|
| 1 | +import os |
| 2 | + |
| 3 | +from ramalama.common import get_accel_env_vars |
| 4 | + |
| 5 | +def create_yaml(template_str, params): |
| 6 | + return(template_str.format(**params)) |
| 7 | + |
| 8 | + |
| 9 | +KSERVE_RUNTIME_TMPL = """ |
| 10 | +apiVersion: serving.kserve.io/v1alpha1 |
| 11 | +kind: ServingRuntime |
| 12 | +metadata: |
| 13 | + name: {runtime}-runtime |
| 14 | +spec: |
| 15 | + annotations: |
| 16 | + prometheus.io/port: '{port}' |
| 17 | + prometheus.io/path: '/metrics' |
| 18 | + multiModel: false |
| 19 | + supportedModelFormats: |
| 20 | + - autoSelect: true |
| 21 | + name: vLLM |
| 22 | + containers: |
| 23 | + - name: kserve-container |
| 24 | + image: {image} |
| 25 | + command: ["python", "-m", "vllm.entrypoints.openai.api_server"] |
| 26 | + args: ["--port={port}", "--model=/mnt/models", "--served-model-name={name}"] |
| 27 | + env: |
| 28 | + - name: HF_HOME |
| 29 | + value: /tmp/hf_home |
| 30 | + ports: |
| 31 | + - containerPort: {port} |
| 32 | + protocol: TCP |
| 33 | +""" |
| 34 | + |
| 35 | +KSERVE_MODEL_SERVICE = """\ |
| 36 | +# RamaLama {name} AI Model Service |
| 37 | +# kubectl create -f to import this kserve file into Kubernetes. |
| 38 | +# |
| 39 | +apiVersion: serving.kserve.io/v1beta1 |
| 40 | +kind: InferenceService |
| 41 | +metadata: |
| 42 | + name: huggingface-{name} |
| 43 | +spec: |
| 44 | + predictor: |
| 45 | + model: |
| 46 | + modelFormat: |
| 47 | + name: vLLM |
| 48 | + storageUri: "oci://{model}" |
| 49 | + resources: |
| 50 | + limits: |
| 51 | + cpu: "6" |
| 52 | + memory: 24Gi{gpu} |
| 53 | + requests: |
| 54 | + cpu: "6" |
| 55 | + memory: 24Gi{gpu} |
| 56 | +""" |
| 57 | + |
| 58 | + |
| 59 | +class Kserve: |
| 60 | + def __init__(self, model, chat_template_path, image, args, exec_args): |
| 61 | + self.ai_image = model |
| 62 | + if hasattr(args, "MODEL"): |
| 63 | + self.ai_image = args.MODEL |
| 64 | + self.ai_image = self.ai_image.removeprefix("oci://") |
| 65 | + if args.name: |
| 66 | + self.name = args.name |
| 67 | + else: |
| 68 | + self.name = os.path.basename(self.ai_image) |
| 69 | + |
| 70 | + self.model = model.removeprefix("oci://") |
| 71 | + self.args = args |
| 72 | + self.exec_args = exec_args |
| 73 | + self.image = image |
| 74 | + self.runtime = args.runtime |
| 75 | + |
| 76 | + def generate(self): |
| 77 | + env_var_string = "" |
| 78 | + for k, v in get_accel_env_vars().items(): |
| 79 | + env_var_string += f"Environment={k}={v}\n" |
| 80 | + |
| 81 | + _gpu = "" |
| 82 | + if os.getenv("CUDA_VISIBLE_DEVICES") != "": |
| 83 | + _gpu = 'nvidia.com/gpu' |
| 84 | + elif os.getenv("HIP_VISIBLE_DEVICES") != "": |
| 85 | + _gpu = 'amd.com/gpu' |
| 86 | + |
| 87 | + outfile = self.name + "-kserve-runtime.yaml" |
| 88 | + outfile = outfile.replace(":", "-") |
| 89 | + print(f"Generating kserve runtime file: {outfile}") |
| 90 | + |
| 91 | + # In your generate() method: |
| 92 | + yaml_content = create_yaml( |
| 93 | + KSERVE_RUNTIME_TMPL, |
| 94 | + { |
| 95 | + 'runtime' : self.runtime, |
| 96 | + 'model' : self.model, |
| 97 | + 'gpu' : _gpu if _gpu else "", |
| 98 | + 'port' : self.args.port, |
| 99 | + 'image' : self.image, |
| 100 | + 'name' : self.name, |
| 101 | + } |
| 102 | + ) |
| 103 | + with open(outfile, 'w') as c: |
| 104 | + c.write(yaml_content) |
| 105 | + |
| 106 | + outfile = self.name + "-kserve.yaml" |
| 107 | + outfile = outfile.replace(":", "-") |
| 108 | + print(f"Generating kserve file: {outfile}") |
| 109 | + yaml_content = create_yaml( |
| 110 | + KSERVE_MODEL_SERVICE, |
| 111 | + { |
| 112 | + 'name': self.name, |
| 113 | + 'model': self.model, |
| 114 | + 'gpu':_gpu if _gpu else "", |
| 115 | + } |
| 116 | + ) |
| 117 | + with open(outfile, 'w') as c: |
| 118 | + c.write(yaml_content) |
0 commit comments