Optimize model loading performance

[DePasqualeOrg]

Parallel loading of tokenizer and weights

Tokenizer loading now runs concurrently with weight loading. For vision models, preprocessor_config.json is also loaded in parallel.

Single config.json read

config.json was being read from disk twice during model loading. Now it's read once and the data is reused for both the base config and model-specific config decoding.

Benchmark results

I've added model loading benchmarks in a separate test target that won't run in CI.

To show the total improvement of all my optimizations, I ran the benchmark with #33 in this repo, including pending optimizations #302, #303, and #304 in swift-transformers and the Hub client in offline mode to avoid a network call. After swift-transformers migrates to swift-huggingface for the Hub API, specify a model repo revision instead to avoid the network call:

Model	Before	This PR	Improvement
LLM (Qwen3-0.6B-4bit)	345 ms	297 ms	48 ms
VLM (Qwen2-VL-2B-Instruct-4bit)	465 ms	361 ms	104 ms

These are the results of the improvement in this PR with swift-transformers in its current, unoptimized state:

Model	Before	This PR	Improvement
LLM (Qwen3-0.6B-4bit)	3912 ms	3879 ms	33 ms
VLM (Qwen2-VL-2B-Instruct-4bit)	4509 ms	4480 ms	29 ms

The cumulative result of all my improvements is that model loading time goes from ~3900–4500 ms to ~300–360 ms.

[davidkoski]

Great idea! I will give this a look in the first week of Jan when I am back

[DePasqualeOrg]

This is now ready for review.

[davidkoski]

Changes look great, thank you!