Edge-Computing-LLM · waqasm86 · Jul 17, 2026 · Jul 8, 2026 · Jul 17, 2026 · Jul 17, 2026
diff --git a/.github/ISSUE_TEMPLATE/config.yml b/.github/ISSUE_TEMPLATE/config.yml
@@ -1,5 +1,5 @@
 blank_issues_enabled: true
 contact_links:
   - name: Support Guide
-    url: https://github.com/waqasm86/llm-observability-stack/blob/main/SUPPORT.md
+    url: https://github.com/Edge-Computing-LLM/llm-observability-stack/blob/main/SUPPORT.md
     about: Check support and troubleshooting guidance before opening an issue.
diff --git a/.github/workflows/helm-validate.yaml b/.github/workflows/helm-validate.yaml
@@ -4,14 +4,17 @@ on:
   push:
   pull_request:
 
+permissions:
+  contents: read
+
 jobs:
   lint-and-render:
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v7
 
       - name: Set up Helm
-        uses: azure/setup-helm@v4
+        uses: azure/setup-helm@v5
         with:
           version: v3.15.4
 
@@ -39,18 +42,8 @@ jobs:
             --set open-webui.webuiSecret.existingSecretName= \
             >/tmp/rendered-full-stack-nvidia.yaml
 
-      - name: Build and validate monitoring chart
-        run: |
-          if [ ! -d monitoring ]; then
-            echo "monitoring chart is not present in this repository; skipping optional monitoring chart validation"
-            exit 0
-          fi
-          helm dependency build monitoring
-          helm lint monitoring
-          helm template llm-monitoring monitoring >/tmp/rendered-monitoring.yaml
-
       - name: Set up Python
-        uses: actions/setup-python@v5
+        uses: actions/setup-python@v6
         with:
           python-version: "3.11"
 

diff --git a/.gitignore b/.gitignore
@@ -9,6 +9,8 @@ values.local-*.yaml
 *.key
 *.crt
 *.kubeconfig
+kubeconfig
+kubeconfig*
 
 # Generated render/debug artifacts
 .generated/
@@ -43,6 +45,9 @@ pictures/
 charts/*.tgz
 charts/**/*.tgz
 monitoring/charts/*.tgz
+llm-observability-stack-*.tgz
+*.tar.gz
+*.zip
 # Large model/data artifacts (keep out of git)
 *.gguf
 *.safetensors
@@ -84,6 +89,7 @@ rendered-with-etcd.yaml
 *.private.yaml
 jupyter-notebooks/python*-pip-list.txt
 local-run-output/
+validation-output/
 *.kube-report.json
 *.kubeconfig.bak
 evidence-local-run/
diff --git a/.helmignore b/.helmignore
@@ -61,6 +61,9 @@ python-toolbox/examples/service_dns_check.py
 
 # Vendored chart docs/assets are not needed at install time
 charts/*.tgz
+charts/gpu-operator/
+charts/nvidia-device-plugin/
+charts/dcgm-exporter/
 charts/ollama/README.md
 charts/ollama/LICENSE
 charts/ollama/artifacthub-repo.yml

