PyPI · #2932 most downloaded on PyPI

fastsafetensors

0.4.01.9M downloads/mofoundation-model-stack/fastsafetensors ↗New releases in your daily digest, changelog included.

High-performance safetensors model loader

Last release 27 days ago
07 Sep 2026
Ships fairly regularly
a new release about every 2 months
Some releases are documented
notes for 9 of 21 stable releases
1 version withdrawn
withdrawn after publishing
2 years old
22 releases · first in 2024

Release timeline

22 releases since 2024

202420252026

Releases

0.4.07 Sept 2026
Sub-file chunk loading bounded by max_batch_bytes by @gitbisector in #90
What's Changed

Sub-file chunk loading bounded by max_batch_bytes by @gitbisector in #90

Add Windows ARM64 wheel builds by @chinazhangchao in #97

Fit planner: deterministic per-file chunk budgets (device_memory_budget) by @gitbisector in #91

Improve DirectStorage fallback and CUDA runtime discovery by @chinazhangchao in #98

fix(tests): keep the 3FS mock reader on host memory by @gitbisector in #100

fix: skip cuFileDriverOpen() when /dev/nvidia-fs0 is missing to prevent fd 0 corruption by @yinjuncheng in #102

Account for yield clones in fit planner by @takeshi-yoshimura in #103

Reuse chunk budgets for stable buffer allocations by @takeshi-yoshimura in #104

Fix GDS capability checks before library initialization by @takeshi-yoshimura in #105

Clamp queue_size to the device memory budget instead of failing by @takeshi-yoshimura in #106

Per-tensor residency predicate for the fit planner by @gitbisector in #108

Prepare the 0.4.0 release by @takeshi-yoshimura in #107

New Contributors

@chinazhangchao made their first contribution in #97

@yinjuncheng made their first contribution in #102

Full Changelog: 0.3.3...0.4.0

0.3.37 Jul 2026

Add aarch64 CI and wheel publishing targets by @takeshi-yoshimura in #83

What's Changed

Add aarch64 CI and wheel publishing targets by @takeshi-yoshimura in #83
Fixes by Fable and Codex by @takeshi-yoshimura in #84
[ROCm] Add a ROCm direct-I/O (hipFile) copier backend by @kurapov-peter in #85
Add framework-hinted runtime library discovery to address runtime mis… by @kurapov-peter in #86
Add caller-side byte-range selection (sub-file reads) + expert-parallel filter by @gitbisector in #81
Fix KeyError on empty metadata in safetensors headers by @noobHappylife in #89
Fall back to nogds at runtime when cuFile handle registration fails by @gitbisector in #87
0.3.3 by @takeshi-yoshimura in #92

New Contributors

@kurapov-peter made their first contribution in #85
@noobHappylife made their first contribution in #89

Full Changelog: 0.3.2...0.3.3

Open source →

0.3.222 May 2026

[FEAT] Windows support through DirectX 12 DirectStorage interop by @SystemPanic in #72

What's Changed

[FEAT] Windows support through DirectX 12 DirectStorage interop by @SystemPanic in #72
fix workflows for Windows and multi GPUs by @takeshi-yoshimura in #77
feat: add UnifiedLoader with config-driven loader dispatch by @ABNER-1 in #70
Add support for F4 and F8_E8M0 dtypes by @takeshi-yoshimura in #79
Universal CUDA/ROCm wheel via runtime platform detection by @wjabbour in #78
fix for 0.3.2 by @takeshi-yoshimura in #80

New Contributors

@SystemPanic made their first contribution in #72

Full Changelog: 0.3.1...0.3.2

Open source →

0.3.16 May 2026

fix publish by @takeshi-yoshimura in #64

What's Changed

fix publish by @takeshi-yoshimura in #64
[ROCm]Use vendor-agnostic check is nogds copier by @ZhiweiYan-96 in #66
[ROCm] Remove hipify-perl build dependency, enable manylinux wheel builds by @wjabbour in #67
add python 3.14 by @takeshi-yoshimura in #75

New Contributors

@ZhiweiYan-96 made their first contribution in #66
@wjabbour made their first contribution in #67

