feat: set vLLM engine_args defaults for gpu_memory_utilization

[ansh-info]

• Summary

Issue: #472

• Prevent vLLM from computing a zero KV-cache budget by aligning engine defaults with Unsloth init defaults.
• Set gpu_memory_utilization and max_model_len in engine_args to the corresponding Unsloth init_args values so large-GPU setups don’t stall.

Motivation

• Training with Qwen2.5-14B + Unsloth/vLLM on H100 was stalling at 0/steps; logs showed “vLLM’s KV Cache can use up to 0.0 GB” despite ample VRAM. Missing engine defaults led vLLM to size the cache to zero.

Details

• In get_model_config, initialize vLLM engine_args with:
  • gpu_memory_utilization = init_args["gpu_memory_utilization"]
  • max_model_len = init_args["max_seq_length"]
• Still allow user overrides via _internal_config["engine_args"].