From 0069769d127a6b495f4beb895646e9d8305ba2d5 Mon Sep 17 00:00:00 2001
From: biondizzle <biondizzle@gmail.com>
Date: Sat, 16 May 2026 19:38:31 +0000
Subject: [PATCH] debug: print global scales

---
 cutedsl/moe_pipeline.py | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/cutedsl/moe_pipeline.py b/cutedsl/moe_pipeline.py
index 63a9dfec..899a4708 100644
--- a/cutedsl/moe_pipeline.py
+++ b/cutedsl/moe_pipeline.py
@@ -187,6 +187,9 @@ def run_nvfp4_moe(
     # Global scales: alpha = igs * weight_gs for each expert
     l1_global_scale_a = torch.tensor([x_igs] * num_experts, dtype=torch.float32, device=device)
     l1_global_scale_b = torch.tensor(weights['l1_gs'], dtype=torch.float32, device=device)
+    print(f"  L1 global_scale_a: {l1_global_scale_a.tolist()}", flush=True)
+    print(f"  L1 global_scale_b: {l1_global_scale_b.tolist()}", flush=True)
+    print(f"  alpha (a*b): {(l1_global_scale_a * l1_global_scale_b).tolist()}", flush=True)
     
     # Run L1 GEMM
     l1_out = run_nvfp4_grouped_gemm(