During chunked prefill, seqlen_k was set to len(seq) (the full sequence
length), causing the attention kernel to access uninitialized KV slots
for tokens not yet scheduled in the current chunk.
Fix: reorder so that end = start + seqlen_q is computed first, then
set seqlen_k = end — limiting attention to the current chunk boundary.
Fixes#212
Tai An
25794a1f29