Add fake Slurm MCP demo for Openclaw sandbox

slurm-mcp-demo/.env.template

@@ -0,0 +1,9 @@
+#refer to this https://docs.nvidia.com/nemoclaw/latest/inference/switch-inference-providers.html#switch-to-a-different-model to set different provider 
+# ── Inference API credentials ─────────────────────────────────────
+INFERENCE_API_KEY=
+
+# ── Inference provider configuration ─────────────────────────────
+INFERENCE_PROVIDER_TYPE=
+INFERENCE_PROVIDER_NAME=
+INFERENCE_BASE_URL=
+INFERENCE_MODEL=

slurm-mcp-demo/.gitignore

@@ -0,0 +1,6 @@
+.venv/
+.env
+fake_host_key
+__pycache__/
+*.pyc
+

slurm-mcp-demo/fake_cluster_mcp_server.py

@@ -0,0 +1,203 @@
+#!/usr/bin/env python3
+"""
+Fake Slurm HPC headnode — MCP server exposing raw Slurm tools.
+
+The OpenClaw agent (which has its own LLM) decides which tool to call.
+No secondary LLM or NL dispatcher runs here.
+
+Tools exposed:
+  get_hostname  — return headnode hostname
+  sinfo         — show partitions / node states
+  srun          — launch a fake interactive training job
+  sbatch        — submit a fake batch job
+  squeue        — show job queue
+  sacctmgr      — show account associations
+  sreport       — show utilisation report
+
+Run:
+  python fake_cluster_mcp_server.py              # streamable-http on 0.0.0.0:9000/mcp
+  python fake_cluster_mcp_server.py --port 9000
+"""
+from __future__ import annotations
+
+import argparse
+import os
+import random
+
+from colorama import Fore, init as colorama_init
+from fastmcp import FastMCP
+
+colorama_init(autoreset=True)
+
+mcp = FastMCP("fake-slurm-cluster")
+
+# ---------------------------------------------------------------------------
+# In-memory job table (persists for the lifetime of the server process)
+# ---------------------------------------------------------------------------
+_jobs: dict[int, dict] = {}
+_next_job_id = 42001
+
+# ---------------------------------------------------------------------------
+# Tool implementations
+# ---------------------------------------------------------------------------
+
+@mcp.tool()
+def get_hostname() -> str:
+    """Return the cluster headnode hostname."""
+    return "dlcluster-headnode"
+
+
+@mcp.tool()
+def sinfo() -> str:
+    """Show available Slurm partitions and node states."""
+    return (
+        "PARTITION    AVAIL  TIMELIMIT   NODES  STATE  NODELIST\n"
+        "gpu-a100*    up     infinite        4  idle   node[01-04]\n"
+        "gpu-h100     up     2-00:00:00      8  idle   node[05-12]\n"
+        "gpu-gb200    up     4-00:00:00      2  idle   node[13-14]\n"
+        "cpu-general  up     infinite       16  idle   node[15-30]\n"
+    )
+
+
+@mcp.tool()
+def srun(
+    gpus: int = 1,
+    time_limit: str = "01:00:00",
+    epochs: int = 5,
+    model: str = "resnet50",
+) -> str:
+    """Launch a fake interactive training job via srun.
+
+    Args:
+        gpus: Number of GPUs to allocate.
+        time_limit: Wall-time limit HH:MM:SS.
+        epochs: Training epochs to simulate.
+        model: Model name printed in the epoch log.
+    """
+    global _next_job_id
+    job_id = _next_job_id
+    _next_job_id += 1
+
+    lines = [
+        f"srun: job {job_id} queued and waiting for resources",
+        f"srun: job {job_id} has been allocated resources",
+        f"Allocated {gpus} GPU(s) on node01 | time_limit={time_limit}",
+        "",
+    ]
+    random.seed(job_id)
+    loss, acc = 3.2, 0.05
+    for epoch in range(1, epochs + 1):
+        loss -= random.uniform(0.2, 0.5)
+        acc  += random.uniform(0.05, 0.12)
+        lines.append(
+            f"Epoch [{epoch}/{epochs}]  loss={loss:.4f}  "
+            f"acc={min(acc, 1.0):.4f}  lr=1e-4  gpu_util=94%  model={model}"
+        )
+    lines += [
+        "",
