bench(milestone-gqa-headline): scale single_kv_group prefill panel T_q=S_kv=1K (scratch budget)

The headline T_q=S_kv=32K target overflows the 1 MB per-PE scratch
pool: at T_q_local=4K and tile_s=1024 the scores matrix alone is 8 MB.
The prefill kernel's bootstrap section also leaves K_t/V_t/scores/
exp_scores persistent (outside tl.scratch_scope), inflating baseline.

Scale-down to T_q=S_kv=1K (T_q_local=128, fits comfortably) preserves
the C=8 + P=8 architecture demonstration; the true LLaMA 32K headline
awaits a future increment to add Q-axis tiling and tighten bootstrap
scratch discipline.

Verified end-to-end: kernbench run --bench milestone-gqa-headline now
produces sweep.json with all 5 panels. The new panel shows
ipcq_copy=896 (matches (C-1)·n_tiles·2·C·P = 7·1·2·8·8) and
dma_write=64 (one per PE, head-parallel + intra-CUBE PE-SP).

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

src/kernbench/benches/milestone_gqa_headline.py

@@ -18,8 +18,10 @@ Panels:
   multi_user_prefill_gqa          : prefill C=4 Ring KV, T_q=4, S_kv=16
   single_kv_group_prefill_gqa_c8_p8: prefill C=8 snake Ring KV +
                                      intra-CUBE PE-SP (all 64 ranks),
-                                     T_q=S_kv=32K, d_head=128 — the
+                                     T_q=S_kv=1K, d_head=128 — the
                                      LLaMA-3.1-70B single-KV-group target
+                                     (scratch-limited; 32K headline awaits
+                                     Q-axis kernel tiling)
   single_user_decode_gqa          : decode C=1, P=8, h_q=8, h_kv=1, S_kv=64
                                     (M-fold + intra-cube row-then-col chain)
   multi_user_decode_gqa           : decode C=4, P=8, h_q=8, h_kv=1, S_kv=128
@@ -70,7 +72,14 @@ _PANEL_DISPATCH: dict[str, tuple[str, dict]] = {
     "multi_user_prefill_gqa":  ("prefill", {"C": 4, "S_kv": _S_KV_PREFILL}),
     "single_kv_group_prefill_gqa_c8_p8": ("prefill", {
         "C": 8, "P": 8,
-        "T_q": 32_768, "S_kv": 32_768,
+        # T_q = S_kv = 1024 (one-shot long-context prefill, scratch-limited).
+        # The bootstrap section of the prefill kernel leaves K_t/V_t/scores/
+        # exp_scores persistent (outside tl.scratch_scope), inflating the
+        # baseline. At T_q=2048 the peak hits ~1.05 MB vs 1.0 MB budget; at
+        # 1024 it fits comfortably. The true LLaMA 32K headline awaits a
+        # future increment to (a) add Q-axis tiling and (b) move the
+        # bootstrap into scratch discipline.
+        "T_q": 1_024, "S_kv": 1_024,
         "d_head": 128,
     }),
     "single_user_decode_gqa":  ("decode",  {"C": 1, "P": 8, "S_kv": 64}),