Pull requests: google/tunix
Pull requests list
Plumb memory and checkpoint concurrency flags
#2280
opened Sep 17, 2026 by
niting
Contributor
Loading…
6 tasks done
Add a production-stack distributed RL smoke test to TPU CI
ready-to-submit
#2279
opened Sep 17, 2026 by
igorts-git
Collaborator
•
Draft
Projects
WriteTask payloads (metadata and step) to base ATIF in WriteTask.__post_init__ via .to_atif_metadata() and .to_atif_step() prior to deep copying.
#2277
opened Sep 17, 2026 by
copybara-service
Bot
Loading…
Adds
from_atif_step() and from_atif_metadata() rehydration functions in converter.py to reconstruct TunixTrajectoryMetadata, TunixAgentStep, and TunixEnvStep instances from base ATIF models by unpacking subclass fields from extra["_atif_ext"].
#2276
opened Sep 17, 2026 by
copybara-service
Bot
Loading…
[DO NOT SUBMIT] Test run_core, run_dev, and tpu-nightly-regression on v5e-8 (linux-x86-ct5lp-224-8tpu) runners.
ready-to-submit
#2274
opened Sep 17, 2026 by
copybara-service
Bot
Loading…
[Tunix] Replace AbstractTrainer.per_token_logps with a scoped model_scope.
#2273
opened Sep 17, 2026 by
copybara-service
Bot
Loading…
ci(tpu): Achieve ~2x TPU CI speedup (
run_dev: 21m 06s -> 9m 34s, run_core: 14m 14s -> 8m 51s)
#2272
opened Sep 17, 2026 by
copybara-service
Bot
Loading…
Jfacevedo/trellis grpo recipe
#2270
opened Sep 17, 2026 by
entrpn
Collaborator
Loading…
6 tasks done
[Tunix] Support exact token continuity across agentic rollout turns.
ready-to-submit
#2269
opened Sep 17, 2026 by
copybara-service
Bot
Loading…
Deliver config that was accepted but never passed
#2265
opened Sep 17, 2026 by
entrpn
Collaborator
Loading…
6 tasks done
Adds
.to_atif_step() projection to Step and .to_atif_metadata() projection to TrajectoryMetadata, Trajectory, and TunixTrajectory in trajectory.py. This allows Tunix RL subclasses (TunixAgentStep, TunixEnvStep, TunixTrajectoryMetadata, TunixTrajectory) to project cleanly to base ATIF models by packing subclass-specific fields into extra and stripping step history when projecting metadata.
#2263
opened Sep 16, 2026 by
copybara-service
Bot
Loading…
* Added
QwenVtcChatTemplateParser subclassing QwenChatTemplateParser to format system/user turns with Qwen ChatML control tokens (<|im_start|>, <|im_end|>) and prefill <reasoning>\n into the assistant generation prompt (_init_generation_prompt) and _parse_assistant.
#2262
opened Sep 16, 2026 by
copybara-service
Bot
Loading…
[docker] Pin jax/jaxlib/libtpu to the ABI tpu-inference resolved
#2261
opened Sep 16, 2026 by
khatwanimohit
Collaborator
Loading…
6 tasks done
Pin sglang-jax to LKG commit in TPU CI workflows.
#2260
opened Sep 16, 2026 by
copybara-service
Bot
Loading…
ci: halve per-PR TPU slice demand and add job timeouts
ready-to-submit
#2254
opened Sep 16, 2026 by
shz-google
Loading…
[Tunix] Support DeepSWE evaluation pipeline with MaxText and vLLM on TPU
#2249
opened Sep 16, 2026 by
susanbao
Collaborator
Loading…
6 tasks done
Add design and architecture documentation (
experimental/docs/clip_ratio_and_raw_length_metrics.md) detailing how response-length truncation (generation/completions/clip_ratio) and raw response length (generation/completions/{mean,max,min}_raw_length) metrics are computed across the producer-consumer rollout and trainer pipeline:
#2245
opened Sep 16, 2026 by
copybara-service
Bot
Loading…
Add CPU preemption drill integration test for AgenticRLLearner.
#2242
opened Sep 16, 2026 by
copybara-service
Bot
Loading…
[Tunix] Make GradientAccumulator an immutable pytree value instead of an nnx.Module.
#2241
opened Sep 16, 2026 by
copybara-service
Bot
Loading…
[Tunix] Replace AbstractTrainer.per_token_logps with a generic fwd_only method.
#2235
opened Sep 15, 2026 by
copybara-service
Bot
Loading…
Implement CodeActAgent and OpenHands Agent Sandbox integration in DeepSWE:
ready-to-submit
#2222
opened Sep 14, 2026 by
copybara-service
Bot
Loading…
Adds background worker (
AsyncSqlWriter) and model serialization codecs
#2208
opened Sep 11, 2026 by
copybara-service
Bot
Loading…
ProTip!
Updated in the last three days: updated:>2026-09-14.