diff --git a/Chart.lock b/Chart.lock
@@ -1,13 +1,4 @@
 dependencies:
-- name: gpu-operator
-  repository: file://charts/gpu-operator
-  version: v1.0.0-devel
-- name: nvidia-device-plugin
-  repository: file://charts/nvidia-device-plugin
-  version: 0.19.2
-- name: dcgm-exporter
-  repository: file://charts/dcgm-exporter
-  version: 4.8.2
 - name: ollama
   repository: file://charts/ollama
   version: 1.50.0
@@ -23,5 +14,5 @@ dependencies:
 - name: opentelemetry-operator
   repository: file://charts/opentelemetry-operator
   version: 0.115.0
-digest: sha256:7a8de8212f1051cbddc881e9d7d865424f83be907ccd7d5da91491038579eba5
-generated: "2026-06-16T14:22:44.167855506+05:00"
+digest: sha256:9a3e5dc71273d49fd4f3e8a46e8142ca5c83ffb8163c9666c97f4a558af01480
+generated: "2026-07-08T19:18:56.195921069+05:00"
diff --git a/Chart.yaml b/Chart.yaml
@@ -13,9 +13,6 @@ kubeVersion: ">=1.27.0-0"
 home: https://github.com/Edge-Computing-LLM/llm-observability-stack
 sources:
   - https://github.com/prometheus-community/helm-charts
-  - https://docs.nvidia.com/datacenter/cloud-native/gpu-operator/latest/
-  - https://github.com/NVIDIA/k8s-device-plugin
-  - https://github.com/NVIDIA/dcgm-exporter
   - https://github.com/open-telemetry/opentelemetry-helm-charts
   - https://github.com/otwld/ollama-helm
   - https://github.com/open-webui/helm-charts
@@ -37,18 +34,6 @@ keywords:
   - opentelemetry
   - genai
 dependencies:
-  - name: gpu-operator
-    version: v1.0.0-devel
-    repository: file://charts/gpu-operator
-    condition: gpu-operator.enabled
-  - name: nvidia-device-plugin
-    version: 0.19.2
-    repository: file://charts/nvidia-device-plugin
-    condition: nvidia-device-plugin.enabled
-  - name: dcgm-exporter
-    version: 4.8.2
-    repository: file://charts/dcgm-exporter
-    condition: dcgm-exporter.enabled
   - name: ollama
     version: 1.50.0
     repository: file://charts/ollama

diff --git a/Modelfile.qwen-1.8b-chat-q4_K_M b/Modelfile.qwen-1.8b-chat-q4_K_M
@@ -0,0 +1,27 @@
+# Modelfile for the locally retained Ollama Qwen 1.8B Chat Q4_K_M GGUF.
+#
+# Upstream: https://ollama.com/library/qwen:1.8b-chat-q4_K_M
+# GGUF SHA-256: ef0125bcc77278420b64229fc29e235435c517f67ab7aa8546a9f5f7be644cef
+# License: Tongyi Qianwen Research License (non-commercial use unless separately licensed).
+#
+# The GeForce 940M has 1 GiB VRAM. Live tuning on this host showed that 24 GPU
+# layers use 860 MiB, while 23 GPU layers use 824 MiB. Keep num_batch at 1:
+# larger batches make this Qwen architecture reserve a CUDA graph buffer that
+# does not fit on this GPU.
+
+FROM /models/gguf/qwen-1.8b-chat-q4_K_M.gguf
+
+TEMPLATE """<|im_start|>system
+{{ .System }}<|im_end|>
+<|im_start|>user
+{{ .Prompt }}<|im_end|>
+<|im_start|>assistant
+"""
+
+PARAMETER stop "<|im_start|>"
+PARAMETER stop "<|im_end|>"
+PARAMETER num_gpu 23
+PARAMETER num_ctx 256
+PARAMETER num_batch 1
+PARAMETER num_predict 128
+PARAMETER temperature 0
diff --git a/README.md b/README.md
@@ -2,11 +2,12 @@
 
 Kubernetes-native observability, benchmarking, and operations tooling for private LLM inference on local edge systems.
 
-Go CLI documentation: [docs/cli.md](docs/cli.md).
+Preferred organization CLI: [`edge-cli`](https://github.com/Edge-Computing-LLM/edge-cli).
+Repo-local legacy/helper CLI documentation: [docs/cli.md](docs/cli.md).
 
-This repository packages a Helm-based stack for k3s and Kubernetes with Ollama/GGUF model serving, Open WebUI, an OpenTelemetry GenAI-instrumented FastAPI proxy, Prometheus, Grafana, OpenTelemetry Collector, blackbox probes, benchmark metrics, and optional NVIDIA GPU monitoring through the NVIDIA runtime, device plugin, GPU Operator, and DCGM-compatible dashboards.
+This repository packages a Helm-based application and observability stack for k3s and Kubernetes with Ollama/GGUF model serving, Open WebUI, an OpenTelemetry GenAI-instrumented FastAPI proxy, Prometheus, Grafana, OpenTelemetry Collector, blackbox probes, benchmark metrics, and NVIDIA/DCGM-compatible dashboards.
 
