Fixes #15949.

Problem

When importing a safetensors model with an unsupported architecture, Ollama copies all blobs into a staging directory first and only then attempts conversion — surfacing the unsupported architecture error after potentially many gigabytes of unnecessary I/O:

converting model
Error: unsupported architecture "MistralForCausalLM"

A user reporting this issue had already copied 77 GB of safetensors files before getting the error.

Fix

convert.CheckArchitecture(configPath string) error reads config.json from the given path, parses ModelParameters.Architectures, and returns a descriptive error for any architecture that is not handled by LoadModelMetadata's switch statement.

convertFromSafetensors in server/create.go calls this before os.MkdirTemp and before any blob links are created. If config.json is missing, unreadable, or has no architectures field, CheckArchitecture returns nil so that ConvertModel can surface those errors with full context as before.

Tests

TestCheckArchitecture in convert/convert_test.go covers:

• Supported architectures (LlamaForCausalLM, Gemma3ForCausalLM) → no error
• Known-unsupported architecture (MistralForCausalLM) → error with architecture name
• Unknown/future architecture → error
• Empty architectures list → nil (let ConvertModel handle)
• Missing config.json → nil (let ConvertModel handle)
• Invalid JSON → nil (let ConvertModel handle)