diff --git a/README.md b/README.md
index 42cd427..bb59956 100644
--- a/README.md
+++ b/README.md
@@ -1,3 +1,35 @@
-# atc-GPU
+# atc-GPU — GPU Lab Infrastructure
 
-GPU Lab infrastructure — VM303 model-manager, gpu-lab Docker stacks, Dockhand
\ No newline at end of file
+Infrastructure-as-code and documentation for the ATC GPU Lab environment on **VM303 (GPU-Dev)**.
+
+## Quick links
+
+| Service | URL |
+|---------|-----|
+| GPU Lab UI | http://10.0.20.106:9000/ |
+| Inference API (active runtime) | http://10.0.20.106:8010/v1/models |
+| OpenAI gateway | http://10.0.20.106:8001/v1/models |
+| Dockhand (GPU-Dev env) | http://atc-docker01.dell-atc.lan:8082/ |
+
+## Repository layout
+
+```
+atc-GPU/
+├── README.md
+├── docs/           # Landscape, models, runbooks, Dockhand mapping
+├── config/         # gpu-lab compose + model-manager catalog (no secrets)
+├── compose/        # Dockhand-safe compose variants
+└── scripts/        # Helper scripts (activate-model)
+```
+
+## Active target (2026-06)
+
+- **Runtime:** SGLang (`sglang-modern`) via `compose.sglang.yml`
+- **Model:** Qwen 3.6 27B AWQ (`qwen36-27b-awq` → served as `qwen3.6-27b-awq`)
+- **Previous:** Llama3-70b-GPTQ on vLLM legacy (to be decommissioned after cutover)
+
+See [docs/runbook-vm303.md](docs/runbook-vm303.md) for activate/rollback steps.
+
+## Related repos
+
+- [mo/Lakehouse](http://atc-mgt01.dell-atc.lan:3001/mo/Lakehouse) — Spark, Trino, Debezium, Airflow, S3 pipeline
diff --git a/config/gpu-lab/RUNBOOK.md b/config/gpu-lab/RUNBOOK.md
new file mode 100644
index 0000000..13537ac
--- /dev/null
+++ b/config/gpu-lab/RUNBOOK.md
@@ -0,0 +1,61 @@
+# GPU Lab Runbook — VM303 (atc-gpu-dev)
+
+## Runtimes (Docker)
+
+| Runtime ID | Compose | Image |
+|---|---|---|
+| `vllm-legacy-040` | `compose.vllm-legacy.yml` | `vllm/vllm-openai:v0.4.0.post1` |
+| `vllm-modern-080` | `compose.vllm-modern.yml` | `vllm/vllm-openai:v0.8.5.post1` |
+| `sglang-modern` | `compose.sglang.yml` | `lmsysorg/sglang:v0.5.12-cu129-runtime` |
+
+Only **one** inference container runs at a time (port **8010**).
+
+## Activate (UI)
+
+GPU Lab → pick model → **Activate**. Writes `/root/gpu-lab/active/current.env`, stops all gpu-lab containers, starts the matching compose stack.
+
+## Manual activate
+
+```bash
+cd /root/gpu-lab
+# edit active/current.env (MODEL_PATH, TP_SIZE, VLLM_EXTRA_ARGS, …)
+docker compose -f compose.vllm-legacy.yml down
+docker compose -f compose.vllm-legacy.yml up -d
+curl -s http://127.0.0.1:8010/v1/models
+```
+
+## Rollback to systemd vLLM
+
+```bash
+cd /root/gpu-lab
+docker compose -f compose.vllm-legacy.yml down
+docker compose -f compose.vllm-modern.yml down
+docker compose -f compose.sglang.yml down
+systemctl start vllm.service
+systemctl restart vllm-gateway.service litellm-proxy.service
+```
+
+## Logs
+
+```bash
+docker logs -f gpu-lab-vllm-legacy
+docker logs -f gpu-lab-vllm-modern
+docker logs -f gpu-lab-sglang
+journalctl -u model-manager -f
+```
+
+## SGLang on V100
+
+Always use `--disable-flashinfer` and `--disable-flashinfer-sampling` (already in compose.sglang.yml).
+
+## Dockhand
+
+- Environment **GPU-Dev** → `10.0.20.106:2375`
+- Stacks: `gpu-lab-vllm-legacy`, `gpu-lab-vllm-modern`, `gpu-lab-sglang` (compose in Dockhand data volume)
+- Running workload appears as stack **`gpu-lab`** (compose project on VM303)
+- API: `GET /api/stacks?env=8` (not `environmentId`)
+- UI: http://atc-docker01.dell-atc.lan:8082/
+
+## Docker data
+
+Images stored under `/root/docker-data` (not `/var` — small partition).
