[DeepSpeed] Fix evaluate()/predict() before train()

[roycho96]

What does this PR do?

Calling trainer.evaluate() before trainer.train() with DeepSpeed is broken in three ways:

1. ZeRO-3 stale state crash: evaluate() creates an inference engine. train() starts with accelerator.free_memory() which destroys it and sets deepspeed_engine_wrapped = None, but Trainer's model_wrapped/deepspeed still hold stale refs, causing AttributeError: 'NoneType' has no attribute 'backward'.

2. ZeRO-3 shared config mutation: deepspeed_init(inference=True) mutates the shared config in-place: trainer_config_finalize(num_training_steps=0) bakes scheduler "auto" values to 0 (causing mismatch on subsequent train()), and del_config_sub_tree("optimizer") removes the optimizer config (fallback to HF optimizer).

3. ZeRO-1/2 evaluate blocked: evaluation_loop() unconditionally calls deepspeed_init(inference=True) which raises ValueError for non-ZeRO-3 stages. ZeRO-1/2 have full parameters on each GPU and don't need an inference engine.

evaluate() before train() is a valid pattern. For example, baseline metric collection, conditional training based on eval results, or data pipeline validation. eval_on_start=True covers some cases but doesn't help when the user needs eval metrics before deciding whether to train.

Fix

• Reset stale Trainer refs at start of _prepare_for_training() when deepspeed_engine_wrapped is already None but model_wrapped still points to a destroyed engine. Follows existing pattern in _update_auto_batch_size().

• In evaluation_loop(), gate deepspeed_init(inference=True) on is_zero3() only. For ZeRO-1/2, use prepare_model(evaluation_mode=True) to bypass DS engine creation entirely. Protect shared config with backup/restore in try/finally.

eval_on_start=True is unaffected.

Related issues

• try eval befor train gives ValueError with deepspeed Zero2 #30205
• deepspeed multi-gpu inference #26874

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline, Pull Request section?
• Was this discussed/approved via a Github issue or the forum? Please add a link to it if that's the case.
• Did you make sure to update the documentation with your changes?
• Did you write any new necessary tests?

Who can review?

@SunMarc @3outeille

[github-actions]

View the CircleCI Test Summary for this PR:

https://huggingface.co/spaces/transformers-community/circle-ci-viz?pr=44889&sha=c1f161