Federate serving-llms-on-epyc from amd/ZenDNN-pytorch-plugin

.github/federation.json

@@ -58,6 +58,16 @@
           "path": "skills/lemonade-router-builder"
         }
       ]
+    },
+    {
+      "repo": "amd/ZenDNN-pytorch-plugin",
+      "license": "Apache-2.0",
+      "skills": [
+        {
+          "path": "skills/serving-llms-on-epyc",
+          "as": "serving-llms-on-epyc"
+        }
+      ]
     }
   ]
 }

skills/serving-llms-on-epyc/.federated.json

@@ -0,0 +1,10 @@
+{
+  "source": "amd-zendnn-pytorch-plugin",
+  "repo": "amd/ZenDNN-pytorch-plugin",
+  "ref": "main",
+  "commit": "6b2db6f2a47cc9b603570f6b4a93c37e5d04dba7",
+  "path": "skills/serving-llms-on-epyc",
+  "license": "Apache-2.0",
+  "content_hash": "eaa5899ac5e840f765fb855097d86ba9748f6dc6d0892a55e2394aaf37599733",
+  "imported_at": "2026-10-09T10:09:25Z"
+}

skills/serving-llms-on-epyc/reference.md

@@ -15,7 +15,7 @@
 ## Hardware support
 
 This recipe supports the **AMD EPYC 9000 server series** for now: Genoa (9004),
-Turin (9005), and 6th Gen [Venice (9006)](https://ir.amd.com/news-events/press-releases/detail/1294/aai-2026-amd-delivers-full-stack-compute-for-the-agentic-ai-era)
+Turin (9005), and 6th Gen Venice (9006)
 (launched at Advancing AI 2026). `scripts/detect.py` reports only these three
 generations as `is_supported_epyc: true`.
 

skills/serving-llms-on-epyc/scripts/check_model.py

@@ -203,18 +203,18 @@ def main():
         # Cannot read the config (gated/offline) -- do not positively block; the
         # gating check and launch will catch real problems.
         print(json.dumps({"model_id": a.model_id, "supported": None, "kind": "undetermined",
-                           "message": f"Could not read architectures ({aerr or 'none declared'}); support unverified. "
-                                      "If gated, set HF_TOKEN. This does not bypass the gating/launch checks."}, indent=2))
+                          "message": f"Could not read architectures ({aerr or 'none declared'}); support unverified. "
+                                     "If gated, set HF_TOKEN. This does not bypass the gating/launch checks."}, indent=2))
         sys.exit(0)
 
     reg, source = registry_from_github(a.vllm_version)
     if reg is None:
         reg, source = registry_from_local()
     if reg is None:
         print(json.dumps({"model_id": a.model_id, "architectures": archs, "supported": None,
-                           "kind": "undetermined",
-                           "message": "Could not load vLLM's model registry (no network and no importable vllm); "
-                                      "support unverified. vLLM confirms support at load (no-retry rule applies)."}, indent=2))
+                          "kind": "undetermined",
+                          "message": "Could not load vLLM's model registry (no network and no importable vllm); "
+                                     "support unverified. vLLM confirms support at load (no-retry rule applies)."}, indent=2))
         sys.exit(0)
 
     kinds = [reg.get(arch) for arch in archs]

skills/serving-llms-on-epyc/scripts/cpu_tune.py

@@ -40,9 +40,9 @@
 import argparse
 import json
 import re
+import shlex
 import shutil
 import subprocess
-import sys
 import time
 
