[No Merge] Debug autoparallel test ci

.github/workflows/integration_test_8gpu_autoparallel.yaml

@@ -64,10 +64,35 @@ jobs:
           rocminfo || true
         fi
 
+        # DEBUG: HIP Device name
+        if [[ "${{ matrix.gpu-arch-type }}" == "rocm" ]]; then
+          echo "::group::DEBUG HIP device-name"
+          # Search for hipcc
+          HIPCC="$(find / -type f -name hipcc -perm -111 -print -quit 2>/dev/null || true)"
+          if [[ -z "$HIPCC" ]]; then
+            echo "ERROR: hipcc not found on filesystem"
+            exit 1
+          fi
+          echo "Found hipcc at: $HIPCC"
+          # Compile and run the checked-in C++ program
+          OUTDIR="./hip_tmp"
+          sudo mkdir -p "$OUTDIR"
+          sudo chown -R $(id -u):$(id -g) "$OUTDIR"
+          "$HIPCC" -O2 torchtitan/experiments/autoparallel/debug_ci_script/debug_hip_device_name_ci.cpp -o "$OUTDIR/hip_device_name"
+          "$OUTDIR/hip_device_name"
+          echo "::endgroup::"
+        fi
+
         pip config --user set global.progress_bar off
 
         python -m pip install --force-reinstall --pre torch --index-url ${{ matrix.index-url }}
 
+        sudo curl -L -o /usr/share/libdrm/amdgpu.ids   https://gitlab.freedesktop.org/mesa/libdrm/-/raw/main/data/amdgpu.ids
+        sudo mkdir -p /opt/amdgpu/share/libdrm
+        sudo ln -sf /usr/share/libdrm/amdgpu.ids /opt/amdgpu/share/libdrm/amdgpu.ids
+        echo "Getting device name:"
+        python -c "import torch; print(torch.cuda.get_device_name(None))"
+
         # Install autoparallel - required dependency for autoparallel experiment
         python -m pip install git+https://github.com/meta-pytorch/autoparallel.git
 

torchtitan/experiments/autoparallel/debug_ci_script/debug_hip_device_name_ci.cpp

@@ -0,0 +1,41 @@
+// How to compile and run:
+// hipcc hip_device_name.cpp -o hip_device_name
+// ./hip_device_name
+#include <hip/hip_runtime.h>
+#include <stdio.h>
+
+#define HIP_CHECK(call)                                                         \
+    do {                                                                        \
+        hipError_t err = call;                                                  \
+        if (err != hipSuccess) {                                                \
+            fprintf(stderr, "HIP error at %s:%d: %s\n", __FILE__, __LINE__,     \
+                    hipGetErrorString(err));                                    \
+            return 1;                                                           \
+        }                                                                        \
+    } while (0)
+
+// Simulates torch.cuda.get_device_name(None) behavior
+void print_current_device_name() {
+    int current_device;
+    if (hipGetDevice(&current_device) != hipSuccess) {  // This is what torch.cuda.current_device() calls
+        fprintf(stderr, "Failed to get current device\n");
+        return;
+    }
+
+    hipDeviceProp_t prop;
+    if (hipGetDeviceProperties(&prop, current_device) != hipSuccess) {
+        fprintf(stderr, "Failed to get device properties\n");
+        return;
+    }
+
+    printf("Current device %d: %s\n", current_device, prop.name);
+}
+
+int main(int argc, char* argv[]) {
+    // Demonstrate hipGetDevice / hipSetDevice
+    // This is what torch.cuda.get_device_name(None) does internally
+    printf("=== Simulating torch.cuda.get_device_name(None) ===\n");
+    print_current_device_name();
+
+    return 0;
+}