Test fix: quantize slot_hidden, scatter FP4, pass slot_x_sf

tests/test_pipeline_real_weights.py

@@ -160,10 +160,16 @@ def main():
     l1_gsb = torch.tensor(l1_gs_list, dtype=torch.float32, device=DEVICE)
     
     # Quantize activation (dynamic gs, not warmup)
+    # KEY FIX: quantize slot_hidden (sorted tokens), NOT padded_hidden.
+    # padded_hidden has zeros in padding rows; quantizing it gives wrong x_sf layout.
     print("\n--- L1 GEMM (dynamic gs) ---")
-    x_fp4, x_sf, l1_gs = quantize_to_nvfp4(padded_hidden)
+    slot_x_fp4, slot_x_sf, l1_gs = quantize_to_nvfp4(slot_hidden)
     print(f"  L1 gs (dynamic): {l1_gs:.6f}")
     
+    # Scatter x_fp4 into padded layout
+    padded_x_fp4 = torch.zeros(total_padded, HIDDEN_SIZE, dtype=torch.uint8, device=DEVICE).view(torch.float4_e2m1fn_x2)
+    padded_x_fp4[padded_dst] = slot_x_fp4
+    
     # For scale_a, we need to use the runner's assembly approach.
     # Use the same _assemble_scales_cudagraph_safe function
     from vllm.nvfp4_cutedsl import CuTeDSLMoERunner
@@ -192,13 +198,13 @@ def main():
     # Just use the runner's scale assembly
     l1_gsa = torch.full((NUM_EXPERTS,), l1_gs, dtype=torch.float32, device=DEVICE)
     l1_scale_a = runner._assemble_scales_cudagraph_safe(
-        x_sf[:num_slots], expert_offsets[:NUM_EXPERTS+1],
+        slot_x_sf, expert_offsets[:NUM_EXPERTS+1],
         padded_expert_offsets,
         runner._padded_x_sf_buf_l1, runner._per_expert_scale_bufs_l1
     )
     
     l1_out = run_nvfp4_grouped_gemm(
-        mat_a=x_fp4, mat_b=l1_mat_b,
+        mat_a=padded_x_fp4, mat_b=l1_mat_b,
         scale_a=l1_scale_a, scale_b=l1_scale_b,
         expert_offsets=padded_expert_offsets[1:NUM_EXPERTS+1],
         global_scale_a=l1_gsa, global_scale_b=l1_gsb,