[Performance] Make _to_consolidated compatible with compile

ghstack-source-id: c0f6116
Pull Request resolved: #1041

Author: Vincent Moens

tensordict/base.py

@@ -3521,9 +3521,10 @@ def _reduce_vals_and_metadata(self, *, dtype=NO_DEFAULT, requires_metadata):
 
         flat_size = []
         start = 0
+        sorting_index = 0
 
         def add_single_value(value, key, metadata_dict, dtype, shape, flat_size):
-            nonlocal start
+            nonlocal start, sorting_index
             n = value.element_size() * value.numel()
             if need_padding:
                 pad = n % 8
@@ -3541,7 +3542,10 @@ def add_single_value(value, key, metadata_dict, dtype, shape, flat_size):
                     start,
                     stop,
                     pad,
+                    flat_size[-1],
+                    sorting_index,
                 )
+            sorting_index = sorting_index + 1
             start = stop
 
         def assign(
@@ -10395,6 +10399,7 @@ def to(self, *args, **kwargs) -> T:
                 pin_memory=non_blocking_pin,
                 num_threads=num_threads,
                 non_blocking=non_blocking,
+                compilable=is_dynamo_compiling(),
             )
 
         if non_blocking is None:
@@ -10452,14 +10457,42 @@ def to_pinmem(tensor, _to=to):
             self._sync_all()
         return result
 
-    def _to_consolidated(self, *, device, pin_memory, num_threads, non_blocking):
+    def _to_consolidated(self, *, device, pin_memory, num_threads, non_blocking, compilable):
         if num_threads is None:
             # unspecified num_threads should mean 0
             num_threads = 0
+
         storage = self._consolidated["storage"]
-        if pin_memory:
-            storage = storage.pin_memory()
-        storage_cast = storage.to(device, non_blocking=True)
+
+        @torch.compiler.disable()
+        def to(storage):
+            if pin_memory:
+                storage = storage.pin_memory()
+            storage_cast = storage.to(device, non_blocking=True)
+            return storage_cast
+        storage_cast = to(storage)
+
+        if compilable:
+            result = self._to_consolidated_compile(device=device, num_threads=num_threads, storage_cast=storage_cast)
+        else:
+            result = self._to_consolidated_eager(device=device, num_threads=num_threads, storage_cast=storage_cast)
+
+        if non_blocking in (False, None):
+            if device.type == "cuda" and non_blocking is False:
+                # sending to CUDA force sync
+                cuda_device = device
+            elif storage.device.type == "cuda":
+                # sending from cuda: need sync unless intentionally not asked for
+                cuda_device = storage.device.type
+            else:
+                cuda_device = None
+            if cuda_device is not None:
+                torch.cuda.current_stream(cuda_device).synchronize()
+
+        return result
+
+    def _to_consolidated_eager(self, *, device, num_threads, storage_cast):
+
         untyped_storage = storage_cast.untyped_storage()
 
         def set_(x):
@@ -10528,20 +10561,111 @@ def copy_dict(d):
                 }
 
             result._consolidated["metadata"] = copy_dict(self._consolidated["metadata"])
-        if non_blocking in (False, None):
-            if device.type == "cuda" and non_blocking is False:
-                # sending to CUDA force sync
-                cuda_device = device
-            elif storage.device.type == "cuda":
-                # sending from cuda: need sync unless intentionally not asked for
-                cuda_device = storage.device.type
-            else:
-                cuda_device = None
-            if cuda_device is not None:
-                torch.cuda.current_stream(cuda_device).synchronize()
-
         return result
 
