Multi-model inference built for ultra-low latency at scale

1import truss_chains as chains
2from truss import truss_config
3
4MISTRAL_HF_MODEL = "mistralai/Mistral-7B-Instruct-v0.2"
5MISTRAL_CACHE = truss_config.ModelRepo(
6    repo_id=MISTRAL_HF_MODEL, allow_patterns=["*.json", "*.safetensors", ".model"]
7)
8HF_ACCESS_TOKEN_NAME = "hf_access_token"
9
10class MistralLLM(chains.ChainletBase):
11    remote_config = chains.RemoteConfig(
12        docker_image=chains.DockerImage(
13            pip_requirements=[
14                "transformers==4.38.1",
15                "torch==2.0.1",
16            ]
17        ),
18        compute=chains.Compute(cpu_count=2, gpu="A10G"),
19        assets=chains.Assets(cached=[MISTRAL_CACHE], secret_keys=[HF_ACCESS_TOKEN_NAME]),
20    )
21
22    def __init__(
23        self,
24        # Adding the `context` to the init arguments, allows us to access the
25        # huggingface token.
26        context: chains.DeploymentContext = chains.depends_context(),
27    ) -> None:
28        # Note the imports of the *specific* python requirements are pushed down to
29        # here. This code will only be executed on the remotely deployed chainlet,
30        # not in the local environment, so we don't need to install these packages
31        # in the local dev environment.
32        import torch
33        import transformers
34
35        self._model = transformers.AutoModelForCausalLM.from_pretrained(
36            MISTRAL_HF_MODEL,
37            torch_dtype=torch.float16,
38            device_map="auto",
39            use_auth_token=context.secrets[HF_ACCESS_TOKEN_NAME],
40        )
41
42        self._tokenizer = transformers.AutoTokenizer.from_pretrained(
43            MISTRAL_HF_MODEL,
44            device_map="auto",
45            torch_dtype=torch.float16,
46            use_auth_token=context.secrets[HF_ACCESS_TOKEN_NAME],
47        )
48
49        self._generate_args = {
50            "max_new_tokens": 512,
51            "temperature": 1.0,
52            "top_p": 0.95,
53            "top_k": 50,
54            "repetition_penalty": 1.0,
55            "no_repeat_ngram_size": 0,
56            "use_cache": True,
57            "do_sample": True,
58            "eos_token_id": self._tokenizer.eos_token_id,
59            "pad_token_id": self._tokenizer.pad_token_id,
60        }
61
62    def run_remote(self, prompt: str) -> str:
63        import torch
64
65        formatted_prompt = f"[INST] {prompt} [/INST]"
66        input_ids = self._tokenizer(
67            formatted_prompt, return_tensors="pt"
68        ).input_ids.cuda()
69        with torch.no_grad():
70            output = self._model.generate(inputs=input_ids, **self._generate_args)
71            result = self._tokenizer.decode(output[0])
72        return result
73
74class PoemGenerator(chains.ChainletBase):
75    def __init__(self, phi_llm: PhiLLM = chains.depends(PhiLLM)) -> None:
76        self._phi_llm = phi_llm
77
78    def run_remote(self, words: list[str]) -> list[str]:
79        results = []
80        for word in words:
81          messages = Messages(
82              messages=[
83                  {"role": "system", "content": "You are poet"},
84                  {"role": "user", "content": f"Write a poem about {word}"},
85              ]
86          )
87          poem = self._phi_llm.run_remote(messages)
88          results.append(poem)
89        return results
90
91class PhiLLM(chains.ChainletBase):
92    remote_config = chains.RemoteConfig(
93        docker_image=chains.DockerImage(
94            pip_requirements=[
95                "transformers==4.41.2",
96                "torch==2.3.0",
97            ]
98        ),
99        compute=chains.Compute(cpu_count=2, gpu="T4"),
100    )