returnn 1.20250220.200053__tar.gz → 1.20250223.154045__tar.gz

{returnn-1.20250220.200053/returnn.egg-info → returnn-1.20250223.154045}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20250220.200053
+Version: 1.20250223.154045
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer

returnn-1.20250223.154045/_setup_info_generated.py

@@ -0,0 +1,2 @@
+version = '1.20250223.154045'
+long_version = '1.20250223.154045+git.354cf31'

{returnn-1.20250220.200053 → returnn-1.20250223.154045}/returnn/frontend/_backend.py

@@ -350,6 +350,11 @@ class Backend(Generic[T]):
         """
         raise NotImplementedError
 
+    @staticmethod
+    def stop_gradient_scope() -> Any:
+        """stop gradient scope"""
+        raise NotImplementedError
+
     @staticmethod
     def scaled_gradient(tensor: Tensor, scale: Union[float, Tensor]) -> Tensor:
         """
@@ -905,9 +910,9 @@ class Backend(Generic[T]):
     ) -> Tensor:
         """
         Scatters into new zero-tensor.
-        If entries in indices are duplicated, the corresponding values in source will be added together
-        (scatter_add in PyTorch).
-        (TF segment_sum can be implemented via this.)
+        If entries in indices are duplicated, with mode="sum", the corresponding values in source will be added together
+        (``scatter_add`` in PyTorch), otherwise min/max.
+        (segment_sum can be implemented via this.)
 
         :param source: [batch_dims..., indices_dim(s)..., feature_dims...]
         :param indices: [batch_dims..., indices_dim(s)...] -> out_dim

{returnn-1.20250220.200053 → returnn-1.20250223.154045}/returnn/frontend/array_.py

@@ -40,6 +40,8 @@ __all__ = [
     "gather",
     "scatter",
     "scatter_argmax",
+    "scatter_logsumexp",
+    "scatter_logmeanexp",
     "slice",
     "shift_right",
     "shift_left",
@@ -758,19 +760,20 @@ def scatter(
 ) -> Tensor:
     """
     Scatters into new zero-tensor.
-    If entries in indices are duplicated, the corresponding values in source will be added together
-    (scatter_add in PyTorch)
-    with mode=="sum",
-    or otherwise it will take the max/min.
+    If entries in indices are duplicated, with ``mode=="sum"``,
+    the corresponding values in source will be added together
+    (``scatter_add`` in PyTorch),
+    or otherwise it will take the respective reduction.
 
     ``scatter`` is the inverse of :func:`gather`.
 
-    (TF segment_sum can be implemented via this.)
+    (segment_sum can be implemented via this.)
 
     :param source: [batch_dims..., indices_dim(s)..., feature_dims...]
     :param indices: [batch_dims..., indices_dim(s)...] -> out_dim
     :param indices_dim:
-    :param mode: "sum" or "max" or "min". also see :func:`scatter_argmax`.
+    :param mode: "sum", "max", "min", "logsumexp", "logmeanexp", "argmax".
+        (Note: If you ever need mean, argmin, etc, please open an issue/PR.)
     :param fill_value:
     :param out_dim: The indices target dim.
         If not given, will be automatically determined as the sparse_dim from indices.
@@ -778,6 +781,16 @@ def scatter(
         and then we use :func:`rf.split_dims` afterwards.
     :return: [batch_dims..., out_dim(s)..., feature_dims...]
     """
+    if mode == "logsumexp":
+        return scatter_logsumexp(
+            source, indices=indices, indices_dim=indices_dim, fill_value=fill_value, out_dim=out_dim
+        )
+    if mode == "logmeanexp":
+        return scatter_logmeanexp(
+            source, indices=indices, indices_dim=indices_dim, fill_value=fill_value, out_dim=out_dim
+        )
+    if mode == "argmax":
+        return scatter_argmax(source, indices=indices, indices_dim=indices_dim, invalid_idx=fill_value, out_dim=out_dim)
     if not out_dim:
         assert isinstance(indices, Tensor) and indices.sparse_dim
         out_dim = indices.sparse_dim
@@ -858,6 +871,76 @@ def scatter_argmax(
     return out
 