diff --git a/config/gpu-lab/compose.sglang.yml b/config/gpu-lab/compose.sglang.yml
new file mode 100644
index 0000000..d6ae859
--- /dev/null
+++ b/config/gpu-lab/compose.sglang.yml
@@ -0,0 +1,38 @@
+name: gpu-lab-sglang
+services:
+  sglang:
+    image: lmsysorg/sglang:v0.5.12-cu129-runtime
+    container_name: gpu-lab-sglang
+    network_mode: host
+    ipc: host
+    shm_size: "32gb"
+    env_file:
+      - active/current.env
+    volumes:
+      - /root/models:/models:ro
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              count: all
+              capabilities: [gpu]
+    entrypoint: ["python3", "-m", "sglang.launch_server"]
+    command:
+      - --model-path
+      - ${MODEL_PATH}
+      - --served-model-name
+      - ${SERVED_NAME}
+      - --host
+      - "0.0.0.0"
+      - --port
+      - "8010"
+      - --tp-size
+      - ${TP_SIZE}
+      - --mem-fraction-static
+      - ${GPU_MEM_UTIL}
+      - --context-length
+      - ${MAX_MODEL_LEN}
+      - --disable-flashinfer
+      - --disable-flashinfer-sampling
+    restart: "no"
diff --git a/config/gpu-lab/compose.vllm-legacy.yml b/config/gpu-lab/compose.vllm-legacy.yml
new file mode 100644
index 0000000..7d90ad5
--- /dev/null
+++ b/config/gpu-lab/compose.vllm-legacy.yml
@@ -0,0 +1,33 @@
+name: gpu-lab-vllm-legacy
+services:
+  vllm-legacy:
+    image: vllm/vllm-openai:v0.4.0.post1
+    container_name: gpu-lab-vllm-legacy
+    network_mode: host
+    ipc: host
+    shm_size: "32gb"
+    env_file:
+      - active/current.env
+    volumes:
+      - /root/models:/models:ro
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              count: all
+              capabilities: [gpu]
+    entrypoint: ["/bin/sh", "-c"]
+    command:
+      - >
+        exec python3 -m vllm.entrypoints.openai.api_server
+        --model "$$MODEL_PATH"
+        --served-model-name "$$SERVED_NAME"
+        --host 0.0.0.0 --port 8010
+        --tensor-parallel-size "$$TP_SIZE"
+        --dtype "$$DTYPE"
+        --max-model-len "$$MAX_MODEL_LEN"
+        --gpu-memory-utilization "$$GPU_MEM_UTIL"
+        --max-num-seqs "$$MAX_NUM_SEQS"
+        $$VLLM_EXTRA_ARGS
+    restart: "no"
diff --git a/config/gpu-lab/compose.vllm-modern.yml b/config/gpu-lab/compose.vllm-modern.yml
new file mode 100644
index 0000000..bff14f7
--- /dev/null
+++ b/config/gpu-lab/compose.vllm-modern.yml
@@ -0,0 +1,34 @@
+name: gpu-lab-vllm-modern
+services:
+  vllm-modern:
+    image: vllm/vllm-openai:v0.8.5.post1
+    container_name: gpu-lab-vllm-modern
+    network_mode: host
+    ipc: host
+    shm_size: "32gb"
+    env_file:
+      - active/current.env
+    volumes:
+      - /root/models:/models:ro
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              count: all
+              capabilities: [gpu]
+    entrypoint: ["/bin/sh", "-c"]
+    command:
+      - >
+        exec python3 -m vllm.entrypoints.openai.api_server
+        --model "$$MODEL_PATH"
+        --served-model-name "$$SERVED_NAME"
+        --host 0.0.0.0 --port 8010
+        --tensor-parallel-size "$$TP_SIZE"
+        --dtype "$$DTYPE"
+        --max-model-len "$$MAX_MODEL_LEN"
+        --gpu-memory-utilization "$$GPU_MEM_UTIL"
+        --max-num-seqs "$$MAX_NUM_SEQS"
+        --trust-remote-code --enforce-eager
+        $$VLLM_EXTRA_ARGS
+    restart: "no"
diff --git a/config/gpu-lab/runtimes.json b/config/gpu-lab/runtimes.json
new file mode 100644
index 0000000..8b5eaae
--- /dev/null
+++ b/config/gpu-lab/runtimes.json
@@ -0,0 +1,24 @@
+{
+  "vllm-legacy-040": {
+    "engine": "vllm",
+    "compose_file": "compose.vllm-legacy.yml",
+    "image": "vllm/vllm-openai:v0.4.0.post1",
+    "v100_ok": true,
+    "label": "vLLM 0.4"
+  },
+  "vllm-modern-080": {
+    "engine": "vllm",
+    "compose_file": "compose.vllm-modern.yml",
+    "image": "vllm/vllm-openai:v0.8.5.post1",
+    "v100_ok": true,
+    "label": "vLLM 0.8"
+  },
+  "sglang-modern": {
+    "engine": "sglang",
+    "compose_file": "compose.sglang.yml",
+    "image": "lmsysorg/sglang:v0.5.12-cu129-runtime",
+    "v100_ok": true,
+    "label": "SGLang",
+    "extra_args": ["--disable-flashinfer", "--disable-flashinfer-sampling"]
+  }
+}
diff --git a/config/model-manager/catalog.json b/config/model-manager/catalog.json