-The repository also includes a Go CLI named `llm-observability`. It reuses `github.com/Edge-Computing-LLM/k3s-nvidia-edge/pkg/edgebase` for base-layer checks and keeps application-specific Helm, validation, and benchmark workflows in this repository.
+The repository also includes a Go CLI named `llm-observability` for repo-local helper workflows. New end-to-end installs should use `edge-cli`, which deploys `k3s-nvidia-edge` first and then this chart.
 
 GitHub repository: <https://github.com/Edge-Computing-LLM/llm-observability-stack>
 
@@ -16,6 +17,8 @@ For local NVIDIA GPU deployments, deploy [`k3s-nvidia-edge`](https://github.com/
 
 `k3s-nvidia-edge` owns k3s, k3s containerd NVIDIA runtime wiring, GPU Operator, NVIDIA device plugin, DCGM exporter, Node Feature Discovery, `RuntimeClass/nvidia`, and the allocatable `nvidia.com/gpu` resource. `llm-observability-stack` then deploys Ollama, Open WebUI, OpenTelemetry, dashboards, benchmarks, and application-level observability on top of that base layer.
 
+An empty local k3s cluster with only CoreDNS, local-path-provisioner, and other default k3s system components is a valid starting point before the base layer is deployed. Run `edge install infra` or validate `k3s-nvidia-edge` before installing GPU profiles from this repository.
+
 Read the full dependency guide before installing GPU profiles:
 
 - [k3s-nvidia-edge dependency](docs/K3S-NVIDIA-EDGE-DEPENDENCY.md)
@@ -25,7 +28,7 @@ Read the full dependency guide before installing GPU profiles:
 - Local private LLM serving through Ollama and legally obtained GGUF models.
 - Kubernetes deployment through Helm with k3s-friendly profiles.
 - NVIDIA GPU scheduling with `runtimeClassName: nvidia` and `nvidia.com/gpu` when a GPU is available.
-- CPU-only deployment profiles for MacOS/minikube and k3s systems without NVIDIA GPUs.
+- Optional CPU validation profiles for development clusters without NVIDIA GPUs.
 - Open WebUI for browser-based interaction with local models.
 - A FastAPI proxy with LLM request metrics for TTFT, latency, tokens per second, prompt tokens, generated tokens, active requests, and errors.
 - Prometheus, Grafana, Alertmanager, kube-state-metrics, node exporter, ServiceMonitors, probes, and alert rules.
@@ -41,30 +44,44 @@ The current local deployment target is a single-node Xubuntu 24 system running k
 - k3s node: combined control-plane and worker.
 - NVIDIA device plugin resource: `nvidia.com/gpu: 1`.
 - RuntimeClass: `nvidia`.
-- Model: Gemma 3 1B IT Q4_K_M GGUF.
+- Model: Qwen 1.8B Chat Q4_K_M GGUF through the local alias
+  `qwen-1-8b-chat-q4-k-m-local`.
 
-Measured after one warmup and three streaming benchmark requests:
+Measured after deployment, warmup, and exact-response, arithmetic, and
+translation inference checks:
 
 | Metric | Result |
 |---|---:|
-| TTFT p50 | 0.377 s |
-| TTFT p95 | 0.381 s |
-| Mean throughput | 11.69 tokens/s |
-| End-to-end p95 | 6.97 s |
-| Peak GPU utilization | 52% |
-| VRAM usage | 554 MiB |
+| Model size | Approximately 1.2 GB |
+| CUDA layers | 23/25 |
+| Processor split | 27% CPU / 73% GPU |
+| Context / batch | 256 / 1 |
+| Observed throughput | 9.75-15.78 tokens/s |
+| VRAM usage | 824 MiB used / 152 MiB free |
+| Residency | `Forever` |
 
 Evidence and reproduction:
 
