Bump the root-dependencies group with 3 updates

[dependabot]

Updates the requirements on sentence-transformers, numkong and setuptools to permit the latest version.
Updates sentence-transformers to 5.7.0

Release notes

Sourced from sentence-transformers's releases.

v5.7.0 - GradCache Overhaul, torch.compile Inference, and a Large Batch of Correctness Fixes

This minor version is a correctness and performance-focused release. It rebuilds all gradient-cached losses on one shared engine, fixing several silently wrong gradients and adding token-based mini-batching for up to 3.9x faster cached-loss training. It also makes model.compile() actually speed up inference, and brings a long list of fixes across embedding quantization, evaluators, hard-negative mining, community detection, and multimodal inputs.

Two changes are marked breaking (🚨): int8/uint8 embedding quantization now clips out-of-range values and floors bucket values, so int8 outputs are no longer bit-identical with earlier versions, and AdaptiveLayerLoss/Matryoshka2dLoss now weight prior-layer losses uniformly by default. There's also a forward-looking deprecation: loading models whose modules import classes from outside sentence_transformers will require trust_remote_code=True from v6.0.

Install this version with

# Training + Inference
pip install sentence-transformers[train]==5.7.0
Inference only, use one of:
pip install sentence-transformers==5.7.0
pip install sentence-transformers[onnx-gpu]==5.7.0
pip install sentence-transformers[onnx]==5.7.0
pip install sentence-transformers[openvino]==5.7.0
Multimodal dependencies (optional):
pip install sentence-transformers[image]==5.7.0
pip install sentence-transformers[audio]==5.7.0
pip install sentence-transformers[video]==5.7.0
Or combine as needed:
pip install sentence-transformers[train,onnx,image]==5.7.0

GradCache overhaul: correct gradients and token-based mini-batching (#3862)

The gradient-cached losses (CachedMultipleNegativesRankingLoss, CachedGISTEmbedLoss, CachedSpladeLoss, the Cross Encoder CachedMultipleNegativesRankingLoss, and MegaBatchMarginLoss) train with large batch sizes at constant memory by embedding in mini-batches and replaying them with cached gradients. Each loss carried its own diverged copy of that machinery. They are now all rebuilt on one shared engine, which fixed several bugs that silently corrupted gradients:

• Cross Encoder CachedMultipleNegativesRankingLoss on GPU: the backward pass used different dropout masks than the forward pass, silently biasing gradients for every reranker trained with dropout active on CUDA or MPS. CPU training was unaffected.
• Running two forward passes before a backward pass (e.g. in custom training loops) made CachedGISTEmbedLoss and the Cross Encoder loss backpropagate the wrong batch's gradients, because the cache was stored on the loss module. The cache now travels with each forward pass's backward hook (the .cache and .random_states loss attributes are gone as a result).
• Pooling(include_prompt=False) (e.g. Instructor models) mutated the attention mask in place, so the backward re-embedding of every cached loss ran with a different mask than the forward pass.
• MatryoshkaLoss(GISTEmbedLoss(...)): the guide model overwrote the cached embeddings, so only the largest Matryoshka dimension was actually trained.

Along the way, this also fixed an autocast dtype crash in the backward pass and the trainer retaining autograd graphs between logging steps when tracking loss components.

MegaBatchMarginLoss's default mini-batched version is rebuilt on the engine as well. It crashed outright on recent releases, and underneath that, its historical implementation only applied the last mini-batch's gradients. It now trains on the full batch (results will differ, for the better), works with MatryoshkaLoss, evaluates under torch.no_grad, and raises for a third input column instead of silently ignoring it.

