Repository navigation

All pull requests

Shard TPU Pathways integration tests across 2 runners.
#6241copybara-service[bot] opened Nov 10, 2026Bot
Re-pin Gemma4 RL scripts to cloud-pathways 20261110-jax_0.10.1 digests.
#6240copybara-service[bot] opened Nov 10, 2026Bot
Re-pin and bump JAX versions of Pathways image references to cloud-pathways.
#6239copybara-service[bot] opened Nov 10, 2026Bot
1
Use cloud-pathways sidecar image repo in MaxText build script.
#6238copybara-service[bot] opened Nov 10, 2026Bot
1
Update default Pathways images in configs, scripts, and wikis to cloud-pathways.
#6237copybara-service[bot] opened Nov 10, 2026Bot
1
Make Profiler.deactivate() idempotent for nsys mode
#6236xd-nv opened Nov 10, 20264 tasks done
2
Keep weight decay off frozen vision and audio encoders
#6235nuinashco-ks opened Nov 10, 20264 tasks
1
test(vllm): check the direct MaxText-to-vLLM weight sync on real model trees
#6234chiajunglien opened Nov 10, 2026Collaborator4 tasks done
7
Register colocated-python handlers before v0 checkpoint managers
#6233yrchen-tw opened Nov 10, 2026Collaborator4 tasks done
6
Skip Orbax ArrayMetadata for colocated-python checkpoints to fix the elastic scale-up hang
#6232yrchen-tw opened Nov 10, 2026Collaborator4 tasks done
5
Adds run_trainer_benchmark.py and README.md under src/maxtext/experimental/rl_trainer_benchmark/ to benchmark and profile the Tunix RL trainer program on a single trainer slice without an orchestrator or rollout slices.
#6229copybara-service[bot] opened Nov 10, 2026Bot
1
fix(goodput): flush GoodputRecorder after recording job_end_time
#6228KevinYenky opened Nov 10, 2026Collaborator4 tasks done
6
fix(vllm): return the vLLM KV cache from more scanned decoder layers
#6226chiajunglien opened Nov 10, 2026Collaborator4 tasks done
4
fix(vllm): interleave prefused MoE wi only for the TP fused kernel
#6225chiajunglien opened Nov 10, 2026Collaborator4 tasks done
6
Add UntrainableParam for weights that are never trained
#6224nuinashco-ks opened Nov 10, 20264 tasks
2
Stream Hugging Face checkpoints from GCS into MaxText (combined #5593-#5806)
#6223snehalv2002 opened Nov 10, 2026Collaborator4 tasks done
3
Introduce root AGENTS.md and the /maxtext-self-review local skill so coding
#6222copybara-service[bot] opened Nov 9, 2026Bot
Add GKE runbook for MaxText in-memory snapshotting with Pathways.
#6221copybara-service[bot] opened Nov 9, 2026Bot
Bring up Ming-Image-0.1-Design text-to-image in Maxtext-M3
#6220lydhr opened Nov 9, 2026Collaborator4 tasks done
8
Pin FSDP prefetch collectives to SparseCore 1 and register async overlap transform
#6219darisoy opened Nov 9, 2026Collaborator7 tasks done
4
Bring up Wan 2.1 I2V 14B 480P image-to-video in MaxText-M3
#6217lydhr opened Nov 9, 2026Collaborator4 tasks done
3
ci: Bump MAX_CPU_WORKERS from 4 to 5 shards
#6216darisoy opened Nov 9, 2026Collaborator7 tasks done
3
Take the colocated host count from the hosts' shard count rows
#6215ecnal-cienet opened Nov 9, 2026Collaborator4 tasks done
4
Add a standalone pure-JAX + Pallas trainer for the open-weight RecursiveSI/flint 40B-A2.5B MoE model on 64-chip TPU v7x (tpu7x:4x4x4) under src/maxtext/experimental/flint40b/, with 27 tuned recipes from 8K to 256K context; the Python entrypoints default to the optimized 256K recipe.
#6214copybara-service[bot] opened Nov 9, 2026Bot
1
Remove the temporary hasattr(..., "with_fwd_only_fn") checks and legacy fwd_only(fn, ...) fallback from TrainerWorker (initialize and per_token_logps) and the MaxText per-token log-probability E2E test now that MaxTextTrainingEngine implements with_fwd_only_fn.
#6212copybara-service[bot] opened Nov 9, 2026Bot
1

Navigation Menu

Repository navigation

All pull requests

Shard TPU Pathways integration tests across 2 runners.

Re-pin Gemma4 RL scripts to cloud-pathways 20261110-jax_0.10.1 digests.

Re-pin and bump JAX versions of Pathways image references to cloud-pathways.

Use cloud-pathways sidecar image repo in MaxText build script.

Update default Pathways images in configs, scripts, and wikis to cloud-pathways.

Make Profiler.deactivate() idempotent for nsys mode

Keep weight decay off frozen vision and audio encoders

test(vllm): check the direct MaxText-to-vLLM weight sync on real model trees

Register colocated-python handlers before v0 checkpoint managers

Skip Orbax ArrayMetadata for colocated-python checkpoints to fix the elastic scale-up hang

Adds `run_trainer_benchmark.py` and `README.md` under `src/maxtext/experimental/rl_trainer_benchmark/` to benchmark and profile the Tunix RL trainer program on a single trainer slice without an orchestrator or rollout slices.

fix(goodput): flush GoodputRecorder after recording job_end_time

fix(vllm): return the vLLM KV cache from more scanned decoder layers

fix(vllm): interleave prefused MoE wi only for the TP fused kernel

Add UntrainableParam for weights that are never trained

Stream Hugging Face checkpoints from GCS into MaxText (combined #5593-#5806)

Introduce root `AGENTS.md` and the `/maxtext-self-review` local skill so coding

Add GKE runbook for MaxText in-memory snapshotting with Pathways.

Bring up Ming-Image-0.1-Design text-to-image in Maxtext-M3

Pin FSDP prefetch collectives to SparseCore 1 and register async overlap transform

Bring up Wan 2.1 I2V 14B 480P image-to-video in MaxText-M3

ci: Bump MAX_CPU_WORKERS from 4 to 5 shards

Take the colocated host count from the hosts' shard count rows

Add a standalone pure-JAX + Pallas trainer for the open-weight `RecursiveSI/flint` 40B-A2.5B MoE model on 64-chip TPU v7x (`tpu7x:4x4x4`) under `src/maxtext/experimental/flint40b/`, with 27 tuned recipes from 8K to 256K context; the Python entrypoints default to the optimized 256K recipe.

Remove the temporary `hasattr(..., "with_fwd_only_fn")` checks and legacy `fwd_only(fn, ...)` fallback from `TrainerWorker` (`initialize` and `per_token_logps`) and the MaxText per-token log-probability E2E test now that `MaxTextTrainingEngine` implements `with_fwd_only_fn`.

Pull requests