| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Validates the loading of a model saved with the sharded_state format. |
| 4 | This script demonstrates how to load a model that was previously saved |
| 5 | using save_sharded_state_offline.py and validates it by running inference. |
| 6 | Example usage: |
| 7 | (First need to save a sharded_state mode) |
| 8 | |
| 9 | python save_sharded_state_offline.py \ |
| 10 | --model /path/to/load \ |
| 11 | --tensor-parallel-size 8 \ |
| 12 | --output /path/to/save/sharded/model |
| 13 | |
| 14 | python load_sharded_state_offline.py \ |
| 15 | --model /path/to/saved/sharded/model \ |
| 16 | --load-format sharded_state \ |
| 17 | --tensor-parallel-size 8 \ |
| 18 | --prompt "Hello, my name is" \ |
| 19 | --max-tokens 50 |
| 20 | """ |
| 21 | |
| 22 | from vllm import LLM, EngineArgs, SamplingParams |
| 23 | from vllm.utils.argparse_utils import FlexibleArgumentParser |
| 24 | |
| 25 | |
| 26 | def parse_args(): |
| 27 | parser = FlexibleArgumentParser() |
| 28 | # Add engine arguments |
| 29 | EngineArgs.add_cli_args(parser) |
| 30 | |
| 31 | # Override default load_format for clarity |
| 32 | parser.set_defaults(load_format="sharded_state") |
| 33 | |
| 34 | # Add validation arguments |
| 35 | parser.add_argument( |
| 36 | "--prompt", type=str, default="Hello, world!", help="Prompt for validation" |
| 37 | ) |
| 38 | parser.add_argument( |
| 39 | "--max-tokens", |
| 40 | type=int, |
| 41 | default=100, |
| 42 | help="Maximum number of tokens to generate", |
| 43 | ) |
| 44 | parser.add_argument( |
| 45 | "--temperature", type=float, default=0.7, help="Sampling temperature" |
| 46 | ) |
| 47 | parser.add_argument( |
| 48 | "--top-p", type=float, default=1.0, help="Top-p sampling parameter" |
| 49 | ) |
| 50 | |
| 51 | return parser.parse_args() |
| 52 | |
| 53 | |
| 54 | def main(): |
| 55 | args = parse_args() |
| 56 | engine_args = EngineArgs.from_cli_args(args) |
| 57 | |
| 58 | print( |
| 59 | f"Loading model from {engine_args.model} using format {engine_args.load_format}" |
| 60 | ) |
| 61 | print(f"Tensor parallel size: {engine_args.tensor_parallel_size}") |
| 62 | |
| 63 | # Load the model using engine args |
| 64 | llm = LLM.from_engine_args(engine_args) |
| 65 | |
| 66 | # Prepare sampling parameters |
| 67 | sampling_params = SamplingParams( |
| 68 | temperature=args.temperature, |
| 69 | top_p=args.top_p, |
| 70 | max_tokens=args.max_tokens, |
| 71 | ) |
| 72 | |
| 73 | print("\nRunning inference:") |
| 74 | print(f"Prompt: {args.prompt}") |
| 75 | |
| 76 | # Generate completion |
| 77 | outputs = llm.generate(args.prompt, sampling_params) |
| 78 | |
| 79 | # Display generated text |
| 80 | print("\nGenerated outputs:") |
| 81 | for output in outputs: |
| 82 | generated_text = output.outputs[0].text |
| 83 | print("-" * 50) |
| 84 | print(f"Full output: {args.prompt}{generated_text}") |
| 85 | print("-" * 50) |
| 86 | |
| 87 | |
| 88 | if __name__ == "__main__": |
| 89 | main() |