TensorRT-LLMs/tests/hlapi/test_llm_quant.py
2024-04-16 19:40:08 +08:00

49 lines
1.5 KiB
Python

import os
import sys
from tensorrt_llm.hlapi.llm import LLM, ModelConfig
from tensorrt_llm.quantization import QuantAlgo
sys.path.append(os.path.join(os.path.dirname(__file__), '..'))
from utils.llm_data import llm_models_root
from utils.util import skip_pre_ampere, skip_pre_hopper
llama_model_path = str(llm_models_root() / "llama-models/llama-7b-hf")
@skip_pre_ampere
def test_llm_int4_awq_quantization():
config = ModelConfig(llama_model_path)
config.quant_config.quant_algo = QuantAlgo.W4A16_AWQ
assert config.quant_config.quant_mode.has_any_quant()
llm = LLM(config)
sampling_config = llm.get_default_sampling_config()
sampling_config.max_new_tokens = 6
for output in llm.generate(["A B C"], sampling_config=sampling_config):
print(output)
assert output.text == "<s> A B C D E F G H I"
@skip_pre_hopper
def test_llm_fp8_quantization():
config = ModelConfig(llama_model_path)
config.quant_config.quant_algo = QuantAlgo.FP8
config.quant_config.kv_cache_quant_algo = QuantAlgo.FP8
config.quant_config.exclude_modules = ["lm_head"]
assert config.quant_config.quant_mode.has_any_quant()
llm = LLM(config)
sampling_config = llm.get_default_sampling_config()
sampling_config.max_new_tokens = 6
for output in llm.generate(["A B C"], sampling_config=sampling_config):
print(output)
assert output.text == "<s> A B C D E F G H I"
if __name__ == "__main__":
test_llm_int4_awq_quantization()
test_llm_fp8_quantization()