Phi-4-mini-instruct-W4A16

Phi-4-mini-instruct — W4A16 (GPTQ, int4 weight-only)

4-bit weight-only (W4A16) quantization of microsoft/Phi-4-mini-instruct, produced with LLM Compressor's GPTQModifier. All Linear layers are quantized except lm_head, which is kept at full precision.

Quantization method

Recipe used (recipe.yaml, included in this repo):

default_stage:
  default_modifiers:
    GPTQModifier:
      targets: [Linear]
      ignore: [lm_head]
      scheme: W4A16
      block_size: 128
      dampening_frac: 0.01
      actorder: static
      requires_calibration_data: true

Calibration set

256 packed sequences × 2048 tokens = 524,288 calibration tokens, sampled (seed=42) from a mixed Korean/English instruction + function-calling corpus, targeting the following source composition:

• Raw pool before filtering/dedup: 49,268 examples (49,179 after removing 88 duplicates)
• Quota-sampled for calibration: 1,311 examples → 1,225 packed into the final 256×2048 blocks
• Tool-calling trajectories in the pool: 8,995 (2,854 multi-tool)
• Contamination check: gsm8k-ko: train split only used; no BFCL/AgentDojo/tau-bench/SafeCommit-eval sources included

safecommit_synth is unreleased in-house synthetic data from the SafeCommit project, not a public HF dataset.

How to run

vLLM (recommended — required for the compressed-tensors W4A16 kernels used here)

vllm serve minjaechoi/Phi-4-mini-instruct-W4A16 --served-model-name phi4-mini

curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "phi4-mini",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

Transformers (requires the compressed-tensors package for int4 dequant kernels)

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "minjaechoi/Phi-4-mini-instruct-W4A16"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")

msgs = [{"role": "user", "content": "Hello!"}]
inputs = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(inputs, max_new_tokens=256)
print(tok.decode(out[0], skip_special_tokens=True))

Benchmarks

Evaluated on this W4A16 checkpoint (not compared against an FP16 baseline run in this project yet):

• AgentDojo (banking suite, utility-only, 16 tasks): 5/16 tasks passed (31.2%) — utility-only scoring, no prompt-injection attacks included.
• BFCL v4: attempted, but the evaluation harness used had a request-routing bug (--local-model-path was not substituted into the API request, so every call 404'd against the vLLM server) — the resulting scores are invalid and are intentionally not published here. Will be updated after a corrected re-run.

Files

• model.safetensors — quantized weights (compressed-tensors pack-quantized format)
• config.json — includes the quantization_config (compressed-tensors) needed by vLLM/transformers to load this checkpoint
• recipe.yaml — the exact LLM Compressor recipe used to produce this checkpoint
• tokenizer.json, tokenizer_config.json, chat_template.jinja — tokenizer/chat template, copied unmodified from the base model
• LICENSE — base model license, included per its terms

License

This checkpoint is a derivative of microsoft/Phi-4-mini-instruct and is distributed under the same license (mit, see LICENSE in this repo). No additional restrictions are added beyond the base model's license.