 
+def scatter_logsumexp(
+    source: Tensor,
+    *,
+    indices: Tensor,
+    indices_dim: Union[Dim, Sequence[Dim]],
+    fill_value: Optional[Union[int, float]] = None,
+    out_dim: Optional[Union[Dim, Sequence[Dim]]] = None,
+) -> Tensor:
+    """
+    Scatters into new zero-tensor.
+    If entries in indices are duplicated, the corresponding values in source will be log-sum-exp'ed together.
+    This is like :func:`scatter` with ``mode="logsumexp"``.
+
+    :param source: [batch_dims..., indices_dim(s)..., feature_dims...]
+    :param indices: [batch_dims..., indices_dim(s)...] -> out_dim
+    :param indices_dim:
+    :param fill_value:
+    :param out_dim: The indices target dim.
+        If not given, will be automatically determined as the sparse_dim from indices.
+        If multiple out dims, use indices into the merged out dims,
+        and then we use :func:`rf.split_dims` afterwards.
+    :return: [batch_dims..., out_dim(s)..., feature_dims...]
+    """
+    if not out_dim:
+        assert isinstance(indices, Tensor) and indices.sparse_dim
+        out_dim = indices.sparse_dim
+    with rf.stop_gradient_scope():
+        max_x = rf.scatter(source, indices=indices, indices_dim=indices_dim, mode="max", out_dim=out_dim)  # [D_out,...]
+        max_x_ = rf.gather(max_x, indices=indices, axis=out_dim)  # [D_src,...]
+    src_ = rf.exp(source - max_x_)
+    if fill_value is not None:
+        fill_value = rf.exp(fill_value - max_x_)
+    tensor = rf.scatter(
+        src_, indices=indices, indices_dim=indices_dim, mode="sum", fill_value=fill_value, out_dim=out_dim
+    )
+    tensor = rf.log(tensor)
+    tensor = rf.where(rf.is_neg_infinite(max_x), rf.zeros((), dtype=source.dtype, device=source.device), tensor)
+    tensor += max_x
+    return tensor
+
+
+def scatter_logmeanexp(
+    source: Tensor,
+    *,
+    indices: Tensor,
+    indices_dim: Union[Dim, Sequence[Dim]],
+    fill_value: Optional[Union[int, float]] = None,
+    out_dim: Optional[Union[Dim, Sequence[Dim]]] = None,
+) -> Tensor:
+    """
+    Scatters into new zero-tensor.
+    If entries in indices are duplicated, the corresponding values in source will be log-mean-exp'ed together.
+    This is like :func:`scatter` with ``mode="logmeanexp"``.
+
+    :param source: [batch_dims..., indices_dim(s)..., feature_dims...]
+    :param indices: [batch_dims..., indices_dim(s)...] -> out_dim
+    :param indices_dim:
+    :param fill_value:
+    :param out_dim: The indices target dim.
+        If not given, will be automatically determined as the sparse_dim from indices.
+        If multiple out dims, use indices into the merged out dims,
+        and then we use :func:`rf.split_dims` afterwards.
+    :return: [batch_dims..., out_dim(s)..., feature_dims...]
+    """
+    ones = rf.ones(dims=indices.dims, dtype=source.dtype, device=source.device)
+    counts = rf.scatter(ones, indices=indices, indices_dim=indices_dim, fill_value=1, out_dim=out_dim)
+    y = scatter_logsumexp(source, indices=indices, indices_dim=indices_dim, fill_value=fill_value, out_dim=out_dim)
+    return y - rf.log(counts)
+
+
 # noinspection PyShadowingBuiltins
 def slice(
     source: Tensor,

{returnn-1.20250220.200053 → returnn-1.20250223.154045}/returnn/frontend/gradient.py

@@ -11,6 +11,7 @@ __all__ = [
     "set_requires_gradient",
     "gradient",
     "stop_gradient",
+    "stop_gradient_scope",
     "scaled_gradient",
     "scaled_gradient_ext",
     "gradient_checkpoint_scope",
@@ -42,6 +43,28 @@ def stop_gradient(source: Tensor) -> Tensor:
     return source._raw_backend.stop_gradient(source)
 
 
+def stop_gradient_scope():
+    """
+    Create a stop gradient scope.
+    All tensors created within this scope will have their gradient stopped.
+
+    Example::
+
+        a = ...
+        b = ...
+        with stop_gradient_scope():
+            x = a + b
+        y = x * c
+
+    In this example, the tensor ``x`` will have its gradient stopped,
+    i.e. the gradient of ``x`` w.r.t. ``a`` and ``b`` will be zero.
+
+    :return: context manager which enables stopping the gradient. It supports __enter__ and __exit__,
+        and the intended usage is with the `with` statement.
+    """
+    return global_backend.stop_gradient_scope()
+
+
 def scaled_gradient(source: Tensor, scale: Union[float, Tensor]) -> Tensor:
     """
     :param source:

{returnn-1.20250220.200053 → returnn-1.20250223.154045}/returnn/torch/distributed.py

@@ -3,10 +3,11 @@ torch.distributed utils
 """
 
 from __future__ import annotations
-from typing import Optional, Any, Dict
+import ast
+import logging
 import os
 import socket
-import logging
+from typing import Optional, Any, Dict
 
 import torch
 from torch.nn.parallel import DistributedDataParallel
@@ -23,19 +24,31 @@ class DistributedContext:
     """
 
     def __init__(self, options: Dict[str, Any]):
-        import torch.distributed as dist
-
         self._opts = CollectionReadCheckCovered(options)
 
