feat(autotuner): autonomous kernel and inference configuration tuning for AMD GPUs

.github/benchmark/vllm-models.json

@@ -0,0 +1,32 @@
+[
+  {
+    "display": "DeepSeek-R1-0528",
+    "path": "deepseek-ai/DeepSeek-R1-0528",
+    "prefix": "deepseek-r1-0528",
+    "args": "--kv-cache-dtype fp8 --tensor-parallel-size 8",
+    "bench_args": "",
+    "suffix": "",
+    "runner": "atom-mi355-8gpu.predownload",
+    "env_vars": ""
+  },
+  {
+    "display": "GLM-5-FP8",
+    "path": "zai-org/GLM-5-FP8",
+    "prefix": "glm-5-fp8",
+    "args": "--kv-cache-dtype fp8 --tensor-parallel-size 8",
+    "bench_args": "",
+    "suffix": "",
+    "runner": "atom-mi355-8gpu.predownload",
+    "env_vars": ""
+  },
+  {
+    "display": "Kimi-K2-Thinking-MXFP4",
+    "path": "amd/Kimi-K2-Thinking-MXFP4",
+    "prefix": "kimi-k2-thinking-mxfp4",
+    "args": "--trust-remote-code --kv-cache-dtype fp8 --tensor-parallel-size 4 --enable-expert-parallel",
+    "bench_args": "",
+    "suffix": "",
+    "runner": "atom-mi355-8gpu.predownload",
+    "env_vars": ""
+  }
+]

.github/workflows/atom-test.yaml

@@ -11,6 +11,15 @@ on:
       - 'docs/**'
       - 'LICENSE'
       - '.gitignore'
+      - 'scripts/**'
+      - '.github/dashboard/**'
+      - '.github/benchmark/vllm*'
+      - '.github/benchmark/oot_*'
+      - '.github/workflows/vllm-benchmark.yaml'
+      - '.github/workflows/atom-vllm-oot-benchmark.yaml'
+      - '.github/workflows/atom-benchmark.yaml'
+      - '.github/workflows/docker-release.yaml'
+      - '.github/workflows/gpu-load-test.yaml'
   schedule:
     # Nightly at 00:00 Beijing time (16:00 UTC)
     - cron: '0 16 * * *'

.github/workflows/atom-vllm-oot-test.yaml

@@ -9,6 +9,14 @@ on:
       - 'docs/**'
       - 'LICENSE'
       - '.gitignore'
+      - 'scripts/**'
+      - '.github/dashboard/**'
+      - '.github/benchmark/vllm*'
+      - '.github/benchmark/oot_*'
+      - '.github/workflows/vllm-benchmark.yaml'
+      - '.github/workflows/atom-benchmark.yaml'
+      - '.github/workflows/docker-release.yaml'
+      - '.github/workflows/gpu-load-test.yaml'
   schedule:
     # Nightly at 02:00 Beijing time (18:00 UTC on the previous day)
     - cron: '0 18 * * *'