fix: provide TensorBoard logging dir in finetuning script

[ElhamDevelopmentStudio]

Summary

Fix the TensorBoard logging crash in finetuning/sft_12hz.py by providing Accelerate with a valid log directory.

Fixes #286

Problem

The fine-tuning script initializes Accelerate like this:

Accelerator(
    gradient_accumulation_steps=4,
    mixed_precision="bf16",
    log_with="tensorboard",
)

With accelerate==1.12.0, this raises:

ValueError: Logging with `tensorboard` requires a `logging_dir` to be passed in.

That causes sft_12hz.py to fail before training starts.

Changes

• Add a new optional CLI argument: --logging_dir
• Default logging_dir to <output_model_path>/logs when not provided
• Create the log directory before constructing Accelerator
• Pass the directory to Accelerator(..., project_dir=logging_dir)

Why This Approach

• Keeps TensorBoard logging enabled instead of disabling it as a workaround
• Preserves current script behavior for existing users
• Gives users an explicit override when they want logs written elsewhere

Example

Existing command still works:

python sft_12hz.py \
  --init_model_path Qwen/Qwen3-TTS-12Hz-1.7B-Base \
  --output_model_path output \
  --train_jsonl train_with_codes.jsonl \
  --batch_size 32 \
  --lr 2e-6 \
  --num_epochs 10 \
  --speaker_name speaker_test

TensorBoard logs will now default to:

output/logs

Optional custom path:

python sft_12hz.py \
  --init_model_path Qwen/Qwen3-TTS-12Hz-1.7B-Base \
  --output_model_path output \
  --train_jsonl train_with_codes.jsonl \
  --logging_dir ./logs \
  --batch_size 32 \
  --lr 2e-6 \
  --num_epochs 10 \
  --speaker_name speaker_test

Validation

Validated in a clean temporary environment with accelerate==1.12.0 and tensorboard:

• Reproduced the original ValueError without project_dir
• Confirmed Accelerator(..., project_dir=...) succeeds
• Confirmed patched sft_12hz.py gets past Accelerator initialization
• Confirmed default and custom log directories are created as expected
• Ran python -m py_compile finetuning/sft_12hz.py