The headline feature is mini_batch_num_tokens, available on CachedMultipleNegativesRankingLoss, CachedMultipleNegativesSymmetricRankingLoss, CachedGISTEmbedLoss, CachedSpladeLoss, and MegaBatchMarginLoss. Instead of a fixed number of sequences per mini-batch, mini-batches are greedily packed by total non-padding token count, giving near-constant work per mini-batch on variable-length data:

from sentence_transformers import SentenceTransformer
from sentence_transformers.sentence_transformer.losses import CachedMultipleNegativesRankingLoss
model = SentenceTransformer("microsoft/mpnet-base")
loss = CachedMultipleNegativesRankingLoss(model, mini_batch_num_tokens=16384)

On the PR's Natural Questions benchmark, cached-loss training with flash attention and a tuned token budget dropped from 715 to 182 seconds (3.9x) versus the previous release, with unchanged quality. The engine also trims trailing padding from each mini-batch, which alone is worth about 26% throughput on the default padded path. The updated training efficiency documentation recommends the smallest token budget that saturates your GPU. mini_batch_size keeps working everywhere as before.

... (truncated)

Commits

• b2a9529 [fix] Name the lone suggested parameter in the mine_hard_negatives missing-ne...
• 00e696c Release v5.7.0
• b69081a [fix] Treat a 1-dimensional query embedding as a single query in semantic_sea...
• c9c4250 Keep queries aligned in semantic_search_seismic when a query matches nothing ...
• 6d331a5 [fix] Support mixed sparse/dense inputs in euclidean and manhattan similari...
• 68c8cda [fix] Reject num_negatives larger than the range_min/range_max window in mine...
• cfb6d43 [trainer] Fix eval DataLoader worker leak with dataloader_persistent_worker...
• f6d3a40 Expand the community window on ties with the threshold (#3900)
• 0b3fcbd Correct the corpus_precision values documented for semantic_search_faiss ...
• 62724e7 Drop FAISS padding placeholders from semantic_search_faiss results (#3887)
• Additional commits viewable in compare view

Updates numkong from 7.7.0 to 7.8.0

Release notes

Sourced from numkong's releases.

Release v7.8.0

Release: v7.8.0 [skip ci]

Minor

• Add: Native thread pools for Python and JavaScript bindings (0354de0b)

Patch

• Fix: Self-contained ISA headers (4383f444)
• Make: Unbreak Windows, RISC-V, and SDE jobs (b9570af1)
• Make: Pin cibuildwheel to 3.x instead of 4.x (bfa76018)
• Make: Unblock wheel builds on current cibuildwheel (96319ad1)
• Fix: Scatter cast writeback into a strided out (6500d1e3)
• Make: Wire WASI to CTest (648092d3)
• Improve: Allocation-free nk.astype for Python (d058a275)
• Make: CMake option to enable AddressSanitizer (#370) (b7a0d2f5)

Release v7.7.1

Release: v7.7.1 [skip ci]

Patch

• Improve: Report Umeyama scale as NaN for degenerate source (895742d8)
• Fix: Fast-math-safe float classification (b25666da)
• Fix: Accumulate i4 angular bias correction in i64 on Icelake (f59c9156)

Commits

• 1a26d12 Release: v7.8.0 [skip ci]
• 4383f44 Fix: Self-contained ISA headers
• b9570af Make: Unbreak Windows, RISC-V, and SDE jobs
• 0354de0 Add: Native thread pools for Python and JavaScript bindings
• bfa7601 Make: Pin cibuildwheel to 3.x instead of 4.x
• 96319ad Make: Unblock wheel builds on current cibuildwheel
• 6500d1e Fix: Scatter cast writeback into a strided out
• 648092d Make: Wire WASI to CTest
• d058a27 Improve: Allocation-free nk.astype for Python
• b7a0d2f Make: CMake option to enable AddressSanitizer (#370)
• Additional commits viewable in compare view

Updates setuptools to 84.0.0

Changelog

Sourced from setuptools's changelog.

v84.0.0

Features

• Newline-separated keywords and platformspypa/setuptools#4887old specification <https://peps.python.org/pep-0345/>_ separated items with spaces and the current one uses commas. (#4887)
• Extensionpypa/distutils#373#5022)
• The C compiler modules now emit log messages through their own compilers.C.* loggers instead of the distutils root logger, part of decoupling the compilers package from distutils. The logger names are normalized to a stable compilers.C.* prefix so they remain constant as the package migrates toward a standalone compilers.C distribution. (#5266)
• The C compilers gained a Compiler.call method -- a thin wrapper over subprocess.check_call (with macOS deployment-target env injection) that is the modern replacement for Compiler.spawn. The compilers no longer depend on distutils.spawn, distutils.dir_util, distutils.file_util, distutils._modified, or distutils.util.execute/split_quoted: the generic newer/newer_group and split_quoted helpers are vendored into the compilers package, and Compiler.mkpath/move_file/execute are implemented directly on the standard library (os.makedirs/shutil.move). The methods are retained for backward compatibility. (#5267)
• The compilers no longer depend on distutils.util, distutils.version, distutils.compat, or distutils._macos_compat. The platform-identification helpers (get_platform/get_host_platform/is_mingw) now live in distutils.compilers.platform.detect and the macOS deployment-target logic and compiler_fixup in distutils.compilers.platform.macos; CygwinCCompiler.gcc_version returns a packaging.version.Version. distutils.util re-exports the platform/macOS helpers from their new homes for backward compatibility rather than keeping duplicate copies. (sysconfig lookups still route through distutils pending its own decoupling.) (#5268)
• The compilers now read their build configuration from the standard library's sysconfig instead of distutils.sysconfig. Per-compiler customization -- previously distutils.sysconfig.customize_compiler -- has moved into Compiler.configure_system(): a no-op on the base class, with UnixCCompiler applying the compiler/flag/archiver settings CPython recorded in sysconfig (and the usual CC/CFLAGS/LDSHARED/… environment overrides). distutils.sysconfig.customize_compiler is retained as a thin wrapper that calls compiler.configure_system(). (#5269)

Bugfixes

• The MSVC linker now passes its arguments through a response file when the command line would exceed the Windows maximum length, fixing failures when linking a large number of objects. (#4177)
• The Cygwin and MinGW compilers now pass -O1 instead of a bare -O. The two are equivalent to GCC, but cc1 rejected the bare form when building 32-bit extensions with -m32. -- by :user:dchaudhari7177 (#4873)
• copy_filepypa/distutils#379#5079)
• Setuptools wheels no longer bundled the project's own test modules. -- by :user:itscloud0 (#5212)
• build_ext no longer fails when cross-compiling with a compiler other than MSVC (such as MinGW). Compiler now provides a no-op initialize()pypa/distutils#399

Improved Documentation

• Clarified what "correspond exactly to the directory structure" means in the packages section of the Package Discovery user guide. (#4109)
• Documented how bdist_wheel's py_limited_api option controls abi3 wheel tagging for extension modules -- by :user:Himanshuagrawal4 (#4741)

Deprecations and Removals

• Compiler.spawn is deprecated in favor of the new Compiler.call. call raises native subprocess exceptions; spawn remains as a shim that emits a DeprecationWarning and translates them to DistutilsExecError. The MSVC spawn compatibility shim for third-party monkeypatches predating the env argument (numpy.distutils before 1.19, per pypa/distutils#15) has been removed. distutils.spawn.spawn is likewise reduced to a thin wrapper around subprocess.check_call: it no longer resolves cmd[0] via shutil.which (subprocess searches PATH itself) nor injects MACOSX_DEPLOYMENT_TARGET (that now lives with the compilers, the only callers to which it applied). (#5267)
• Building an extension with a MACOSX_DEPLOYMENT_TARGET lower than the interpreter's configured value now raises compilers.errors.PlatformError instead of distutils.errors.DistutilsPlatformError (the macOS deployment-target check moved into the compilers package). CygwinCCompiler.gcc_version returns a packaging.version.Version rather than the removed distutils.version.LooseVersion. Completing the transition begun in pypa/distutils#246, UnixCCompiler.runtime_library_dir_option now returns the ["-Wl,--enable-new-dtags", "-Wl,-rpath,<dir>"] list directly for GNU ld rather than collapsing it into a single string, and the temporary distutils.compat.consolidate_linker_args shim has been removed. (#5268)
• The compilers now define their own exception vocabulary instead of borrowing distutils' framework errors. Language-agnostic exceptions (Error, UnknownFileType, and a new PlatformError) live at distutils.compilers.errors, leaving room for future compilers.<language> siblings; the C/C++-specific CompileError/LinkError/LibError/PreprocessError remain in distutils.compilers.C.errors. The compilers now raise compilers.errors.PlatformError where they previously raised distutils.errors.DistutilsPlatformError/DistutilsModuleError, and compilers._modified.newer raises the stdlib FileNotFoundError. distutils.errors keeps its own framework exceptions and re-exports the compiler ones (CCompilerError, CompileError, etc.) for backward compatibility; because CCompilerError is compilers.errors.Error, code catching it (as distutils' top-level handlers do) still catches the new PlatformError. (#5270)
• customize_compiler now asserts that the compiler-related config variables (CC, CXX, CFLAGS, etc.) resolve to strings, raising AssertionError if any are unexpectedly Nonepypa/distutils#363

v83.0.0

Features

• Require Python 3.10 or later.

... (truncated)

Commits

• 72e919a Merge pull request #5293 from pypa/bugfix/integration-pip-flit-backend
• 1b29701 Select the top-level pyproject.toml when reading build requirements
• bb1b381 Bump version: 83.0.0 → 84.0.0
• ee6fdd7 Sync with distutils @ e8eb87855 (#5292)
• 2a4a9e4 Merge remote-tracking branch 'origin/main' into distutils-e8eb87855
• cbd1195 Merge https://github.com/jaraco/skeleton
• bd3594e Merge pull request #5287 from Avasam/Configuring-lint.flake8-comprehensions.a...
• f02e90a Configure C408 to allow dict(a=1) rather than disabling it
• c55f52b Configuring lint.flake8-comprehensions.allow-dict-calls-with-keyword-argument...
• e9904b0 Match the distutils sdist base type for the user_options override
• Additional commits viewable in compare view

Dependabot will resolve any conflicts with this PR as long as you don't alter it yourself. You can also trigger a rebase manually by commenting @dependabot rebase.

---

Dependabot commands and options

You can trigger Dependabot actions by commenting on this PR:

• @dependabot rebase will rebase this PR
• @dependabot recreate will recreate this PR, overwriting any edits that have been made to it
• @dependabot show <dependency name> ignore conditions will show all of the ignore conditions of the specified dependency
• @dependabot ignore <dependency name> major version will close this group update PR and stop Dependabot creating any more for the specific dependency's major version (unless you unignore this specific dependency's major version or upgrade to it yourself)
• @dependabot ignore <dependency name> minor version will close this group update PR and stop Dependabot creating any more for the specific dependency's minor version (unless you unignore this specific dependency's minor version or upgrade to it yourself)
• @dependabot ignore <dependency name> will close this group update PR and stop Dependabot creating any more for the specific dependency (unless you unignore this specific dependency or upgrade to it yourself)
• @dependabot unignore <dependency name> will remove all of the ignore conditions of the specified dependency
• @dependabot unignore <dependency name> <ignore condition> will remove the ignore condition of the specified dependency and ignore conditions

[Copilot]

Pull request overview

This PR updates the project’s declared Python dependencies (runtime + build) in pyproject.toml to allow newer upstream releases, aligning the package metadata with the latest versions targeted by Dependabot.

Changes:

• Bump sentence-transformers requirement from >=5.4 to >=5.7.0.
• Bump numkong from ==7.7.0 to ==7.8.0.
• Bump build requirement setuptools from >=68 to >=84.0.0.

---

💡 Add a code-review agent skill for context-aware, tailored reviews. Learn more in the docs.