LLM benchmarking: replace vLLM with SGLang

cloud-infrastructure/compute-including-hpc/ai-infra-gpu/ai-infrastructure/llm-benchmark-docker/files/compose.yaml

@@ -1,17 +1,18 @@
 # Copyright (c) 2025 Oracle and/or its affiliates.
-version: "3"
 services:
   llm:
-    image: vllm/vllm-openai:v0.8.5.post1
+    image: lmsysorg/sglang:latest
     container_name: llm
     runtime: nvidia
+    ipc: host
     volumes:
       - "$HOME/.cache/huggingface:/huggingface"
       - "$PWD:/appli"
     ports:
-      - "127.0.0.1:8000:8000"
+      - "30000:30000"
     environment:
       "HF_HOME": "/huggingface"
+      "NCCL_IB_DISABLE": "1"
     working_dir: "/appli"
     entrypoint:
       - "/appli/scripts/startllm.py"
@@ -29,9 +30,11 @@ services:
     command:
       - "wait-for-it.sh"
       - "--timeout=300"
-      - "llm:8000"
+      - "llm:30000"
       - "--"
       - "/appli/scripts/benchmark.py"
+      - "--scenario"
+      - "gen_heavy"
   plot:
     build: plot
     container_name: plot

cloud-infrastructure/compute-including-hpc/ai-infra-gpu/ai-infrastructure/llm-benchmark-docker/files/config.json

@@ -1,7 +1,9 @@
 {
-    "model": "meta-llama/Llama-3.1-8B-Instruct",
-    "gpu_memory_utilization": 0.98,
-    "tensor_parallel_size": 1,
-    "max_model_len": 8192,
-    "max_num_batched_tokens": 8192
+    "model": "openai/gpt-oss-120b",
+    "tensor-parallel-size": 2,
+    "genai-perf": {
+	"inputs": {
+	    "reasoning_effort": "low"
+	}
+    }
 }

cloud-infrastructure/compute-including-hpc/ai-infra-gpu/ai-infrastructure/llm-benchmark-docker/files/scripts/benchmark.py

@@ -85,7 +85,7 @@ def get_shape():
 parser.add_argument(
     "--port",
     type=int,
-    default=8000,
+    default=30_000,
     help="Port to use for vLLM.",
 )
 parser.add_argument(
@@ -106,6 +106,10 @@ def get_shape():
 model_name = server_configuration["model"].rsplit("/", 1)[-1]
 model_stub = model_name.lower()
 
+extra_inputs = []
+for k, v in server_configuration.get("genai-perf", {}).get("inputs", {}).items():
+    extra_inputs.extend(["--extra-inputs", f"{k}:{v}"])
+
 benchmark_command = [
     "genai-perf",
     "profile",
@@ -151,7 +155,7 @@ def get_shape():
             f"--extra-inputs=min_tokens:{tokens_mean - tokens_stddev}",
         ] + [
             f"--{k}={v}" for k, v in scfg.items()
-        ]
+        ] + extra_inputs
 
         logging.warning("running: %s", " ".join(cmd))
         subprocess.check_call(cmd)

cloud-infrastructure/compute-including-hpc/ai-infra-gpu/ai-infrastructure/llm-benchmark-docker/files/scripts/startllm.py

@@ -11,8 +11,8 @@
 parser.add_argument(
     "--port",
     type=int,
-    default=8000,
-    help="Port to use for vLLM.",
+    default=30000,
+    help="Port to use for sglang.",
 )
 parser.add_argument(
     "server_configuration",
@@ -26,12 +26,19 @@
 with args.server_configuration.open() as fd:
     configuration = json.load(fd)
 
+# Remove performance measurement parameters
+configuration.pop("genai-perf", None)
+
 model = configuration.pop("model")
 server_command = [
-    "vllm",
-    "serve",
-    "--disable-log-requests",
+    "python3",
+    "-m",
+    "sglang.launch_server",
+    "--log-requests-level",
+    "0",
+    "--host=0.0.0.0",
     f"--port={args.port}",
+    "--model-path",
     model,
 ] + [
     f"--{k}={v}" for k, v in configuration.items()