Pull requests: google/tunix
Pull requests list
Fix Qwen2 KV-cache dtype mismatch in cached decoding
#1322
opened Mar 27, 2026 by
skwh54
Loading…
4 of 6 tasks
[Tunix] Add support for loading models from Orbax checkpoints for evaluation of individual trained ckpts.
#1318
opened Mar 27, 2026 by
copybara-service
bot
Loading…
add logging of example rewards with snipped output, controlled under
debug=True flag from GRPOConfig
#1317
opened Mar 27, 2026 by
copybara-service
bot
Loading…
updating qwen2 and toy transformer with seqence packing
#1316
opened Mar 27, 2026 by
copybara-service
bot
Loading…
[Tunix] Always log kl regardless if it's enabled.
#1313
opened Mar 26, 2026 by
copybara-service
bot
Loading…
improve rl_cluster init corner case handling and test coverage
#1311
opened Mar 26, 2026 by
copybara-service
bot
Loading…
fix file path for window, : is reserved name on window
#1307
opened Mar 26, 2026 by
HugoTian
Loading…
Add
valid_traj to AgenticRL TrainExample.
#1283
opened Mar 23, 2026 by
copybara-service
bot
Loading…
security: replace eval() with safe AST math evaluator in calculate reward
#1279
opened Mar 23, 2026 by
RaymondSeven
Loading…
[Tunix] Snapshot gradient norm and Log the clipped gradient norm during training.
#1264
opened Mar 19, 2026 by
copybara-service
bot
Loading…
fix: guard against division by zero in masked_var Bessel correction
#1262
opened Mar 18, 2026 by
kbhujbal
Loading…
6 tasks done
[Tunix] Update qwix dependency and TPU test image.
#1260
opened Mar 18, 2026 by
copybara-service
bot
Loading…
[Tunix] add advantage/mean and advantage/std metrics logging and match grpo_learner implementation with agentic_grpo_learner
#1258
opened Mar 17, 2026 by
copybara-service
bot
Loading…
add support for repeating kv head tensors during weight sync.
#1251
opened Mar 16, 2026 by
NicoGrande
Loading…
6 tasks done
allow passing eval_ds parameters through tunix cli, plus bugfixes
#1238
opened Mar 12, 2026 by
andytwigg
Loading…
ProTip!
Find all pull requests that aren't related to any open issues with -linked:issue.