fix(train): initialize DeepSpeed mode from top-level config

[YuCeong-May]

Summary

Type of change

• Bug fix
• Documentation
• Example or demo
• Runtime or deployment
• Benchmark or evaluation
• Model/training change

Validation

• python -m compileall funasr examples tests
• Docs or links checked
• Runtime/deployment command tested

User impact

Notes for reviewers

[LauraGPT]

Thanks for fixing the DeepSpeed mode propagation. There is one blocking compatibility issue before this can merge.

All current DeepSpeed training examples put both settings under train_conf, for example:

++train_conf.use_deepspeed=false
++train_conf.deepspeed_config=...

This patch also passes use_deepspeed= and deepspeed_config= explicitly and then expands **kwargs.get("train_conf"). Python therefore raises before entering Trainer:

TypeError: Trainer() got multiple values for keyword argument use_deepspeed

Please normalize the effective values once, with top-level values taking precedence and train_conf retained as a backwards-compatible fallback. Then pass a copied train_conf with the consumed distributed-mode keys removed, so use_ddp, use_fsdp, use_deepspeed, and deepspeed_config cannot be supplied twice.

Please also add focused regression coverage for:

1. the existing nested train_conf form;
2. the new top-level form overriding nested defaults;
3. rejecting simultaneous FSDP and DeepSpeed.

compileall passes because this is a call-time configuration failure, so a behavioral test is needed.

[LauraGPT]

Re-reviewed head 241dea6. The blocking duplicate-key path is resolved.

Verified:

• existing nested train_conf DeepSpeed settings remain supported
• top-level values override nested defaults
• consumed distributed keys are removed before Trainer receives **trainer_conf
• FSDP plus DeepSpeed is rejected
• DDP is disabled when either FSDP or DeepSpeed is selected
• focused tests: 3 passed
• additional precedence/invalid-combination matrix passed
• py_compile and diff checks passed
• synthetic merge with current main d410a56 is clean and limited to the intended two files

Approved.