mirror of
https://github.com/NVIDIA/TensorRT-LLM.git
synced 2026-01-25 05:02:59 +08:00
9dbc5b38ba
* Update TensorRT-LLM --------- Co-authored-by: Marks101 <markus.schnoes@gmx.de> Co-authored-by: lkm2835 <lkm2835@gmail.com>
48 lines
1.4 KiB
Python
48 lines
1.4 KiB
Python
import os
|
|
import sys
|
|
|
|
from tensorrt_llm.hlapi.llm import LLM, SamplingParams
|
|
from tensorrt_llm.hlapi.llm_utils import QuantAlgo, QuantConfig
|
|
|
|
sys.path.append(os.path.join(os.path.dirname(__file__), '..'))
|
|
from utils.util import skip_pre_ampere, skip_pre_hopper
|
|
|
|
try:
|
|
from .test_llm import llama_model_path
|
|
except ImportError:
|
|
from test_llm import llama_model_path
|
|
|
|
|
|
@skip_pre_ampere
|
|
def test_llm_int4_awq_quantization():
|
|
quant_config = QuantConfig()
|
|
quant_config.quant_algo = QuantAlgo.W4A16_AWQ
|
|
assert quant_config.quant_mode.has_any_quant()
|
|
|
|
llm = LLM(llama_model_path, quant_config=quant_config)
|
|
|
|
sampling_params = SamplingParams(max_new_tokens=6)
|
|
for output in llm.generate(["A B C"], sampling_params=sampling_params):
|
|
print(output)
|
|
assert output.outputs[0].text == "D E F G H I"
|
|
|
|
|
|
@skip_pre_hopper
|
|
def test_llm_fp8_quantization():
|
|
quant_config = QuantConfig()
|
|
quant_config.quant_algo = QuantAlgo.FP8
|
|
quant_config.kv_cache_quant_algo = QuantAlgo.FP8
|
|
|
|
assert quant_config.quant_mode.has_any_quant()
|
|
|
|
llm = LLM(llama_model_path, quant_config=quant_config)
|
|
sampling_params = SamplingParams(max_new_tokens=6)
|
|
for output in llm.generate(["A B C"], sampling_params=sampling_params):
|
|
print(output)
|
|
assert output.outputs[0].text == "D E F G H I"
|
|
|
|
|
|
if __name__ == "__main__":
|
|
test_llm_int4_awq_quantization()
|
|
test_llm_fp8_quantization()
|