mirror of
https://github.com/NVIDIA/TensorRT-LLM.git
synced 2026-01-31 08:11:27 +08:00
db4edea1e1
* Update TensorRT-LLM --------- Co-authored-by: Kota Tsuyuzaki <bloodeagle40234@gmail.com> Co-authored-by: Pzzzzz <hello-cd.plus@hotmail.com> Co-authored-by: Patrick Reiter Horn <patrick.horn@gmail.com>
52 lines
1.5 KiB
Python
52 lines
1.5 KiB
Python
import os
|
|
import sys
|
|
|
|
from tensorrt_llm.hlapi.llm import LLM, ModelConfig, SamplingParams
|
|
from tensorrt_llm.quantization import QuantAlgo
|
|
|
|
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():
|
|
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_params = SamplingParams(max_new_tokens=6)
|
|
for output in llm.generate(["A B C"], sampling_params=sampling_params):
|
|
print(output)
|
|
assert output.text == "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', 'router', 'vocab_embedding', 'position_embedding',
|
|
'block_embedding'
|
|
]
|
|
|
|
assert config.quant_config.quant_mode.has_any_quant()
|
|
|
|
llm = LLM(config)
|
|
sampling_params = SamplingParams(max_new_tokens=6)
|
|
for output in llm.generate(["A B C"], sampling_params=sampling_params):
|
|
print(output)
|
|
assert output.text == "D E F G H I"
|
|
|
|
|
|
if __name__ == "__main__":
|
|
test_llm_int4_awq_quantization()
|
|
test_llm_fp8_quantization()
|