[Doc] Add better clarity for tensorizer usage (#4090)

Co-authored-by: Roger Wang <136131678+ywang96@users.noreply.github.com>

docs/source/models/engine_args.rst

@@ -45,7 +45,7 @@ Below, you can find an explanation of every engine argument for vLLM:
     * "safetensors" will load the weights in the safetensors format.
     * "npcache" will load the weights in pytorch format and store a numpy cache to speed up the loading.
     * "dummy" will initialize the weights with random values, mainly for profiling.
-    * "tensorizer" will load serialized weights using `CoreWeave's Tensorizer model deserializer. <https://github.com/coreweave/tensorizer>`_. See `tensorized_vllm_model.py` in the examples folder to serialize a vLLM model, and for more information. Tensorizer support for vLLM can be installed with `pip install vllm[tensorizer]`.
+    * "tensorizer" will load serialized weights using `CoreWeave's Tensorizer model deserializer. <https://github.com/coreweave/tensorizer>`_ See `examples/tensorize_vllm_model.py <https://github.com/vllm-project/vllm/blob/main/examples/tensorize_vllm_model.py>`_ to serialize a vLLM model, and for more information.
 
 .. option:: --dtype {auto,half,float16,bfloat16,float,float32}
 

examples/tensorize_vllm_model.py

@@ -23,14 +23,16 @@ from vllm.model_executor.tensorizer_loader import TensorizerArgs
 # yapf: disable
 """
 tensorize_vllm_model.py is a script that can be used to serialize and 
-deserialize vLLM models. These models can be loaded using tensorizer directly 
-to the GPU extremely quickly. Tensor encryption and decryption is also 
-supported, although libsodium must be installed to use it. Install
-vllm with tensorizer support using `pip install vllm[tensorizer]`.
+deserialize vLLM models. These models can be loaded using tensorizer 
+to the GPU extremely quickly over an HTTP/HTTPS endpoint, an S3 endpoint,
+or locally. Tensor encryption and decryption is also supported, although 
+libsodium must be installed to use it. Install vllm with tensorizer support 
+using `pip install vllm[tensorizer]`.
 
-To serialize a model, you can run something like this:
+To serialize a model, install vLLM from source, then run something 
+like this from the root level of this repository:
 
-python tensorize_vllm_model.py \
+python -m examples.tensorize_vllm_model \
    --model EleutherAI/gpt-j-6B \
    --dtype float16 \
    serialize \
@@ -38,31 +40,57 @@ python tensorize_vllm_model.py \
    --suffix vllm
    
 Which downloads the model from HuggingFace, loads it into vLLM, serializes it,
-and saves it to your S3 bucket. A local directory can also be used.
+and saves it to your S3 bucket. A local directory can also be used. This
+assumes your S3 credentials are specified as environment variables
+in the form of `S3_ACCESS_KEY_ID`, `S3_SECRET_ACCESS_KEY`, and `S3_ENDPOINT`.
+To provide S3 credentials directly, you can provide `--s3-access-key-id` and 
+`--s3-secret-access-key`, as well as `--s3-endpoint` as CLI args to this 
+script.
 
 You can also encrypt the model weights with a randomly-generated key by 
 providing a `--keyfile` argument.
 
-To deserialize a model, you can run something like this:
+To deserialize a model, you can run something like this from the root 
+level of this repository:
 
-python tensorize_vllm_model.py \
+python -m examples.tensorize_vllm_model \
    --model EleutherAI/gpt-j-6B \
    --dtype float16 \
    deserialize \
    --path-to-tensors s3://my-bucket/vllm/EleutherAI/gpt-j-6B/vllm/model.tensors
 
 Which downloads the model tensors from your S3 bucket and deserializes them.
-To provide S3 credentials, you can provide `--s3-access-key-id` and 
-`--s3-secret-access-key`, as well as `--s3-endpoint` as CLI args to this script,
-the OpenAI entrypoint, as arguments for LLM(), or as environment variables
-in the form of `S3_ACCESS_KEY_ID`, `S3_SECRET_ACCESS_KEY`, and `S3_ENDPOINT`.
-
 
 You can also provide a `--keyfile` argument to decrypt the model weights if 
 they were serialized with encryption.
 
-For more information on the available arguments, run 
-`python tensorize_vllm_model.py --help`.
+For more information on the available arguments for serializing, run 
+`python -m examples.tensorize_vllm_model serialize --help`.
+
+Or for deserializing:
+
+`python -m examples.tensorize_vllm_model deserialize --help`.
+
+Once a model is serialized, it can be used to load the model when running the
+OpenAI inference client at `vllm/entrypoints/openai/api_server.py` by providing
+the `--tensorizer-uri` CLI argument that is functionally the same as the
+`--path-to-tensors` argument in this script, along with `--vllm-tensorized`, to
+signify that the model to be deserialized is a vLLM model, rather than a 
+HuggingFace `PreTrainedModel`, which can also be deserialized using tensorizer
+in the same inference server, albeit without the speed optimizations. To
+deserialize an encrypted file, the `--encryption-keyfile` argument can be used
+to provide the path to the keyfile used to encrypt the model weights. For
+information on all the arguments that can be used to configure tensorizer's
+deserialization, check out the tensorizer options argument group in the
+`vllm/entrypoints/openai/api_server.py` script with `--help`.
+
+Tensorizer can also be invoked with the `LLM` class directly to load models:
+
+    llm = LLM(model="facebook/opt-125m",
+              load_format="tensorizer",
+              tensorizer_uri=path_to_opt_tensors,
+              num_readers=3,
+              vllm_tensorized=True)
 """
 
 

vllm/model_executor/tensorizer_loader.py

@@ -126,7 +126,6 @@ class TensorizerArgs:
             "s3_endpoint": self.s3_endpoint,
         }
 
-        # Omitting self.dtype and self.device as this behaves weirdly
         self.deserializer_params = {
             "verify_hash": self.verify_hash,
             "encryption": self.encryption_keyfile,
@@ -145,7 +144,7 @@ class TensorizerArgs:
             parser: argparse.ArgumentParser) -> argparse.ArgumentParser:
         """Tensorizer CLI arguments"""
 
-        # Create the argument group
+        # Tensorizer options arg group
         group = parser.add_argument_group(
             'tensorizer options',
             description=('Options for configuring the behavior of the'
@@ -205,9 +204,7 @@ class TensorizerArgs:
 
     @classmethod
     def from_cli_args(cls, args: argparse.Namespace) -> "TensorizerArgs":
-        # Get the list of attributes of this dataclass.
         attrs = [attr.name for attr in dataclasses.fields(cls)]
-        # Set the attributes from the parsed arguments.
         tensorizer_args = cls(**{
             attr: getattr(args, attr)
             for attr in attrs if hasattr(args, attr)
@@ -291,7 +288,6 @@ class TensorizerAgent:
             nn.Module: The deserialized model.
         """
         before_mem = get_mem_usage()
-        # Lazy load the tensors from S3 into the model.
         start = time.perf_counter()
         with open_stream(
                 self.tensorizer_args.tensorizer_uri,