diff --git a/templates/llm-a10.yaml b/templates/llm-a10.yaml new file mode 100644 index 0000000..59a534a --- /dev/null +++ b/templates/llm-a10.yaml @@ -0,0 +1,199 @@ +--- +apiVersion: apps/v1 +kind: Deployment +metadata: + name: llm-a10 + namespace: api + labels: + app: llm + base: huggingface-hub + gpu: nvidia-a10 + model: llama-3-70b +spec: + selector: + matchLabels: + app: llm + base: huggingface-hub + gpu: nvidia-a10 + model: llama-3-70b + strategy: + type: Recreate # we have to reuse the gpus, not double-allocating + template: + metadata: + annotations: + instrumentation.opentelemetry.io/inject-sdk: "true" + labels: + app: llm + base: huggingface-hub + gpu: nvidia-a10 + model: llama-3-70b + spec: + affinity: + nodeAffinity: + requiredDuringSchedulingIgnoredDuringExecution: + nodeSelectorTerms: + - matchExpressions: + - key: nvidia.com/gpu.product + operator: In + values: + - NVIDIA-A10 + containers: + - name: tgi + image: ghcr.io/huggingface/text-generation-inference:2.3.0 + imagePullPolicy: Always + args: + - --otlp-endpoint + - $(OTEL_EXPORTER_OTLP_ENDPOINT) + env: + # - name: CORS_ALLOW_ORIGIN + # value: "*" + # Specify the batch sizes to compute cuda graphs for. + # Use "0" to disable. + # Default = "1,2,4,8,16,32" + # - name: CUDA_GRAPHS + # value: "0" + # - name: DTYPE + # value: float16 + - name: MAX_TOTAL_TOKENS + # value: "131072" + value: "4096" + - name: MODEL_ID + # value: yanolja/EEVE-Korean-Instruct-10.8B-v1.0 + # value: beomi/Llama-3-Open-Ko-8B-Instruct-preview + value: meta-llama/Meta-Llama-3.1-70B-Instruct + - name: PORT + value: "8080" + # Possible values: + # - awq: 4 bit quantization. Requires a specific AWQ quantized model: . Should replace GPTQ models wherever possible because of the better latency + # - eetq: 8 bit quantization, doesn't require specific model. Should be a drop-in replacement to bitsandbytes with much better performance. Kernels are from + # - gptq: 4 bit quantization. Requires a specific GTPQ quantized model: . llm-inference will use exllama (faster) kernels wherever possible, and use triton kernel (wider support) when it's not. AWQ has faster kernels + # - bitsandbytes: Bitsandbytes 8bit. Can be applied on any model, will cut the memory requirement in half, but it is known that the model will be much slower to run than the native f16 + # - bitsandbytes-nf4: Bitsandbytes 4bit. Can be applied on any model, will cut the memory requirement by 4x, but it is known that the model will be much slower to run than the native f16 + # - bitsandbytes-fp4: Bitsandbytes 4bit. nf4 should be preferred in most cases but maybe this one has better perplexity performance for you model + # - fp8: [FP8](https://developer.nvidia.com/blog/nvidia-arm-and-intel-publish-fp8-specification-for-standardization-as-an-interchange-format-for-ai/) (e4m3) works on H100 and above This dtype has native ops should be the fastest if available. This is currently not the fastest because of local unpacking + padding to satisfy matrix multiplication limitations + # - name: QUANTIZE + # value: eetq + - name: RUST_LOG + value: INFO + - name: SHARDED + value: "true" + # - name: TRUST_REMOTE_CODE + # value: "false" + resources: + limits: + # cpu: "8" + # memory: "300Gi" + nvidia.com/gpu: "8" + # FIXME: add readiness & liveness check + # FIXME: See https://github.com/huggingface/text-generation-inference/blob/8511669cb29115bdf0bc2da5328e69d041030996/router/src/server.rs#L128 + ports: + - name: http + protocol: TCP + containerPort: 8080 + volumeMounts: + - name: cache + mountPath: /data + - name: dshm + mountPath: /dev/shm + - name: huggingface-hub-token + mountPath: /root/.cache/huggingface/token + subPath: token + volumes: + - name: cache + persistentVolumeClaim: + claimName: huggingface-hub-cache + - name: dshm + emptyDir: + medium: Memory + sizeLimit: 16Gi + - name: huggingface-hub-token + secret: + secretName: huggingface-hub + items: + - key: HUGGING_FACE_HUB_TOKEN + path: token +--- +apiVersion: v1 +kind: Service +metadata: + name: llm + namespace: api + labels: + app: llm + model: llama-3-70b +spec: + type: ClusterIP + selector: + app: llm + model: llama-3-70b + ports: + - name: http + protocol: TCP + port: 80 + targetPort: 8080 +--- +apiVersion: networking.k8s.io/v1 +kind: Ingress +metadata: + name: llm + namespace: api + annotations: + cert-manager.io/cluster-issuer: mobilex.kr + nginx.ingress.kubernetes.io/enable-cors: "true" + nginx.ingress.kubernetes.io/proxy-body-size: 20M + nginx.ingress.kubernetes.io/proxy-read-timeout: "3600" + nginx.ingress.kubernetes.io/proxy-send-timeout: "3600" +spec: + ingressClassName: mobilex.kr + tls: + - hosts: + - api.mobilex.kr + secretName: api.mobilex.kr-certs + rules: + - host: api.mobilex.kr + http: + paths: + - path: /generate + pathType: Prefix + backend: + service: + name: llm + port: + number: 80 + - path: /v1/chat + pathType: Prefix + backend: + service: + name: llm + port: + number: 80 +--- +apiVersion: networking.k8s.io/v1 +kind: Ingress +metadata: + name: llm-chatgpt + namespace: api + annotations: + cert-manager.io/cluster-issuer: mobilex.kr + nginx.ingress.kubernetes.io/enable-cors: "true" + nginx.ingress.kubernetes.io/proxy-body-size: 20M + nginx.ingress.kubernetes.io/proxy-read-timeout: "3600" + nginx.ingress.kubernetes.io/proxy-send-timeout: "3600" + nginx.ingress.kubernetes.io/rewrite-target: /v1/chat/$2 +spec: + ingressClassName: mobilex.kr + tls: + - hosts: + - api.mobilex.kr + secretName: api.mobilex.kr-certs + rules: + - host: api.mobilex.kr + http: + paths: + - path: /chat(/|$)(.*) + pathType: ImplementationSpecific + backend: + service: + name: llm + port: + number: 80 diff --git a/templates/llm.yaml b/templates/llm.yaml index 2b817c3..497b617 100644 --- a/templates/llm.yaml +++ b/templates/llm.yaml @@ -16,6 +16,8 @@ spec: base: huggingface-hub gpu: nvidia-a100 model: llama-3-70b + strategy: + type: Recreate # we have to reuse the gpus, not double-allocating template: metadata: annotations: @@ -37,7 +39,7 @@ spec: - NVIDIA-A100-SXM4-40GB containers: - name: tgi - image: ghcr.io/huggingface/text-generation-inference:2.2.0 + image: ghcr.io/huggingface/text-generation-inference:2.3.0 imagePullPolicy: Always args: - --otlp-endpoint @@ -53,7 +55,8 @@ spec: # - name: DTYPE # value: float16 - name: MAX_TOTAL_TOKENS - value: "131072" + # value: "131072" + value: "4096" - name: MODEL_ID # value: yanolja/EEVE-Korean-Instruct-10.8B-v1.0 # value: beomi/Llama-3-Open-Ko-8B-Instruct-preview