · 0 reactions · 1 assignee · Indexed Sep 13, 2026
GCS upload errors should not be silent Open

AI-Hypercomputer/maxtext#1817 · 2 comments · 0 reactions · 1 assignee · Indexed Sep 13, 2026
[OOM] Huge memory usage in MoE Open

AI-Hypercomputer/maxtext#1828 · 1 comment · 0 reactions · 1 assignee · Indexed Sep 13, 2026
[REQUEST] Qwen3 support Open

AI-Hypercomputer/maxtext#1831 · 1 comment · 1 reaction · 1 assignee · Indexed Sep 13, 2026
Regression: HF JSONL training halts early with Open

bug

AI-Hypercomputer/maxtext#2147 · 4 comments · 0 reactions · 1 assignee · Indexed Sep 13, 2026
Conversion to unscanned format failed Open

bug

AI-Hypercomputer/maxtext#2176 · 2 comments · 0 reactions · 2 assignees · Indexed Sep 13, 2026
Gemma3 Model Token Logits Mismatch with Different max_prefill_predict_length Open

bug

AI-Hypercomputer/maxtext#2179 · 2 comments · 0 reactions · 1 assignee · Indexed Sep 13, 2026
Run pretraining of llama 7b/8b models on GPUs using a3/1vm.sh Open

bug

AI-Hypercomputer/maxtext#2193 · 4 comments · 0 reactions · 1 assignee · Indexed Sep 13, 2026
QAT SFT Training Fails with Qwen3 MOE Open

bug

AI-Hypercomputer/maxtext#2305 · 0 comments · 0 reactions · 1 assignee · Indexed Sep 13, 2026
[feature request] Support of GLM4.5V and Deepseek V3.1 Open

feature request

AI-Hypercomputer/maxtext#2330 · 3 comments · 0 reactions · 1 assignee · Indexed Sep 13, 2026
Memory Leak when initialising Qwen model training with multihost runner Open

bug

AI-Hypercomputer/maxtext#2374 · 1 comment · 0 reactions · 1 assignee · Indexed Sep 13, 2026
wandb logging support Open

feature request

AI-Hypercomputer/maxtext#2434 · 5 comments · 0 reactions · 1 assignee · Indexed Sep 13, 2026
gemma2-9b regressed by introduction of OSS-GPT Open

bug

AI-Hypercomputer/maxtext#2542 · 4 comments · 0 reactions · 2 assignees · Indexed Sep 13, 2026
Model conversion on pathways cluster not working Open

bug

AI-Hypercomputer/maxtext#2578 · 1 comment · 0 reactions · 1 assignee · Indexed Sep 13, 2026
Documentation: base.yml comments don't match actual quantization option names in code Open

AI-Hypercomputer/maxtext#2593 · 0 comments · 0 reactions · 1 assignee · Indexed Sep 13, 2026
Dependency conflicts between tpu-commons, tokamax, and google-tunix Open

bug

AI-Hypercomputer/maxtext#2664 · 1 comment · 0 reactions · 1 assignee · Indexed Sep 13, 2026
Robustness against network errors Open

feature request

AI-Hypercomputer/maxtext#2861 · 1 comment · 0 reactions · 1 assignee · Indexed Sep 13, 2026
Gemma3 to hf mapping error Open

bug

AI-Hypercomputer/maxtext#2867 · 5 comments · 0 reactions · 1 assignee · Indexed Sep 13, 2026
Grain pool is exiting Open

bug

AI-Hypercomputer/maxtext#2878 · 3 comments · 0 reactions · 1 assignee · Indexed Sep 13, 2026
SFT Model Convert to HuggingFace Error Open

bug

AI-Hypercomputer/maxtext#3003 · 2 comments · 0 reactions · 1 assignee · Indexed Sep 13, 2026
Qwen3 MoE Conversion / Pre-Training Error Open

bug

AI-Hypercomputer/maxtext#3231 · 2 comments · 0 reactions · 1 assignee · Indexed Sep 13, 2026
[Bug] Tunix GRPO VllmRollout.get_per_token_logps API mismatch and JAX concatenate type error Open

bug

AI-Hypercomputer/maxtext#3363 · 3 comments · 0 reactions · 1 assignee · Indexed Sep 13, 2026
Issue with gradient accumulation and expansion_factor_real_data>1 Open

bug

AI-Hypercomputer/maxtext#3381 · 0 comments · 0 reactions · 1 assignee · Indexed Sep 13, 2026
Conversion scripts fail on standard TPU VMs for large models Open

bug

AI-Hypercomputer/maxtext#3418 · 3 comments · 0 reactions · 1 assignee · Indexed Sep 13, 2026
[Bug] Qwix Quantization Qwen3-14B: jax._src.dtypes.InvalidInputException Open

bug

AI-Hypercomputer/maxtext#3520 · 2 comments · 0 reactions · 1 assignee · Indexed Sep 13, 2026
Refactor: Decouple GMM kernel backend dispatch from quantization logic in MoE sparse_matmul Open

feature request

AI-Hypercomputer/maxtext#3646 · 1 comment · 0 reactions · 1 assignee · Indexed Sep 13, 2026

AI-Hypercomputer/maxtext

Get new issues in your inbox