new file mode 100644
index 0000000..6951fba
--- /dev/null
+++ b/config/model-manager/catalog.json
@@ -0,0 +1,515 @@
+{
+  "models": [
+    {
+      "id": "llama3-70b-gptq",
+      "name": "Llama 3 70B GPTQ",
+      "description": "Meta Llama 3 open weights. Strong general coder (lab default). 4\u00d7 V100.",
+      "tags": [
+        "coding",
+        "opensource",
+        "v100_ready"
+      ],
+      "license": "Llama 3 Community License",
+      "repo_id": null,
+      "local_path": "/root/models/llama3-70b-quantized",
+      "served_name": "llama3-70b-gptq",
+      "size_gb_estimate": 38,
+      "vllm_min_version": "0.4.0",
+      "vllm": {
+        "tensor_parallel_size": 4,
+        "quantization": "gptq",
+        "dtype": "float16",
+        "max_model_len": 8192,
+        "gpu_memory_utilization": 0.88,
+        "max_num_seqs": 4,
+        "extra_args": [
+          "--disable-custom-all-reduce",
+          "--enforce-eager"
+        ]
+      },
+      "runtime_id": "vllm-legacy-040",
+      "inference_backend": "vllm",
+      "v100_activate": true,
+      "v100_block_reason": null
+    },
+    {
+      "id": "qwen25-coder-32b-gptq",
+      "name": "Qwen 2.5 Coder 32B GPTQ",
+      "description": "Alibaba Qwen open weights. Top open-source code LLM (GPT-4o class). 4\u00d7 V100.",
+      "tags": [
+        "coding",
+        "opensource",
+        "qwen",
+        "v100_ready"
+      ],
+      "license": "Apache-2.0",
+      "repo_id": "Qwen/Qwen2.5-Coder-32B-Instruct-GPTQ-Int4",
+      "local_path": "/root/models/qwen25-coder-32b-gptq",
+      "served_name": "qwen2.5-coder-32b-gptq",
+      "size_gb_estimate": 20,
+      "vllm_min_version": "0.4.0",
+      "vllm": {
+        "tensor_parallel_size": 4,
+        "quantization": "gptq",
+        "dtype": "float16",
+        "max_model_len": 8192,
+        "gpu_memory_utilization": 0.9,
+        "max_num_seqs": 4,
+        "extra_args": [
+          "--trust-remote-code",
+          "--enforce-eager"
+        ]
+      },
+      "runtime_id": "vllm-legacy-040",
+      "inference_backend": "vllm",
+      "v100_activate": true,
+      "v100_block_reason": null
+    },
+    {
+      "id": "qwen25-32b-gptq",
+      "name": "Qwen 2.5 32B GPTQ",
+      "description": "Alibaba Qwen open weights (general instruct, strong coder).",
+      "tags": [
+        "coding",
+        "opensource",
+        "qwen",
+        "v100_ready"
+      ],
+      "license": "Apache-2.0",
+      "repo_id": "Qwen/Qwen2.5-32B-Instruct-GPTQ-Int4",
+      "local_path": "/root/models/qwen25-32b-gptq",
+      "served_name": "qwen2.5-32b-gptq",
+      "size_gb_estimate": 20,
+      "vllm_min_version": "0.4.0",
+      "vllm": {
+        "tensor_parallel_size": 4,
+        "quantization": "gptq",
+        "dtype": "float16",
+        "max_model_len": 8192,
+        "gpu_memory_utilization": 0.9,
+        "max_num_seqs": 4,
+        "extra_args": [
+          "--trust-remote-code",
+          "--enforce-eager"
+        ]
+      },
+      "runtime_id": "vllm-legacy-040",
+      "inference_backend": "vllm",
+      "v100_activate": true,
+      "v100_block_reason": null
+    },
+    {
+      "id": "deepseek-coder-33b-gptq",
+      "name": "DeepSeek Coder 33B GPTQ",
+      "description": "DeepSeek open weights (TheBloke GPTQ). Top-tier coder, TP=4.",
+      "tags": [
+        "coding",
+        "deepseek",
+        "opensource",
+        "v100_ready"
+      ],
+      "license": "DeepSeek License",
+      "repo_id": "TheBloke/deepseek-coder-33B-instruct-GPTQ",
+      "local_path": "/root/models/deepseek-coder-33b-gptq",
+      "served_name": "deepseek-coder-33b-gptq",
+      "size_gb_estimate": 18,
+      "vllm_min_version": "0.4.0",
+      "vllm": {
+        "tensor_parallel_size": 4,
+        "quantization": "gptq",
+        "dtype": "float16",
+        "max_model_len": 4096,
+        "gpu_memory_utilization": 0.88,
+        "max_num_seqs": 4,
+        "extra_args": [
+          "--trust-remote-code",
+          "--enforce-eager"
+        ]
+      },
+      "runtime_id": "vllm-legacy-040",
+      "inference_backend": "vllm",
+      "v100_activate": true,
+      "v100_block_reason": null
+    },
+    {
