Add unified DeepSpeed finetune demo with MMLU/GSM8K benchmarks

[delock]

Summary

This PR adds a new standalone finetuning example under training/deepspeed_finetune_demo/ that demonstrates DeepSpeed's philosophy: use different training features via config files with no code change needed.

The example is extracted and extended from DeepSpeed-ZenFlow/finetuning.

Key Features

• Dataset registry with automatic format detection (Alpaca, Magicoder, MMLU, etc.) and sample_rate support for downsampling large datasets
• MMLU and GSM8K benchmark evaluation scripts (vLLM-based generation + accuracy scoring)
• Auto-detection of flash_attn availability (falls back gracefully if not installed)
• DistributedSampler for correct multi-GPU data sharding
• Checkpoint conversion (convert_ds_to_hf.py) from DeepSpeed format to HuggingFace, with AutoEP (expert parallelism) support
• 9 DeepSpeed config variants: ZeRO-2/3, Offload, ZenFlow, SuperOffload, Muon optimizer, AutoTP

Tested Configurations

Verified on Qwen2.5-0.5B with 2x RTX 4090 (AutoDL):

Benchmark	Baseline	After 10-step finetune
GSM8K	28.43%	—
MMLU	33.12%	38.01% (+4.89%)

Full pipeline validated: train → convert checkpoint → vLLM eval.

File Structure

training/deepspeed_finetune_demo/
├── README.md                    # Documentation
├── finetune_llama.py            # Main training script
├── finetune.sh / benchmark.sh / profile.sh
├── run_and_evaluate.sh          # End-to-end: train + convert + eval
├── convert_ds_to_hf.py          # DeepSpeed → HuggingFace checkpoint conversion
├── requirements.txt
├── configs/                     # 9 DeepSpeed config variants
│   ├── z2_config.json           # ZeRO Stage 2 with AdamW
│   ├── z3_config.json           # ZeRO Stage 3 with AdamW
│   ├── zo_config.json           # ZeRO Offload, stage 2
│   ├── ...
│   └── z2_muon.json             # ZeRO 2 with Muon optimizer
└── evaluate/
    ├── mmlu/                    # MMLU eval (gen + score)
    └── gsm8k/                   # GSM8K eval (gen + score)

Dataset Support

Dataset	Format	Use Case
tatsu-lab/alpaca	Alpaca	General instruction tuning
sahil2801/CodeAlpaca-20k	Alpaca	Code instruction tuning
meta-math/MetaMathQA	Alpaca (sample_rate=0.1)	Math reasoning (GSM8K downstream)
cais/mmlu	MMLU MCQ	Knowledge (MMLU downstream)
Any HF dataset	Auto-detect	Magicoder or Alpaca format