debug: print global scales

2026-05-16 19:38:31 +00:00
parent 84589fe984
commit 0069769d12
1 changed files with 3 additions and 0 deletions
--- 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(