[fix] original size computation (#1037)

* flip per_tensor's default

* fixed original size computation
Co-authored-by: Min Xu <min.xu.public@gmail.com>

fairscale/experimental/wgit/cli.py

@@ -36,6 +36,11 @@ def main(argv: List[str] = None) -> None:
         metavar="FILE_PATH",
         help="add a file to the staged changeset (default: none)",
     )
+    add_parser.add_argument(
+        "--no_per_tensor",
+        action="store_true",
+        help="Disable per-tensor adding of a file",
+    )
 
     commit_parser = subparsers.add_parser("commit", description="Commits the staged changes")
     commit_parser.add_argument("commit", action="store_true", help="Commit the staged changes")
@@ -76,7 +81,7 @@ def main(argv: List[str] = None) -> None:
 
     if args.command == "add":
         repo = Repo(Path.cwd())
-        repo.add(args.add)
+        repo.add(args.add, per_tensor=not args.no_per_tensor)
 
     if args.command == "status":
         repo = Repo(Path.cwd())

fairscale/experimental/wgit/repo.py

@@ -192,7 +192,7 @@ class Repo:
     def add(
         self,
         in_file_path: str,
-        per_tensor: bool = False,
+        per_tensor: bool = True,
         gzip: bool = True,
         sparsify: bool = False,
         sparsify_policy: Any = None,
@@ -209,7 +209,7 @@ class Repo:
                 Add a file in a per-tensor fashion. This enables more deduplication
                 due to tensors being identical. Deduplication cannot be disabled
                 completely because we use a content addressable SHA1_Store class.
-                Default: False
+                Default: True
             gzip (bool, optional):
                 Enable gzip based lossless compression on the object being added.
                 Default: True

fairscale/experimental/wgit/sha1_store.py

@@ -273,7 +273,10 @@ class SHA1_Store:
 
         # Update the sizes for this entry.
         entry = _get_json_entry(self._json_dict[sha1_hash])
-        o_diff = orig_size if ref_count == 1 else entry[ENTRY_OS_KEY]
+        assert (
+            ref_count == 1 or entry[ENTRY_OS_KEY] % (ref_count - 1) == 0
+        ), f"incorrect size: {entry[ENTRY_OS_KEY]} and {ref_count}"
+        o_diff = orig_size if ref_count == 1 else (entry[ENTRY_OS_KEY] // (ref_count - 1))
         d_diff = orig_size if ref_count == 1 else 0
         c_diff = comp_size if ref_count == 1 else 0
         entry[ENTRY_OS_KEY] += o_diff

tests/experimental/wgit/test_api.py

@@ -85,7 +85,8 @@ def test_api_commit(capsys, repo):
     assert line[0].rstrip().split()[-1] == commit_msg
 
 
-def test_api_status(capsys, repo):
+@pytest.mark.parametrize("per_tensor", [True, False])
+def test_api_status(capsys, repo, per_tensor):
     # delete the repo and initialize a new one:
     shutil.rmtree(".wgit")
     repo = Repo(Path.cwd(), init=True)
@@ -96,7 +97,7 @@ def test_api_status(capsys, repo):
 
     # check status before after a file is added but not committed
     chkpt0 = f"checkpoint_{random.randint(0, 1)}.pt"
-    repo.add(chkpt0)
+    repo.add(chkpt0, per_tensor=per_tensor)
     out = repo.status()
     key_list = list(repo._get_metdata_files().keys())
     assert out == {key_list[0]: RepoStatus.CHANGES_ADDED_NOT_COMMITED}
@@ -107,18 +108,18 @@ def test_api_status(capsys, repo):
     assert out == {key_list[0]: RepoStatus.CLEAN}
 
     # check status after a new change has been made to the file
-    torch.save(nn.Linear(1, int(15e5)), chkpt0)
+    torch.save(nn.Linear(1, int(15e5)).state_dict(), chkpt0)
     out = repo.status()
     assert out == {key_list[0]: RepoStatus.CHANGES_NOT_ADDED}
 
     # add the new changes made to weigit
-    repo.add(chkpt0)
+    repo.add(chkpt0, per_tensor=per_tensor)
     out = repo.status()
     assert out == {key_list[0]: RepoStatus.CHANGES_ADDED_NOT_COMMITED}
 
     # check status after a new different file is added to be tracked by weigit
     chkpt3 = "checkpoint_3.pt"
-    repo.add(chkpt3)
+    repo.add(chkpt3, per_tensor=per_tensor)
     key_list = list(repo._get_metdata_files().keys())
     out = repo.status()
     assert out == {

tests/experimental/wgit/test_cli.py

@@ -35,7 +35,7 @@ def create_test_dir():
     # create random checkpoints
     size_list = [30e5, 35e5, 40e5]
     for i, size in enumerate(size_list):
-        torch.save(nn.Linear(1, int(size)), f"checkpoint_{i}.pt")
+        torch.save(nn.Linear(1, int(size)).state_dict(), f"checkpoint_{i}.pt")
 
     # Test init.
     cli.main(["init"])
@@ -53,7 +53,7 @@ def test_cli_init(create_test_dir, capsys):
 
 def test_cli_add(create_test_dir, capsys):
     chkpt0 = "checkpoint_0.pt"
-    cli.main(["add", chkpt0])
+    cli.main(["add", "--no_per_tensor", chkpt0])
 
     sha1_store = SHA1_Store(
         Path.cwd().joinpath(".wgit", "sha1_store"),