fix: correct Nvfp4Linear dimension inference from checkpoint weights

Weight shape (N_packed, K_packed) means:
- out_features = N_packed (GEMM output dim in BF16)
- in_features = K_packed * 2 (BF16 input dim, for activation buffer)

single_shot_inference.py

@@ -131,12 +131,12 @@ def make_nvfp4_linear(in_features, out_features, device, all_w, pfx, proj_name):
     d = device
     weight, ws, ws2, isc = get_nvfp4_weight(all_w, pfx, proj_name)
     assert weight is not None, f"{pfx}.{proj_name}.weight not found"
-    # Infer actual dimensions from checkpoint weight shape
-    # weight is (O, I//2) uint8 where O = out_features//2
-    actual_out = weight.shape[0] * 2
-    actual_in = weight.shape[1] * 2
-    if actual_out != out_features or actual_in != in_features:
-        log.warning(f"{pfx}.{proj_name}: declared ({in_features},{out_features}) but weight is ({actual_in},{actual_out})")
+    # Checkpoint weight is (N_packed, K_packed) uint8
+    # NVFP4 GEMM output dim = N_packed BF16 elements
+    # Activation buffer needs K_packed FP4 columns = in_features BF16
+    # So: in_features = K_packed * 2, out_features = N_packed
+    actual_out = weight.shape[0]  # N_packed = GEMM output dimension
+    actual_in = weight.shape[1] * 2  # K_packed * 2 = BF16 input dim (for buffer allocation)
     lin = Nvfp4Linear(actual_in, actual_out, max_num_tokens=8192, device=d)
     lin.fp4 = [weight.to(d)]; lin.sf = [ws.to(d)]
     lin.ws2 = [ws2.to(d) if ws2 is not None else None]  # weight_scale_2