Auto-extract context size from GGUF and Safetensors models at package time

[Copilot]

Context size is only visible after running inference via docker model requests | jq '.[0].config'. This change extracts it directly from model files during packaging, making it immediately available in docker model ls.

Changes

GGUF models (pkg/distribution/internal/gguf/create.go)

• Extract llama.context_length from metadata and populate Config.ContextSize
• Reuse metadata extraction to avoid duplicate parsing

Safetensors models (pkg/distribution/internal/safetensors/create.go)

• Parse config.json for common context size fields: max_position_embeddings, n_positions, max_length, n_ctx
• Gracefully handle missing/invalid configs

Tests (pkg/distribution/internal/safetensors/context_test.go)

• Coverage for all supported field names and error cases

Example

// GGUF: extract once and reuse
metadata := extractGGUFMetadata(&gguf.Header)
if contextLengthStr, ok := metadata["llama.context_length"]; ok {
    if parsed, err := strconv.ParseUint(contextLengthStr, 10, 64); err == nil {
        cfg.ContextSize = &parsed
    }
}

// Safetensors: parse config.json if present
configPath := filepath.Join(filepath.Dir(paths[0]), "config.json")
if contextSize := extractContextSizeFromConfig(configPath); contextSize > 0 {
    cfg.ContextSize = &contextSize
}

Backward compatible: models without context size metadata continue to work as before.

Original prompt

---

This section details on the original issue you should resolve

<issue_title>Make it easy to see/verify the a model configuration</issue_title>
<issue_description>Currently, unless the model has the context size embedded in the OCI artifact so it's displayed in docker model ls , you can only see the full config after running at least one inference using:

docker model requests ai/gemma3-qat | jq '.[0].config'
{
  "context-size": 1024
}

This is not friendly and should be improved.</issue_description>

Comments on the Issue (you are @copilot in this section)

@ericcurtin if it's not present parse the gguf or safetensors file for the value I guess, to be implemented.

Might make sense to solve this at model package time. If no value present make package parse the value and add it anyway.

Both techniques work for us I guess.</comment_new>

• Fixes Make it easy to see/verify the a model configuration docker/model-runner#398

---

✨ Let Copilot coding agent set things up for you — coding agent works faster and does higher quality work when set up for your repo.