 OS_HEADROOM_GB = 16
@@ -111,14 +111,17 @@ def socket_busy_pct(cpus, interval=0.5):
     """Mean CPU-busy% across `cpus` over `interval` seconds, from /proc/stat."""
     def snap():
         d = {}
-        for ln in open("/proc/stat"):
-            if ln.startswith("cpu") and len(ln) > 3 and ln[3].isdigit():
-                p = ln.split()
-                vals = list(map(int, p[1:]))
-                idle = vals[3] + (vals[4] if len(vals) > 4 else 0)
-                d[int(p[0][3:])] = (idle, sum(vals))
+        with open("/proc/stat") as f:
+            for ln in f:
+                if ln.startswith("cpu") and len(ln) > 3 and ln[3].isdigit():
+                    p = ln.split()
+                    vals = list(map(int, p[1:]))
+                    idle = vals[3] + (vals[4] if len(vals) > 4 else 0)
+                    d[int(p[0][3:])] = (idle, sum(vals))
         return d
-    a = snap(); time.sleep(interval); b = snap()
+    a = snap()
+    time.sleep(interval)
+    b = snap()
     di = sum(b[c][0] - a[c][0] for c in cpus if c in a and c in b)
     dt = sum(b[c][1] - a[c][1] for c in cpus if c in a and c in b)
     return round(100 * (1 - di / dt), 1) if dt else 0.0
@@ -157,8 +160,10 @@ def main():
         else:
             chosen = min(sids, key=lambda s: busy[s])
             reason = f"all sockets busy (>={args.busy_threshold}%) -> least-busy"
+            # Dict is built outside the f-string: "{ {...} }" needs spaces that flake8 rejects (E201/E202).
+            busy_map = {s: busy[s] for s in sids}
             warn = (f"all {len(sids)} sockets are busy (>= {args.busy_threshold}%): "
-                    f"{ {s: busy[s] for s in sids} }. Proceeding on the least-busy socket "
+                    f"{busy_map}. Proceeding on the least-busy socket "
                     f"{chosen}; performance may suffer. Pass --socket N to override.")
 
     sock = socks[chosen]
@@ -213,7 +218,7 @@ def main():
         print(json.dumps(result, indent=2))
         return
 
-    print(f'export VLLM_CPU_OMP_THREADS_BIND="{bind}"')
+    print(f"export VLLM_CPU_OMP_THREADS_BIND={shlex.quote(bind)}")
     print(f"export VLLM_CPU_KVCACHE_SPACE={kv}")
     print(f"# socket {chosen} ({reason}); per-socket busy%: {busy}")
     print(f"#   container: {container_cpuset}")

skills/serving-llms-on-epyc/scripts/validate.py

@@ -109,7 +109,7 @@ def _probe_stack(run_prefix, source):
     labels which path was probed. Returns the parsed stack dict (with `source`)
     or (None, message) if the probe could not run or produced no JSON.
     """
-    cmd = f'{run_prefix}python -c "{PROBE}"'
+    cmd = f"{run_prefix}python -c {shlex.quote(PROBE)}"
     rc, out, err = _sh(cmd, timeout=120)
     if rc != 0 or not out:
         return None, (err or "probe failed")[:200]
@@ -288,8 +288,10 @@ def main():
         missing = [lib for lib in ("libtcmalloc", "libiomp") if lib not in ld]
         if missing:
             issues.append({"check": "perf_libs", "severity": "advisory",
-                           "message": f"LD_PRELOAD is missing {', '.join(missing)}; vLLM CPU warns about this and throughput suffers without them (host/conda path).",
-                           "fix": "export LD_PRELOAD=/usr/lib/x86_64-linux-gnu/libtcmalloc_minimal.so.4:$CONDA_PREFIX/lib/libiomp5.so:$LD_PRELOAD"})
+                           "message": f"LD_PRELOAD is missing {', '.join(missing)}; vLLM CPU warns about this and "
+                                      "throughput suffers without them (host/conda path).",
+                           "fix": "export LD_PRELOAD=/usr/lib/x86_64-linux-gnu/libtcmalloc_minimal.so.4:"
+                                  "$CONDA_PREFIX/lib/libiomp5.so:$LD_PRELOAD"})
 
     # 8. HF cache mount (container path). The default bind-mounts the HF cache into
     #    the container; on NFS homes / symlinked caches / root-squash that fails at

skills/serving-llms-on-epyc/skill-card.md

@@ -10,4 +10,4 @@ AMD
 
 ## License
 
-MIT
+Apache-2.0