+      "id": "qwen25-coder-7b",
+      "name": "Qwen 2.5 Coder 7B",
+      "description": "Alibaba Qwen open weights. Fast code model, 1 GPU.",
+      "tags": [
+        "coding",
+        "opensource",
+        "qwen",
+        "v100_ready"
+      ],
+      "license": "Apache-2.0",
+      "repo_id": "Qwen/Qwen2.5-Coder-7B-Instruct",
+      "local_path": "/root/models/qwen25-coder-7b",
+      "served_name": "qwen2.5-coder-7b",
+      "size_gb_estimate": 15,
+      "vllm_min_version": "0.4.0",
+      "vllm": {
+        "tensor_parallel_size": 1,
+        "dtype": "float16",
+        "max_model_len": 8192,
+        "gpu_memory_utilization": 0.9,
+        "max_num_seqs": 8,
+        "extra_args": [
+          "--trust-remote-code",
+          "--enforce-eager"
+        ]
+      },
+      "runtime_id": "vllm-legacy-040",
+      "inference_backend": "vllm",
+      "v100_activate": true,
+      "v100_block_reason": null
+    },
+    {
+      "id": "deepseek-coder-6.7b",
+      "name": "DeepSeek Coder 6.7B",
+      "description": "DeepSeek open weights. Fast lightweight coder, 1 GPU.",
+      "tags": [
+        "coding",
+        "deepseek",
+        "opensource",
+        "v100_ready"
+      ],
+      "license": "DeepSeek License",
+      "repo_id": "deepseek-ai/deepseek-coder-6.7b-instruct",
+      "local_path": "/root/models/deepseek-coder-6.7b",
+      "served_name": "deepseek-coder-6.7b",
+      "size_gb_estimate": 14,
+      "vllm_min_version": "0.4.0",
+      "vllm": {
+        "tensor_parallel_size": 1,
+        "dtype": "float16",
+        "max_model_len": 8192,
+        "gpu_memory_utilization": 0.9,
+        "max_num_seqs": 8,
+        "extra_args": [
+          "--trust-remote-code",
+          "--enforce-eager"
+        ]
+      },
+      "runtime_id": "vllm-legacy-040",
+      "inference_backend": "vllm",
+      "v100_activate": true,
+      "v100_block_reason": null
+    },
+    {
+      "id": "deepseek-coder-7b",
+      "name": "DeepSeek Coder 7B v1.5",
+      "description": "DeepSeek open weights. Balanced 7B coder.",
+      "tags": [
+        "coding",
+        "deepseek",
+        "opensource",
+        "v100_ready"
+      ],
+      "license": "DeepSeek License",
+      "repo_id": "deepseek-ai/deepseek-coder-7b-instruct-v1.5",
+      "local_path": "/root/models/deepseek-coder-7b",
+      "served_name": "deepseek-coder-7b",
+      "size_gb_estimate": 15,
+      "vllm_min_version": "0.4.0",
+      "vllm": {
+        "tensor_parallel_size": 1,
+        "dtype": "float16",
+        "max_model_len": 8192,
+        "gpu_memory_utilization": 0.9,
+        "max_num_seqs": 8,
+        "extra_args": [
+          "--trust-remote-code",
+          "--enforce-eager"
+        ]
+      },
+      "runtime_id": "vllm-legacy-040",
+      "inference_backend": "vllm",
+      "v100_activate": true,
+      "v100_block_reason": null
+    },
+    {
+      "id": "starcoder2-15b",
+      "name": "StarCoder2 15B",
+      "description": "BigCode/Hugging Face. Apache-2.0 code specialist, 1\u20132 GPU.",
+      "tags": [
+        "coding",
+        "opensource",
+        "v100_ready"
+      ],
+      "license": "Apache-2.0",
+      "repo_id": "bigcode/starcoder2-15b",
+      "local_path": "/root/models/starcoder2-15b",
+      "served_name": "starcoder2-15b",
+      "size_gb_estimate": 30,
+      "vllm_min_version": "0.4.0",
+      "vllm": {
+        "tensor_parallel_size": 2,
+        "dtype": "float16",
+        "max_model_len": 8192,
+        "gpu_memory_utilization": 0.9,
+        "max_num_seqs": 4,
+        "extra_args": [
+          "--trust-remote-code",
+          "--enforce-eager"
+        ]
+      },
+      "runtime_id": "vllm-legacy-040",
+      "inference_backend": "vllm",
+      "v100_activate": true,
+      "v100_block_reason": null
+    },
+    {
+      "id": "granite-code-8b",
+      "name": "IBM Granite Code 8B",
+      "description": "IBM open weights. Enterprise-friendly Apache-2.0 coder.",
+      "tags": [
+        "coding",
+        "opensource",
+        "v100_ready"
+      ],
+      "license": "Apache-2.0",
+      "repo_id": "ibm-granite/granite-3.3-8b-instruct",
+      "local_path": "/root/models/granite-code-8b",
