feat: update sglang_runner

.gitignore

@@ -21,4 +21,7 @@ configs/runner_configs/*.yaml
 # Local-only benchmark artifacts (not needed for submission)
 accuracy_outputs.jsonl
 run.log
-samples.jsonl
+samples.jsonl.ipynb_checkpoints/
+*_backup/
+backup/
+.ipynb_checkpoints/

runners/nvidia_sglang_c43a8309/meta.json

@@ -0,0 +1,11 @@
+{
+  "id": "nvidia_sglang_c43a8309",
+  "platform": "nvidia",
+  "name": "SGLang on NVIDIA",
+  "framework": "SGLang",
+  "submitted_by": "Gong-K",
+  "description": "AccelMark runner for NVIDIA GPUs using SGLang. Supports all suites (A–G). Enables direct throughput and latency comparison between SGLang and vLLM on the same hardware.",
+  "supersedes_chain": [],
+  "notes": "Decouple runners from suite and scenario knowledge — load_model() uses use_async from parallelism dict instead of checking scenario name.",
+  "created": "2026-04-03"
+}

runners/nvidia_sglang_c43a8309/requirements.txt

@@ -0,0 +1,38 @@
+# AccelMark — NVIDIA SGLang runner dependencies
+# Tested combination: torch 2.9.1 + sglang 0.4.x + CUDA 12.1
+#
+# SGLang install — use the official release wheel:
+#   pip install sglang[all]
+# or for a pinned version:
+#   pip install "sglang[all]==0.5.6"
+#
+# See https://sgl-project.github.io/start/install.html for full options.
+
+# Core
+torch==2.9.1
+torchvision==0.24.1
+torchaudio==2.9.1
+
+# LLM inference
+sglang[all]>=0.5.6
+
+# Transformers (for tokenizer)
+transformers==4.57.1
+tokenizers>=0.22.2
+huggingface-hub>=0.36.2
+accelerate>=1.13.0
+
+# AccelMark dependencies
+numpy>=2.2.6
+jsonschema>=4.26.0
+psutil>=7.2.2
+tqdm>=4.67.3
+
+# NVIDIA monitoring
+nvidia-ml-py>=13.595.45
+
+# Async support
+aiohttp>=3.13.5
+
+# Config file parsing
+pyyaml>=6.0