 - [Single-node k3s GeForce 940M guide](docs/SINGLE-NODE-K3S-GEFORCE-940M.md)
 - [Local k3s NVIDIA deployment report - 2026-07-02](docs/LOCAL-K3S-NVIDIA-REPORT-2026-07-02.md)
+- [Live layered validation - 2026-07-08](docs/LIVE-VALIDATION-2026-07-08.md)
 - [Verified local GPU results](docs/VERIFIED-LOCAL-GPU-RESULTS.md)
 - [Xubuntu k3s NVIDIA runbook](docs/XUBUNTU-K3S-NVIDIA-RUNBOOK.md)
 - [Sanitized benchmark artifact](artifacts/geforce-940m-benchmark.json)
 - [GeForce 940M Helm profile](values.geforce-940m-k3s.yaml)
+- [Qwen GGUF runtime evidence companion](https://github.com/Edge-Computing-LLM/qwen-gguf-observability)
 
 These numbers prove constrained local edge feasibility. They do not claim enterprise load, concurrency, fleet reliability, or production readiness.
 
+The companion repository performs read-only runtime contract checks and
+captures sanitized evidence. This chart remains the source of truth for the
+Modelfile, Helm values, model lifecycle, and workload configuration.
+
+The separate
+[`Frontend-Edge-LLM-Observability`](https://github.com/Edge-Computing-LLM/Frontend-Edge-LLM-Observability)
+project provides the TypeScript/Vue dashboard. It consumes controlled DCGM and
+future telemetry endpoints; frontend source is not duplicated in this chart.
+
 ## Who This Is For
 
 - Developers running private LLMs on local Linux systems.
@@ -82,7 +99,7 @@ These numbers prove constrained local edge feasibility. They do not claim enterp
 
 ## Platform Components
 
-- Vendored Helm charts for Ollama, Open WebUI, NVIDIA GPU Operator, NVIDIA device plugin, DCGM exporter, kube-prometheus-stack, OpenTelemetry Collector, and OpenTelemetry Operator.
+- Vendored Helm charts for Ollama, Open WebUI, kube-prometheus-stack, OpenTelemetry Collector, and OpenTelemetry Operator.
 - FastAPI OpenTelemetry GenAI-instrumented proxy with Prometheus metrics.
 - TTFT, latency, token, throughput, active-request, HTTP, and error telemetry.
 - Optional kube-prometheus-stack, Grafana, Alertmanager, node exporter, and kube-state-metrics from the root umbrella chart.
@@ -115,7 +132,7 @@ Prometheus + Grafana + Alertmanager
         +-- ServiceMonitors, probes, benchmark Pushgateway, Kubernetes metrics
 ```
 
-The verified laptop profile uses Ollama/GGUF. The same observability contract can be used on larger RTX workstations, GPU Operator/DCGM clusters, NIM endpoints, or cloud GPU clusters.
+The verified laptop profile uses Ollama/GGUF. The same observability contract can be used on larger local RTX workstations with the NVIDIA substrate prepared by `k3s-nvidia-edge`.
 
 ## Repository Layout
 
@@ -149,7 +166,14 @@ Build the CLI:
 go build -o bin/llm-observability ./cmd/llm-observability
 ```
 
-Recommended local CLI flow when `k3s-nvidia-edge` is already healthy:
+Preferred local CLI flow from the organization control plane:
+
+```bash
+edge install all --accelerator auto --yes
+edge validate observability
+```
+
+Repo-local helper flow when `k3s-nvidia-edge` is already healthy:
 
 ```bash
 bin/llm-observability doctor
@@ -163,7 +187,7 @@ bin/llm-observability validate
 - Helm 3 or 4.
 - For local NVIDIA k3s GPU profiles: `k3s-nvidia-edge` deployed and validated first.
 - NVIDIA driver and NVIDIA Container Toolkit for GPU profiles.
-- NVIDIA device plugin or GPU Operator exposing `nvidia.com/gpu` for GPU mode.
+- `RuntimeClass/nvidia` and `nvidia.com/gpu` provided by `k3s-nvidia-edge` for GPU mode.
 - A legally obtained GGUF model available on node storage.
 - Python 3.11 for tests and benchmark tooling.
 
@@ -180,6 +204,11 @@ helm version
 
 The local bootstrap helper detects the Kubernetes runtime before installing. It uses NVIDIA mode when Kubernetes advertises `nvidia.com/gpu`; otherwise it writes a CPU-only overlay and runs the same edge LLM observability path without NVIDIA runtime or GPU resource requests.
 
+The organization CLI exposes the same policy directly. The full installer uses
+host detection, while `edge install observability --accelerator auto --yes` uses
+Kubernetes allocatable GPU capacity. Explicit `cpu` and `nvidia` modes are also
+available for deterministic automation.
+
 ## Quick Start
 
 ### A. Minimal validation profile
@@ -193,18 +222,22 @@ helm template llm-observability-stack . \
 
 Review the machine-specific model host path before using this profile on another system. The profile schedules on nodes with `nvidia.com/gpu.present=true`, which supports a single-node k3s control-plane/worker laptop without requiring a separate worker label.
 
-Deploy and validate `k3s-nvidia-edge` first:
+This profile uses the locally retained Qwen 1.8B Chat Q4_K_M GGUF and
+`Modelfile.qwen-1.8b-chat-q4_K_M`. On the 1 GiB GeForce 940M it pins 23/25
+layers to CUDA, limits batch size to 1, uses a 256-token context, and keeps the
+model loaded indefinitely. The measured steady-state allocation is 824 MiB VRAM.
+
+Preferred: deploy and validate the base layer through `edge-cli` first:
 
