docs: clarify Qwen3 model metadata path

[zcxGGmu]

Summary

• Clarify the Qwen3 AMX tutorial's <model_dir> requirement.
• Explain that the directory still needs tokenizer/config metadata, including tokenizer_config.json and the chat template, even when large safetensor weights are not required for GGUF serving.

Refs #1226

Test Plan

• python documentation guard check for the new Qwen3 model metadata guidance
• git diff --check HEAD~1..HEAD

[greptile-apps]

Greptile Summary

This PR adds a short clarifying paragraph to the Qwen3 AMX tutorial explaining the <model_dir> requirement: the directory must contain tokenizer and config metadata files even when serving from GGUF weights, because the large .safetensors weight files are not needed but the tokenizer files still are.

• Clarifies that config.json, tokenizer.json, and tokenizer_config.json (which embeds the chat template) must be present in <model_dir> for chat requests to resolve tokenizer.chat_template correctly.
• Addresses a known confusion point (issue Cannot use chat template functions because tokenizer.chat_template is not set and no template argument was passed! #1226) where users may skip downloading the model metadata since the weights come from GGUF.

Confidence Score: 5/5

Documentation-only change with no runtime impact; safe to merge.

The change adds a single clarifying paragraph to a markdown tutorial. The listed files (config.json, tokenizer.json, tokenizer_config.json) are the correct minimal set needed by AutoTokenizer, and the explanation of why .safetensors weights can be omitted is accurate for GGUF-based serving. No code, config, or behavior is affected.

Files Needing Attention: No files require special attention.

Important Files Changed

Filename	Overview
doc/en/AMX.md	Adds a clarifying paragraph explaining that <model_dir> must contain tokenizer/config metadata files (config.json, tokenizer.json, tokenizer_config.json) even when serving from GGUF weights, and that the large .safetensors files are not required.

Sequence Diagram

sequenceDiagram
    participant User
    participant KTransformers
    participant model_dir as model_dir (metadata only)
    participant gguf_dir as gguf_dir (weights)

    User->>KTransformers: start server (--model_path, --gguf_path)
    KTransformers->>model_dir: load config.json (architecture)
    KTransformers->>model_dir: load tokenizer.json + tokenizer_config.json (chat template)
    KTransformers->>gguf_dir: load BF16 GGUF weights
    Note over model_dir: .safetensors NOT required here
    User->>KTransformers: chat request
    KTransformers->>model_dir: resolve tokenizer.chat_template
    KTransformers-->>User: response

Loading

_{Reviews (2): Last reviewed commit: "[docs]: clarify Qwen3 metadata files" | Re-trigger Greptile}

[zcxGGmu]

Thanks for the review — I pushed 84ec1899d007 addressing the feedback.

Changes made:

• Folded the chat-template note into the tokenizer_config.json metadata description.

Validation:

• git diff --check — passed.
• Content guard — passed: Qwen3 metadata text folds chat template into tokenizer_config.json.
• Diff scan — passed: no secret/debug-marker patterns found.