cld2labs/llama-3.1-8b-instruct

third_party/Dell/model-deployment/llama-3.1-8b-instruct/deployment.md

@@ -0,0 +1,101 @@
+## Step 1: Prerequisites to Deploy Llama-3.1-8B-Instruct Model on Xeon with Keycloak
+
+Ensure the Enterprise Inference stack with Keycloak is already deployed before proceeding.
+
+Edit `core/scripts/generate-token.sh` and set your values before sourcing it:
+
+| Variable                  | Description                                                              |
+| ------------------------- | ------------------------------------------------------------------------ |
+| `BASE_URL`                | Hostname of your cluster (e.g. `api.example.com`), without `https://`   |
+| `KEYCLOAK_ADMIN_USERNAME` | Keycloak admin username                                                  |
+| `KEYCLOAK_PASSWORD`       | Keycloak admin password                                                  |
+| `KEYCLOAK_CLIENT_ID`      | Keycloak client ID configured during EI deployment                       |
+
+Then run:
+
+```bash
+export HUGGING_FACE_HUB_TOKEN="your_token_here"
+
+cd ~/Enterprise-Inference
+source core/scripts/generate-token.sh
+```
+
+This exports: `BASE_URL`, `KEYCLOAK_CLIENT_ID`, `KEYCLOAK_CLIENT_SECRET`, and `TOKEN`.
+
+## Step 2: Deploy Llama-3.1-8B-Instruct Model
+
+```bash
+helm install vllm-llama-8b ./core/helm-charts/vllm \
+  --values ./core/helm-charts/vllm/xeon-values.yaml \
+  --set LLM_MODEL_ID="meta-llama/Llama-3.1-8B-Instruct" \
+  --set global.HUGGINGFACEHUB_API_TOKEN="$HUGGING_FACE_HUB_TOKEN" \
+  --set ingress.enabled=true \
+  --set ingress.secretname="${BASE_URL}" \
+  --set ingress.host="${BASE_URL}" \
+  --set oidc.client_id="$KEYCLOAK_CLIENT_ID" \
+  --set oidc.client_secret="$KEYCLOAK_CLIENT_SECRET" \
+  --set apisix.enabled=true \
+  --set tensor_parallel_size="1" \
+  --set pipeline_parallel_size="1"
+```
+
+## Step 3: Verify the Deployment
+
+```bash
+kubectl get pods
+kubectl get apisixroutes
+```
+
+Expected Output:
+
+```
+NAME                                    READY   STATUS    RESTARTS
+keycloak-0                              1/1     Running   0
+keycloak-postgresql-0                   1/1     Running   0
+vllm-llama-8b-<hash>-<hash>             1/1     Running   0
+```
+
+> Note: The pod name suffix `<hash>-<hash>` is auto-generated by Kubernetes and will differ on each deployment. Ensure all pods show `1/1 Running`.
+
+```
+NAME                         HOSTS
+vllm-llama-8b-apisixroute    api.example.com
+```
+
+## Step 4: Test the Deployed Model
+
+```bash
+curl -k https://${BASE_URL}/Llama-3.1-8B-Instruct-vllmcpu/v1/completions \
+  -X POST \
+  -H "Content-Type: application/json" \
+  -H "Authorization: Bearer $TOKEN" \
+  -d '{
+    "model": "meta-llama/Llama-3.1-8B-Instruct",
+    "prompt": "What is Deep Learning?",
+    "max_tokens": 25,
+    "temperature": 0
+  }'
+```
+
+If successful, the model will return a completion response.
+
+## To undeploy the model
+
+```bash
+helm uninstall vllm-llama-8b
+```
+
+## Parameters
+
+| Parameter                                                  | Description                                                                                       |
+| ---------------------------------------------------------- | ------------------------------------------------------------------------------------------------- |
+| `--set LLM_MODEL_ID="meta-llama/Llama-3.1-8B-Instruct"`   | Defines the target model from **Hugging Face** to deploy.                                         |
+| `--set global.HUGGINGFACEHUB_API_TOKEN="..."`              | Authenticates access to gated or private Hugging Face models. Replace with your own secure token. |
+| `--set ingress.enabled=true`                               | Enables Kubernetes **Ingress** to expose the model service externally.                            |
+| `--set ingress.host="${BASE_URL}"`                         | Public hostname or FQDN for the inference endpoint (maps to your Ingress controller IP).          |
+| `--set ingress.secretname="${BASE_URL}"`                   | Kubernetes **TLS Secret** used for HTTPS termination at the ingress layer.                        |
+| `--set oidc.client_id="..."`                               | Keycloak OIDC client ID used for token-based authentication.                                      |
+| `--set oidc.client_secret="..."`                           | Keycloak OIDC client secret corresponding to the client ID.                                       |
+| `--set apisix.enabled=true`                                | Enables **APISIX** as the API gateway for routing and authentication.                             |
+| `--set tensor_parallel_size="1"`                           | Number of tensor parallel workers. Set to the number of available Gaudi cards per node.           |
+| `--set pipeline_parallel_size="1"`                         | Number of pipeline parallel stages. Typically `1` for single-node deployments.                   |