 ```bash
-cd /media/waqasm86/External1/Waqas-Projects/Project-Edge-Computing-LLM/k3s-nvidia-edge
-bin/k3s-nvidia-edge install --yes --sudo=false --use-local-chart --skip-base-package-install --skip-toolkit-install --skip-k3s-install
-bin/k3s-nvidia-edge validate --yes
+edge install infra --yes
+edge validate infra
 ```
 
 Then deploy the LLM stack:
 
 ```bash
-cd /media/waqasm86/External1/Waqas-Projects/Project-Edge-Computing-LLM/llm-observability-stack
+cd /media/waqasm86/External1/Waqas-Projects/Project-Linux-Kubernetes-Nvidia/Project-Edge-Computing-LLM/llm-observability-stack
 
 helm upgrade --install llm-observability-stack . \
   -n llm-observability --create-namespace \
@@ -213,15 +246,18 @@ helm upgrade --install llm-observability-stack . \
 ./hack/test-geforce-940m-inference.sh
 ```
 
-### C. Full-stack NVIDIA profile
+### C. Full observability NVIDIA profile
 
 ```bash
 helm upgrade --install llm-observability-stack . \
   -n llm-observability --create-namespace \
   -f values.full-stack-nvidia.example.yaml
 ```
 
-Use private values files or existing Kubernetes Secrets for OpenTelemetry and Open WebUI secrets. Never commit secrets.
+This installs the LLM and observability layer only. It does not install GPU
+Operator, NVIDIA device plugin, or DCGM exporter. Use private values files or
+existing Kubernetes Secrets for OpenTelemetry and Open WebUI secrets. Never
+commit secrets.
 
 ### D. Local full-stack k3s profile
 
@@ -272,7 +308,7 @@ Run the public benchmark from another terminal:
 
 ```bash
 ./benchmarks/ollama_benchmark.py \
-  --model gemma3-1b-it-gguf-local \
+  --model qwen-1-8b-chat-q4-k-m-local \
   --warmup-runs 1 \
   --runs 10 \
   --output artifacts/benchmark-local.json

diff --git a/SUPPORT.md b/SUPPORT.md
@@ -4,8 +4,8 @@
 
 Use GitHub Issues for setup, debugging, and enhancement questions:
 
-- Repository: https://github.com/waqasm86/llm-observability-stack
-- New issue: https://github.com/waqasm86/llm-observability-stack/issues/new
+- Repository: https://github.com/Edge-Computing-LLM/llm-observability-stack
+- New issue: https://github.com/Edge-Computing-LLM/llm-observability-stack/issues/new
 
 ## What to Include
 

diff --git a/benchmarks/ollama_benchmark.py b/benchmarks/ollama_benchmark.py
@@ -1,4 +1,4 @@
-#!/usr/bin/env python3
+#!/usr/bin/env python3.11
 """Run repeatable Ollama streaming benchmarks and optionally publish summary metrics."""
 
 from __future__ import annotations