Fix: size the ring task window for host_build_graph PA Case1/Case2 - #2164
Fix: size the ring task window for host_build_graph PA Case1/Case2#2164noabauma wants to merge 1 commit into
Conversation
Both cases are whole-graph-resident and build more tasks than the 16384-task default window, so bind fails with -1000 before any kernel runs. Case1 orchestrates ~65,792 tasks and Case2 ~32,832; each case now requests the next power of two above its own count via CallConfig.runtime_env.ring_task_window. Only the task window needs a knob -- the graph heap is committed to its measured size after orchestration and takes no configuration.
|
No actionable comments were generated in the recent review. 🎉 ℹ️ Recent review info⚙️ Run configurationConfiguration used: Organization UI Review profile: CHILL Plan: Advanced Run ID: 📒 Files selected for processing (1)
Included review availability: Your plan provides up to 1 included review per hour; 0 remain after this review. 📝 WalkthroughWalkthroughTwo paged-attention host build graph test cases now set explicit ChangesPaged attention task window configuration
Estimated code review effort: 1 (Trivial) | ~2 minutes Merge Risk: ⚪ Minimal · up to The paged-attention host build graph cases now request sufficiently sized task windows, preventing pre-kernel bind failures for their resident task graphs. No current merge-readiness risk remains. 🚥 Pre-merge checks | ✅ 5✅ Passed checks (5 passed)
Warning Some tools did not complete. Review the errors below. 🔧 Ruff (0.16.3)tests/st/a2a3/host_build_graph/paged_attention/test_paged_attention.py�[1;31mruff failed�[0m Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. A rabbit tunes the task table bright Comment |
Both cases are whole-graph-resident and build more tasks than the 16384-task default window, so bind fails with -1000 before any kernel runs. Case1 orchestrates ~65,792 tasks and Case2 ~32,832; each case now requests the next power of two above its own count via CallConfig.runtime_env.ring_task_window.
Only the task window needs a knob -- the graph heap is committed to its measured size after orchestration and takes no configuration.