jsbaicenter/r1-1776-distill-llama-70b-FP8-Dynamic

The deepseek-ai/DeepSeek-R1-Distill-Llama-70B model quantized to fp8.

quantization using llm_compressor

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

# Define the model ID for the model you want to quantize
MODEL_ID = "perplexity-ai/r1-1776-distill-llama-70b"

# Load the model and tokenizer with appropriate parameters
model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID, 
    device_map="auto", 
    torch_dtype="auto",
    trust_remote_code=True,  # Add this to automatically trust remote code
    low_cpu_mem_usage=True,  # Help with memory issues during loading
    offload_folder="offload"  # Use disk offloading for large models
)

tokenizer = AutoTokenizer.from_pretrained(
    MODEL_ID,
    trust_remote_code=True  # Also need this for tokenizer
)

# Configure the quantization recipe
recipe = QuantizationModifier(targets="Linear", scheme="FP8_DYNAMIC", ignore=["lm_head"])

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

# Define the directory to save the quantized model
SAVE_DIR = MODEL_ID.split("/")[1] + "-FP8-Dynamic"

# Save the quantized model and tokenizer
model.save_pretrained(SAVE_DIR)
tokenizer.save_pretrained(SAVE_DIR)

print(f"Quantized model saved to {SAVE_DIR}")