Remove the Horovod backend, horovod utils, Ray 1.12 compat shim,
and all Horovod-related tests. Ray Train is now the sole distributed
training backend.

Remove neuropod utils, export commands, and tests. Neuropod is no
longer maintained upstream.

Bump to Python 3.12, PyTorch 2.6, Ray 2.54, transformers 5.x,
torchaudio 2.x, NumPy 2.x, Dask 2026.1.2, MLflow 3.10.
Update Dockerfiles, CI workflow, pytest config, setup.py, and
requirements files accordingly.

# Conflicts:
#	.github/workflows/pytest.yml
#	README.md
#	docker/ludwig-gpu/Dockerfile
#	docker/ludwig-ray-gpu/Dockerfile
#	docker/ludwig-ray/Dockerfile
#	pytest.ini
#	requirements.txt
#	requirements_distributed.txt
#	requirements_hyperopt.txt
#	requirements_serve.txt
#	requirements_test.txt
#	requirements_viz.txt
#	setup.cfg
#	setup.py

…3.10

- Use F.scaled_dot_product_attention instead of custom matmul
- Replace torch.bmm with element-wise multiply in combiners
- Profiler API: start_us/duration_us -> start_ns/duration_ns
- NumPy: np.bool -> bool, np.int16 -> np.int32 for date overflow
- Pandas: fillna(method=) -> bfill()/ffill()
- torchaudio: sox_io_backend.load() -> torchaudio.load()
- matplotlib: fix _get_coord_info monkey-patch
- Various cleanup of deprecated APIs

- Remove output_attentions support from image encoders (SDPA default)
- Fix HuggingFace tokenizer dispatch for albert/roberta/distilbert
- Simplify tokenizer class hierarchy

- Replace DatasetPipeline with ray.data.Dataset (lazy execution)
- Train/eval functions save results to checkpoint (result.metrics
  is None without Checkpoint in Ray Train 2.54)
- Fix Dask-expr breaking changes: read-only divisions, PyArrow
  string defaults, concat API, repartition kwargs
- Update RayBackend, DaskEngine, datasource, sampler, predictor
- Use ActorPoolStrategy instead of compute="actors"

- tune.report() -> tune.report(metrics=..., checkpoint=...)
- tune.get_trial_id() -> tune.get_context().get_trial_id()
- local_dir -> storage_path, keep_checkpoints_num -> CheckpointConfig
- Adapt BOHB config space for ConfigSpace 1.x API
- Fix best_trial.logdir -> best_trial.local_path

- Rewrite log_model() to save locally then use mlflow.log_artifacts()
  (Model.log() in MLflow 3.x logs to model registry, not run artifacts)
- Add FileNotFoundError handling in _log_artifacts()
- Use setup_mlflow instead of removed mlflow_mixin

- Remove GBM/Horovod references from backward compatibility
- Update calibration utils for new API
- Clean up imports and remove dead code paths across api, automl,
  collect, datasets, evaluate, experiment, predict, preprocess, train

- Add device= to tensor creation across test files
- Use tiny-random HF models instead of @slow full models
- Update backward compatibility tests for removed GBM/Horovod
- Fix metric module tests, tokenizer tests, calibration tests
- Various test adjustments for PyTorch 2.6, transformers 5.x, Ray 2.54

- Remove GBM/Horovod test references
- Fix test_explain regex for Python 3.12 error messages
- xfail TorchScript audio/HF tests (upstream incompatibilities)
- Add importorskip for whylogs
- Fix class imbalance test ray fixtures
- Update visualization tests for removed formats

- Reduce default num_examples (100->25), image sizes (12x12->8x8)
- Remove redundant csv/parquet parametrizations
- Fix GPU sanity check to use ray.cluster_resources()
- Add num_gpus=0 to test cluster fixtures, reduce object_store_memory
- Widen eval metric tolerance for small datasets (rtol=0.1)
- Fix hyperopt ray backend: 1 train worker, cpu_resources_per_trial=1
- Use temp dirs for predict output (test isolation)

for more information, see https://pre-commit.ci

- Add setuptools to pip install (marshmallow-jsonschema needs pkg_resources
  which is no longer bundled with Python 3.12 by default)
- Fix syntax error in get_model_type_jsonschema: missing if before elif
  (leftover from GBM removal during rebase)

for more information, see https://pre-commit.ci

- Add docstrings to train_fn/eval_fn Ray Train workers
- Remove horovod-related comments from trainer
- Clean unused typing imports across data/backend/trainer modules
- Update Ray DatasetPipeline → modern ray.data.Dataset APIs

