Phi-4 mini

pllm.phi4_mini.v1 recognizes the phi3 remote-code configuration used by microsoft/Phi-4-mini-instruct revision cfbefacb99257ffa30c83adab238a50856ac3083. The raw configuration SHA-256 is ac65d86061d3d0d704ee2511fd0eb8713ef19eb6eedba17c3080a4165d5b933b.

The adapter emits fused QKV and gate/up projections as linear operations followed by explicit slices. LongRoPE records original context, short and long factor vectors, scaling ratio, and partial rotary width. The tied output head references the token embedding matrix.

The upstream model name and adapter identity are Phi-4 mini even though the checked configuration discriminator is phi3. This page does not claim checkpoint import or runtime execution.

Python SDK example

from pathlib import Path

from pllm import lower_model

fixture = Path("crates/pllm-models/tests/fixtures/Phi-4-mini-instruct-cfbefac-config.json")
plan = lower_model(fixture.read_bytes(), batch=1, max_input_tokens=16, max_new_tokens=4)
operations = {operation["id"]: operation for operation in plan.prefill["operations"]}
assert operations["layer.0.qkv_linear"]["output_shape"] == (1, 16, 5120)
assert operations["layer.0.rope_q"]["attributes"]["rope_type"] == "longrope"

API: Python SDK objects and signatures

Run this example from a PLLM source checkout; package installations do not include the pinned fixture. Lowering does not import or execute its checkpoint.