Tensorize vLLM Model#

Source vllm-project/vllm.

  1import argparse
  2import dataclasses
  3import json
  4import os
  5import uuid
  6
  7from vllm import LLM
  8from vllm.engine.arg_utils import EngineArgs
  9from vllm.model_executor.model_loader.tensorizer import (TensorizerArgs,
 10                                                         TensorizerConfig,
 11                                                         tensorize_vllm_model)
 12
 13# yapf conflicts with isort for this docstring
 14# yapf: disable
 15"""
 16tensorize_vllm_model.py is a script that can be used to serialize and 
 17deserialize vLLM models. These models can be loaded using tensorizer 
 18to the GPU extremely quickly over an HTTP/HTTPS endpoint, an S3 endpoint,
 19or locally. Tensor encryption and decryption is also supported, although 
 20libsodium must be installed to use it. Install vllm with tensorizer support 
 21using `pip install vllm[tensorizer]`. To learn more about tensorizer, visit
 22https://github.com/coreweave/tensorizer
 23
 24To serialize a model, install vLLM from source, then run something 
 25like this from the root level of this repository:
 26
 27python -m examples.tensorize_vllm_model \
 28   --model facebook/opt-125m \
 29   serialize \
 30   --serialized-directory s3://my-bucket \
 31   --suffix v1
 32   
 33Which downloads the model from HuggingFace, loads it into vLLM, serializes it,
 34and saves it to your S3 bucket. A local directory can also be used. This
 35assumes your S3 credentials are specified as environment variables
 36in the form of `S3_ACCESS_KEY_ID`, `S3_SECRET_ACCESS_KEY`, and 
 37`S3_ENDPOINT_URL`. To provide S3 credentials directly, you can provide 
 38`--s3-access-key-id` and `--s3-secret-access-key`, as well as `--s3-endpoint` 
 39as CLI args to this script.
 40
 41You can also encrypt the model weights with a randomly-generated key by 
 42providing a `--keyfile` argument.
 43
 44To deserialize a model, you can run something like this from the root 
 45level of this repository:
 46
 47python -m examples.tensorize_vllm_model \
 48   --model EleutherAI/gpt-j-6B \
 49   --dtype float16 \
 50   deserialize \
 51   --path-to-tensors s3://my-bucket/vllm/EleutherAI/gpt-j-6B/v1/model.tensors
 52
 53Which downloads the model tensors from your S3 bucket and deserializes them.
 54
 55You can also provide a `--keyfile` argument to decrypt the model weights if 
 56they were serialized with encryption.
 57
 58To support distributed tensor-parallel models, each model shard will be
 59serialized to a separate file. The tensorizer_uri is then specified as a string
 60template with a format specifier such as '%03d' that will be rendered with the
 61shard's rank. Sharded models serialized with this script will be named as
 62model-rank-%03d.tensors
 63
 64For more information on the available arguments for serializing, run 
 65`python -m examples.tensorize_vllm_model serialize --help`.
 66
 67Or for deserializing:
 68
 69`python -m examples.tensorize_vllm_model deserialize --help`.
 70
 71Once a model is serialized, tensorizer can be invoked with the `LLM` class 
 72directly to load models:
 73
 74    llm = LLM(model="facebook/opt-125m",
 75              load_format="tensorizer",
 76              model_loader_extra_config=TensorizerConfig(
 77                    tensorizer_uri = path_to_tensors,
 78                    num_readers=3,
 79                    )
 80              )
 81            
 82A serialized model can be used during model loading for the vLLM OpenAI
 83inference server. `model_loader_extra_config` is exposed as the CLI arg
 84`--model-loader-extra-config`, and accepts a JSON string literal of the
 85TensorizerConfig arguments desired.
 86
 87In order to see all of the available arguments usable to configure 
 88loading with tensorizer that are given to `TensorizerConfig`, run:
 89
 90`python -m examples.tensorize_vllm_model deserialize --help`
 91
 92under the `tensorizer options` section. These can also be used for
 93deserialization in this example script, although `--tensorizer-uri` and
 94`--path-to-tensors` are functionally the same in this case.
 95"""
 96
 97
 98def parse_args():
 99    parser = argparse.ArgumentParser(
100        description="An example script that can be used to serialize and "
101        "deserialize vLLM models. These models "
102        "can be loaded using tensorizer directly to the GPU "
103        "extremely quickly. Tensor encryption and decryption is "
104        "also supported, although libsodium must be installed to "
105        "use it.")
