JamAndTeaStudios/DeepSeek-R1-Distill-Llama-70B-FP8-Dynamic

Model Overview

• Model Optimizations:
  • Weight quantization: FP8
  • Activation quantization: FP8
• Release Date: 1/28/2025

Quantized version of deepseek-ai/DeepSeek-R1-Distill-Llama-70B to FP8 data type, ready for inference with SGLang >= 0.3 or vLLM >= 0.5.2. This optimization reduces the number of bits per parameter from 16 to 8, reducing the disk size and GPU memory requirements by approximately 50%. Only the weights and activations of the linear operators within transformers blocks are quantized.

License

We notice that deepseek-ai/DeepSeek-R1-Distill-Llama-70B is licensed under MIT, however the original LLama model is licensed under llama-3. We are adopting llama-3 for this model to be consistent.

Deployment

Use with SGLang

python -m sglang.launch_server --model-path JamAndTeaStudios/DeepSeek-R1-Distill-Llama-70B-FP8-Dynamic \
--port 30000 --host 0.0.0.0

Creation

This model was created with llm-compressor by running the code snippet below.

Model Creation Code

from transformers import AutoModelForCausalLM, AutoTokenizer

from llmcompressor.modifiers.quantization import QuantizationModifier
from llmcompressor.transformers import oneshot

MODEL_ID = "google/gemma-2-27b-it"

# 1) Load model.
model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID, device_map="auto", torch_dtype="auto"
)
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)

# 2) Configure the quantization algorithm and scheme.
# In this case, we:
#   * quantize the weights to fp8 with per channel via ptq
#   * quantize the activations to fp8 with dynamic per token
recipe = QuantizationModifier(
    targets="Linear", scheme="FP8_DYNAMIC", ignore=["lm_head"]
)

# 3) Apply quantization and save in compressed-tensors format.
OUTPUT_DIR = MODEL_ID.split("/")[1] + "-FP8-Dynamic"
oneshot(
    model=model,
    recipe=recipe,
    tokenizer=tokenizer,
    output_dir=OUTPUT_DIR,
)

# Confirm generations of the quantized model look sane.
print("========== SAMPLE GENERATION ==============")
input_ids = tokenizer("Hello my name is", return_tensors="pt").input_ids.to("cuda")
output = model.generate(input_ids, max_new_tokens=20)
print(tokenizer.decode(output[0]))
print("==========================================")

Evaluation

TBA

Play Retail Mage

Retail Mage (Steam) is an immersive sim that uses online LLM inference in almost all features in the gameplay!

Reviews

“A true to life experience detailing how customer service really works.” 10/10 – kpolupo

“I enjoyed how many things were flammable in the store.” 5/5 – mr_srsbsns

“I've only known that talking little crow plushie in MageMart for a day and a half but if anything happened to him I would petrify everyone in this store and then myself.” 7/7 – neondenki