Provide a more flexible API to specify exclude / include file patterns

[imeoer]

Background: previously, in the model-csi-driver, we provided an HTTP API to mount models into pod:

curl --unix-socket $volume_dir/csi/csi.sock \
  -H "Content-Type: application/json" \
  -X POST http://localhost/api/v1/volumes/$volume_name/mounts \
  -d '{
    "mount_id": "$mount_id",
    "reference": "$reference",
    "exclude_model_weights": true
  }'

When exclude_model_weights = true, the mount skips fetching model weight files. This is intended for scenarios like rfork: in the target sglang instance, the model mount only needs to provide the non-weights type files before weight loading, while the weights are later loaded from a seed sglang instance via GPU-Direct RDMA.

Initially, exclude_model_weights = true worked by excluding files whose mediaType is application/vnd.cncf.model.weight.v1.*. However, for models like kimi k2, sglang loads tiktoken.model to initialize the tokenizer before loading weights, but since tiktoken.model is identified as a weight-type file from modelspec, it gets excluded, causing sglang to fail to boot.

One option is to improve modctl’s weight-file detection (it cannot rely only on file extensions and may need smarter heuristics), but this cannot fix already-built model images.

Another option is to adjust the model csi API parameters as follows:

1. Keep exclude_model_weights = true unchanged: it still excludes files with mediaType: application/vnd.cncf.model.weight.v1.*.
2. Add exclude_file_patterns: [] (.gitignore compatible syntax) to let users exclude/include specific files by filename patterns.

This enables usage like:

{
  exclude_model_weights: true,
  exclude_file_patterns: ["model.safetensors.index.json", "!tiktoken.model"]
}

or fully controlled by the user, such as:

{
  exclude_file_patterns: ["*.safetensors", "model.safetensors.index.json"]
}

Define the precedence as exclude_file_patterns > exclude_model_weights, so users can flexibly control which files are included or excluded during mounting, addressing on-demand loading issues like the kimi k2 case.

WDYT?

[aagumin]

Hi! Help wanted?