Update

Author: JewelRoam

graph_net/analysis.py

@@ -66,10 +66,10 @@ def analysis(args):
     data = {"Compiler": [], "log2(speedup)": []}
 
     # A: CINN (Simulate)
-    data["log2(speedup)"].extend(
-        np.random.normal(loc=0.35, scale=0.2, size=num_samples_per_compiler)
-    )
-    data["Compiler"].extend(["CINN"] * num_samples_per_compiler)
+    # data["log2(speedup)"].extend(
+    #     np.random.normal(loc=0.35, scale=0.2, size=num_samples_per_compiler)
+    # )
+    # data["Compiler"].extend(["CINN"] * num_samples_per_compiler)
 
     # B: torch.inductor
     # inductor_log = os.path.join(args.test_compiler_log_file)
@@ -80,37 +80,37 @@ def analysis(args):
     data["Compiler"].extend(["torch.inductor"] * len(log2_speedups))
 
     # C: tvm (Simulate)
-    data["log2(speedup)"].extend(
-        np.random.normal(loc=0.3, scale=0.15, size=num_samples_per_compiler)
-    )
-    data["Compiler"].extend(["tvm"] * num_samples_per_compiler)
+    # data["log2(speedup)"].extend(
+    #     np.random.normal(loc=0.3, scale=0.15, size=num_samples_per_compiler)
+    # )
+    # data["Compiler"].extend(["tvm"] * num_samples_per_compiler)
 
     # D: XLA (Simulate)
-    data["log2(speedup)"].extend(
-        np.concatenate(
-            [
-                np.random.normal(
-                    loc=-0.5, scale=0.1, size=int(num_samples_per_compiler * 0.6)
-                ),
-                np.random.normal(
-                    loc=0.2, scale=0.2, size=int(num_samples_per_compiler * 0.4)
-                ),
-            ]
-        )
-    )
-    data["Compiler"].extend(["XLA"] * num_samples_per_compiler)
+    # data["log2(speedup)"].extend(
+    #     np.concatenate(
+    #         [
+    #             np.random.normal(
+    #                 loc=-0.5, scale=0.1, size=int(num_samples_per_compiler * 0.6)
+    #             ),
+    #             np.random.normal(
+    #                 loc=0.2, scale=0.2, size=int(num_samples_per_compiler * 0.4)
+    #             ),
+    #         ]
+    #     )
+    # )
+    # data["Compiler"].extend(["XLA"] * num_samples_per_compiler)
 
     # E: TensorRT (Simulate)
-    data["log2(speedup)"].extend(
-        np.random.normal(loc=0.5, scale=0.1, size=num_samples_per_compiler)
-    )
-    data["Compiler"].extend(["TensorRT"] * num_samples_per_compiler)
+    # data["log2(speedup)"].extend(
+    #     np.random.normal(loc=0.5, scale=0.1, size=num_samples_per_compiler)
+    # )
+    # data["Compiler"].extend(["TensorRT"] * num_samples_per_compiler)
 
     # F: BladeDISC (Simulate)
-    data["log2(speedup)"].extend(
-        np.random.normal(loc=0.05, scale=0.3, size=num_samples_per_compiler)
-    )
-    data["Compiler"].extend(["BladeDISC"] * num_samples_per_compiler)
+    # data["log2(speedup)"].extend(
+    #     np.random.normal(loc=0.05, scale=0.3, size=num_samples_per_compiler)
+    # )
+    # data["Compiler"].extend(["BladeDISC"] * num_samples_per_compiler)
 
     df = pd.DataFrame(data)
     df["Compiler"] = pd.Categorical(df["Compiler"], categories=compilers, ordered=True)

graph_net/benchmark_demo.sh

@@ -1,20 +1,41 @@
 #!/bin/bash
+benchmark_dir="/work/GraphNet/benchmark_logs"
+mkdir -p "${benchmark_dir}"
 
-for i in /work/GraphNet/samples/torchvision/*/; do
-    dir_name=$(basename "${i%/}")
-    
-    echo "Processing ${dir_name}"
-    
-    python -m graph_net.torch.test_compiler \
-        --model-path "${i}" \
-        --compiler "inductor" \
-        --warmup 3 \
-        --trials 10 \
-        --device cuda \
-        --output-dir "/work/GraphNet/benchmark_logs"
-       
-done  > torchvision_cuda.log 2>&1
+global_log="${benchmark_dir}/global_0830.log"
+> "$global_log"
+
+echo "[$(date)] Script started in background (PID: $$)" | tee -a "$global_log"
+{
+    valid_packages=("timm" "torchaudio" "torchgeometric" "torchvision" "transformers-auto-model" "ultralytics")
+    for i in /work/GraphNet/samples/*/; do
+        package_name=$(basename "${i%/}")
+        if [[ " ${valid_packages[*]} " == *" ${package_name} "* ]]; then
+            echo "[$(date)] Processing package: ${package_name}" | tee -a "$global_log"
+            for j in "$i"*/; do
+                model_name=$(basename "${j%/}")
+                
+                echo "Processing model: ${model_name}"
+                
+                python -m graph_net.torch.test_compiler \
+                    --model-path "${j}" \
+                    --compiler "inductor" \
+                    --warmup 3 \
+                    --trials 10 \
+                    --device cuda \
+                    --output-dir "${benchmark_dir}/${package_name}"
+                
+            done
+        else
+            echo "[$(date)] Skipping package (not in valid list): ${package_name}" | tee -a "$global_log"
+        fi
+    done
+    echo "[$(date)] Script completed" | tee -a "$global_log"
+} >> "$global_log" 2>&1
+
+
+# nohup bash /work/GraphNet/graph_net/benchmark_demo.sh > /dev/ 2>&1 &
 
 # python3 -m graph_net.analysis --test-compiler-log-file /work/GraphNet/torchvision_cuda.log
 
-python3 -m graph_net.analysis --benchmark-path /work/GraphNet/benchmark
+# python3 -m graph_net.analysis --benchmark-path "${benchmark_dir}"