[Bug] SPMD Paged Attention Highperf Fails For >= 8192 sequences #1022

Platform

a2a3

Runtime Variant

tensormap_and_ringbuffer

Description

The spmd_paged_attention_highperf scene test works for shorter decode
sequence lengths, but fails when extended to longer paged-attention sequences.
The failure is reproducible for the 8192-token case added in
hw-native-sys/simpler#986.

The compared PyPTO SPMD implementation can run the same logical GQA paged
attention problem for 8192 and 16384 tokens, so the long-sequence failure
appears specific to the simpler highperf kernel/runtime path rather than the
input shape itself.

Steps to Reproduce

1. Check out the code from
   hw-native-sys/simpler#986.
2. Build/install the project as usual for onboard a2a3 testing.
3. Run the highperf SPMD paged-attention long sequence case on an a2a3 device:

   python tests/st/a2a3/tensormap_and_ringbuffer/spmd_paged_attention_highperf/test_spmd_paged_attention_highperf.py \
     -p a2a3 -d 6 --case b1_h32_kv8_s8192_bs128_fp16

Expected Behavior

The highperf SPMD paged-attention kernel should complete successfully for the
8192-token GQA decode case and produce output matching the golden reference.

The expected shape is:

batch = 1
num_heads = 32
num_kv_heads = 8
head_dim = 128
kv_seq = 8192
block_size = 128
dtype = fp16

Actual Behavior

The 8192-token highperf case fails with an AICore timeout. The observed error
code is:

507018

Shorter cases such as 128, 512, and 4096 tokens complete successfully. The
failure has been reproduced for the 8192-token highperf case.

Git Commit ID

afb5c5a

CANN Version

9.0.0

Driver Version

25.5.1

Host Platform

Linux (aarch64)

Additional Context

No response