Add KV cache length diagnostic during decode

single_shot_inference.py

@@ -64,7 +64,8 @@ NUM_GPUS = 8
 SKIP_ROUTED_MOE = _args.skip_moe  # If True, only use shared expert (debug)
 INVERSE_ROPE = not _args.no_inverse_rope  # If False, skip inverse RoPE on attention output (diagnostic)
 SKIP_MHC = _args.skip_mhc   # If True, bypass mHC and use simple residual connections (diagnostic)
-MHC_DIAG = True        # If True, print per-layer mHC diagnostics (B_l row/col sums, C_l values)
+MHC_DIAG = False       # If True, print per-layer mHC diagnostics (B_l row/col sums, C_l values)
+GROWTH_DIAG = True     # If True, print per-layer residual growth analysis
 # When True: applies inverse RoPE at query position → converts absolute→relative
 # When False: leaves relative position encoding intact for output projection
 # DSV4 partial RoPE only affects last 64/512 dims; first 448 are always un-RoPE'd
@@ -1055,6 +1056,11 @@ def main():
         decode_pos = len(all_tokens) - 1
         positions = torch.tensor([decode_pos], dtype=torch.long, device='cuda:0')
 
+        # Diag: KV cache length
+        if step < 3 or step % 10 == 0:
+            k0, _ = kv_caches[0].get()
+            print(f"  KV: {k0.shape[1]} entries, pos={decode_pos}", flush=True)
+
         emb = embed(tid)  # (1, H) on gpu0
         X = mHCBlock.init_state(emb, n_hc)  # (1, n_hc, H)