LMStudio: select GPU runtime in CI, harden ping test

playbooks/core/lmstudio-rocm-llms/README.md

@@ -73,28 +73,63 @@ Learn how to start chatting with a ChatGPT-grade LLM completely locally.
 7. If not in the chat window, press `Ctrl + 1` or click on the 👾 button on the top left of the screen.
 8. Send a message and start interacting with the model!
 
+<!-- @os:windows -->
+<!-- @test:id=lmstudio-select-gpu-runtime-windows timeout=120 hidden=True -->
+```powershell
+# CI: pin a GPU (Vulkan) runtime so tests don't fall back to the CPU engine.
+lms runtime ls
+$rt = ((lms runtime ls) -match 'vulkan' | Select-Object -First 1)
+if ($rt) {
+  lms runtime select (($rt.Trim() -split '\s+')[0])
+  lms runtime ls | Select-String 'ENGINE|✓'
+} else {
+  Write-Output "WARNING: no Vulkan runtime installed; GPU acceleration unavailable. Install with: lms get <vulkan-runtime>"
+}
+```
+<!-- @test:end -->
+<!-- @os:end -->
+
 <!-- @os:windows -->
 <!-- @test:id=lmstudio-load-model-windows timeout=1200 hidden=True -->
 ```powershell
 lms unload --all
 lms ps
 $ID = "${lms_model}-$env:GITHUB_RUN_ID"
 Set-Content -Path "$env:TEMP\lmstudio_model_id.txt" -Value $ID -Encoding utf8
+# retry once: large-model loads can transiently fail under memory pressure
 lms load ${lms_model} --context-length 32768 --gpu max --identifier "$ID" -y
+if ($LASTEXITCODE -ne 0) { lms unload --all; Start-Sleep 5; lms load ${lms_model} --context-length 32768 --gpu max --identifier "$ID" -y }
 lms ps
 lms chat "$ID" -p "Reply with exactly: OK"
 ```
 <!-- @test:end -->
 <!-- @os:end -->
 
+<!-- @os:linux -->
+<!-- @test:id=lmstudio-select-gpu-runtime-linux timeout=120 hidden=True -->
+```bash
+# CI: pin a GPU (Vulkan) runtime so tests don't fall back to the CPU engine.
+lms runtime ls
+GPU_RT="$(lms runtime ls 2>/dev/null | awk '/vulkan/{print $1; exit}')"
+if [ -n "$GPU_RT" ]; then
+  lms runtime select "$GPU_RT"
+  lms runtime ls | grep -E 'ENGINE|✓'
+else
+  echo "WARNING: no Vulkan runtime installed; GPU acceleration unavailable. Install with: lms get <vulkan-runtime>"
+fi
+```
+<!-- @test:end -->
+<!-- @os:end -->
+
 <!-- @os:linux -->
 <!-- @test:id=lmstudio-load-model-linux timeout=1200 hidden=True -->
 ```bash
 lms unload --all || true
 lms ps
 ID="${lms_model}-${GITHUB_RUN_ID}"
 echo "$ID" > /tmp/lmstudio_model_id.txt
-lms load ${lms_model} --context-length 32768 --gpu max --identifier "$ID" -y
+# retry once: large-model loads can transiently fail under memory pressure
+lms load ${lms_model} --context-length 32768 --gpu max --identifier "$ID" -y || { lms unload --all; sleep 5; lms load ${lms_model} --context-length 32768 --gpu max --identifier "$ID" -y; }
 lms ps # Verify model is really loaded
 lms chat "$ID" -p "Reply with exactly: OK"
 ```
@@ -265,12 +300,12 @@ req = urllib.request.Request(
    "model": model_id,
    "messages": [{"role":"user","content":"What is 2 + 2? Reply with only the number."}],
    "temperature": 0,
-   "max_tokens": 500
+   "max_tokens": 64
  }).encode("utf-8"),
  headers={"Content-Type":"application/json"},
  method="POST",
 )
-with urllib.request.urlopen(req, timeout=60) as r:
+with urllib.request.urlopen(req, timeout=120) as r:
  print(r.read().decode("utf-8", "replace"))
 ```
 <!-- @test:end --> 
@@ -290,12 +325,12 @@ req = urllib.request.Request(
    "model": model_id,
    "messages": [{"role":"user","content":"What is 47 + 42? Reply with only the number in words."}],
    "temperature": 0,
-   "max_tokens": 500
+   "max_tokens": 64
  }).encode("utf-8"),
  headers={"Content-Type":"application/json"},
  method="POST",
 )
-with urllib.request.urlopen(req, timeout=60) as r:
+with urllib.request.urlopen(req, timeout=120) as r:
  print(r.read().decode("utf-8", "replace"))
 ```
 <!-- @test:end -->