returnn 1.20250904.193235__tar.gz → 1.20250904.204457__tar.gz

{returnn-1.20250904.193235/returnn.egg-info → returnn-1.20250904.204457}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20250904.193235
+Version: 1.20250904.204457
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer

returnn-1.20250904.204457/_setup_info_generated.py

@@ -0,0 +1,2 @@
+version = '1.20250904.204457'
+long_version = '1.20250904.204457+git.178d1fb'

{returnn-1.20250904.193235 → returnn-1.20250904.204457}/returnn/torch/frontend/_backend.py

@@ -1641,6 +1641,8 @@ class TorchBackend(Backend[torch.Tensor]):
                 name=f"random_{distribution}", dims=dims, dtype=dtype, sparse_dim=sparse_dim, feature_dim=feature_dim
             )
             out.raw_tensor = torch.empty(shape, dtype=dtype_, device=device or rf.get_default_device())
+        if out.raw_tensor.device.type == "meta":
+            return out  # nothing more to do
         assert explicit_state is None  # not implemented otherwise
         generator = None  # using the global default from PT
         assert isinstance(static, bool)

{returnn-1.20250904.193235 → returnn-1.20250904.204457/returnn.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20250904.193235
+Version: 1.20250904.204457
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer

{returnn-1.20250904.193235 → returnn-1.20250904.204457}/tests/test_rf_encoder_conformer.py

@@ -218,19 +218,22 @@ def test_e_branchformer():
     #         espnet2.asr.layers.cgmlp.ConvolutionalGatingMLP.forward
     #         espnet.nets.pytorch_backend.transformer.positionwise_feed_forward.PositionwiseFeedForward.forward
     #   (CTC loss, Att-decoder)
-    with PyTracer(
-        [
-            ESPnetASRModel.forward,
-            ESPnetASRModel.encode,
-            EBranchformerEncoder.forward,
-            EBranchformerEncoderLayer.forward,
-            RelPositionMultiHeadedAttention.forward,
-            RelPositionalEncoding.forward,
-            ConvolutionalGatingMLP.forward,
-            ConvolutionalSpatialGatingUnit.forward,
-        ],
-        torch.Tensor,
-    ) as trace_espnet, torch.no_grad():
+    with (
+        PyTracer(
+            [
+                ESPnetASRModel.forward,
+                ESPnetASRModel.encode,
+                EBranchformerEncoder.forward,
+                EBranchformerEncoderLayer.forward,
+                RelPositionMultiHeadedAttention.forward,
+                RelPositionalEncoding.forward,
+                ConvolutionalGatingMLP.forward,
+                ConvolutionalSpatialGatingUnit.forward,
+            ],
+            torch.Tensor,
+        ) as trace_espnet,
+        torch.no_grad(),
+    ):
         loss, stats, weight = model(
             speech=raw_audio.raw_tensor,
             speech_lengths=raw_audio_spatial_dim.dyn_size,
@@ -292,16 +295,19 @@ def test_e_branchformer():
         assert isinstance(layer_rf, EBranchformerLayer)
         import_params_espnet_e_branchformer_layer_to_rf(layer, layer_rf)
 
-    with PyTracer(
-        [
-            ConformerEncoder.__call__,
-            EBranchformerLayer.__call__,
-            rf.RelPosSelfAttention.__call__,
-            rf.relative_positional_encoding,
-            FeedForwardConvGated.__call__,
-        ],
-        Tensor,
-    ) as trace_rf, torch.no_grad():
+    with (
+        PyTracer(
+            [
+                ConformerEncoder.__call__,
+                EBranchformerLayer.__call__,
+                rf.RelPosSelfAttention.__call__,
+                rf.relative_positional_encoding,
+                FeedForwardConvGated.__call__,
+            ],
+            Tensor,
+        ) as trace_rf,
+        torch.no_grad(),
+    ):
         # ESPnet E-Branchformer does not use masking properly. Keep it disabled here as well.
         with global_config_ctx(Config({"rf_use_mask": False})):
             enc_out, _ = model_rf(enc_in, in_spatial_dim=enc_spatial_dim)
@@ -434,3 +440,59 @@ def test_e_branchformer_custom_ff():
     x = rf.random_normal((time_dim, encoder.in_dim))
     y, _ = encoder(x, in_spatial_dim=time_dim)
     y.verify_out_shape({time_dim, encoder.out_dim})
+
+
+def test_e_branchformer_meta_dev_num_params():
+    from returnn.frontend.encoder.conformer import ConformerEncoder, ConformerEncoderLayer
+    from returnn.frontend.encoder.e_branchformer import EBranchformerLayer
+    from returnn.torch.frontend.bridge import rf_module_to_pt_module
+
+    rf.select_backend_torch()
+
+    in_dim = Dim(80, name="logmel")
+    num_layers = 16
+    dim = 1024
+
+    with rf.set_default_device_ctx("meta"):
+        conformer = ConformerEncoder(
+            in_dim=in_dim,
+            input_layer=None,
+            num_layers=num_layers,
+            out_dim=dim,
+            encoder_layer=rf.build_dict(
+                ConformerEncoderLayer,
+                ff=rf.build_dict(
+                    rf.encoder.conformer.ConformerPositionwiseFeedForward,
+                    activation=rf.build_dict(rf.relu_square),
+                    with_bias=False,
+                ),
+                num_heads=8,
+            ),
+        )
+        conformer_pt = rf_module_to_pt_module(conformer)
+        assert all(parameter.device.type == "meta" for parameter in conformer_pt.parameters())
+        num_params = sum(parameter.numel() for parameter in conformer_pt.parameters())
+        print(f"Conformer {num_layers}x{dim} net params #: {num_params}")
+        assert num_params == 403619840
+
+        dim_ = dim // 8 * 7
+        e_branchformer = ConformerEncoder(
+            in_dim=in_dim,
+            input_layer=None,
+            num_layers=num_layers,
+            out_dim=dim_,
+            encoder_layer=rf.build_dict(
+                EBranchformerLayer,
+                ff=rf.build_dict(
+                    rf.encoder.conformer.ConformerPositionwiseFeedForward,
+                    activation=rf.build_dict(rf.relu_square),
+                    with_bias=False,
+                ),
+                num_heads=8,
+            ),
+        )
+        e_branchformer_pt = rf_module_to_pt_module(e_branchformer)
+        assert all(parameter.device.type == "meta" for parameter in e_branchformer_pt.parameters())
+        num_params_ = sum(parameter.numel() for parameter in e_branchformer_pt.parameters())
+        print(f"E-Branchformer {num_layers}x{dim_} net params #: {num_params_}")
+        assert num_params_ == 413034496

returnn-1.20250904.193235/_setup_info_generated.py

@@ -1,2 +0,0 @@
-version = '1.20250904.193235'
-long_version = '1.20250904.193235+git.fd139c1'