+      "served_name": "granite-code-8b",
+      "size_gb_estimate": 16,
+      "vllm_min_version": "0.4.0",
+      "vllm": {
+        "tensor_parallel_size": 1,
+        "dtype": "float16",
+        "max_model_len": 8192,
+        "gpu_memory_utilization": 0.9,
+        "max_num_seqs": 8,
+        "extra_args": [
+          "--trust-remote-code",
+          "--enforce-eager"
+        ]
+      },
+      "runtime_id": "vllm-legacy-040",
+      "inference_backend": "vllm",
+      "v100_activate": true,
+      "v100_block_reason": null
+    },
+    {
+      "id": "yi-coder-9b",
+      "name": "Yi-Coder 9B",
+      "description": "01.AI open weights. Strong 9B code model, 1 GPU.",
+      "tags": [
+        "coding",
+        "opensource",
+        "v100_ready"
+      ],
+      "license": "Apache-2.0",
+      "repo_id": "01-ai/Yi-Coder-9B-Chat",
+      "local_path": "/root/models/yi-coder-9b",
+      "served_name": "yi-coder-9b",
+      "size_gb_estimate": 18,
+      "vllm_min_version": "0.4.0",
+      "vllm": {
+        "tensor_parallel_size": 1,
+        "dtype": "float16",
+        "max_model_len": 8192,
+        "gpu_memory_utilization": 0.9,
+        "max_num_seqs": 8,
+        "extra_args": [
+          "--trust-remote-code",
+          "--enforce-eager"
+        ]
+      },
+      "runtime_id": "vllm-legacy-040",
+      "inference_backend": "vllm",
+      "v100_activate": true,
+      "v100_block_reason": null
+    },
+    {
+      "id": "deepseek-coder-v2-lite",
+      "name": "DeepSeek Coder V2 Lite",
+      "description": "DeepSeek MoE coder (16B/2.4B active). vLLM 0.8 on V100.",
+      "tags": [
+        "coding",
+        "deepseek",
+        "opensource",
+        "v100_ready"
+      ],
+      "license": "DeepSeek License",
+      "repo_id": "deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct",
+      "local_path": "/root/models/deepseek-coder-v2-lite",
+      "served_name": "deepseek-coder-v2-lite",
+      "size_gb_estimate": 32,
+      "vllm_min_version": "0.6.0",
+      "vllm": {
+        "tensor_parallel_size": 2,
+        "dtype": "bfloat16",
+        "max_model_len": 8192,
+        "gpu_memory_utilization": 0.9,
+        "max_num_seqs": 4,
+        "extra_args": [
+          "--trust-remote-code",
+          "--enforce-eager"
+        ]
+      },
+      "runtime_id": "vllm-modern-080",
+      "inference_backend": "vllm",
+      "v100_activate": true,
+      "v100_block_reason": null
+    },
+    {
+      "id": "qwen3-coder-next",
+      "name": "Qwen3 Coder Next",
+      "description": "Alibaba agentic coder (80B MoE, 3B active). vLLM 0.8 on V100.",
+      "tags": [
+        "coding",
+        "opensource",
+        "qwen",
+        "v100_ready"
+      ],
+      "license": "Apache-2.0",
+      "repo_id": "Qwen/Qwen3-Coder-Next",
+      "local_path": "/root/models/qwen3-coder-next",
+      "served_name": "qwen3-coder-next",
+      "size_gb_estimate": 45,
+      "vllm_min_version": "0.8.0",
+      "vllm": {
+        "tensor_parallel_size": 4,
+        "dtype": "bfloat16",
+        "max_model_len": 8192,
+        "gpu_memory_utilization": 0.9,
+        "max_num_seqs": 4,
+        "extra_args": [
+          "--trust-remote-code",
+          "--enforce-eager"
+        ]
+      },
+      "runtime_id": "vllm-modern-080",
+      "inference_backend": "vllm",
+      "v100_activate": true,
+      "v100_block_reason": null
+    },
+    {
+      "id": "qwen3-coder-30b",
+      "name": "Qwen3 Coder 30B A3B",
+      "description": "Alibaba MoE agentic coder. vLLM 0.8 on V100.",
+      "tags": [
+        "coding",
+        "opensource",
+        "qwen",
+        "v100_ready"
+      ],
+      "license": "Apache-2.0",
+      "repo_id": "Qwen/Qwen3-Coder-30B-A3B-Instruct",
+      "local_path": "/root/models/qwen3-coder-30b",
+      "served_name": "qwen3-coder-30b",
+      "size_gb_estimate": 60,
+      "vllm_min_version": "0.8.0",
+      "vllm": {
+        "tensor_parallel_size": 4,
+        "dtype": "bfloat16",
+        "max_model_len": 8192,
+        "gpu_memory_utilization": 0.9,
+        "max_num_seqs": 4,
+        "extra_args": [
+          "--trust-remote-code",
+          "--enforce-eager"
+        ]
+      },
+      "runtime_id": "vllm-modern-080",
