Skip to content

fp8-example.md

Latest commit

 

History

History
71 lines (55 loc) · 1.72 KB

fp8-example.md

File metadata and controls

71 lines (55 loc) · 1.72 KB

Llama4 FP8 Example

Code Walkthrough

Let's walk through the main steps of the quantization process:

  1. Load model
  2. Configure quantization algorithm and scheme
  3. Apply quantization
  4. Confirm generations of the quantized model look sane
  5. Save to disk in compressed-tensors format

1. Load Model

Load the model using AutoModelForCausalLM:

from compressed_tensors.offload import dispatch_model
from transformers import AutoModelForCausalLM, AutoTokenizer

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

MODEL_ID = "meta-llama/Llama-4-Scout-17B-16E-Instruct"

model = AutoModelForCausalLM.from_pretrained(MODEL_ID, dtype="auto")
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)

2. Configure the Quantization Algorithm and Scheme

recipe = QuantizationModifier(
    targets="Linear",
    scheme="FP8_BLOCK",
    ignore=[
        "re:.*lm_head",
        "re:.*self_attn",
        "re:.*router",
        "re:.*vision_model.*",
        "re:.*multi_modal_projector.*",
        "Llama4TextAttention",
    ],
)

3. Apply Quantization

oneshot(model=model, recipe=recipe)

4. Confirm Generations of the Quantized Model Look Sane

print("========== SAMPLE GENERATION ==============")
dispatch_model(model)
input_ids = tokenizer("Hello my name is", return_tensors="pt").input_ids.to(
    model.device
)
output = model.generate(input_ids, max_new_tokens=20)
print(tokenizer.decode(output[0]))
print("==========================================")

5. Save to Disk in Compressed-Tensors Format

SAVE_DIR = MODEL_ID.split("/")[1] + "-FP8-BLOCK"
model.save_pretrained(SAVE_DIR)
tokenizer.save_pretrained(SAVE_DIR)