communitylabs/cascadia-phi-4-mini-int4

Phi-4-mini Instruct — Cascadia int4 shards

Pre-exported Cascadia inference shards for microsoft/Phi-4-mini-instruct, so operators can deploy without running an export themselves.

These are OpenVINO IR artifacts for Cascadia's ov-runtime engine. They are not loadable by transformers — see Using these below.

Contents

Preset	Path	Stages	Quant	Size
1	int4/stages-1	1	int4	1.9 GB

int4/stages-1/
  pipeline_config.json     # model geometry: layers, heads, rope, arch tag
  stage_0/
    openvino_model.xml     # IR graph
    openvino_model.bin     # int4 weights
    stage_config.json      # layer_start, layer_end, has_embed, has_head
  tokenizer/               # tokenizer.json + configs

cascadia.json at the repo root is the machine-readable index (sizes, checksums, export version) that Cascadia's model registry reads.

Using these

hf download communitylabs/cascadia-phi-4-mini-int4 --local-dir ./phi-4-mini
cascadia worker --model ./phi-4-mini/int4/stages-1 --engine ov-runtime

The worker takes a local path, never a HuggingFace id — Cascadia workers never download or convert models at serve time.

Provenance

Exported with Cascadia's tools/export_shards.py, export_version: v5_canonical_inputs.

Source	microsoft/Phi-4-mini-instruct
Layers	32
Hidden size	3072
Attention heads	24 (8 KV)
Vocab	200064
RoPE theta	10000.0 (partial rotary factor 0.75)
Quantization	int4 (NNCF, group size 128)

Quantization is lossy. For anything quality-sensitive, benchmark against the source model rather than assuming parity.

License

MIT, inherited from microsoft/Phi-4-mini-instruct. These shards are a derivative work; the upstream license and its terms apply.