+      "inference_backend": "vllm",
+      "v100_activate": true,
+      "v100_block_reason": null
+    },
+    {
+      "id": "qwen36-27b",
+      "name": "Qwen 3.6 27B",
+      "description": "Qwen open weights (dense). On V100 via SGLang \u2014 prefer AWQ variant.",
+      "tags": [
+        "coding",
+        "opensource",
+        "qwen",
+        "v100_via_sglang"
+      ],
+      "license": "Apache-2.0",
+      "repo_id": "Qwen/Qwen3.6-27B",
+      "local_path": "/root/models/qwen36-27b",
+      "served_name": "qwen3.6-27b",
+      "size_gb_estimate": 54,
+      "vllm_min_version": "0.19.0",
+      "vllm": {
+        "tensor_parallel_size": 4,
+        "dtype": "bfloat16",
+        "max_model_len": 8192,
+        "gpu_memory_utilization": 0.9,
+        "max_num_seqs": 4,
+        "extra_args": [
+          "--trust-remote-code",
+          "--enforce-eager"
+        ]
+      },
+      "runtime_id": "sglang-modern",
+      "inference_backend": "sglang",
+      "v100_activate": true,
+      "v100_block_reason": null
+    },
+    {
+      "id": "qwen36-27b-awq",
+      "name": "Qwen 3.6 27B AWQ",
+      "description": "Qwen open weights (AWQ). Primary Qwen 3.6 path on V100 via SGLang.",
+      "tags": [
+        "coding",
+        "opensource",
+        "qwen",
+        "v100_via_sglang"
+      ],
+      "license": "Apache-2.0",
+      "repo_id": "QuantTrio/Qwen3.6-27B-AWQ",
+      "local_path": "/root/models/qwen36-27b-awq",
+      "served_name": "qwen3.6-27b-awq",
+      "size_gb_estimate": 16,
+      "vllm_min_version": "0.19.0",
+      "vllm": {
+        "tensor_parallel_size": 4,
+        "quantization": "awq",
+        "dtype": "float16",
+        "max_model_len": 8192,
+        "gpu_memory_utilization": 0.9,
+        "max_num_seqs": 8,
+        "extra_args": [
+          "--trust-remote-code",
+          "--enforce-eager"
+        ]
+      },
+      "runtime_id": "sglang-modern",
+      "inference_backend": "sglang",
+      "v100_activate": true,
+      "v100_block_reason": null
+    },
+    {
+      "id": "codellama-34b",
+      "name": "Code Llama 34B",
+      "description": "Meta open weights (custom license). Classic code LLM \u2014 downloadable via HF.",
+      "tags": [
+        "coding",
+        "opensource",
+        "v100_ready"
+      ],
+      "license": "Llama 3 Community License",
+      "repo_id": "codellama/CodeLlama-34b-Instruct-hf",
+      "local_path": "/root/models/codellama-34b",
+      "served_name": "codellama-34b",
+      "size_gb_estimate": 68,
+      "vllm_min_version": "0.4.0",
+      "vllm": {
+        "tensor_parallel_size": 4,
+        "dtype": "float16",
+        "max_model_len": 4096,
+        "gpu_memory_utilization": 0.9,
+        "max_num_seqs": 2,
+        "extra_args": [
+          "--enforce-eager"
+        ]
+      },
+      "runtime_id": "vllm-legacy-040",
+      "inference_backend": "vllm",
+      "v100_activate": true,
+      "v100_block_reason": null
+    }
+  ]
+}
diff --git a/config/model-manager/vm-specs.json b/config/model-manager/vm-specs.json
new file mode 100644
index 0000000..ce4c337
--- /dev/null
+++ b/config/model-manager/vm-specs.json
@@ -0,0 +1,14 @@
+{
+  "vmid": 303,
+  "name": "atc-gpu-dev",
+  "proxmox_host": "atc-gpu",
+  "ip": "10.0.20.106",
+  "cpu_model": "Intel Xeon Gold 6148 @ 2.40GHz",
+  "vcpu": 32,
+  "sockets": 2,
+  "cores_per_socket": 16,
+  "memory_gib": 128,
+  "disk_gib": 250,
+  "gpu": "4× Tesla V100 PCIe 32GB",
+  "os": "Rocky Linux 9"
+}
diff --git a/docs/dockhand.md b/docs/dockhand.md
new file mode 100644
index 0000000..3f35932
--- /dev/null
+++ b/docs/dockhand.md
@@ -0,0 +1,12 @@
+# Dockhand — GPU Lab stacks
+
+| Dockhand env | ID | Docker host | Stacks |
+|--------------|-----|-------------|--------|
+| GPU-Dev | 8 | 10.0.20.106:2375 | gpu-lab-vllm-legacy, gpu-lab-vllm-modern, gpu-lab-sglang |
+| Spark+Trino / Lakehouse | 9 | 10.0.21.50 | lakehouse-spark-trino (Trino, Spark, kafka-connect, s3-kafka-consumer) |
+| Airflow | 10 | 10.0.21.55 | airflow |
+
+- Dockhand UI: http://atc-docker01.dell-atc.lan:8082/
+- API stacks: `GET /api/stacks?env=8`
+
+Only one GPU inference stack should be **running** at a time; others appear as stopped compose projects.
