docs(examples): demo workload_calculator in vllm and train_mnist

Use max_tokens as the workload for the vllm generate() endpoint (the
LLM-standard metric) and input pixel count for the train_mnist infer()
endpoint. Left off the square example since its cost is constant.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: vastzuby

examples/deployments/train_mnist/deploy.py

@@ -64,7 +64,13 @@ async def __aexit__(self, *exc):
         pass
 
 
-@app.remote(benchmark_dataset=[{"pixel_values": [[0.0] * 28] * 28}])
+# Optional: workload = input pixel count, so bigger images count as more work.
+@app.remote(
+    benchmark_dataset=[{"pixel_values": [[0.0] * 28] * 28}],
+    workload_calculator=lambda pixel_values: float(
+        len(pixel_values) * len(pixel_values[0])
+    ),
+)
 async def infer(pixel_values: list[list[float]]) -> dict:
     """Classify a 28x28 grayscale MNIST image.
 

examples/deployments/vllm/deploy.py

@@ -24,7 +24,11 @@ async def __aexit__(self, *exc):
         self.engine.shutdown_background_loop()
 
 
-@app.remote(benchmark_dataset=[{"prompt": "Hello"}])
+# Optional: workload = max_tokens, so the autoscaler sizes by total tokens/sec.
+@app.remote(
+    benchmark_dataset=[{"prompt": "Hello"}],
+    workload_calculator=lambda prompt, max_tokens=128: float(max_tokens),
+)
 async def generate(prompt: str, max_tokens: int = 128) -> str:
     from vllm import SamplingParams
     import uuid