Merge pull request #181 from coreweave/eta/1d-multibyte-slicing

fix(serialization): Slice 1-D multibyte data as bytes for `pwrite`

Author: wbrown

CHANGELOG.md

@@ -5,10 +5,12 @@ All notable changes to this project will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
-## [Unreleased]
+## [2.9.1] - 2024-11-27
 
 ### Fixed
 
+- `TensorSerializer` no longer sometimes fails to serialize very large
+  1-dimensional tensors with multibyte `dtype`s
 - `RedisStreamFile.readable()` and `RedisStreamFile.seekable()` now correctly
   return `True`
 
@@ -407,7 +409,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
   - `get_gpu_name`
   - `no_init_or_tensor`
 
-[Unreleased]: https://github.com/coreweave/tensorizer/compare/v2.9.0...HEAD
+[2.9.1]: https://github.com/coreweave/tensorizer/compare/v2.9.0...v2.9.1
 [2.9.0]: https://github.com/coreweave/tensorizer/compare/v2.8.1...v2.9.0
 [2.8.1]: https://github.com/coreweave/tensorizer/compare/v2.8.0...v2.8.1
 [2.8.0]: https://github.com/coreweave/tensorizer/compare/v2.7.2...v2.8.0

tensorizer/_version.py

@@ -1 +1 @@
-__version__ = "2.9.0"
+__version__ = "2.9.1"

tensorizer/serialization.py

@@ -3510,7 +3510,7 @@ def _mv_suffix(data: "collections.abc.Buffer", start: int):
         if not isinstance(data, memoryview):
             data = memoryview(data)
         try:
-            if data.ndim != 1:
+            if data.ndim != 1 or data.format != "B":
                 data = data.cast("B")
             return data[start:]
         finally:
@@ -3526,13 +3526,15 @@ def _pwrite_syscall(
             verify if isinstance(verify, int) else self._buffer_size(data)
         )
         bytes_just_written: int = os.pwrite(self._fd, data, offset)
-        bytes_written += bytes_just_written
+        if bytes_just_written > 0:
+            bytes_written += bytes_just_written
         while bytes_written < expected_bytes_written and bytes_just_written > 0:
             # Writes larger than ~2 GiB may not complete in a single pwrite call
             offset += bytes_just_written
             with self._mv_suffix(data, bytes_written) as mv:
                 bytes_just_written = os.pwrite(self._fd, mv, offset)
-            bytes_written += bytes_just_written
+            if bytes_just_written > 0:
+                bytes_written += bytes_just_written
         if isinstance(verify, int) or verify:
             self._verify_bytes_written(bytes_written, expected_bytes_written)
         return bytes_written