diff --git a/docs/landscape.md b/docs/landscape.md
new file mode 100644
index 0000000..37820c3
--- /dev/null
+++ b/docs/landscape.md
@@ -0,0 +1,30 @@
+# GPU Lab Landscape
+
+## Nodes
+
+| Host | IP | Role |
+|------|-----|------|
+| VM303 GPU-Dev | 10.0.20.106 | GPU Lab UI (:9000), inference (:8010), gateway (:8001) |
+| atc-docker01 | 10.0.21.45 | Homepage, Dockhand (:8082) |
+| Bart-GPU | — | Reference node (documented in model-manager) |
+| Mo-GPU | — | Reference node |
+
+## VM303 specs
+
+See `config/model-manager/vm-specs.json`: 4× Tesla V100 32GB, Docker with NVIDIA runtime.
+
+## Inference runtimes
+
+Only one runtime is active at a time on port **8010**:
+
+| Runtime ID | Engine | Compose file |
+|------------|--------|--------------|
+| `vllm-legacy-040` | vLLM 0.4.x | `compose.vllm-legacy.yml` |
+| `vllm-modern-080` | vLLM 0.8.x | `compose.vllm-modern.yml` |
+| `sglang-modern` | SGLang 0.5.12 | `compose.sglang.yml` |
+
+Models live under `/root/models/` on VM303 (not in git).
+
+## Docker data
+
+Images stored under `/root/docker-data` (avoid small `/var` partition).
diff --git a/docs/models.md b/docs/models.md
new file mode 100644
index 0000000..0326b66
--- /dev/null
+++ b/docs/models.md
@@ -0,0 +1,20 @@
+# Model catalog (VM303)
+
+| ID | Name | Runtime | Backend | V100 |
+|---|---|---|---|---|
+| `llama3-70b-gptq` | Llama 3 70B GPTQ | vllm-legacy-040 | vllm | yes |
+| `qwen25-coder-32b-gptq` | Qwen 2.5 Coder 32B GPTQ | vllm-legacy-040 | vllm | yes |
+| `qwen25-32b-gptq` | Qwen 2.5 32B GPTQ | vllm-legacy-040 | vllm | yes |
+| `deepseek-coder-33b-gptq` | DeepSeek Coder 33B GPTQ | vllm-legacy-040 | vllm | yes |
+| `qwen25-coder-7b` | Qwen 2.5 Coder 7B | vllm-legacy-040 | vllm | yes |
+| `deepseek-coder-6.7b` | DeepSeek Coder 6.7B | vllm-legacy-040 | vllm | yes |
+| `deepseek-coder-7b` | DeepSeek Coder 7B v1.5 | vllm-legacy-040 | vllm | yes |
+| `starcoder2-15b` | StarCoder2 15B | vllm-legacy-040 | vllm | yes |
+| `granite-code-8b` | IBM Granite Code 8B | vllm-legacy-040 | vllm | yes |
+| `yi-coder-9b` | Yi-Coder 9B | vllm-legacy-040 | vllm | yes |
+| `deepseek-coder-v2-lite` | DeepSeek Coder V2 Lite | vllm-modern-080 | vllm | yes |
+| `qwen3-coder-next` | Qwen3 Coder Next | vllm-modern-080 | vllm | yes |
+| `qwen3-coder-30b` | Qwen3 Coder 30B A3B | vllm-modern-080 | vllm | yes |
+| `qwen36-27b` | Qwen 3.6 27B | sglang-modern | sglang | yes |
+| `qwen36-27b-awq` | Qwen 3.6 27B AWQ | sglang-modern | sglang | yes |
+| `codellama-34b` | Code Llama 34B | vllm-legacy-040 | vllm | yes |
diff --git a/docs/runbook-vm303-source.md b/docs/runbook-vm303-source.md
new file mode 100644
index 0000000..13537ac
--- /dev/null
+++ b/docs/runbook-vm303-source.md
@@ -0,0 +1,61 @@
+# GPU Lab Runbook — VM303 (atc-gpu-dev)
+
+## Runtimes (Docker)
+
+| Runtime ID | Compose | Image |
+|---|---|---|
+| `vllm-legacy-040` | `compose.vllm-legacy.yml` | `vllm/vllm-openai:v0.4.0.post1` |
+| `vllm-modern-080` | `compose.vllm-modern.yml` | `vllm/vllm-openai:v0.8.5.post1` |
+| `sglang-modern` | `compose.sglang.yml` | `lmsysorg/sglang:v0.5.12-cu129-runtime` |
+
+Only **one** inference container runs at a time (port **8010**).
