8000 Workflow runs · google-deepmind/kfac-jax · GitHub
[go: up one dir, main page]
More Web Proxy on the site http://driver.im/
Skip to content

Actions: google-deepmind/kfac-jax

Actions

All workflows

Actions

Loading...
Loading

Showing runs from all workflows
1,403 workflow runs
1,403 workflow runs

Filter by Event

Filter by Status

Filter by Branch

Filter by Actor

No longer using "iterator_on_device" in dataset functions in examples code, as this is 1) arguably pointless since data transfers cannot occur in parallel with other ops (at least on TPU), and 2) leads to a nasty error on TPUs now that xla_tpu_use_enhanced_launch_barrier=True by default. The error happens because Jaxline's py_prefetch function runs in a separate thread, and this leads to random timing of the step function vs the data transfer function, which is interpreted as an "out of order" error by the "launch barrier" when JAX is running in multi-process mode on TPUs. ci #1669: Pull request #334 synchronize by copybara-service bot
May 9, 2025 12:31 3m 35s test_756114230
test_756114230
May 9, 2025 12:31 3m 35s
Push on main CodeQL #940: by copybara-service bot
May 9, 2025 12:31 1m 2s main
main
May 9, 2025 12:31 1m 2s
No longer using "iterator_on_device" in dataset functions in examples… ci #1668: Commit 0a21460 pushed by copybara-service bot
May 9, 2025 12:31 3m 48s main
main
May 9, 2025 12:31 3m 48s
PR #334 CodeQL #939: by copybara-service bot
May 9, 2025 12:31 1m 7s refs/pull/334/head
refs/pull/334/head
May 9, 2025 12:31 1m 7s
No longer using "iterator_on_device" in dataset functions in examples code, as this is 1) arguably pointless since data transfers cannot occur in parallel with other ops (at least on TPU), and 2) leads to a nasty error on TPUs now that xla_tpu_use_enhanced_launch_barrier=True by default. The error happens because Jaxline's py_prefetch function runs in a separate thread, and this leads to random timing of the step function vs the data transfer function, which is interpreted as an "out of order" error by the "launch barrier" when JAX is running in multi-process mode on TPUs. ci #1667: Pull request #334 synchronize by copybara-service bot
May 9, 2025 12:24 3m 44s test_756114230
test_756114230
May 9, 2025 12:24 3m 44s
PR #334 CodeQL #938: by copybara-service bot
May 9, 2025 12:24 1m 7s refs/pull/334/head
refs/pull/334/head
May 9, 2025 12:24 1m 7s
No longer using "iterator_on_device" in dataset functions in examples code, as this is 1) arguably pointless since data transfers cannot occur in parallel with other ops (at least on TPU), and 2) leads to a nasty error on TPUs now that xla_tpu_use_enhanced_launch_barrier=True by default. The error happens because Jaxline's py_prefetch function runs in a separate thread, and this leads to random timing of the step function vs the data transfer function, which is interpreted as an "out of order" error by the "launch barrier" when JAX is running in multi-process mode on TPUs. ci #1666: Pull request #334 synchronize by copybara-service bot
May 8, 2025 21:20 3m 36s test_756114230
test_756114230
May 8, 2025 21:20 3m 36s
PR #334 CodeQL #937: by copybara-service bot
May 8, 2025 21:20 1m 3s refs/pull/334/head
refs/pull/334/head
May 8, 2025 21:20 1m 3s
Changing should_update_damping to include check of self._use_adaptive_damping. This ensures that when self._use_adaptive_damping is False, we don't needlessly compile the step function twice for the two different values of ((step_counter + 1) % self._damping_adaptation_interval == 0). ci #1665: Pull request #335 synchronize by copybara-service bot
May 8, 2025 11:37 3m 38s test_752807192
test_752807192
May 8, 2025 11:37 3m 38s
Push on main CodeQL #936: by github-advanced-security bot
May 8, 2025 11:37 1m 23s main
main
May 8, 2025 11:37 1m 23s
Changing should_update_damping to include check of self._use_adaptive… ci #1664: Commit 320d025 pushed by copybara-service bot
May 8, 2025 11:37 3m 42s main
main
May 8, 2025 11:37 3m 42s
PR #335 CodeQL #935: by copybara-service bot
May 8, 2025 11:37 1m 2s refs/pull/335/head
refs/pull/335/head
