[Feature] Sfa/MRV2 pcp support (#14026)
### What this PR does / why we need it? SFA PCP Implementation: Introduced AscendSFACPMetadata and AscendSFACPImpl to handle Prefill Context Parallelism (PCP) logic for Sparse Flash Attention. Indexer Backend Updates: Updated AscendSFAIndexerBackend to support PCP and marked it as a cache-only backend. PCP Manager Integration: Added a new patch for the PCP manager to ensure compatibility with the vLLM worker and proper integration. Graph Capture and Partitioning: Modified aclgraph_utils and pcp_manager to support PCP-local input partitioning and ensure CUDA graph capture binds to correct buffers. Configuration Validation: Added strict configuration validation in pcp_manager to enforce graph-safe PCP execution on Ascend NPUs. ### Does this PR introduce _any_ user-facing change? NO ### How was this patch tested? - vLLM version: v0.27.1 - vLLM main: https://github.com/vllm-project/vllm/commit/ba07e4a48fc951300d97eb506217dd530583dea3 --------- Signed-off-by: joiandjoe <garancemr@icloud.com> Signed-off-by: Your Name <you@example.com> Signed-off-by: jayQH <qinhuaisheng@h-partners.com> Signed-off-by: Your <you@example.com> Signed-off-by: ZhangwenTaoHW <zhangwentao101@huawei.com> Signed-off-by: LeeWenquan <83354342+SunnyLee151064@users.noreply.github.com> Signed-off-by: chenaoxuan <cax1165@163.com> Signed-off-by: leolee <83689697+li1how@users.noreply.github.com> Signed-off-by: Bonan Wang <bryanwang20@gmail.com> Signed-off-by: SunnyLee151064 <83354342+SunnyLee151064@users.noreply.github.com> Co-authored-by: joiandjoe <garancemr@icloud.com> Co-authored-by: ZhangwenTaoHW <zhangwentao101@huawei.com> Co-authored-by: Your Name <you@example.com> Co-authored-by: chenaoxuan <cax1165@163.com> Co-authored-by: jayQH <qinhuaisheng@h-partners.com> Co-authored-by: leolee <83689697+li1how@users.noreply.github.com> Co-authored-by: Bonan Wang <bryanwang20@gmail.com>
L
LeeWenquan committed
2617fb6abf541145ea7f12d1dd39c757f1180b66
Parent: 247c9f4
Committed by GitHub <noreply@github.com>
on 8/28/2026, 6:25:23 AM