Eliminate storage views. #9466

ezyang · 2018-07-16T16:13:09Z

Storage views were previously used to implement CUDA IPC sharing,
but they weren't necessary. The new strategy is described in
Note [CUDA IPC and the caching allocator].

This also fixes an unrelated bug, where we weren't actually using
the Tensor forking pickler, because we didn't register a pickler
for torch.Tensor.

Fixes #9447. Fixes #46.

Signed-off-by: Edward Z. Yang [email protected]

CC @apaszke

Storage views were previously used to implement CUDA IPC sharing, but they weren't necessary. The new strategy is described in Note [CUDA IPC and the caching allocator]. This also fixes an unrelated bug, where we weren't actually using the Tensor forking pickler, because we didn't register a pickler for torch.Tensor. Fixes pytorch#9447. Fixes pytorch#46. Signed-off-by: Edward Z. Yang <[email protected]>

facebook-github-bot

@ezyang has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

Signed-off-by: Edward Z. Yang <[email protected]>

facebook-github-bot

@ezyang has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

torch/csrc/generic/StorageMethods.cpp

 }

 #ifndef THD_GENERIC_FILE
 PyObject * THPStorage_(_rootStorage)(THPStorage *self)


torch/csrc/generic/StorageSharing.cpp

+      && (_handle == Py_None || PyBytes_Check(_handle)))) {
    THPUtils_invalidArguments(args, NULL, "_new_shared in CUDA mode", 1,
-        "(int device, bytes handle, int storage_size, int offset, int view_size");
+        "(int device, bytes handle, int storage_size)");


torch/csrc/generic/StorageSharing.cpp


    _handle = PyBytes_FromStringAndSize((char *)&handle, CUDA_IPC_HANDLE_SIZE);
-    _offset = PyLong_FromSsize_t((Py_ssize_t)offset);
+    _offset = PyLong_FromSsize_t((Py_ssize_t)offset / sizeof(real));


torch/csrc/generic/StorageSharing.cpp

-  if (offset != 0 || view_size != storage_size) {
-    return THPStorage_(newTHView)(base.get(), offset, view_size);
-  }
+  base->flag = TH_STORAGE_REFCOUNTED;  // NB: Not resizable


torch/serialization.py

-                    deserialized_objects[target_cdata] = root[offset:offset + size]
+                    if offset != 0 or size != root.size():
+                        warnings.warn("Detected storage view in legacy serialized data: "
+                                      "storages are no longer natively supported, so we are making "


torch/serialization.py

+                        tensor = torch._utils._rebuild_tensor(root, offset, (size,), (1,))
+                        obj = tensor.clone().storage()
+                    else:
+                        obj = root[offset:offset + size]


Signed-off-by: Edward Z. Yang <[email protected]>

facebook-github-bot

@ezyang has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

Signed-off-by: Edward Z. Yang <[email protected]>

facebook-github-bot

@ezyang has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

Summary: Storage views were previously used to implement CUDA IPC sharing, but they weren't necessary. The new strategy is described in Note [CUDA IPC and the caching allocator]. This also fixes an unrelated bug, where we weren't actually using the Tensor forking pickler, because we didn't register a pickler for torch.Tensor. Fixes #9447. Fixes #46. Signed-off-by: Edward Z. Yang <[email protected]> CC apaszke Pull Request resolved: pytorch/pytorch#9466 Reviewed By: apaszke Differential Revision: D8859698 Pulled By: ezyang fbshipit-source-id: 3362cb92f6ae4aa37084c57d79b31004bd0b4a97

Summary: Storage views were previously used to implement CUDA IPC sharing, but they weren't necessary. The new strategy is described in Note [CUDA IPC and the caching allocator]. This also fixes an unrelated bug, where we weren't actually using the Tensor forking pickler, because we didn't register a pickler for torch.Tensor. Fixes pytorch#9447. Fixes pytorch#46. Signed-off-by: Edward Z. Yang <[email protected]> CC apaszke Pull Request resolved: pytorch#9466 Reviewed By: apaszke Differential Revision: D8859698 Pulled By: ezyang fbshipit-source-id: 3362cb92f6ae4aa37084c57d79b31004bd0b4a97

In pytorch#9466 I got rid of storage views and eliminated all places where they were used... OR SO I THOUGHT. In actuality, under certain conditions (specifically, if you trained a CUDA multiprocessing model shared over CUDA IPC and then serialized your parameters), you could also serialize storage slices to the saved model format. In pytorch#9466, I "fixed" the case when you loaded the legacy model format (really, just unshared the storages--not strictly kosher but if you aren't updating the parameters, shouldn't matter), but NOT the modern model format, so such models would fail. So, I could have applied the legacy model format fix too, but hyperfraise remarked that he had applied a fix that was effectively the same as unsharing the storages, but it had caused his model to behave differently. So I looked into it again, and realized that using a custom deleter, I could simulate the same behavior as old storage slices. So back they come. In principle, I could also reimplement storage views entirely using our allocators, but I'm not going to do that unless someone really really wants it. Fixes pytorch#10120. Signed-off-by: Edward Z. Yang <[email protected]>