+    def _to_consolidated_compile(self, *, device, num_threads, storage_cast):
+
+        def get_tensors_length(metadata, lengths=None, pos=None, keys=None, prefix=()):
+            root = False
+            if lengths is None:
+                lengths = []
+                pos = []
+                keys = []
+                root = True
+            for k, v in metadata["leaves"].items():
+                lengths.append(v[-2])
+                pos.append(v[-1])
+                keys.append(prefix + (k,))
+            for k, d in metadata.items():
+                if "leaves" in d:
+                    get_tensors_length(d, lengths=lengths, pos=pos, keys=keys, prefix=prefix + (k,))
+            if root:
+                # l = torch.empty(len(lengths), dtype=torch.long)
+                # l[torch.as_tensor(pos)] = torch.as_tensor(lengths)
+                out0 = [None, ] * len(pos)
+                out1 = [None, ] * len(pos)
+                for p, l, k in zip(pos, lengths, keys):
+                    out0[p] = k
+                    out1[p] = l
+                return out0, out1
+
+        def split_storage(consolidated):
+            keys, splits = get_tensors_length(consolidated["metadata"])
+            return dict(zip(keys, consolidated["storage"].split(splits)))
+
+        if num_threads is None:
+            # unspecified num_threads should mean 0
+            num_threads = 0
+
+        _consolidated = {"storage": storage_cast}
+        if "metadata" in self._consolidated:
+            # faster than deepcopy
+            def copy_dict(d):
+                return {
+                    k: v if not isinstance(v, dict) else copy_dict(v)
+                    for k, v in d.items()
+                }
+
+            _consolidated["metadata"] = copy_dict(self._consolidated["metadata"])
+
+        slice_map = split_storage(_consolidated)
+
+        def set_(name, x):
+            if not isinstance(name, tuple):
+                name = (name,)
+            if x.is_nested:
+                from torch._subclasses.fake_tensor import FakeTensor
+                from torch._subclasses.functional_tensor import FunctionalTensor
+                from torch.nested._internal.nested_tensor import (
+                    _tensor_symint_registry,
+                    NestedTensor,
+                )
+                from torch.nested._internal.ops import extract_kwargs
+
+                if x.layout != torch.jagged:
+                    raise RuntimeError(
+                        "to(device) with nested tensors that do not have a jagged layout is not implemented yet. "
+                        "Please raise an issue on GitHub."
+                    )
+                kwargs = extract_kwargs(x)
+                values = x._values
+                lengths = x._lengths
+                offsets = x._offsets
+                storage_offsets = slice_map[(*name[:-1], "<NJT_OFFSETS>"+name[-1],)]
+                kwargs["offsets"] = storage_offsets.view(offsets.dtype).view(offsets.shape)
+                if lengths is not None:
+                    storage_lengths = slice_map[(*name[:-1], "<NJT_LENGTHS>"+name[-1],)]
+                    kwargs["lengths"] = storage_lengths.view(lengths.dtype).view(lengths.shape)
+                    ragged_source = lengths
+                else:
+                    ragged_source = offsets
+                new_thing = kwargs.get("lengths", kwargs.get("offsets"))
+                if isinstance(new_thing, (FakeTensor, FunctionalTensor)):
+                    from torch._subclasses.functional_tensor import (
+                        mb_unwrap_functional_tensor,
+                    )
+
+                    # Temporary hack until we have the union find
+                    tgt = mb_unwrap_functional_tensor(new_thing)
+                    src = mb_unwrap_functional_tensor(ragged_source)
+                    tgt.nested_int_memo = src.nested_int_memo
+                else:
+                    _tensor_symint_registry[new_thing] = _tensor_symint_registry[
+                        ragged_source
+                    ]
+
+                storage_values = slice_map[(*name[:-1], "<NJT_VALUES>"+name[-1],)]
+                return NestedTensor(
+                    storage_values.view(values.dtype).view(values.shape),
+                    **kwargs,
+                )
+            return slice_map[name].view(x.dtype).view(x.shape)
+
+        result = self._fast_apply(
+            set_, device=torch.device(device), num_threads=num_threads, named=True, nested_keys=True,
+        )
+        result._consolidated = _consolidated
+        return result
     def _sync_all(self):
         if _has_cuda:
             # TODO: dynamo doesn't like torch.cuda.is_initialized