All workflows

Workflow runs

• Qwen3.5-397B RL trainer on 128 v7x: the tuned recipe turns on the fp8 fake-quant hoist, fused SwiGLU backward, attention savings, the 2048 wi forward tile and the GDN context-parallel m scan, and pins the MaxText that defines them

  completed successfully:
  Check the Documentation Build#8145:Pull request#2948synchronizebyparambole
  parambole/rl-4600-final-tunix

  Oct 19, 20261m 46s
  More actions for workflow run 8145

• [main] Tunix Package Tests — ba95987a5a3f1d3cf8a680238c1600a0db6453af

  completed successfully:
  Tunix Package Tests#11550:Scheduled
  main

  Oct 19, 202627m 27s
  More actions for workflow run 11550

• [main] Tunix Package Tests — move splash attention out of experimental ops

  completed successfully:
  Tunix Package Tests#11549:Pull request#2962openedbycopybara-service[bot]Bot
  test_1001052780

  Oct 18, 202627m 28s
  More actions for workflow run 11549

• move splash attention out of experimental ops

  completed successfully:
  Check the Documentation Build#8144:Pull request#2962openedbycopybara-service[bot]Bot
  test_1001052780

  Oct 18, 20262m 3s
  More actions for workflow run 8144

• move splash attention out of experimental ops

  completed successfully:
  Auto-assign issues and PRs#1864:Pull request#2962openedbycopybara-service[bot]Bot
  Oct 18, 20269s
  More actions for workflow run 1864

• move splash attention out of experimental ops

  completed successfully:
  GitHub Actions Scan#4447:Pull request#2962openedbycopybara-service[bot]Bot
  Oct 18, 202618s
  More actions for workflow run 4447

• [main] Tunix Package Tests — ba95987a5a3f1d3cf8a680238c1600a0db6453af

  completed successfully:
  Tunix Package Tests#11548:Scheduled
  main

  Oct 18, 202639m 56s
  More actions for workflow run 11548

• Tunix Nightly Regression Tests

  completed successfully:
  Tunix Nightly Regression Tests#429:Scheduled
  main

  Oct 18, 20261h 44m 3s
  More actions for workflow run 429

• The library side of router replay is on main: the rollout collector encodes the MoE experts each token was routed to, the orchestrator aligns them to the training batch, and the trainer passes them to MaxText as forced_routed_experts. Nothing in the distributed examples turns it on: the rollout node never enables capture in vLLM, and the encoding stays at auto with no expert count. For a model with 256 to 1023 experts per layer, such as Qwen3.5-35B-A3B with 256, auto without a count picks the tier from the highest expert id it sees: a batch whose ids all happen to be below 255 is encoded as uint8 instead of uint10, and the trainer step recompiles for the other shape.

  completed successfully:
  Check the Documentation Build#8143:Pull request#2888synchronizebycopybara-service[bot]Bot
  test_999564282

  Oct 18, 20261m 19s
  More actions for workflow run 8143

• [main] Tunix Package Tests — The library side of router replay is on main: the rollout collector encodes the MoE experts each token was routed to, the orchestrator aligns them to the training batch, and the trainer passes them to MaxText as forced_routed_experts. Nothing in the distributed examples turns it on: the rollout node never enables capture in vLLM, and the encoding stays at auto with no expert count. For a model with 256 to 1023 experts per layer, such as Qwen3.5-35B-A3B with 256, auto ...

  completed successfully:
  Tunix Package Tests#11547:Pull request#2888synchronizebycopybara-service[bot]Bot
  test_999564282

  Oct 18, 202627m 40s
  More actions for workflow run 11547

• The library side of router replay is on main: the rollout collector encodes the MoE experts each token was routed to, the orchestrator aligns them to the training batch, and the trainer passes them to MaxText as forced_routed_experts. Nothing in the distributed examples turns it on: the rollout node never enables capture in vLLM, and the encoding stays at auto with no expert count. For a model with 256 to 1023 experts per layer, such as Qwen3.5-35B-A3B with 256, auto without a count picks the tier from the highest expert id it sees: a batch whose ids all happen to be below 255 is encoded as uint8 instead of uint10, and the trainer step recompiles for the other shape.

  completed successfully:
  GitHub Actions Scan#4446:Pull request#2888synchronizebycopybara-service[bot]Bot
  Oct 18, 202618s
  More actions for workflow run 4446

• [main] Tunix Package Tests — The library side of router replay is on main: the rollout collector encodes the MoE experts each token was routed to, the orchestrator aligns them to the training batch, and the trainer passes them to MaxText as forced_routed_experts. Nothing in the distributed examples turns it on: the rollout node never enables capture in vLLM, and the encoding stays at auto with no expert count. For a model with 256 to 1023 experts per layer, such as Qwen3.5-35B-A3B with 256, auto ...