third_party/Dell/model-deployment/llama-3.1-8b-instruct/model-card.md

@@ -0,0 +1,64 @@
+# Llama-3.1-8B-Instruct
+
+This model uses Llama-3.1-8B-Instruct, a 8 billion-parameter instruction-tuned model from Meta Platforms, Inc. (Meta AI). It belongs to the Llama 3.1 model family and is optimized for multilingual dialogue, code tasks, and general instruction-following across a large context window.
+
+For full details including model specifications, licensing, intended use, safety guidance, and example prompts, please visit the official Hugging Face page: **Official Hugging Face Page**
+
+https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct
+
+This model provides inference services only; weights are hosted by Hugging Face under Meta’s license.
+
+Ensure compliance with the Llama 2 Community License Agreement before using this model.
+
+### Model Attribution
+
+**Developer:**	Meta Platforms, Inc. (Meta AI)
+
+**purpose:** Instruction-following model for dialogue, code generation/completion, multilingual tasks
+
+**Sizes/Variants:**	8 B parameters (instruction tuned); the Llama 3.1 family also includes 70 B and 405 B parameter variants
+
+**Modalities:**	Text input → Text (including code) output
+
+**Parameter Size:** ~8 billion
+
+**Max Context:**	Up to ~128 k tokens (for the 3.1 family)
+
+**License:** Llama 3.1 Community License (custom commercial license)
+
+**Minimum required CPU Cores:** 157
+
+**Minimum required PCIe Cards:** 1
+
+### Usage Notice
+
+**By using this model, you agree that:**
+
+- Inputs and outputs are processed through Llama-3.1-8B-Instruct under Meta’s Community License.
+- You will comply with Meta’s licensing terms, including restrictions on redistribution, commercial scale-use thresholds, attribution (“Built with Llama”), and acceptable use policy.
+- All generated content (text or code) must be reviewed for accuracy, compliance, and safety before deployment.
+- The model should not be used for generating malicious content, disallowed content, or automating decisions in high-risk or regulated systems without appropriate safeguards.
+
+### Intended Applications
+
+- Instruction-following chatbots and assistants (multilingual)
+- Code generation, completion, refactoring tasks (Python, Java, JavaScript, etc.)
+- Multilingual support (English, German, French, Italian, Portuguese, Hindi, Spanish, Thai) and potentially others with fine-tuning.
+- Large-context tasks: summarization of long documents, dialog over long history, RAG (retrieve-and-generate) over extended context.
+- Research, prototyping, and commercial workflows (subject to license terms).
+
+### Limitations
+
+- Although capable, the 8 B size still has trade-offs: accuracy and depth of reasoning may lag behind much larger models.
+- As with all large language models, risk of hallucinations (incorrect statements), biases, or unsafe outputs remains.
+- The custom license restricts certain uses (e.g., if your product has > 700 million monthly active users you may require a special license) as described in Meta’s license terms.
+- The model does not guarantee tool-use, vision/multimodal input (unless you fine-tune or wrap appropriately) – it is primarily text → text.
+- Running it efficiently still requires significant hardware/resources for full context and best performance
+
+### References
+
+“Introducing Llama 3.1: Our most capable models to date”. https://ai.meta.com/blog/meta-llama-3-1
+
+Hugging Face Model Card: https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct
+
+Meta Llama GitHub Repository & License Details. https://github.com/meta-llama/llama3

third_party/Dell/model-deployment/troubleshooting.md

@@ -0,0 +1,79 @@
+# Troubleshooting Guide
+
+This section provides common issues observed when running inference against models deployed via Helm commands on Intel® AI for Enterprise Inference, along with step-by-step resolutions.
+
+**Issues:**
+  1. [Gateway Timeout (504) on Inference Requests](#1-gateway-timeout-504-on-inference-requests)
+
+---
+
+### 1. Gateway Timeout (504) on Inference Requests
+
+**Context:** Model deployed via Helm commands. Inference request sent through the ingress stack (ingress-nginx -> APISIX -> vLLM service).
+
+**Error:** Inference requests return `504 Gateway Timeout` after 60 seconds:
+
+```
+"POST /<model-name>/v1/completions HTTP/2.0" 504
+upstream timed out (110: Operation timed out) ... 60.001
+```
+
+**Cause:**
+
+CPU-based model inference (`vllm-cpu`) generates tokens at ~0.3-0.4 tokens/s. Responses requiring more than ~24 tokens exceed the default 60s upstream timeout enforced by ingress-nginx and APISIX.
+
+**Fix:**
+
+**Step 1 - Increase the nginx ingress timeout**
+
+Apply to both the `default` and `auth-apisix` namespaces. To find ingress names:
+
+```bash
+kubectl get ingress -A | grep <model-name>
+```
+
+Then annotate each ingress:
+
+```bash
+kubectl annotate ingress <ingress-name> -n <namespace> \
+  nginx.ingress.kubernetes.io/proxy-read-timeout="300" \
+  nginx.ingress.kubernetes.io/proxy-send-timeout="300" \
+  nginx.ingress.kubernetes.io/proxy-connect-timeout="60" \
+  --overwrite
+```
+
+**Step 2 - Increase the APISIX route timeout**
+
+To find the route name:
+
+```bash
+kubectl get apisixroute -n auth-apisix | grep <model-name>
+```
+
+Edit the route:
+
+```bash
+kubectl edit apisixroute <route-name> -n auth-apisix
+```
+
+Update the timeout section under the route:
+
+```yaml
+spec:
+  http:
+    - name: <route-name>
+      timeout:
+        connect: 60s
+        send: 300s
+        read: 300s
+```
+
+**Verification:**
+
+Re-run the inference request and confirm a `200 OK` response is returned within the new timeout window.
+
+**Notes:**
+
+- The nginx ingress annotation takes effect immediately; no pod restart required.
+- For GPU-based deployments this timeout is rarely needed as throughput is significantly higher (30-50 tokens/s vs 0.3-0.4 tokens/s on CPU).
+- If requests still time out after increasing both timeouts, reduce `max_tokens` in the request payload to limit response length.