@tduncan / Nvidia-Optimization-Stack-VLLM
Code Commits Issues Pulls Cloud Agent CI/CD
main
main
md 102 lines 4.42 KB
Rendered Raw
1 # Loading models with CoreWeave's Tensorizer
2
3 vLLM supports loading models with [CoreWeave's Tensorizer](https://docs.coreweave.com/coreweave-machine-learning-and-ai/inference/tensorizer).
4 vLLM model tensors that have been serialized to disk, an HTTP/HTTPS endpoint, or S3 endpoint can be deserialized
5 at runtime extremely quickly directly to the GPU, resulting in significantly
6 shorter Pod startup times and CPU memory usage. Tensor encryption is also supported.
7
8 vLLM fully integrates Tensorizer in to its model loading machinery. The following will give a brief overview on how to get started with using Tensorizer on vLLM.
9
10 ## Installing Tensorizer
11
12 To install `tensorizer`, run `pip install vllm[tensorizer]`.
13
14 ## The basics
15
16 To load a model using Tensorizer, the model first needs to be serialized by
17 Tensorizer. [The example script](../../../examples/features/tensorize_vllm_model.py) takes care of this process.
18
19 Let's walk through a basic example by serializing `facebook/opt-125m` using the script, and then loading it for inference.
20
21 ## Serializing a vLLM model with Tensorizer
22
23 To serialize a model with Tensorizer, call the example script with the necessary
24 CLI arguments. The docstring for the script itself explains the CLI args
25 and how to use it properly in great detail, and we'll use one of the examples from the docstring directly, assuming we want to serialize and save our model at our S3 bucket example `s3://my-bucket`:
26
27 ```bash
28 python examples/features/tensorize_vllm_model.py \
29 --model facebook/opt-125m \
30 serialize \
31 --serialized-directory s3://my-bucket \
32 --suffix v1
33 ```
34
35 This saves the model tensors at `s3://my-bucket/vllm/facebook/opt-125m/v1`. If you intend on applying a LoRA adapter to your tensorized model, you can pass the HF id of the LoRA adapter in the above command, and the artifacts will be saved there too:
36
37 ```bash
38 python examples/features/tensorize_vllm_model.py \
39 --model facebook/opt-125m \
40 --lora-path <lora_id> \
41 serialize \
42 --serialized-directory s3://my-bucket \
43 --suffix v1
44 ```
45
46 ## Serving the model using Tensorizer
47
48 Once the model is serialized where you want it, you can load the model using `vllm serve` or the `LLM` entrypoint. You can pass the directory where you saved the model to the `model` argument for `LLM()` and `vllm serve`. For example, to serve the tensorized model saved previously with the LoRA adapter, you'd do:
49
50 ```bash
51 vllm serve s3://my-bucket/vllm/facebook/opt-125m/v1 \
52 --load-format tensorizer \
53 --enable-lora
54 ```
55
56 Or, with `LLM()`:
57
58 ```python
59 from vllm import LLM
60 llm = LLM(
61 "s3://my-bucket/vllm/facebook/opt-125m/v1",
62 load_format="tensorizer",
63 enable_lora=True,
64 )
65 ```
66
67 ## Options for configuring Tensorizer
68
69 `tensorizer`'s core objects that serialize and deserialize models are `TensorSerializer` and `TensorDeserializer` respectively. In order to pass arbitrary kwargs to these, which will configure the serialization and deserialization processes, you can provide them as keys to `model_loader_extra_config` with `serialization_kwargs` and `deserialization_kwargs` respectively. Full docstrings detailing all parameters for the aforementioned objects can be found in `tensorizer`'s [serialization.py](https://github.com/coreweave/tensorizer/blob/main/tensorizer/serialization.py) file.
70
71 As an example, CPU concurrency can be limited when serializing with `tensorizer` via the `limit_cpu_concurrency` parameter in the initializer for `TensorSerializer`. To set `limit_cpu_concurrency` to some arbitrary value, you would do so like this when serializing:
72
73 ```bash
74 python examples/features/tensorize_vllm_model.py \
75 --model facebook/opt-125m \
76 --lora-path <lora_id> \
77 serialize \
78 --serialized-directory s3://my-bucket \
79 --serialization-kwargs '{"limit_cpu_concurrency": 2}' \
80 --suffix v1
81 ```
82
83 As an example when customizing the loading process via `TensorDeserializer`, you could limit the number of concurrency readers during deserialization with the `num_readers` parameter in the initializer via `model_loader_extra_config` like so:
84
85 ```bash
86 vllm serve s3://my-bucket/vllm/facebook/opt-125m/v1 \
87 --load-format tensorizer \
88 --enable-lora \
89 --model-loader-extra-config '{"deserialization_kwargs": {"num_readers": 2}}'
90 ```
91
92 Or with `LLM()`:
93
94 ```python
95 from vllm import LLM
96 llm = LLM(
97 "s3://my-bucket/vllm/facebook/opt-125m/v1",
98 load_format="tensorizer",
99 enable_lora=True,
100 model_loader_extra_config={"deserialization_kwargs": {"num_readers": 2}},
101 )
102 ```