June 2026 monthly summary for vllm-project/vllm-ascend. Focused on stabilizing high-load inference with flashcomm1 in a multi-request/concurrency scenario. Delivered a targeted bug fix that eliminates FIA error 561002 by padding seq_lens and block_table to the derived batch size, centralizing the fix in AscendAttentionMetadataBuilder.build() to cover all code paths (chunked prefill, decode-only, prefill-cache-hit). The patch preserves user-facing behavior while removing crashes under heavy concurrency. The fix was validated on NPU with tensor-parallel-size 2, full batch, and stress traffic (≥ max_num_seqs), ensuring reliable operation. No user-visible changes; improved system reliability and throughput under peak load. Technologies demonstrated include buffer sizing, attention metadata construction, NPU FIA operations, and stress testing. Business value: reduced downtime, higher throughput, and improved SLA adherence for real-time inference.