Repository navigation
Reuse one TPU slice across cases in the Torch DMA and weight sync benchmarks: - #1385
Merged
Merged
Conversation
|
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
Bot
force-pushed
the
test_1008715198
branch
2 times, most recently
from
November 3, 2026 18:51
9ece227 to
78773f9
Compare
…chmarks: - `examples/microbenchmarks/torch_dma_kv_cache_benchmark.py`: Run every test case on one set of long-lived per-device worker processes (`_BenchmarkWorkers`) instead of a fresh `mp.spawn` per case, so the multi-process TPU slice is created and torn down once per run. A failed case stops its workers and the next case starts a fresh group. - `examples/microbenchmarks/torch_weight_sync_benchmark_runner.py`: The destination drives the whole `--parallelism_list` ladder from one set of processes, as the source already does, with fresh tensors and a fresh `WeightSynchronizer` per rung. `--controller_address` becomes `--controller_host` (each rung's controller port is derived the same way as on the source), and `--src_log_dir` becomes `--results_dir`, which both roles use for their per-rung result files. PiperOrigin-RevId: 1013149750
copybara-service
Bot
force-pushed
the
test_1008715198
branch
from
November 3, 2026 19:06
78773f9 to
5f59e6a
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Reuse one TPU slice across cases in the Torch DMA and weight sync benchmarks:
examples/microbenchmarks/torch_dma_kv_cache_benchmark.py: Run every test case on one set of long-lived per-device worker processes (_BenchmarkWorkers) instead of a freshmp.spawnper case, so the multi-process TPU slice is created and torn down once per run. A failed case stops its workers and the next case starts a fresh group.examples/microbenchmarks/torch_weight_sync_benchmark_runner.py: The destination drives the whole--parallelism_listladder from one set of processes, as the source already does, with fresh tensors and a freshWeightSynchronizerper rung.--controller_addressbecomes--controller_host(each rung's controller port is derived the same way as on the source), and--src_log_dirbecomes--results_dir, which both roles use for their per-rung result files.