update the S3 path to include model name

Author: namanlalitnyu

.github/scripts/run_vllm_profiling.sh

@@ -71,6 +71,13 @@ run_profiling_tests() {
         exit 1
     fi
 
+    # Get S3 path components from environment
+    local s3_date="${S3_UPLOAD_DATE:-}"
+    local s3_repo="${S3_REPOSITORY:-}"
+    local s3_sha="${S3_HEAD_SHA:-}"
+    local s3_run_id="${S3_GITHUB_RUN_ID:-}"
+    local s3_job="${S3_GITHUB_JOB:-}"
+
     # Iterate over profiling tests
     jq -c '.[]' "$profiling_test_file" | while read -r params; do
         # Get the test name
@@ -99,36 +106,58 @@ run_profiling_tests() {
         kill_gpu_processes
 
         # Create a profiling sub-directory for each test case to isolate the
-        # generated traces (e.g. using the model name hierarchy)
-        local sanitized_test_name="${TEST_NAME// /_}"
-        local test_name_directory="${base_profiler_dir}/${sanitized_test_name}"
-        mkdir -p "${test_name_directory}"
-        chmod 755 "${test_name_directory}"
+        # generated traces using the S3 path structure
+        MODEL_NAME=$(echo "$server_params" | jq -r '.model')
+        local sanitized_model_name="${MODEL_NAME// /_}"
+        
+        # Build the directory path following S3 structure
+        local model_name_directory="${base_profiler_dir}"
+        
+        # Check if all S3 path components are available
+        if [[ -n "${s3_date}" && -n "${s3_repo}" && -n "${s3_sha}" && -n "${s3_run_id}" && -n "${s3_job}" ]]; then
+            # Build the complete S3 path structure
+            model_name_directory="${model_name_directory}/${s3_date}/${s3_repo}/${sanitized_model_name}/${s3_sha}/${s3_run_id}/${s3_job}"
+        else
+            # Error out if any S3 variable is missing
+            echo "ERROR: Required S3 path variables are missing. Cannot proceed with profiling."
+            echo "Missing variables:"
+            [[ -z "${s3_date}" ]] && echo "  - S3_UPLOAD_DATE"
+            [[ -z "${s3_repo}" ]] && echo "  - S3_REPOSITORY"
+            [[ -z "${s3_sha}" ]] && echo "  - S3_HEAD_SHA"
+            [[ -z "${s3_run_id}" ]] && echo "  - S3_GITHUB_RUN_ID"
+            [[ -z "${s3_job}" ]] && echo "  - S3_GITHUB_JOB"
+            echo "Please ensure all required environment variables are set."
+            exit 1
+        fi
+        
+        echo "Creating profiling directory: ${model_name_directory}"
+        mkdir -p "${model_name_directory}"
+        chmod 755 "${model_name_directory}"
 
         # Override the profiler output directory for this test only
-        export VLLM_TORCH_PROFILER_DIR="${test_name_directory}"
+        export VLLM_TORCH_PROFILER_DIR="${model_name_directory}"
 
         # Run the profiling test
         if start_vllm_server "$server_args"; then
             run_profiling "$client_args"
             cleanup_server
 
             # Debug: Check if profiling files were created
-            echo "DEBUG: Checking profiling directory: $test_name_directory"
-            if [ -d "$test_name_directory" ]; then
-                echo "DEBUG: Profiling directory exists for test $TEST_NAME"
-                ls -la "$test_name_directory" || echo "DEBUG: Directory is empty or inaccessible"
-                find "$test_name_directory" -type f 2>/dev/null | head -10 | while read file; do
+            echo "DEBUG: Checking profiling directory: $model_name_directory"
+            if [ -d "$model_name_directory" ]; then
+                echo "DEBUG: Profiling directory exists for model $MODEL_NAME"
+                ls -la "$model_name_directory" || echo "DEBUG: Directory is empty or inaccessible"
+                find "$model_name_directory" -type f 2>/dev/null | head -10 | while read file; do
                     echo "DEBUG: Found profiling file: ${file}"
                     rename_profiling_file "$file" "vllm"
                 done
             else
-                echo "DEBUG: Profiling directory does not exist for test $TEST_NAME!"
+                echo "DEBUG: Profiling directory does not exist for model $MODEL_NAME!"
             fi
 
-            echo "Profiling test $TEST_NAME completed successfully."
+            echo "Profiling test $MODEL_NAME completed successfully."
         else
-            echo "Failed to start vLLM server for test $TEST_NAME."
+            echo "Failed to start vLLM server for test $MODEL_NAME."
             continue
         fi
     done

.github/workflows/vllm-profiling.yml

@@ -179,6 +179,7 @@ jobs:
           VLLM_TORCH_PROFILER_DIR: /tmp/workspace/vllm-profiling/profiling-results
           CUDA_VISIBLE_DEVICES: 0
           VLLM_USE_V1: 1
+          REPOSITORY: vllm-project/vllm
 
         run: |
           set -eux
@@ -189,6 +190,9 @@ jobs:
             ON_CPU=0
           fi
 
+          # Prepare S3 path components
+          UPLOAD_DATE=$(date -u +"%Y-%m-%d")
+
           container_name=$(docker run \
             ${GPU_FLAG:-} \
             ${SCCACHE_SERVER_PORT_DOCKER_FLAG:-} \
@@ -203,6 +207,11 @@ jobs:
             -e VLLM_USE_V1 \
             -e DEVICE_NAME \
             -e ON_CPU="${ON_CPU}" \
+            -e S3_UPLOAD_DATE="${UPLOAD_DATE}" \
+            -e S3_REPOSITORY="${REPOSITORY}" \
+            -e S3_HEAD_SHA="${HEAD_SHA}" \
+            -e S3_GITHUB_RUN_ID="${GITHUB_RUN_ID}" \
+            -e S3_GITHUB_JOB="${GITHUB_JOB}" \
             --ipc=host \
             --tty \
             --detach \
@@ -214,21 +223,10 @@ jobs:
           )
           docker exec -t "${container_name}" bash -c "cd vllm-profiling && bash ../.github/scripts/run_vllm_profiling.sh"
 
-      - name: Prepare S3 upload metadata
-        id: prepare_s3_upload
-        env:
-          REPOSITORY: vllm-project/vllm
-        run: |
-          set -eux
-
-          UPLOAD_DATE=$(date -u +"%Y-%m-%d")
-          echo "upload-date=${UPLOAD_DATE}" >> "${GITHUB_OUTPUT}"
-          echo "s3-prefix=${UPLOAD_DATE}/${REPOSITORY}/${HEAD_SHA}/${GITHUB_RUN_ID}/${GITHUB_JOB}" >> "${GITHUB_OUTPUT}"
-
       - name: Upload profiling results to S3
         uses: seemethere/upload-artifact-s3@v5
         with:
-          s3-prefix: ${{ steps.prepare_s3_upload.outputs.s3-prefix }}
+          s3-prefix: ""
           retention-days: 180
           path: vllm-profiling/profiling-results
           if-no-files-found: warn