Add multi-model support for Qwen 2.5 and Qwen3 Coder

[martysai]

Summary

Extends PR #8 to support multiple Qwen Coder models with model-specific configurations.

Changes

• Model Configuration Registry (src/training/models.py)

  • SamplingConfig and ModelConfig dataclasses
  • Pre-configured support for Qwen 2.5 Coder (7B, 14B, 32B) and Qwen3 Coder 30B-A3B (MoE)
  • Model-specific sampling parameters (Qwen3 uses temp=1.0, top_p=0.95)

• API Enhancements (src/training/serve.py)

  • Per-request model selection via model parameter
  • New /models endpoint to list available configurations
  • Health endpoint now reports active model info
  • Model-specific keep_alive settings (300s for MoE models)

• CLI Improvements (scripts/run_ollama.py)

  • --model-key for registry-based model selection
  • --list-models to display available models
  • Model-specific sampling parameters in requests

• Cross-Platform Support

  • .gitattributes for consistent line endings
  • Windows PowerShell/CMD examples in documentation

Available Models

Model Key	Architecture	Memory (Q4)	Context
qwen2.5-coder-32b	Dense	~18GB	32K
qwen2.5-coder-14b	Dense	~8GB	32K
qwen2.5-coder-7b	Dense	~4GB	32K
qwen3-coder-30b	MoE (3.3B active)	~18GB	256K

Test plan

• 20 new tests for models module
• 10 new tests for model selection in serve
• All 48 tests passing
• Manual testing with Ollama

🤖 Generated with Claude Code