docs: document Python deps for SafeTensors to GGUF conversion

Add required packages (torch, safetensors, gguf, transformers) and
optional model-specific imports in internal/convert/data/README.md.
Summarize the same install step in the root README under Model Formats.

Made-with: Cursor

Author: ganisback

README.md

@@ -214,6 +214,14 @@ csghub-lite config set server_url https://my-private-csghub.example.com
 | GGUF | Yes | Yes (via llama.cpp) |
 | SafeTensors | Yes | Yes (auto-converted to GGUF) |
 
+SafeTensors checkpoints are converted once using the bundled llama.cpp `convert_hf_to_gguf.py` and **system Python** (PyTorch is not shipped inside the release binary). Install these packages once:
+
+```bash
+pip3 install torch safetensors gguf transformers
+```
+
+Use Python 3.10+ on `PATH` (Windows: `python` or `python3`). Some models may need extra packages (for example `sentencepiece`); see [`internal/convert/data/README.md`](internal/convert/data/README.md) for the full list and troubleshooting (`gguf` version mismatch, optional `CSGHUB_LITE_CONVERTER_URL`).
+
 ## Development
 
 ```bash

internal/convert/data/README.md

@@ -20,3 +20,33 @@ This file is embedded into the `csghub-lite` binary (`go:embed`) so SafeTensors
 3. Update **`BundledConverterLLamacppRef`** and the table above.
 
 Optional: set **`CSGHUB_LITE_CONVERTER_URL`** at runtime to a raw mirror URL instead of using the embedded file.
+
+## Python runtime dependencies
+
+`csghub-lite` materializes this script and runs it with a system **Python 3** interpreter. The binary pre-checks imports before conversion (`internal/convert/convert_python.go`); all of the following must be importable:
+
+| Package | Role |
+|---------|------|
+| `torch` | Load tensors / weights |
+| `safetensors` | Read `.safetensors` checkpoints |
+| `gguf` | Write GGUF; if the script fails with `AttributeError` involving `MODEL_ARCH` or `gguf`, upgrade: `pip3 install -U gguf` |
+| `transformers` | `AutoConfig`, tokenizers, etc. |
+
+One-time install (same as the CLI error text):
+
+```bash
+pip3 install torch safetensors gguf transformers
+```
+
+On macOS/Linux the tool tries `python3.13` … `python3.10`, then `python3` / `python`, plus common Homebrew paths. On Windows it looks for `python` / `python3` on `PATH`.
+
+### Optional / model-specific imports
+
+The upstream script may import extra packages for certain architectures. These are **not** verified up front; install if conversion fails with `ModuleNotFoundError`:
+
+| Package | When needed |
+|---------|----------------|
+| `numpy` | Used by the script; usually already installed with `torch` |
+| `sentencepiece` | Some SentencePiece-based tokenizers |
+| `huggingface_hub` | e.g. `snapshot_download` paths in the converter |
+| `mistral_common` | Some Mistral flows; upstream suggests `pip install mistral-common[image,audio]` |