diff --git a/single_shot_inference.py b/single_shot_inference.py
index 339dca6d..a4cf7966 100644
--- a/single_shot_inference.py
+++ b/single_shot_inference.py
@@ -350,7 +350,7 @@ def forward_attention(x_normed, w, li, cfg, rope_cos, rope_sin,
 
     # 1. Q: q_a (NVFP4 GEMM) → q_a_norm → q_b (NVFP4 GEMM) → q_b_norm
     q_a = prod_lin['q_a'](x_normed)
-    if li < 3:
+    if VERBOSE >= 2 and li < 3:
         # Compare q_a with PyTorch reference
         q_a_ref = do_nvfp4_linear_ref(x_normed, w, pfx, 'q_a_proj')
         if q_a_ref is not None:
@@ -399,7 +399,7 @@ def forward_attention(x_normed, w, li, cfg, rope_cos, rope_sin,
 
     # 6. Production FMHA
     attn_out = _run_production_fmha(q_heads, all_kv, n_h, hd, T, seq_len, scale, dev, li, w, pfx)
-    if li < 3:
+    if VERBOSE >= 2 and li < 3:
         # Compare with PyTorch reference
         k_exp = all_kv.unsqueeze(0).expand(n_h, -1, -1).contiguous()
         v_exp = k_exp.clone()
@@ -419,7 +419,7 @@ def forward_attention(x_normed, w, li, cfg, rope_cos, rope_sin,
         a_grp = a_flat.reshape(T, o_groups, gid); oa_3d = oa_bf.reshape(o_groups, o_rank, gid)
         g_out = torch.bmm(a_grp.permute(1, 0, 2), oa_3d.transpose(1, 2))
         g_flat = g_out.permute(1, 0, 2).reshape(T, o_groups * o_rank)
-        if li < 3:
+        if VERBOSE >= 2 and li < 3:
             print(f"  L{li} wo_a: |g_flat|={g_flat.abs().max().item():.6f} shape={g_flat.shape}", flush=True)
         F_attn = prod_lin['o_b'](g_flat)
     else:
@@ -430,7 +430,7 @@ def forward_attention(x_normed, w, li, cfg, rope_cos, rope_sin,
         else:
             log.warning(f"L{li}: No o_a_proj weight, zero attention output")
             F_attn = torch.zeros(T, cfg["hidden_size"], dtype=torch.bfloat16, device=dev)
-    if li < 3:
+    if VERBOSE >= 2 and li < 3:
         print(f"  L{li} F_attn: |F_attn|={F_attn.abs().max().item():.6f}", flush=True)
     return F_attn, q_a
 
@@ -444,13 +444,13 @@ def moe_forward(x, li, moe_runner, se_runner, router, token_id):
     torch.cuda.synchronize(x.device)
     if topk_ids.max().item() >= 384 or topk_ids.min().item() < 0:
         print(f"  L{li} BAD topk_ids: min={topk_ids.min().item()} max={topk_ids.max().item()}", flush=True)
-    if li < 3:
+    if VERBOSE >= 2 and li < 3:
         print(f"  L{li} MoE input: |x|={x.abs().max().item():.4f} has_nan={torch.isnan(x).any().item()}", flush=True)
     routed_out = moe_runner.run(x, topk_w, topk_ids)
-    if li < 3:
+    if VERBOSE >= 2 and li < 3:
         print(f"  L{li} MoE routed: |out|={routed_out.abs().max().item():.4f} has_nan={torch.isnan(routed_out).any().item()}", flush=True)
     shared_out = se_runner.run(x)
-    if li < 3:
+    if VERBOSE >= 2 and li < 3:
         has_nan = torch.isnan(shared_out).any().item()
         out_max = shared_out.abs().max().item() if not has_nan else float('nan')
         print(f"  L{li} MoE shared: |out|={out_max:.4f} has_nan={has_nan}", flush=True)
@@ -794,7 +794,7 @@ def main():
                 err = torch.cuda.current_stream(gpu).query()
                 print(f"  CRASH at token {pi} layer {li} gpu {gpu}: {e}", flush=True)
                 raise
-            if pi == 0 and li < 3:
+            if VERBOSE >= 2 and pi == 0 and li < 3:
                 torch.cuda.synchronize(gpu)
                 print(f"  Token {pi} L{li}: OK |X|={X.abs().max().item():.1f}", flush=True)
         X = X.to('cuda:0'); torch.cuda.set_device(0)
@@ -829,7 +829,7 @@ def main():
         next_id = torch.argmax(logits, -1).item(); all_tokens.append(next_id)
         dt = time.time() - t1
         has_nan = torch.isnan(logits.float()).any().item()
-        if step % 5 == 0 or has_nan:
+        if step % 1 == 0 or has_nan:
             tv, ti = torch.topk(logits[0], 5)
             top5 = ' '.join(f'{tokenizer.decode([t.item()])}({v.item():.1f})' for t, v in zip(ti[:5], tv[:5]))
             print(f"  Step {step}: {next_id} '{tokenizer.decode([next_id])}' ({dt:.2f}s) "