Skip to content

Pull requests: google/tpu-sync

New pull request New
116 Open 688 Closed
116 Open 688 Closed
Author
Filter by author
Loading
Label
Filter by label
Loading
Use alt + click/return to exclude labels
or ⇧ + click/return for logical OR
Projects
Filter by project
Loading
Milestones
Filter by milestone
Loading
Reviews
Filter by reviews
No reviews Review required Approved review Changes requested
Assignee
Filter by who’s assigned
Assigned to nobody Loading
Sort
Sort by
Newest Oldest Most commented Least commented Recently updated Least recently updated Best match
Most reactions
👍 👎 😄 🎉 😕 ❤️ 🚀 👀

Pull requests list

Fixed an issue about heterogeneous layer transfer size.
#806 opened Aug 28, 2026 by copybara-service Bot Loading…
Enhance the shared-memory recovery e2e tests with production scenes: hybrid multi-array, multi-shard, and MPMD.
#805 opened Aug 28, 2026 by copybara-service Bot Loading…
Fix TSAN data race on vptr during WeightSynchronizerBase teardown
#804 opened Aug 28, 2026 by copybara-service Bot Loading…
Add multi-dimensional label buffering and metric routing to BufferedMetricsExporter.
#800 opened Aug 27, 2026 by copybara-service Bot Loading…
Implement PrometheusShmExporter to expose aggregated shared memory metrics over HTTP on port 9090.
#799 opened Aug 27, 2026 by copybara-service Bot Loading…
Unify JAX WeightSynchronizer API across Pathways FFI and native backends
#794 opened Aug 27, 2026 by copybara-service Bot Loading…
Add multi-dimensional telemetry labels support to TPU Raiden.
#789 opened Aug 27, 2026 by copybara-service Bot Loading…
Allow internal repo to access device_buffer in torch_tpu/eager BUILD.
#783 opened Aug 26, 2026 by copybara-service Bot Loading…
1
Implement low-overhead, memory-mapped ShmWriter for recording TPU Raiden telemetry metrics directly into POSIX shared memory (/dev/shm).
#782 opened Aug 26, 2026 by copybara-service Bot Loading…
Multi-worker TPU training setups require aggregated host-level metrics across multiple isolated worker processes without inter-process lock contention or high CPU serialization overhead.
#777 opened Aug 25, 2026 by copybara-service Bot Loading…
Fix compilation errors in KVCacheStoreMock.
#775 opened Aug 24, 2026 by copybara-service Bot Loading…
Fix PyTorch missing visibility compilation failures
#774 opened Aug 24, 2026 by copybara-service Bot Loading…
[tpu_raiden] Hardening: async recovery dispatch, result snapshotting, clock decoupling, and teardown cancellation
#773 opened Aug 24, 2026 by copybara-service Bot Loading…
Fix compilation errors and test failures across tpu_sync targets
#772 opened Aug 24, 2026 by copybara-service Bot Loading…
Add support for persistent storage backend offloading and recalling in TPU Raiden
#768 opened Aug 24, 2026 by copybara-service Bot Loading…
Transfer staging leases and generation-scoped pool plan cleanup
#761 opened Aug 23, 2026 by yinlin09 Collaborator • Draft
A TP>1 decode engine with replicated KV caches needs the same bytes on all of its workers, but a pool-reshard plan allowed exactly one destination. The changes are for supporting transfer to N destinations.
#744 opened Aug 21, 2026 by copybara-service Bot Loading…
[DRAFT] Stage resharding transfers in plan-allocated host blocks
#736 opened Aug 21, 2026 by yinlin09 Collaborator • Draft
raw_transfer: compile across the XLA PjRtRawBufferInterface->CommonPjRtRawBuffer rename
#732 opened Aug 20, 2026 by lokic233 Loading…
2
Adopt the caller's device block when re-inserting an HBM-only entry.
#728 opened Aug 20, 2026 by copybara-service Bot Loading…
Harden legacy UUID KV transfer leases and teardown pull ready
#723 opened Aug 19, 2026 by wangyang59 Loading…
@juncgu-google
8
[DRAFT] Allocate host staging per transfer instead of fixed slots
#722 opened Aug 19, 2026 by yinlin09 Collaborator • Draft
Demote cold KV blocks to placement targets when free blocks run low.
#712 opened Aug 18, 2026 by copybara-service Bot Loading…
1
[Weight sync destination 4/N] Complete RolloutWorker's destination entry points
#689 opened Aug 17, 2026 by copybara-service Bot Loading…
1
[Weight sync destination 3/N] Close rollout admission during weight sync rounds
#688 opened Aug 17, 2026 by copybara-service Bot Loading…
1
ProTip! Follow long discussions with comments:>50.