+        f"Training complete. Checkpoints → /checkpoint/user/run_{job_id}/",
+    ]
+    _jobs[job_id] = {"state": "COMPLETED", "user": "user",
+                     "partition": "gpu-a100", "name": model}
+    return "\n".join(lines)
+
+
+@mcp.tool()
+def sbatch(script_name: str = "train.sh") -> str:
+    """Submit a fake batch job.
+
+    Args:
+        script_name: Name of the batch script.
+    """
+    global _next_job_id
+    job_id = _next_job_id
+    _next_job_id += 1
+    _jobs[job_id] = {"state": "RUNNING", "user": "user",
+                     "partition": "gpu-a100", "name": script_name}
+    return f"Submitted batch job {job_id}"
+
+
+@mcp.tool()
+def squeue(user: str = "user") -> str:
+    """Show jobs in the Slurm queue.
+
+    Args:
+        user: Filter by username; "all" to see every job.
+    """
+    header = (
+        "             JOBID PARTITION     NAME     USER  ST       TIME  NODES NODELIST\n"
+    )
+    rows = []
+    for jid, info in _jobs.items():
+        if user == "all" or info["user"] == user:
+            st = "R" if info["state"] == "RUNNING" else "CG"
+            rows.append(
+                f"             {jid:>5}  {info['partition']:<10} "
+                f"{info['name']:<8} {info['user']:<8}  {st}  0:01       1 node01"
+            )
+    return header + ("\n".join(rows) if rows else "(no jobs)")
+
+
+@mcp.tool()
+def sacctmgr(user: str = "user") -> str:
+    """Show Slurm account associations for a user.
+
+    Args:
+        user: Username to query.
+    """
+    return (
+        "   Cluster    Account       User  Partition  Share  MaxJobs        QOS\n"
+        "---------- ---------- --------- ---------- ------  -------  ---------\n"
+        "dlcluster        root                            1             normal\n"
+        "dlcluster        root      root                  1             normal\n"
+        f"dlcluster     {user:<10}                    1             normal\n"
+        f"dlcluster     {user:<10} {user:<9}            1      200    normal\n"
+    )
+
+
+@mcp.tool()
+def sreport(user: str = "user") -> str:
+    """Show cluster utilisation report for a user.
+
+    Args:
+        user: Username to query.
+    """
+    return (
+        "-----------------------------------------------------------\n"
+        "Cluster/Account/User Utilization 2024-01-01 - 2024-01-31\n"
+        "Usage reported in CPU Minutes\n"
+        "-----------------------------------------------------------\n"
+        "   Cluster         Account     Login       Used\n"
+        "--------- --------------- --------- ----------\n"
+        "dlcluster            root                12,400\n"
+        f"dlcluster      {user:<12} {user:<10}  298,102\n"
+    )
+
+
+# ---------------------------------------------------------------------------
+# Entry point
+# ---------------------------------------------------------------------------
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(
+        description="Fake Slurm MCP Server",
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter,
+    )
+    parser.add_argument("--host", default=os.environ.get("MCP_HOST", "0.0.0.0"))
+    parser.add_argument("--port", type=int, default=int(os.environ.get("MCP_PORT", "9000")))
+    parser.add_argument("--path", default=os.environ.get("MCP_PATH", "/mcp"))
+    args = parser.parse_args()
+
+    print(
+        Fore.GREEN +
+        f"[mcp-server] fake-slurm-cluster  →  "
+        f"http://{args.host}:{args.port}{args.path}"
+    )
+    print(Fore.YELLOW + "[mcp-server] Reachable from sandbox via host's LAN/bridge IP on that port.")
+    mcp.run(
+        transport="streamable-http",
+        host=args.host,
+        port=args.port,
+        path=args.path,
+        show_banner=False,
+    )