Snowflake 模型注册表——在 Snowflake 数据云与 AI 云自建部署 LLM 的便捷工具 | 技术实践

# Pick your model of choice from HFimport huggingface_hubfrom snowflake.ml.model import HuggingFacePipelineModelfrom snowflake.ml.model import openai_signaturesfrom snowflake.ml.utils import connection_paramsfrom snowflake.snowpark import Sessionfrom snowflake import snowparkfrom snowflake.ml.registry import registry as registry_modulefrom snowflake.ml.utils.connection_params import SnowflakeLoginOptionsmodel = "Qwen/Qwen3-30B-A3B"model_ref = HuggingFacePipelineModel(   model="Qwen/Qwen3-30B-A3B",   task="text-generation",)# Get a reference to model registry session = session = get_active_session()assert session is not Noneregistry = registry_module.Registry(   session=session,)assert registry is not None# Log your model into model registrmodel_name = "Qwen3_30B_A3B_model_final"mv = registry.log_model(   model=model_ref,   model_name=model_name,   signatures=openai_signatures.OPENAI_CHAT_SIGNATURE)mv

model = "Qwen/Qwen3-30B-A3B"cache_dir = "Qwen/Qwen3_30B_A3B"repo_snapshot_dir = huggingface_hub.snapshot_download(   repo_id=model,   cache_dir=cache_dir,)model_ref = HuggingFacePipelineModel(   model="Qwen/Qwen3-30B-A3B",   task="text-generation",   download_snapshot=False,)assert model_ref.repo_snapshot_dir is Nonemodel_ref.repo_snapshot_dir = repo_snapshot_dirassert not model_ref.repo_snapshot_dir is None

## Take Note of model version printed after logging the model and update the model version in next line! #### TO DO #####model_name = "QWEN3_30B_A3B_MODEL_FINAL"version_name = "WICKED_BADGER_1"mv = registry.get_model(model_name).version(version_name)from snowflake.ml.model.inference_engine import InferenceEnginemv.create_service(   service_name="Qwen3_30B_A3B_service_vLLM",      service_compute_pool="MAHDI_GPU_COMPUTE_POOL",  # Pick a compute pool name with GPU   gpu_requests="4",  # 4 if the compute pool has 4 GPUs   ingress_enabled=True,  # set a False if you don't want to create a service endpoint   image_repo="MAHDI_PLAYGROUND.LLMS.Qwen3_30B_A3B_image_repo",  # Optional: specify a container services image repo name   inference_engine_options={       "engine": InferenceEngine.VLLM,       "engine_args_override": [            "--tensor-parallel-size=4",            "--gpu-memory-utilization=0.85",            "--max-model-len=1024",            "--max-num-seqs=2"       ],   },)

import openaiclient = openai.OpenAI(base_url=url, api_key="not_a_real_key")services_df = mv.list_services()services_dfendpoint = services_df.iloc[-1]["inference_endpoint"] # change index if the model has more than once service## v1 is the current version of inference endpoint for Open AI compatible APIs## no relation with model versioning in Snowflake ML Registryurl = f"https://{endpoint}/v1"url## in next line, you don't need an actual key from openai. ## You can pass any string.. ## the Snowflake PAT is what it is being used for authentication. client = openai.OpenAI(base_url=url, api_key="not_a_real_key")pat = "........."#model_name="Qwen3_30B_A3B_model_final"## CASE SENSITIVE!!!!!model_name="QWEN3_30B_A3B_MODEL_FINAL"messages = [   {"role": "system", "content": "You are a helpful assistant. Answer in a concise manner."},   {"role": "user", "content": "Who is your creator??"},]completions = client.chat.completions.create(   model=model_name,   messages=messages,   extra_headers={"Authorization": f'Snowflake Token="{pat}"'},   max_completion_tokens=250,)completions

-- input table with questions in string formatCREATE OR REPLACE TABLE llm_input_table (question VARCHAR);INSERT INTO    llm_input_table (question)VALUES    ('What is the capital of France?'),    ('What is the capital of India?'),    ('What is the capital of Italy?');-- service function callwith chat_inputs as (    select        array_construct(            object_construct(                'role',                'system',                'content',                'You are a helpful assistant.'            ),            object_construct(                'role',                'user',                'content',                question            )        )::ARRAY as messages,        0.9::FLOAT as temperature,        100::NUMBER as max_completion_tokens,        NULL::ARRAY as stop,        1::NUMBER as n,        False::BOOLEAN as stream,        0.9::FLOAT as top_p,        0.9::FLOAT as frequency_penalty,        0.1::FLOAT as presence_penalty    from        llm_input_table)SELECT    <SERVICE_NAME> ! __CALL__(        messages,        temperature,                max_completion_tokens,        stop,        n,        stream,        top_p,        frequency_penalty,        presence_penalty    )from    chat_inputs;