+
+## Activate (UI)
+
+GPU Lab → pick model → **Activate**. Writes `/root/gpu-lab/active/current.env`, stops all gpu-lab containers, starts the matching compose stack.
+
+## Manual activate
+
+```bash
+cd /root/gpu-lab
+# edit active/current.env (MODEL_PATH, TP_SIZE, VLLM_EXTRA_ARGS, …)
+docker compose -f compose.vllm-legacy.yml down
+docker compose -f compose.vllm-legacy.yml up -d
+curl -s http://127.0.0.1:8010/v1/models
+```
+
+## Rollback to systemd vLLM
+
+```bash
+cd /root/gpu-lab
+docker compose -f compose.vllm-legacy.yml down
+docker compose -f compose.vllm-modern.yml down
+docker compose -f compose.sglang.yml down
+systemctl start vllm.service
+systemctl restart vllm-gateway.service litellm-proxy.service
+```
+
+## Logs
+
+```bash
+docker logs -f gpu-lab-vllm-legacy
+docker logs -f gpu-lab-vllm-modern
+docker logs -f gpu-lab-sglang
+journalctl -u model-manager -f
+```
+
+## SGLang on V100
+
+Always use `--disable-flashinfer` and `--disable-flashinfer-sampling` (already in compose.sglang.yml).
+
+## Dockhand
+
+- Environment **GPU-Dev** → `10.0.20.106:2375`
+- Stacks: `gpu-lab-vllm-legacy`, `gpu-lab-vllm-modern`, `gpu-lab-sglang` (compose in Dockhand data volume)
+- Running workload appears as stack **`gpu-lab`** (compose project on VM303)
+- API: `GET /api/stacks?env=8` (not `environmentId`)
+- UI: http://atc-docker01.dell-atc.lan:8082/
+
+## Docker data
+
+Images stored under `/root/docker-data` (not `/var` — small partition).
diff --git a/docs/runbook-vm303.md b/docs/runbook-vm303.md
new file mode 100644
index 0000000..98781a9
--- /dev/null
+++ b/docs/runbook-vm303.md
@@ -0,0 +1,47 @@
+# GPU Lab Runbook — VM303 (atc-gpu-dev)
+
+See also `config/gpu-lab/RUNBOOK.md` (synced from VM303).
+
+## Activate via UI
+
+GPU Lab → select model → **Activate**. Writes `/root/gpu-lab/active/current.env`, stops other gpu-lab containers, starts matching compose stack.
+
+## Manual activate (SGLang + Qwen 3.6 27B AWQ)
+
+```bash
+cd /root/gpu-lab
+cat > active/current.env << 'EOF'
+MODEL_PATH=/models/qwen36-27b-awq
+SERVED_NAME=qwen3.6-27b-awq
+TP_SIZE=4
+GPU_MEM_UTIL=0.9
+MAX_MODEL_LEN=8192
+EOF
+docker compose -f compose.vllm-legacy.yml down
+docker compose -f compose.vllm-modern.yml down
+docker compose -f compose.sglang.yml up -d
+curl -s http://127.0.0.1:8010/v1/models
+systemctl restart vllm-gateway.service
+```
+
+## Rollback to systemd vLLM
+
+```bash
+cd /root/gpu-lab
+docker compose -f compose.vllm-legacy.yml down
+docker compose -f compose.vllm-modern.yml down
+docker compose -f compose.sglang.yml down
+systemctl start vllm.service
+systemctl restart vllm-gateway.service litellm-proxy.service
+```
+
+## Logs
+
+```bash
+docker logs -f gpu-lab-sglang
+journalctl -u model-manager -f
+```
+
+## SGLang on V100
+
+Always use `--disable-flashinfer` and `--disable-flashinfer-sampling` (in `compose.sglang.yml`).
diff --git a/scripts/activate-model.sh b/scripts/activate-model.sh
new file mode 100755
index 0000000..c10ff25
--- /dev/null
+++ b/scripts/activate-model.sh
@@ -0,0 +1,7 @@
+#!/usr/bin/env bash
+# Activate a model via GPU Lab API (run on VM303 or with curl to :9000)
+set -euo pipefail
+MODEL_ID="${1:?Usage: activate-model.sh <model-id>}"
+curl -sf -X POST "http://127.0.0.1:9000/api/activate/${MODEL_ID}"
+echo
+curl -sf "http://127.0.0.1:8010/v1/models" | python3 -m json.tool