106    parser = EngineArgs.add_cli_args(parser)
107    subparsers = parser.add_subparsers(dest='command')
108
109    serialize_parser = subparsers.add_parser(
110        'serialize', help="Serialize a model to `--serialized-directory`")
111
112    serialize_parser.add_argument(
113        "--suffix",
114        type=str,
115        required=False,
116        help=(
117            "The suffix to append to the serialized model directory, which is "
118            "used to construct the location of the serialized model tensors, "
119            "e.g. if `--serialized-directory` is `s3://my-bucket/` and "
120            "`--suffix` is `v1`, the serialized model tensors will be "
121            "saved to "
122            "`s3://my-bucket/vllm/EleutherAI/gpt-j-6B/v1/model.tensors`. "
123            "If none is provided, a random UUID will be used."))
124    serialize_parser.add_argument(
125        "--serialized-directory",
126        type=str,
127        required=True,
128        help="The directory to serialize the model to. "
129        "This can be a local directory or S3 URI. The path to where the "
130        "tensors are saved is a combination of the supplied `dir` and model "
131        "reference ID. For instance, if `dir` is the serialized directory, "
132        "and the model HuggingFace ID is `EleutherAI/gpt-j-6B`, tensors will "
133        "be saved to `dir/vllm/EleutherAI/gpt-j-6B/suffix/model.tensors`, "
134        "where `suffix` is given by `--suffix` or a random UUID if not "
135        "provided.")
136
137    serialize_parser.add_argument(
138        "--keyfile",
139        type=str,
140        required=False,
141        help=("Encrypt the model weights with a randomly-generated binary key,"
142              " and save the key at this path"))
143
144    deserialize_parser = subparsers.add_parser(
145        'deserialize',
146        help=("Deserialize a model from `--path-to-tensors`"
147              " to verify it can be loaded and used."))
148
149    deserialize_parser.add_argument(
150        "--path-to-tensors",
151        type=str,
152        required=True,
153        help="The local path or S3 URI to the model tensors to deserialize. ")
154
155    deserialize_parser.add_argument(
156        "--keyfile",
157        type=str,
158        required=False,
159        help=("Path to a binary key to use to decrypt the model weights,"
160              " if the model was serialized with encryption"))
161
162    TensorizerArgs.add_cli_args(deserialize_parser)
163
164    return parser.parse_args()
165
166
167
168def deserialize():
169    llm = LLM(model=args.model,
170              load_format="tensorizer",
171              tensor_parallel_size=args.tensor_parallel_size,
172              model_loader_extra_config=tensorizer_config
173    )
174    return llm
175
176
177if __name__ == '__main__':
178    args = parse_args()
179
180    s3_access_key_id = (getattr(args, 's3_access_key_id', None)
181                        or os.environ.get("S3_ACCESS_KEY_ID", None))
182    s3_secret_access_key = (getattr(args, 's3_secret_access_key', None)
183                            or os.environ.get("S3_SECRET_ACCESS_KEY", None))
184    s3_endpoint = (getattr(args, 's3_endpoint', None)
185                or os.environ.get("S3_ENDPOINT_URL", None))
186
187    credentials = {
188        "s3_access_key_id": s3_access_key_id,
189        "s3_secret_access_key": s3_secret_access_key,
190        "s3_endpoint": s3_endpoint
191    }
192
193    model_ref = args.model
194
195    model_name = model_ref.split("/")[1]
196
197    keyfile = args.keyfile if args.keyfile else None
198
199    if args.model_loader_extra_config:
200        config = json.loads(args.model_loader_extra_config)
201        tensorizer_args = \
202            TensorizerConfig(**config)._construct_tensorizer_args()
203        tensorizer_args.tensorizer_uri = args.path_to_tensors
204    else:
205        tensorizer_args = None
206
207    if args.command == "serialize":
208        eng_args_dict = {f.name: getattr(args, f.name) for f in
209                        dataclasses.fields(EngineArgs)}
210
211        engine_args = EngineArgs.from_cli_args(
212            argparse.Namespace(**eng_args_dict)
213        )
214
215        input_dir = args.serialized_directory.rstrip('/')
216        suffix = args.suffix if args.suffix else uuid.uuid4().hex
217        base_path = f"{input_dir}/vllm/{model_ref}/{suffix}"
218        if engine_args.tensor_parallel_size > 1:
219            model_path = f"{base_path}/model-rank-%03d.tensors"
220        else:
221            model_path = f"{base_path}/model.tensors"
222
223        tensorizer_config = TensorizerConfig(
224            tensorizer_uri=model_path,
225            encryption_keyfile=keyfile,
226            **credentials)
227
228        tensorize_vllm_model(engine_args, tensorizer_config)
229
230    elif args.command == "deserialize":
231        if not tensorizer_args:
232            tensorizer_config = TensorizerConfig(
233                tensorizer_uri=args.path_to_tensors,
234                encryption_keyfile = keyfile,
235                **credentials
236            )
237        deserialize()
238    else:
239        raise ValueError("Either serialize or deserialize must be specified.")