All pull requests

Pull requests

• Adds the tpu_raiden_host_pool_blocks{host_ip, dp_rank, state} gauge tracking

  #1378·copybara-service[bot] opened Oct 28, 2026Bot·

  ·

• Records the duration of every KV cache recall batch in

  #1375·copybara-service[bot] opened Oct 28, 2026Bot·

  ·

• Internal

  #1374·copybara-service[bot] opened Oct 28, 2026Bot·

  ·

• Implements the kv_save_blocks_total counter metric (tpu_raiden_kv_save_blocks_total{host_ip, dp_rank, tier, outcome}) to track block save outcomes across KV cache offload tiers:

  #1371·copybara-service[bot] opened Oct 27, 2026Bot·

  ·

• Wire tpu_raiden_kv_recall_blocks_total into XEye KV Cache Console

  #1370·copybara-service[bot] opened Oct 27, 2026Bot·

  ·

• Allow golden_commit.py sglang and the sglang_canary workflow job to match and wait on sgl-project/sglang-torchtpu's weekly Monday 12:00 UTC scheduled Nightly Test (TPU - Remote Wheel) run (or an explicit --run-id) read-only via GITHUB_TOKEN when SGLANG_GITHUB_TOKEN is unset.

  #1363·copybara-service[bot] opened Oct 26, 2026Bot·

  ·
  1

• Counts every KV cache block recall attempt in KVCacheStore, by tier and

  #1351·copybara-service[bot] opened Oct 26, 2026Bot·

  ·

• Add host-memory source and destination to the KV transfer protocol: a

  #1348·copybara-service[bot] opened Oct 26, 2026Bot·

  ·
  1

• In KVCacheStore::Create, RecoverFromLocalManifest runs before store_monitor_ is constructed, so its RequestRepublish call was a no-op when recovering from a warm shared-memory metadata table (RAIDEN_SHM_KEY). Record the number of blocks recovered during Create and invoke store_monitor_->RequestRepublish() once the monitor starts so recovered host-resident blocks are advertised to the global registry on startup.

  #1310·copybara-service[bot] opened Oct 22, 2026Bot·

  ·
  1

• Add PyTorch KVCacheManager functional offload/reload tests

  #1289·copybara-service[bot] opened Oct 19, 2026Bot·

  ·
  1

• Add a TPU buffer provider for Mooncake TENT.

  #1279·copybara-service[bot] opened Oct 19, 2026Bot·

  ·
  1

• Bypass DRAM cache residency in KVCacheStore when TDS storage caching is active.

  #1278·copybara-service[bot] opened Oct 19, 2026Bot·

  ·
  1

• ci: add jax_version workflow_dispatch input to wheel workflows

  #1261·linamy85 opened Oct 15, 2026Collaborator·

  ·
  1

• Update install instructions to the tpu-sync-{jax,torch} PyPI packages.

  #1245·copybara-service[bot] opened Oct 13, 2026Bot·

  ·

• Reject an invalid read registration with an error instead of 0

  #1242·yinlin09 opened Oct 13, 2026Collaborator·

  ·

• Release idle data connections per transfer (TPU_RAIDEN_DATA_CONN_REUSE) and decouple streams per push from the rank count (TPU_RAIDEN_TRANSFER_STREAMS)

  pull ready
  #1241·sixiang-google opened Oct 13, 2026Collaborator·

  ·
  1

• Report why a transfer failed in poll_stats_with_details

  #1239·yinlin09 opened Oct 13, 2026Collaborator·

  ·

• Introduce StagingArena base class with LayerArena, SlotArena and BlockArena

  #1232·copybara-service[bot] opened Oct 12, 2026Bot·

  ·

• Add fault injection hooks for tpu_sync DMA dispatch paths.

  #1230·copybara-service[bot] opened Oct 12, 2026Bot·

  ·
  1

• Await already-started copies before reporting partial-start errors.

  #1228·copybara-service[bot] opened Oct 12, 2026Bot·

  ·
  1

• Keep the PJRT client alive until torch raw buffer aliases are destroyed

  #1199·norx1991 opened Oct 6, 2026·

  ·
  1

• Let a pool reshard transfer name any subset of source ranks

  #1180·jinzhenfan opened Oct 5, 2026·

  ·

• Add COMMAND_START_SWARM_SESSION to start a swarm session on a listener

  #1170·copybara-service[bot] opened Oct 5, 2026Bot·

  ·

• Add swarm bundle-pull functional test with a trainer-less seed sampler

  #1169·copybara-service[bot] opened Oct 5, 2026Bot·

  ·

• Add sampler-side bundle pull engine to WeightSynchronizerBase

  #1168·copybara-service[bot] opened Oct 5, 2026Bot·

  ·