Full Changelog: 0.3...0.3.1

Open source →

0.322 Apr 2026

cast FP8 tensors to bf16 for NCCL ops on pre-sm90 GPUs by @bet0x in #54

What's Changed

cast FP8 tensors to bf16 for NCCL ops on pre-sm90 GPUs by @bet0x in #54
add copier registry by @lixin010 in #56
fix: avoid unnecessary copier construction for non-owned files by @lixin010 in #58
Add unified memory copier for DGX Spark / Grace Hopper by @gitbisector in #60
fix memcopy kind by @takeshi-yoshimura in #61
fix memcopy kind again by @takeshi-yoshimura in #62
feat: add 3FS copier support by @ABNER-1 in #59
fix for 0.3 by @takeshi-yoshimura in #63

New Contributors

@bet0x made their first contribution in #54
@lixin010 made their first contribution in #56
@gitbisector made their first contribution in #60

Full Changelog: 0.2.2...0.3

Open source →

0.2.017 Dec 2025

add disable cache in LazyTensorFactory by @ABNER-1 in https://github.com/foundation-model-stack/fastsafetensors/pull/30

What's Changed

add disable cache in LazyTensorFactory by @ABNER-1 in https://github.com/foundation-model-stack/fastsafetensors/pull/30
init gds only if need by @ABNER-1 in https://github.com/foundation-model-stack/fastsafetensors/pull/31
Introduce a Copier interface to support custom scenarios by @ABNER-1 in https://github.com/foundation-model-stack/fastsafetensors/pull/32
[FEAT] [ROCm] Add ROCm support to fastsafetensors by @tjtanaa in https://github.com/foundation-model-stack/fastsafetensors/pull/34
[Bugfix] fix hipMemcpyKind enum by @tjtanaa in https://github.com/foundation-model-stack/fastsafetensors/pull/35
Add pipeline loader by @ABNER-1 in https://github.com/foundation-model-stack/fastsafetensors/pull/33
Check CU_DEVICE_ATTRIBUTE_GPU_DIRECT_RDMA_SUPPORTED by @takeshi-yoshimura in https://github.com/foundation-model-stack/fastsafetensors/pull/39
[WiP] Refactoring before 0.2 by @takeshi-yoshimura in https://github.com/foundation-model-stack/fastsafetensors/pull/41
fix Dockerfile.build by @takeshi-yoshimura in https://github.com/foundation-model-stack/fastsafetensors/pull/44

New Contributors

@ABNER-1 made their first contribution in https://github.com/foundation-model-stack/fastsafetensors/pull/30
@tjtanaa made their first contribution in https://github.com/foundation-model-stack/fastsafetensors/pull/34

Full Changelog: https://github.com/foundation-model-stack/fastsafetensors/compare/0.1.15...0.2.0

Open source →

Additional notes

What's Changed

add disable cache in LazyTensorFactory by @ABNER-1 in #30
init gds only if need by @ABNER-1 in #31
Introduce a Copier interface to support custom scenarios by @ABNER-1 in #32
[FEAT] [ROCm] Add ROCm support to fastsafetensors by @tjtanaa in #34
[Bugfix] fix hipMemcpyKind enum by @tjtanaa in #35
Add pipeline loader by @ABNER-1 in #33
Check CU_DEVICE_ATTRIBUTE_GPU_DIRECT_RDMA_SUPPORTED by @takeshi-yoshimura in #39
[WiP] Refactoring before 0.2 by @takeshi-yoshimura in #41
fix Dockerfile.build by @takeshi-yoshimura in #44

New Contributors

@ABNER-1 made their first contribution in #30
@tjtanaa made their first contribution in #34

Full Changelog: 0.1.15...0.2.0

Open source →

0.1.1518 Jul 2025

[PDPD]fix paddle from_dlpack error by @bukejiyu in https://github.com/foundation-model-stack/fastsafetensors/pull/24

What's Changed

[PDPD]fix paddle from_dlpack error by @bukejiyu in https://github.com/foundation-model-stack/fastsafetensors/pull/24
Track memory usage and fix bugs by @takeshi-yoshimura in https://github.com/foundation-model-stack/fastsafetensors/pull/27

New Contributors

@bukejiyu made their first contribution in https://github.com/foundation-model-stack/fastsafetensors/pull/24

Full Changelog: https://github.com/foundation-model-stack/fastsafetensors/compare/0.1.14...0.1.15

Open source →

Additional notes

What's Changed

[PDPD]fix paddle from_dlpack error by @bukejiyu in #24
Track memory usage and fix bugs by @takeshi-yoshimura in #27

New Contributors

@bukejiyu made their first contribution in #24

Full Changelog: 0.1.14...0.1.15

Open source →