[WIP] feat: Add n_offset optimization for sliding window attention

[zminglei]

Summary

Adapt the n_offset sliding window optimization from vllm for sgl-attn. This optimization shifts KV gmem pointers to the start of the sliding window and reduces seqlen_k, so the kernel only iterates over actual window blocks instead of the full KV sequence length.

Changes

Core optimization (block.h, mainloop_fwd_sm90_tma_gmma_ws.hpp)

• Compute n_offset in get_n_block_min_max() to identify where the sliding window starts
• Return n_offset as part of 3-tuple from get_n_block_min_max()
• In load(): shift KV TMA pointers by n_offset so we start reading from the window start
• In mma(): adjust seqlen_k by subtracting n_offset, inline mask helper functions for compatibility with const seqlen_k in SeqlenInfoQK
• In mma_pv(): handle 3-tuple return from get_n_block_min_max()

Tile size tuning (tile_size.h)

• Use kBlockN=160 with IntraWGOverlap=false for SWA with hdim<=64

API changes (flash_api.cpp)

• Disable PagedKV TMA for is_local (required for n_offset pointer shifting)
• Add PackGQA override for hdim=64, h/h_k=8, is_local

Template infrastructure (epilogue_fwd.hpp, flash_fwd_launch_template.h, static_switch.h)

• Add kBlockH template parameter (infrastructure for future PackGQA_TMA support)
• Add PACK_GQA_BLOCK_SWITCH macro
• kBlockH forced to 1 for now (see Known Limitations)

Performance

Benchmarked on gpt-oss-120b (hdim=64, SWA=128, TP1, H200):

• Before: 9.84ms TPOT
• After: 9.35ms TPOT (~5% improvement)
• Benchmark: 16 concurrent requests, 1K input tokens, 12K output tokens
• Correctness verified: GSM8K accuracy 85.5% (matches baseline)

Known Limitations

• kBlockH is forced to 1, which disables PackGQA_TMA. Enabling kBlockH > 1 requires validating sgl-attn epilogue with TMA-based Q/O access patterns (this caused correctness issues during testing).
• The n_offset optimization only activates for is_local (sliding window) attention layers.

Test Plan

• Build succeeds (ninja -j32, 372 targets)
• GSM8K accuracy: 85.5% (verified correct output)
• TPOT benchmark: 9.35ms (5% improvement over 9.84ms baseline)
• vLLM same-node baseline: 9.12ms (remaining gap ~2.5%)

[zminglei]

close this PR, moving to #37