Zijie Tian 2c2383c786 ⚡️ perf: optimize XAttention estimate with hierarchical block sum ...

Replace slow softmax_fuse_block_sum (block_size=4096) with optimized
hierarchical approach (estimate_block_size=1024):

- Add estimate_block_size parameter to XAttentionBSAPolicy (default 1024)
- Rewrite select_blocks to use hierarchical aggregation:
  1. Fine-grained softmax with small block size (15x faster kernel)
  2. Aggregate to CPU block level via reshape + sum
  3. Score + threshold selection (replaces mask + voting)

Performance improvement (CPU Offload mode):
- softmax_fuse_block_sum: 48% → 1% of total time (44x faster)
- 128K: XAttention now +2.4% faster than Full (was -59%)
- 64K: -3.8% (was -21%)
- 32K: -6.0% (was -14%)

Generated with [Claude Code](https://claude.ai/code)
via [Happy](https://happy.engineering)

Co-Authored-By: Claude <noreply@anthropic.com>
Co-Authored-By: Happy <yesreply@happy.engineering>

2026-01-28 06:47:13 +08:00

..

architecture_guide.md

✨ feat: add comprehensive RULER benchmark testing

2026-01-18 20:34:06 +08:00

bench_offload_results.md

⚡️ perf: optimize XAttention estimate with hierarchical block sum

2026-01-28 06:47:13 +08:00

block_sparse_attn_interface.md

📝 docs: add BSA interface documentation and cleanup temp files

2026-01-20 04:27:19 +08:00

chunked_attention_solutions.md

📝 docs: add chunked attention solutions guide and update doc index

2026-01-20 04:48:20 +08:00

cpu_offload_optimization_strategies.md

📝 docs: add CPU offload optimization strategies guide

2026-01-27 04:44:36 +08:00

cpu_scheduling_latency_analysis.md

📝 docs: add CPU scheduling latency analysis for chunked attention

2026-01-27 03:42:12 +08:00

cuda_graph_memory_guide.md

📝 docs: add CUDA Graph memory mechanism guide

2026-01-21 02:59:21 +08:00

cuda_graph_offload_guide.md

[docs] Added cuda_graph_guide.md

2026-01-21 21:56:24 +08:00

debugging_guide.md

✨ feat: add comprehensive RULER benchmark testing

2026-01-18 20:34:06 +08:00

estimate_block_size_performance.md

⚡️ perf: optimize XAttention estimate with hierarchical block sum

2026-01-28 06:47:13 +08:00

gpu_only_sparse_integration.md

📝 docs: add GPU-only sparse policy integration baseline

2026-01-27 04:36:31 +08:00

gpu_only_xattn_guide.md

📝 docs: add GPU-Only XAttention guide with performance analysis

2026-01-27 07:21:46 +08:00

known_issues.md

✨ feat: add comprehensive RULER benchmark testing

2026-01-18 20:34:06 +08:00

memory_communication_benchmark.md

⚡️ perf: optimize XAttention estimate phase with K-only loading

2026-01-28 06:24:20 +08:00

nsys_wrong_event_order_bug.md

📝 docs: add nsys wrong event order bug investigation

2026-01-24 04:32:05 +08:00

observer_architecture.md

📈 feat: add MemoryObserver for GPU-CPU communication tracking

2026-01-28 04:06:45 +08:00

optimization_guide.md

✨ feat: add comprehensive RULER benchmark testing

2026-01-18 20:34:06 +08:00

ruler_32k_chunked_offload_issue.md

🐛 fix: resolve CPU KV cache state leakage between requests

2026-01-21 01:12:21 +08:00

ruler_benchmark_results_32k.md

✨ feat: add comprehensive RULER benchmark testing

2026-01-18 20:34:06 +08:00

sparse_attention_guide.md

📝 docs: add XAttention algorithm guide based on COMPASS implementation

2026-01-20 02:50:03 +08:00

sparse_policy_architecture.md

♻️ refactor: remove cross-layer pipeline and rename compute_chunked_prefill

2026-01-20 02:10:40 +08:00

sparse_policy_implementation_guide.md

📝 docs: add SparsePolicy implementation guide and update rules

2026-01-20 02:25:46 +08:00

xattention_algorithm_guide.md

📝 docs: add XAttention algorithm guide based on COMPASS implementation

2026-01-20 02:50:03 +08:00

xattention_bsa_test_report.md

[WIP] Before integrate the xattn operator.

2026-01-19 21:19:21 +08:00

xattn_bsa_policy_design.md

📝 docs: update XAttention BSA Policy with benchmarks and memory management

2026-01-23 09:35:18 +08:00

xattn_chunked_prefill.md

feat: add xattn_estimate_chunked for chunked prefill support

2026-01-22 01:13:17 +08:00

xattn_kernels_guide.md

docs: add XAttention kernels guide

2026-01-23 03:22:25 +08:00

xattn_performance_analysis.md

📝 docs: add XAttention performance analysis documentation

2026-01-28 00:57:20 +08:00