-        # when no backend is specified, both gloo and nccl backends will be created
-        # the gloo backend will be used for collectives with CPU tensors and
-        # the nccl backend will be used for collectives with CUDA tensors
-        dist.init_process_group(backend=self._opts.get("backend", None))
+        # Subprocesses have issues initializing torch.distributed process groups.
+        #
+        # We therefore pass rank/size information of the process group via an env
+        # variable that is automatically inherited in any created subprocess.
+        env_var_name = "_RETURNN_TORCH_DISTRIBUTED_INIT_INFO"
+        prev_init_info = os.environ.get(env_var_name)
+        if prev_init_info:
+            self.prev_init_info = ast.literal_eval(prev_init_info)
+            self._rank = self.prev_init_info["rank"]
+            self._size = self.prev_init_info["size"]
+        else:
+            import torch.distributed as dist
+
+            # when no backend is specified, both gloo and nccl backends will be created
+            # the gloo backend will be used for collectives with CPU tensors and
+            # the nccl backend will be used for collectives with CUDA tensors
+            dist.init_process_group(backend=self._opts.get("backend", None))
+            self._rank = dist.get_rank()
+            self._size = dist.get_world_size()
+            os.environ[env_var_name] = repr({"rank": self._rank, "size": self._size})
 
         self._local_rank = int(os.environ["LOCAL_RANK"])
         self._local_size = int(os.environ["LOCAL_WORLD_SIZE"])
-        self._rank = dist.get_rank()
-        self._size = dist.get_world_size()
 
         _logger.info(
             "Torch distributed initialized. Hostname %s, pid %i, rank %i / size %i, local rank %s / local size %s."
@@ -123,9 +136,9 @@ _is_set_up = False
 _ctx = None  # type: Optional[DistributedContext]
 
 
-def get_ctx(config=None) -> Optional[DistributedContext]:
+def get_ctx(config: Optional[Config] = None) -> Optional[DistributedContext]:
     """
-    :param Config|None config:
+    :param config:
     :returns: the global context if Torch distributed is enabled, or None otherwise.
       If we did not setup the context yet, it will automatically create it.
     """

{returnn-1.20250220.200053 → returnn-1.20250223.154045}/returnn/torch/frontend/_backend.py

@@ -3,7 +3,7 @@ Backend for exposing PyTorch-specific functionality.
 """
 
 from __future__ import annotations
-from typing import Optional, Union, Sequence, Tuple, List, Dict, Generator
+from typing import Optional, Union, Any, Sequence, Tuple, List, Dict, Generator
 import contextlib
 import torch
 import numpy
@@ -212,6 +212,11 @@ class TorchBackend(Backend[torch.Tensor]):
         out.raw_tensor = out.raw_tensor.detach()
         return out
 
+    @staticmethod
+    def stop_gradient_scope() -> Any:
+        """stop gradient scope"""
+        return torch.no_grad()
+
     @staticmethod
     def scaled_gradient(tensor: Tensor, scale: Union[float, Tensor]) -> Tensor:
         """scaled gradient"""
@@ -1061,9 +1066,9 @@ class TorchBackend(Backend[torch.Tensor]):
     ) -> Tensor:
         """
         Scatters into new zero-tensor.
-        If entries in indices are duplicated, the corresponding values in source will be added together
-        (scatter_add in PyTorch).
-        (TF segment_sum can be implemented via this.)
+        If entries in indices are duplicated, with mode="sum", the corresponding values in source will be added together
+        (``scatter_add`` in PyTorch), otherwise min/max.
+        (segment_sum can be implemented via this.)
 
         :param source: [batch_dims..., indices_dim(s)..., feature_dims...]
         :param indices: [batch_dims..., indices_dim(s)...] -> out_dim

{returnn-1.20250220.200053 → returnn-1.20250223.154045}/returnn/util/basic.py

@@ -3773,9 +3773,9 @@ def should_write_to_disk(config):
     if config.typed_value("torch_distributed") is not None:
         assert BackendEngine.is_torch_selected(), "torch_distributed assumes PyTorch"
 
-        import torch.distributed
+        import returnn.torch.distributed as torch_distributed
 
-        if torch.distributed.get_rank() != 0:
+        if torch_distributed.get_ctx(config).rank() != 0:
             return False
     elif config.is_true("use_horovod"):
         assert BackendEngine.is_tensorflow_selected(), "use_horovod currently assumes TensorFlow"

{returnn-1.20250220.200053 → returnn-1.20250223.154045/returnn.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: returnn
-Version: 1.20250220.200053
+Version: 1.20250223.154045
 Summary: The RWTH extensible training framework for universal recurrent neural networks
 Home-page: https://github.com/rwth-i6/returnn/
 Author: Albert Zeyer

returnn-1.20250220.200053/_setup_info_generated.py

@@ -1,2 +0,0 @@
-version = '1.20250220.200053'
-long_version = '1.20250220.200053+git.bb5c0aa'