May 8, 2025 11:37 1m 2s
No longer using "iterator_on_device" in dataset functions in examples code, as this is 1) arguably pointless since data transfers cannot occur in parallel with other ops (at least on TPU), and 2) leads to a nasty error on TPUs now that xla_tpu_use_enhanced_launch_barrier=True by default. The error happens because Jaxline's py_prefetch function runs in a separate thread, and this leads to random timing of the step function vs the data transfer function, which is interpreted as an "out of order" error by the "launch barrier" when JAX is running in multi-process mode on TPUs. ci #1663: Pull request #334 synchronize by copybara-service bot
May 8, 2025 11:14 3m 41s test_756114230
test_756114230
May 8, 2025 11:14 3m 41s
PR #334 CodeQL #934: by copybara-service bot
May 8, 2025 11:14 1m 11s refs/pull/334/head
refs/pull/334/head
May 8, 2025 11:14 1m 11s
Changing should_update_damping to include check of self._use_adaptive_damping. This ensures that when self._use_adaptive_damping is False, we don't needlessly compile the step function twice for the two different values of ((step_counter + 1) % self._damping_adaptation_interval == 0). ci #1662: Pull request #335 synchronize by copybara-service bot
May 8, 2025 11:12 3m 39s test_752807192
test_752807192
May 8, 2025 11:12 3m 39s
PR #335 CodeQL #933: by copybara-service bot
May 8, 2025 11:12 1m 9s refs/pull/335/head
refs/pull/335/head
May 8, 2025 11:12 1m 9s
No longer using "iterator_on_device" in dataset functions in examples code, as this is 1) arguably pointless since data transfers cannot occur in parallel with other ops (at least on TPU), and 2) leads to a nasty error on TPUs now that xla_tpu_use_enhanced_launch_barrier=True by default. The error happens because Jaxline's py_prefetch function runs in a separate thread, and this leads to random timing of the step function vs the data transfer function, which is interpreted as an "out of order" error by the "launch barrier" when JAX is running in multi-process mode on TPUs. ci #1661: Pull request #334 synchronize by copybara-service bot
May 8, 2025 04:10 3m 40s test_756114230
test_756114230
May 8, 2025 04:10 3m 40s
DA0E
PR #334 CodeQL #932: by copybara-service bot
May 8, 2025 04:10 1m 7s refs/pull/334/head
refs/pull/334/head
May 8, 2025 04:10 1m 7s
Changing should_update_damping to include check of self._use_adaptive_damping. This ensures that when self._use_adaptive_damping is False, we don't needlessly compile the step function twice for the two different values of ((step_counter + 1) % self._damping_adaptation_interval == 0). ci #1660: Pull request #335 opened by copybara-service bot
May 8, 2025 04:00 3m 43s test_752807192
test_752807192
May 8, 2025 04:00 3m 43s
PR #335 CodeQL #931: by copybara-service bot
May 8, 2025 04:00 1m 6s refs/pull/335/head
refs/pull/335/head
May 8, 2025 04:00 1m 6s
No longer using "iterator_on_device" in dataset functions in examples code, as this is 1) arguably pointless since data transfers cannot occur in parallel with other ops (at least on TPU), and 2) leads to a nasty error on TPUs now that xla_tpu_use_enhanced_launch_barrier=True by default. The error happens because Jaxline's py_prefetch function runs in a separate thread, and this leads to random timing of the step function vs the data transfer function, which is interpreted as an "out of order" error by the "launch barrier" when JAX is running in multi-process mode on TPUs. ci #1659: Pull request #334 opened by copybara-service bot
May 8, 2025 03:58 3m 45s test_756114230
test_756114230
May 8, 2025 03:58 3m 45s
PR #334 CodeQL #930: by copybara-service bot
May 8, 2025 03:58 3m 5s refs/pull/334/head
refs/pull/334/head
May 8, 2025 03:58 3m 5s
Scheduled CodeQL #929: by github-advanced-security bot
May 6, 2025 13:46 1m 21s main
main
May 6, 2025 13:46 1m 21s
Scheduled CodeQL #928: by github-advanced-security bot
April 29, 2025 12:22 1m 24s main
main
April 29, 2025 12:22 1m 24s
CodeQL CodeQL #927: by github-advanced-security bot
April 28, 2025 15:59 1m 57s main
main
April 28, 2025 15:59 1m 57s
0