Generation with Quantization

Source https://github.com/NVIDIA/TensorRT-LLM/tree/main/examples/llm-api/llm_quantization.py.

 1### Generation with Quantization
 2import logging
 3
 4import torch
 5
 6from tensorrt_llm import LLM, SamplingParams
 7from tensorrt_llm.llmapi import CalibConfig, QuantAlgo, QuantConfig
 8
 9major, minor = torch.cuda.get_device_capability()
10enable_fp8 = major > 8 or (major == 8 and minor >= 9)
11enable_nvfp4 = major >= 10
12
13quant_and_calib_configs = []
14
15if not enable_nvfp4:
16    # Example 1: Specify int4 AWQ quantization to QuantConfig.
17    # We can skip specifying CalibConfig or leave a None as the default value.
18    quant_and_calib_configs.append(
19        (QuantConfig(quant_algo=QuantAlgo.W4A16_AWQ), None))
20
21if enable_fp8:
22    # Example 2: Specify FP8 quantization to QuantConfig.
23    # We can create a CalibConfig to specify the calibration dataset and other details.
24    # Note that the calibration dataset could be either HF dataset name or a path to local HF dataset.
25    quant_and_calib_configs.append(
26        (QuantConfig(quant_algo=QuantAlgo.FP8,
27                     kv_cache_quant_algo=QuantAlgo.FP8),
28         CalibConfig(calib_dataset='cnn_dailymail',
29                     calib_batches=256,
30                     calib_max_seq_length=256)))
31else:
32    logging.error(
33        "FP8 quantization only works on post-ada GPUs. Skipped in the example.")
34
35if enable_nvfp4:
36    # Example 3: Specify NVFP4 quantization to QuantConfig.
37    quant_and_calib_configs.append(
38        (QuantConfig(quant_algo=QuantAlgo.NVFP4,
39                     kv_cache_quant_algo=QuantAlgo.FP8),
40         CalibConfig(calib_dataset='cnn_dailymail',
41                     calib_batches=256,
42                     calib_max_seq_length=256)))
43else:
44    logging.error(
45        "NVFP4 quantization only works on Blackwell. Skipped in the example.")
46
47
48def main():
49
50    for quant_config, calib_config in quant_and_calib_configs:
51        # The built-in end-to-end quantization is triggered according to the passed quant_config.
52        llm = LLM(model="TinyLlama/TinyLlama-1.1B-Chat-v1.0",
53                  quant_config=quant_config,
54                  calib_config=calib_config)
55
56        # Sample prompts.
57        prompts = [
58            "Hello, my name is",
59            "The president of the United States is",
60            "The capital of France is",
61            "The future of AI is",
62        ]
63
64        # Create a sampling params.
65        sampling_params = SamplingParams(temperature=0.8, top_p=0.95)
66
67        for output in llm.generate(prompts, sampling_params):
68            print(
69                f"Prompt: {output.prompt!r}, Generated text: {output.outputs[0].text!r}"
70            )
71
72    # Got output like
73    # Prompt: 'Hello, my name is', Generated text: 'Jane Smith. I am a resident of the city. Can you tell me more about the public services provided in the area?'
74    # Prompt: 'The president of the United States is', Generated text: 'considered the head of state, and the vice president of the United States is considered the head of state. President and Vice President of the United States (US)'
75    # Prompt: 'The capital of France is', Generated text: 'located in Paris, France. The population of Paris, France, is estimated to be 2 million. France is home to many famous artists, including Picasso'
76    # Prompt: 'The future of AI is', Generated text: 'an open and collaborative project. The project is an ongoing effort, and we invite participation from members of the community.\n\nOur community is'
77
78
79if __name__ == '__main__':
80    main()