Adjust/Improve pipeline to be compatible with more models

[yuqiannemo]

1. Add more dependencies to support more models
2. Protect sensitive info like openrouter API

[JerryLife]

Thanks for working on broader model compatibility here. Supporting more HF model families is a good direction, and the ModelLoader change is addressing a real misclassification issue.

That said, I don't think this is ready to merge as-is because the dependency changes would make the base package less portable.

Moving these packages into project.dependencies introduces hard platform requirements for every pip install llm-dna user, even if they never use those model families:

• mlx and mlx-lm are Apple Silicon-specific — this will break installation on Linux/Windows, which is where most of our users run the package
• optimum, compressed-tensors, etc. are architecture-specific/transitive runtime deps rather than core package requirements

I think the right direction is:

• Keep the base install minimal
• Move platform/model-specific packages into optional dependency groups (e.g. apple, quantization)
• Document which extras are needed for which model families

On ModelLoader, the current fix works for openai/gpt-oss, but it's brittle because it carves out exceptions from a broad openai/gpt- rule — every new HuggingFace model under the openai/ namespace would need another exclusion. Instead, I'd suggest tightening the OpenRouter prefixes to be more specific:

openrouter_prefixes = [
    "openrouter/",
    "openrouter:",
    "anthropic/claude-",
    "deepseek/",
    "openai/gpt-3",
    "openai/gpt-4",
    "google/gemini-",
    "x-ai/grok-",
    "cohere/command",
    "perplexity/",
]

This way openai/gpt-3.5-turbo and openai/gpt-4o still route to OpenRouter, but openai/gpt-oss-* naturally falls through to HuggingFace without needing any exclusion logic.

One smaller process note: the PR description looks incomplete (2. is empty). Please expand that before merge so the packaging/dependency rationale is clear.

Net: I like the goal of the PR, but the dependency strategy needs to change before this can be merged.

[yuqiannemo]

Thanks for working on broader model compatibility here. Supporting more HF model families is a good direction, and the ModelLoader change is addressing a real misclassification issue.

That said, I don't think this is ready to merge as-is because the dependency changes would make the base package less portable.

Moving these packages into project.dependencies introduces hard platform requirements for every pip install llm-dna user, even if they never use those model families:

• mlx and mlx-lm are Apple Silicon-specific — this will break installation on Linux/Windows, which is where most of our users run the package
• optimum, compressed-tensors, etc. are architecture-specific/transitive runtime deps rather than core package requirements

I think the right direction is:

• Keep the base install minimal
• Move platform/model-specific packages into optional dependency groups (e.g. apple, quantization)
• Document which extras are needed for which model families

On ModelLoader, the current fix works for openai/gpt-oss, but it's brittle because it carves out exceptions from a broad openai/gpt- rule — every new HuggingFace model under the openai/ namespace would need another exclusion. Instead, I'd suggest tightening the OpenRouter prefixes to be more specific:

openrouter_prefixes = [
    "openrouter/",
    "openrouter:",
    "anthropic/claude-",
    "deepseek/",
    "openai/gpt-3",
    "openai/gpt-4",
    "google/gemini-",
    "x-ai/grok-",
    "cohere/command",
    "perplexity/",
]

This way openai/gpt-3.5-turbo and openai/gpt-4o still route to OpenRouter, but openai/gpt-oss-* naturally falls through to HuggingFace without needing any exclusion logic.

One smaller process note: the PR description looks incomplete (2. is empty). Please expand that before merge so the packaging/dependency rationale is clear.

Net: I like the goal of the PR, but the dependency strategy needs to change before this can be merged.

Fixed