[ci] More aggressively tune the GPU and benchmark

.ci/gpu/reset-gcp-h100.sh

@@ -0,0 +1,13 @@
+#!/usr/bin/env bash
+# Script to tune NVIDIA H100 GPU on GCP
+# To reset GPU status
+
+# Reset GPU and Memory clocks
+sudo nvidia-smi -rgc
+sudo nvidia-smi -rmc
+
+# Restore the default power limit (500W)
+sudo nvidia-smi -pl 500
+
+# Disable persistent mode
+sudo nvidia-smi -pm 0

.ci/gpu/tune-gcp-h100.sh

@@ -0,0 +1,16 @@
+#!/usr/bin/env bash
+# Script to tune NVIDIA H100 GPU on GCP
+# To stablize performance
+
+set -ex
+
+# Enable persistent mode
+sudo nvidia-smi -pm 1
+# Lock power limit to 650W
+sudo nvidia-smi -pl 650
+
+# Default Memory Frequency: 2619 MHz
+# Default Graphics Frequency: 1980 MHz
+sudo nvidia-smi -lgc 1980,1980
+sudo nvidia-smi -lmc 2619,2619
+sudo nvidia-smi -ac 2619,1980

.github/workflows/_linux-benchmark-h100.yml

@@ -36,7 +36,7 @@ jobs:
           submodules: recursive
       - name: Tune Nvidia GPU
         run: |
-          sudo nvidia-smi -pm 1
+          bash .ci/gpu/tune-gcp-h100.sh
           sudo ldconfig
           nvidia-smi
       - name: Benchmarking
@@ -52,4 +52,10 @@ jobs:
         run: |
           . "${SETUP_SCRIPT}"
           latest_result_json=$(find ./benchmark-output/ -name "result.json"  | sort -r | head -n 1)
-          python .ci/upload/scribe.py --json ${latest_result_json}
+          python ./.ci/upload/scribe.py --json ${latest_result_json}
+      - name: Restore Nvidia GPU
+        if: always()
+        run: |
+          bash .ci/gpu/reset-gcp-h100.sh
+          sudo ldconfig
+          nvidia-smi

benchmarks/nightly/run.py

@@ -46,6 +46,7 @@ def setup_tritonbench_cwd():
         "latency,gbps",
         "--num-inputs",
         "6",
+        "--cudagraph",
     ],
     "bf16_gemm": [
         "--op",
@@ -58,6 +59,7 @@ def setup_tritonbench_cwd():
         "latency,tflops",
         "--num-inputs",
         "4",
+        "--cudagraph",
     ],
 }