- Remove unused typing imports across all schema modules
- Remove hardcoded version references (v0.7, v0.8)
- Remove unused CATEGORY/NUMBER constant imports from checks.py
- Fix line-too-long in optimizers.py description field

- Remove commented-out TensorFlow attention and EmbedSparse code
- Clean unused typing imports from feature/encoder/module files
- Change augmentation log from info to debug level
- Fix typo: pipline → pipeline in image feature

- Fix broken import: merge_with_defaults moved to schema.model_types.utils
- Remove Python 3.7 cached_property TODO
- Remove outdated version warning TODOs
- Clean unused typing imports across api/hyperopt/utils modules

- Remove GBM/Horovod test references and model type configurations
- Update Ray backend test configs for Ray 2.54 APIs
- Remove dead TestDatasetWindowAutosizing class (old Ray 2.3 APIs)
- Clean unused typing imports across all test files
- Update conftest fixtures for modern Ray/PyTorch

- Update Python requirement: 3.8+ → 3.10+
- Remove version-specific feature references

cublasSgemmStridedBatched has known issues on certain GPU/driver
combinations (e.g., RTX 2080 Ti + CUDA 12.8 + driver 580.x) that
cause CUBLAS_STATUS_INVALID_VALUE for all batched 3D+ matmuls.
Switching to cublasLt resolves this system-wide.

The forced flash attention context managers cause issues when flash
attention is not available. The default SDPA dispatch handles kernel
selection automatically and correctly.

- Use dtype instead of deprecated torch_dtype kwarg
- Load models in float32 by default for numerical stability
- Merge rope_scaling with existing config to preserve rope_theta
- Rename rope_scaling 'type' field to 'rope_type' (transformers 5.x)
- Fix AdaLoRA pretrained config loading (total_step=None → 10000)

In transformers 5.x, batch_decode() on a 1D array treats it as a
single sequence. Decode each token individually to preserve per-token
prediction lists. Also fix idx2response to return a string instead
of a single-element list.

Route LLM models to the LLM ray trainers registry instead of the
ECD-only ray trainers registry.

- Skip quantization tests when bitsandbytes unavailable
- Update rope_scaling test to use rope_type key
- Update expected tokenizer file names for merged LoRA tests

- Add missing to_device() call in batch_collect_activations
- Use device="cpu" for torchscript tests (inputs are always CPU tensors)
- xfail audio torchscript test (upstream torchaudio incompatibility)

- test_visualization.py: use sys.executable instead of bare "python"
- test_cli.py: use full path to ludwig binary via sys.executable
- test_explain.py: shorten abstract class error regex for Python 3.12
- test_preprocessing.py: skip semantic_retrieval when sentence_transformers missing
- test_config_sampling.py: increase timeout to 600s
- test_encoder.py: relax parameter update assertion for frozen embeddings + dropout

- Use ray.train.torch.get_device() instead of get_torch_device() in
  train_fn/eval_fn to respect Ray Train's use_gpu setting
- Fix BatchInferModel to respect num_gpus=0 (force CPU when no GPUs)
- Fix BatchInferModel to use get_predictor_cls() for correct predictor
  class (LlmPredictor for LLM models instead of base Predictor)
- Fix LLM.to_device() to refresh curr_device from actual parameters
  before short-circuit check, preventing stale device tracking
- Fix LLM.generate() to move input_ids to model device
- Fix NoneTrainer to init_dist_strategy("local") for metric sync_context
- Fix device alignment in text_feature.py and llm_utils.py for targets
  vs predictions on different devices

- Fix convert_preds to use orient="list" for DataFrame.to_dict() so
  predictions can be indexed by position (test split has non-zero index)
- Relax ParallelCNN encoder test assertion: with max reduction and
  dropout=0.5, sparse gradients can legitimately result in zero updates