fix: restore torch_dtype when loading layerwise reranker model

[St4r4x]

Root cause

Closes #1564.

In FlagEmbedding/finetune/reranker/decoder_only/layerwise/load_model.py, both calls to LayerWiseMiniCPMForCausalLM.from_pretrained() in get_model() had their torch_dtype argument commented out, while attn_implementation="flash_attention_2" is still passed unconditionally whenever model_args.use_flash_attn is set:

model = LayerWiseMiniCPMForCausalLM.from_pretrained(
    model_args.model_name_or_path,
    trust_remote_code=model_args.trust_remote_code,
    # torch_dtype=torch.float16 if training_args.fp16 else torch.bfloat16,
    attn_implementation = "flash_attention_2" if model_args.use_flash_attn else None,
    ...
)

Without torch_dtype, the model loads in fp32 by default. modeling_minicpm_reranker.py's MiniCPMFlashAttention2.forward() only casts hidden states away from fp32 by falling back to self.q_proj.weight.dtype — which is also fp32 in this case, so the cast is a no-op and the flash-attn kernel then receives fp32 tensors. Flash Attention 2 only supports fp16/bf16 inputs, so this crashes as soon as attention runs during finetuning with --use_flash_attn.

Note the commented-out line already referenced training_args.fp16, but get_model() only ever received model_args as a parameter — training_args was never in scope, so simply uncommenting it would raise a NameError.

Fix

• Thread training_args through get_model() (and its single caller in runner.py).
• Restore torch_dtype=torch.float16 if training_args.fp16 else torch.bfloat16 in both from_pretrained() calls, matching the dtype-selection idiom already used by this repo's inference-side loaders (e.g. FlagEmbedding/inference/reranker/decoder_only/layerwise.py).

Also restored two pre-existing unused imports (AutoModelForCausalLM in load_model.py, os in runner.py) that a local lint auto-fix on my end stripped as an incidental side effect while editing these files — flagging this explicitly for transparency, in case a maintainer would rather those be removed in a separate/dedicated cleanup instead.

Validation

I don't have a GPU + real model weights handy to run this specific finetuning path end-to-end, so I validated by code inspection instead:

• Confirmed AbsRerankerTrainingArguments subclasses transformers.TrainingArguments, so .fp16/.bf16 are valid attributes once training_args is passed in.
• Confirmed via git blame that the torch_dtype line has been commented out since this file's original commit, and that a later commit (19edba7b) switched use_flash_attention_2=True/False to attn_implementation="flash_attention_2" without addressing the dtype issue.
• Confirmed get_model() in this file has exactly one caller (runner.py), so the added training_args parameter doesn't break other call sites.
• Traced MiniCPMFlashAttention2.forward() to confirm the fp32-input crash mechanism described above.
• The exact same commented-out torch_dtype pattern also exists in the sibling loaders finetune/reranker/decoder_only/base/load_model.py, finetune/embedder/decoder_only/base/load_model.py, and finetune/embedder/decoder_only/icl/load_model.py — I scoped this PR to the file named in 这行注释掉会导致微调模型时flashattention2报错(fla2只支持fp16、bf16，关还关不掉)，赶紧修复了吧。 #1564 only, to keep the change minimal and focused.