Phi-4-mini-instruct-FP8-dynamic

Model Overview

• Model Architecture: Phi3ForCausalLM
  • Input: Text
  • Output: Text
• Model Optimizations:
  • Activation quantization: FP8
  • Weight quantization: FP8
• Intended Use Cases: The model is intended for broad multilingual commercial and research use. The model provides uses for general purpose AI systems and applications which require:

1. Memory/compute constrained environments.
2. Latency bound scenarios.
3. Math reasoning and logic.

• Release Date: 03/03/2025
• Version: 1.0
• Model Developers: Red Hat

Model Optimizations

This model was obtained by quantizing activation and weights of Phi-4-mini-instruct to FP8 data type. This optimization reduces the number of bits used to represent weights and activations from 16 to 8, reducing GPU memory requirements (by approximately 50%) and increasing matrix-multiply compute throughput (by approximately 2x). Weight quantization also reduces disk size requirements by approximately 50%.

Only weights and activations of the linear operators within transformers blocks are quantized. Weights are quantized with a symmetric static per-channel scheme, whereas activations are quantized with a symmetric dynamic per-token scheme. The llm-compressor library is used for quantization.

Deployment

This model can be deployed efficiently using the vLLM backend, as shown in the example below.

vllm serve RedHatAI/Phi-4-mini-instruct-FP8-dynamic --max_model_len 131072

from openai import OpenAI
# Set OpenAI's API key and API base to use vLLM's API server.
openai_api_key = "EMPTY"
openai_api_base = "http://localhost:8000/v1"

client = OpenAI(
    api_key=openai_api_key,
    base_url=openai_api_base,
)

generated_text = client.chat.completions.create(
    model="RedHatAI/Phi-4-mini-instruct-FP8-dynamic",
    messages=[
        {"role": "user", "content": "Give me a short introduction to large language model."},
    ],
)
print(generated_text.choices[0].message.content)

Creation

Creation details

This model was created with [llm-compressor](https://github.com/vllm-project/llm-compressor) by running the code snippet below.

from transformers import AutoModelForCausalLM, AutoTokenizer
from llmcompressor.modifiers.quantization import QuantizationModifier
from llmcompressor import oneshot

# Load model
model_stub = "microsoft/Phi-4-mini-instruct"
model_name = model_stub.split("/")[-1]

tokenizer = AutoTokenizer.from_pretrained(model_stub)

model = AutoModelForCausalLM.from_pretrained(
    model_stub,
    device_map="auto",
    torch_dtype="auto",
)

# Configure the quantization algorithm and scheme
recipe = QuantizationModifier(
    targets="Linear",
    scheme="FP8_dynamic",
    ignore=["lm_head"],
)

# Apply quantization
oneshot(
    model=model,
    recipe=recipe,
)

# Save to disk in compressed-tensors format
save_path = model_name + "-FP8-dynamic"
model.save_pretrained(save_path)
tokenizer.save_pretrained(save_path)
print(f"Model and tokenizer saved to: {save_path}")

Evaluation

The model was evaluated on the Mathh 500 benchmarks using lighteval, and on GSM8k-Platinum, MMLU CoT, MMLU-Pro, and IFEval using lm-evaluation-harness. In both cases vLLM is used as the backend

Evaluation commands

Start vLLM server

vllm serve RedHatAI/Phi-4-mini-instruct-FP8-dynamic --max_model_len 131072

lm-evaluation-harness

lm_eval --model local-chat-completions \
  --tasks gsm8k_platinum_cot_llama \
  --model_args "model=RedHatAI/Phi-4-mini-instruct-FP8-dynamic,max_length=131072,base_url=http://0.0.0.0:8000/v1/chat/completions,num_concurrent=128,max_retries=3,tokenized_requests=False,timeout=600,tokenizer_backend=None" \
  --apply_chat_template \
  --num_fewshot 5 \
  --fewshot_as_multiturn \
  --output_path gsm8k_platinum_phi4_mini_instruct_fp8_dynamic \
  --gen_kwargs "do_sample=False,temperature=0.0,max_gen_toks=16000"

lm_eval --model local-chat-completions \
  --tasks mmlu_cot_llama \
  --model_args "model=RedHatAI/Phi-4-mini-instruct-FP8-dynamic,max_length=131072,base_url=http://0.0.0.0:8000/v1/chat/completions,num_concurrent=128,max_retries=3,tokenized_requests=False,timeout=600,tokenizer_backend=None" \
  --apply_chat_template \
  --output_path mmlu_cot_phi4_mini_instruct_fp8_dynamic \
  --gen_kwargs "do_sample=False,temperature=0.0,max_gen_toks=16000"

lm_eval --model local-chat-completions \
  --tasks mmlu_pro_chat \
  --model_args "model=RedHatAI/Phi-4-mini-instruct-FP8-dynamic,max_length=131072,base_url=http://0.0.0.0:8000/v1/chat/completions,num_concurrent=128,max_retries=3,tokenized_requests=False,timeout=600,tokenizer_backend=None" \
  --apply_chat_template \
  --num_fewshot 5 \
  --fewshot_as_multiturn \
  --output_path mmlu_pro_phi4_mini_instruct_fp8_dynamic \
  --gen_kwargs "do_sample=False,temperature=0.0,max_gen_toks=16000"

lm_eval --model local-chat-completions \
  --tasks ifeval \
  --model_args "model=RedHatAI/Phi-4-mini-instruct-FP8-dynamic,max_length=131072,base_url=http://0.0.0.0:8000/v1/chat/completions,num_concurrent=128,max_retries=3,tokenized_requests=False,timeout=600,tokenizer_backend=None" \
  --apply_chat_template \
  --output_path ifeval_phi4_mini_instruct_fp8_dynamic \
  --gen_kwargs "do_sample=False,temperature=0.0,max_gen_toks=16000"

lighteval

litellm_config.yaml

model_parameters:
  provider: "hosted_vllm"
  model_name: "hosted_vllm/RedHatAI/Phi-4-mini-instruct-FP8-dynamic"
  base_url: "http://0.0.0.0:8000/v1"
  api_key: ""
  timeout: 600
  concurrent_requests: 128
  generation_parameters:
    temperature: 0.0
    max_new_tokens: 16000

lighteval endpoint litellm litellm_config.yaml \
    math_500|0 \
    --output-dir phi4_mini_instruct_fp8_dynamic \
    --save-details

Accuracy

Benchmark	Phi-4-mini-instruct	Phi-4-mini-instruct-FP8-dynamic (this model)	Recovery
Math 500	57.60	58.20	101.7%
GSM8k-Platinum	84.12	84.70	100.7%
MMLU CoT	67.01	66.97	99.9%
MMLU-Pro	46.75	45.60	97.5%