  completed successfully:
  Tunix Package Tests#11546:Pull request#2888synchronizebycopybara-service[bot]Bot
  test_999564282

  Oct 18, 202627m 34s
  More actions for workflow run 11546

• The library side of router replay is on main: the rollout collector encodes the MoE experts each token was routed to, the orchestrator aligns them to the training batch, and the trainer passes them to MaxText as forced_routed_experts. Nothing in the distributed examples turns it on: the rollout node never enables capture in vLLM, and the encoding stays at auto with no expert count. For a model with 256 to 1023 experts per layer, such as Qwen3.5-35B-A3B with 256, auto without a count picks the tier from the highest expert id it sees: a batch whose ids all happen to be below 255 is encoded as uint8 instead of uint10, and the trainer step recompiles for the other shape.

  completed successfully:
  Check the Documentation Build#8142:Pull request#2888synchronizebycopybara-service[bot]Bot
  test_999564282

  Oct 18, 20261m 44s
  More actions for workflow run 8142

• The library side of router replay is on main: the rollout collector encodes the MoE experts each token was routed to, the orchestrator aligns them to the training batch, and the trainer passes them to MaxText as forced_routed_experts. Nothing in the distributed examples turns it on: the rollout node never enables capture in vLLM, and the encoding stays at auto with no expert count. For a model with 256 to 1023 experts per layer, such as Qwen3.5-35B-A3B with 256, auto without a count picks the tier from the highest expert id it sees: a batch whose ids all happen to be below 255 is encoded as uint8 instead of uint10, and the trainer step recompiles for the other shape.

  completed successfully:
  GitHub Actions Scan#4445:Pull request#2888synchronizebycopybara-service[bot]Bot
  Oct 18, 202618s
  More actions for workflow run 4445

• [main] Tunix Package Tests — 06f9f308e9293fb3c7bbb9e793442f680857c915

  completed successfully:
  Tunix Package Tests#11545:Scheduled
  main

  Oct 18, 202627m 40s
  More actions for workflow run 11545

• Generate PPO eval rollouts with the eval rollout config

  requires action with the application:
  Tunix Package Tests#11544:Pull request#2961openedbyvineethsaivs
  vineethsaivs:fix/ppo-eval-rollout-config

  Oct 18, 2026Action required
  More actions for workflow run 11544

• Generate PPO eval rollouts with the eval rollout config

  requires action with the application:
  Check the Documentation Build#8141:Pull request#2961openedbyvineethsaivs
  vineethsaivs:fix/ppo-eval-rollout-config

  Oct 18, 2026Action required
  More actions for workflow run 8141

• Generate PPO eval rollouts with the eval rollout config

  completed successfully:
  GitHub Actions Scan#4444:Pull request#2961openedbyvineethsaivs
  Oct 18, 202622s
  More actions for workflow run 4444

• Generate PPO eval rollouts with the eval rollout config

  completed successfully:
  Auto-assign issues and PRs#1863:Pull request#2961openedbyvineethsaivs
  Oct 18, 202610s
  More actions for workflow run 1863

• [main] Tunix Package Tests — 06f9f308e9293fb3c7bbb9e793442f680857c915

  completed successfully:
  Tunix Package Tests#11543:Scheduled
  main

  Oct 17, 202627m 34s
  More actions for workflow run 11543

• Tunix Nightly Regression Tests

  completed successfully:
  Tunix Nightly Regression Tests#428:Scheduled
  main

  Oct 17, 20261h 45m 6s
  More actions for workflow run 428

• [main] Tunix Package Tests — Enforce strict request_id conservation in PoolExecutionSession and propagate rollout worker errors loudly.

  completed successfully:
  Tunix Package Tests#11542:Pull request#2958synchronizebycopybara-service[bot]Bot
  test_1000982500

  Oct 17, 202627m 24s
  More actions for workflow run 11542

• Enforce strict request_id conservation in PoolExecutionSession and propagate rollout worker errors loudly.

  completed successfully:
  Check the Documentation Build#8140:Pull request#2958synchronizebycopybara-service[bot]Bot
  test_1000982500

  Oct 17, 20262m 6s
  More actions for workflow run 8140

• Enforce strict request_id conservation in PoolExecutionSession and propagate rollout worker errors loudly.

  completed successfully:
  GitHub Actions Scan#4443:Pull request#2958synchronizebycopybara-service[bot]Bot
  Oct 17, 202618s
  More actions for workflow run 4443

• Enforce strict request_id conservation in PoolExecutionSession and propagate rollout worker errors loudly.

  completed successfully:
  Check the Documentation Build#8139:Pull request#2958synchronizebycopybara-service[bot]Bot
  test_1000982500

  Oct 17, 20261m 38s
  More actions for workflow run 8139