Summary: In #9466 I got rid of storage views and eliminated all places where they were used... OR SO I THOUGHT. In actuality, under certain conditions (specifically, if you trained a CUDA multiprocessing model shared over CUDA IPC and then serialized your parameters), you could also serialize storage slices to the saved model format. In #9466, I "fixed" the case when you loaded the legacy model format (really, just unshared the storages--not strictly kosher but if you aren't updating the parameters, shouldn't matter), but NOT the modern model format, so such models would fail. So, I could have applied the legacy model format fix too, but hyperfraise remarked that he had applied a fix that was effectively the same as unsharing the storages, but it had caused his model to behave differently. So I looked into it again, and realized that using a custom deleter, I could simulate the same behavior as old storage slices. So back they come. In principle, I could also reimplement storage views entirely using our allocators, but I'm not going to do that unless someone really really wants it. Fixes #10120. Signed-off-by: Edward Z. Yang <[email protected]> Pull Request resolved: #11314 Reviewed By: ailzhang Differential Revision: D9671966 Pulled By: ezyang fbshipit-source-id: fd863783d03b6a6421d6b9ae21ce2f0e44a0dcce

Summary: In pytorch#9466 I got rid of storage views and eliminated all places where they were used... OR SO I THOUGHT. In actuality, under certain conditions (specifically, if you trained a CUDA multiprocessing model shared over CUDA IPC and then serialized your parameters), you could also serialize storage slices to the saved model format. In pytorch#9466, I "fixed" the case when you loaded the legacy model format (really, just unshared the storages--not strictly kosher but if you aren't updating the parameters, shouldn't matter), but NOT the modern model format, so such models would fail. So, I could have applied the legacy model format fix too, but hyperfraise remarked that he had applied a fix that was effectively the same as unsharing the storages, but it had caused his model to behave differently. So I looked into it again, and realized that using a custom deleter, I could simulate the same behavior as old storage slices. So back they come. In principle, I could also reimplement storage views entirely using our allocators, but I'm not going to do that unless someone really really wants it. Fixes pytorch#10120. Signed-off-by: Edward Z. Yang <[email protected]> Pull Request resolved: pytorch#11314 Reviewed By: ailzhang Differential Revision: D9671966 Pulled By: ezyang fbshipit-source-id: fd863783d03b6a6421d6b9ae21ce2f0e44a0dcce

ezyang requested review from apaszke, colesbury, gchanan, soumith and zdevito as code owners July 16, 2018 16:13

facebook-github-bot reviewed Jul 16, 2018

View reviewed changes

Continue to 'support' deserialization of storage views by cloning.

b1e7aaf

Signed-off-by: Edward Z. Yang <[email protected]>

ezyang force-pushed the pr/delete-storage-views branch from 2618def to b1e7aaf Compare July 16, 2018 18:55

facebook-github-bot reviewed Jul 16, 2018

View reviewed changes

apaszke approved these changes Jul 16, 2018

View reviewed changes

CR

1fbf3d0

Signed-off-by: Edward Z. Yang <[email protected]>

facebook-github-bot reviewed Jul 16, 2018

View reviewed changes

Braino bugfix

c08f91f

Signed-off-by: Edward Z. Yang <[email protected]>

facebook-github-bot reviewed Jul 16, 2018

View reviewed changes

facebook-github-bot closed this in 976f925 Jul 16, 2018

ezyang mentioned this pull request Sep 6, 2018

Reimplement storage slicing. #11314

Closed

ezyang added the merged label Jun 26, 2019

Eliminate storage views. #9466

Eliminate storage views. #9466

Uh oh!

Conversation

ezyang commented Jul 16, 2018

Uh oh!

facebook-github-bot left a comment

Choose a reason for hiding this comment

Uh oh!

facebook-github-bot left a comment

Choose a reason for hiding this comment

Uh oh!

This comment was marked as off-topic.

Uh oh!

This comment was marked as off-topic.

Uh oh!

This comment was marked as off-topic.

Uh oh!

This comment was marked as off-topic.

Uh oh!

This comment was marked as off-topic.

Uh oh!

This comment was marked as off-topic.

Uh oh!

This comment was marked as off-topic.

Uh oh!

This comment was marked as off-topic.

Uh oh!

This comment was marked as off-topic.

Uh oh!

This comment was marked as off-topic.

Uh oh!

This comment was marked as off-topic.

Uh oh!

facebook-github-bot left a comment

Choose a reason for hiding this comment

Uh oh!

facebook-github-bot left a comment

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants