BLD: require peft>=0.18.0 for transformers 5.x compatibility

[OliverBryant]

Problem

Launching any rerank model on the sentence_transformers engine fails at load time with:

ImportError: Failed to import module 'sentence-transformers': cannot import name
'HybridCache' from 'transformers' (.../transformers/__init__.py)

Please make sure 'sentence-transformers' is installed. You can install it by
`pip install sentence-transformers`

The suggested remedy is misleading — sentence-transformers is installed and is not the problem.

Root cause

peft<=0.17.1 imports HybridCache at module level in peft/peft_model.py:

from transformers import Cache, DynamicCache, EncoderDecoderCache, HybridCache, PreTrainedModel

HybridCache was removed in transformers 5.0.0 (present in 4.57.1, gone from 5.0.0 onward). Because peft declares an unbounded transformers requirement, pip resolves the incompatible pair without any warning, so images build fine and the failure only appears at runtime.

The failing import chain:

xinference/model/rerank/sentence_transformers/core.py  ->  import sentence_transformers
  -> sentence_transformers/base/model_card.py          ->  from transformers import TrainerCallback
  -> transformers/trainer_utils.py                     ->  from peft import PeftMixedModel, PeftModel
  -> peft/peft_model.py                                ->  ImportError: cannot import name 'HybridCache'

core.py catches the ImportError and re-raises it as "Failed to import module 'sentence-transformers'", which hides the real cause. GGUF LLMs never touch this chain, so on an affected image LLMs work while rerank/embedding models appear broken — which makes this easy to misdiagnose as a rerank bug.

The <=0.17.1 upper bound was introduced in #4249 while transformers was still 4.x. It became self-inconsistent once xinference/deploy/docker/requirements-runtime.txt moved to transformers==5.13.1: the GPU image installs xinference under that constraint, and the CPU image pins both in requirements_cpu-ml.txt (transformers>=4.53.3 + peft<=0.17.1).

peft 0.18.0 moved the HybridCache import inside the function that needs it, so it no longer breaks at import time.

Fix

Require peft>=0.18.0 in the three places the pin appears:

• pyproject.toml — also covers the GPU image, which installs xinference with --constraint requirements-runtime.txt
• xinference/deploy/docker/requirements_cpu/requirements_cpu-ml.txt
• .github/workflows/python.yaml

Comments were added at the first two sites recording the transformers-5.x constraint, so the bound is not lowered again without checking the shipped transformers major version.

Verification

In a clean venv:

• transformers==5.13.1 + peft==0.17.1 installs with no pip warning and reproduces the exact ImportError from the report.
• With peft>=0.18.0 (resolved 0.20.0) against the same transformers==5.13.1: import sentence_transformers, from sentence_transformers.cross_encoder import CrossEncoder, a full CrossEncoder.predict rerank call, and from peft import PeftModel, PeftMixedModel (used by llm/transformers/core.py) all succeed.
• pre-commit run --files <changed files> passes; pyproject.toml parses and resolves to peft>=0.18.0.

Not run locally: rerank tests that require model downloads or GPU, and a full image rebuild — both too costly in this environment. test_auto_detect_type fails on my machine both with and without this change, due to a pre-existing local numpy/scipy ABI mismatch unrelated to the patch.

[gemini-code-assist]

Code Review

This pull request updates the dependency constraint for peft from <=0.17.1 to >=0.18.0 in both pyproject.toml and requirements_cpu-ml.txt to avoid compatibility issues with transformers 5.0.0. Feedback points out that the corresponding update to .github/workflows/python.yaml is missing and should be added to maintain consistency across all configuration files.

[qinxuye]

LGTM