Skip to content

[TPU Raiden] Backend Tier Extension (Part 1/4): Protocols, Interfaces & Abstract Driver Architecture - #880

Merged
copybara-service[bot] merged 1 commit into
mainfrom
test_974774417
Sep 16, 2026
Merged

copybara-service[bot] merged 1 commit into
mainfrom
test_974774417

Conversation

@copybara-service

@copybara-service copybara-service Bot commented Sep 8, 2026

Copy link
Copy Markdown

[TPU Raiden] Backend Tier Extension (Part 1/4): Protocols, Interfaces & Abstract Driver Architecture

Declares the protocol, configuration and driver interfaces that let TPU Raiden
offload and recall KV blocks to a secondary storage tier beyond Host DRAM.
Declarative only: every driver method is a stub returning UnimplementedError,
with the implementations in Part 2.

  • Input: BackendConfig carries serving topology as a ParallelismConfig (tp_size,
    tp_rank); tp_rank is required rather than defaulted, so workers in a sharded
    job cannot collide on one shard directory. PosixBackendOptions is the typed
    parse-and-validate point for a POSIX backend.
  • Output: PosixPathMapper declares the layout
    <root_dir>/<model_name>/tp<tp_size>_r<tp_rank>///<hash_hex>.bin;
    BlockKeyMapper::MapKey requires a hex-encoded, length-bounded hash.
  • Interfaces: the TransferBackendBuffers RPC and BackendTransferSpec in
    worker_service.proto, the pure virtual KVBackend async interface in
    backends/backend.h, and BlockStatus::SHARED_STORAGE.

@google-cla

google-cla Bot commented Sep 8, 2026

Copy link
Copy Markdown

Thanks for your pull request! It looks like this may be your first contribution to a Google open source project. Before we can look at your pull request, you'll need to sign a Contributor License Agreement (CLA).

View this failed invocation of the CLA check for more information.

For the most up to date status, view the checks section at the bottom of the pull request.

@copybara-service
copybara-service Bot force-pushed the test_974774417 branch 3 times, most recently from 93ebf63 to 25e056d Compare September 15, 2026 17:50
… & Abstract Driver Architecture

Declares the protocol, configuration and driver interfaces that let TPU Raiden
offload and recall KV blocks to a secondary storage tier beyond Host DRAM.
Declarative only: every driver method is a stub returning UnimplementedError,
with the implementations in Part 2.

- Input: BackendConfig carries serving topology as a ParallelismConfig (tp_size,
  tp_rank); tp_rank is required rather than defaulted, so workers in a sharded
  job cannot collide on one shard directory. PosixBackendOptions is the typed
  parse-and-validate point for a POSIX backend.
- Output: PosixPathMapper declares the layout
  <root_dir>/<model_name>/tp<tp_size>_r<tp_rank>/<l1>/<l2>/<hash_hex>.bin;
  BlockKeyMapper::MapKey requires a hex-encoded, length-bounded hash.
- Interfaces: the TransferBackendBuffers RPC and BackendTransferSpec in
  worker_service.proto, the pure virtual KVBackend async interface in
  backends/backend.h, and BlockStatus::SHARED_STORAGE.

PiperOrigin-RevId: 982702207
@copybara-service
copybara-service Bot merged commit 1e7fa85 into main Sep 16, 2026
3 of 5 checks passed
@copybara-service
copybara-service Bot deleted the test_974774417 branch September 16, 2026 20:48
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

0 participants