Ithanil/Llama-3_3-Nemotron-Super-49B-v1-FP8-Dynamic

Llama-3.3-Nemotron-Super-49B-v1-FP8-Dynamic

FP8-Dynamic quantization of https://huggingface.co/nvidia/Llama-3_3-Nemotron-Super-49B-v1

Created with llmcompressor using the following code:

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

MODEL_ID = "/models/Llama-3_3-Nemotron-Super-49B-v1"
model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID, device_map="auto", torch_dtype="auto", trust_remote_code=True,
)
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)

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

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

# Save the model.
SAVE_DIR = MODEL_ID + "-FP8-Dynamic"
model.save_pretrained(SAVE_DIR)
tokenizer.save_pretrained(SAVE_DIR)

To run it with vllm, use the latest version (0.8.2 as of now) and add the following PR: https://github.com/vllm-project/vllm/pull/15008

Make sure to read the original model's README for further guidance.