[CI][Benchmark] Optimize performance benchmark workflow (#1039)

### What this PR does / why we need it?

This is a post patch of #1014, for some convenience optimization
- Set cached dataset path for speed
- Use pypi to install escli-tool
- Add benchmark results convert script to have a developer-friendly
result
- Patch the `benchmark_dataset.py` to disable streaming load for
internet
- Add more trigger ways for different purpose, `pr` for debug,
`schedule` for daily test, `dispatch` and `pr-labled` for manual testing
of a single(current) commit
- Disable latency test for `qwen-2.5-vl`, (This script does not support
multi-modal yet)

### Does this PR introduce _any_ user-facing change?
No

### How was this patch tested?
CI passed

---------

Signed-off-by: wangli <wangli858794774@gmail.com>

.github/workflows/nightly_benchmarks.yaml

@@ -15,21 +15,17 @@
 # limitations under the License.
 #
 
-name: 'run benchmarks main'
+name: 'Benchmarks / Performance'
+# This workflow runs nightly benchmarks for vllm-ascend.
 
 on:
   schedule:
+    # Run at 24:00 everyday
     - cron: '00 16 * * *'
   workflow_dispatch:
-  
-  # after merged, secrets will be available
-  # pull_request:
-  #   branches:
-  #       - 'main'
-  #       - '*-dev'
-  #   paths:
-  #       - '.github/workflows/nightly_benchmarks.yaml'
 
+  pull_request:
+    types: [ labeled ]
 
 # Bash shells do not use ~/.profile or ~/.bashrc so these shells need to be explicitly
 # declared as "shell: bash -el {0}" on steps that need to be properly activated.
@@ -38,9 +34,15 @@ defaults:
   run:
     shell: bash -el {0}
 
+concurrency:
+  group: pr-${{ github.event.pull_request.number }}
+  cancel-in-progress: true
+
 jobs:
   test:
-    name: run benchmarks main
+    if: ${{ contains(github.event.pull_request.labels.*.name, 'performance-test') && contains(github.event.pull_request.labels.*.name, 'ready-for-test') || github.event_name == 'schedule' }}
+
+    name: Benchmarks/vLLM=${{ matrix.vllm_branch }}, vLLM-Ascend=${{ matrix.vllm_ascend_branch }}
     runs-on: 'linux-arm64-npu-static-8'
     strategy:
       matrix:
@@ -85,13 +87,10 @@ jobs:
         run: |
           git config --global --add safe.directory "$GITHUB_WORKSPACE"
           git config --global url."https://gh-proxy.test.osinfra.cn/https://github.com/".insteadOf https://github.com/
-    
 
       - name: Checkout vllm-project/vllm-ascend repo
         uses: actions/checkout@v4
-        with:
+
-          ref: ${{ matrix.vllm_ascend_branch }}
-          
       - name: Checkout vllm-project/vllm repo
         uses: actions/checkout@v4
         with:
@@ -109,25 +108,44 @@ jobs:
           pip install -e .
           pip install -r benchmarks/requirements-bench.txt
 
-      - name: Checkout cosdt/elastic-tool
+      - name: Run current commit benchmarks
-        uses: actions/checkout@v4
+        if: github.event_name != 'schedule'
+        run: |
+          # Sometimes we only want to run benchmarks on the current commit
+          # This is useful for debugging or a release benchmark
+          bash benchmarks/scripts/run-performance-benchmarks.sh
+          # Convert the benchmark results to markdown format
+          python3 benchmarks/scripts/convert_json_to_markdown.py
+
+      - name: Generate step summary
+        if: github.event_name != 'schedule'
+        run: |
+          cat ./benchmarks/results/benchmark_results.md >> $GITHUB_STEP_SUMMARY
+
+      - name: Upload benchmark artifacts
+        if: github.event_name != 'schedule'
+        uses: actions/upload-artifact@v4
         with:
-          repository: cosdt/elastic-tool
+          name: "benchmark-performance-${{ matrix.vllm_branch }}-${{ matrix.vllm_ascend_branch }}-report"
-          path: ./elastic_tool
+          path: ./benchmarks/results/benchmark_results.md
-          ref: 0.1.0-dev
+          if-no-files-found: warn
+          retention-days: 90
+          overwrite: true
 
       - name: Install elastic_tool
-        working-directory: ./elastic_tool
+        if: github.event_name == 'schedule'
         run: |
-          pip install -e .
+          pip install escli-tool==0.2.0
           
       - name: Collect pr info from vllm-project/vllm-ascend
+        if: github.event_name == 'schedule'
         run: |
           # Only get the pull request which may influences performance
           git log --pretty=format:"%H %s" -- '**/*.py' ':!docs/*' ':!tests/*' ':!examples/*' > commit_log.txt
           escli check commit_log.txt
 
       - name: Run benchmark iteration
+        if: github.event_name == 'schedule'
         run: |
           while IFS= read -r line || [[ -n "$line" ]]; do
             commit_id=${line%% *}

benchmarks/requirements-bench.txt

@@ -1,3 +1,5 @@
 pandas
 datasets
-modelscope
+modelscope
+libcst
+tabulate

benchmarks/scripts/convert_json_to_markdown.py

@@ -0,0 +1,183 @@
+import argparse
+import json
+import os
+from pathlib import Path
+
+import pandas as pd
+from tabulate import tabulate
+
+CUR_PATH = Path(__file__).parent.resolve()
+# latency results and the keys that will be printed into markdown
+latency_results = []
+latency_column_mapping = {
+    "test_name": "Test name",
+    "avg_latency": "Mean latency (ms)",
+    "P50": "Median latency (ms)",
+    "P99": "P99 latency (ms)",
+}
+
+# throughput tests and the keys that will be printed into markdown
+throughput_results = []
+throughput_results_column_mapping = {
+    "test_name": "Test name",
+    "num_requests": "Num of reqs",
+    "total_num_tokens": "Total num of tokens",
+    "elapsed_time": "Elapsed time (s)",
+    "requests_per_second": "Tput (req/s)",
+    "tokens_per_second": "Tput (tok/s)",
+}
+
+# serving results and the keys that will be printed into markdown
+serving_results = []
+serving_column_mapping = {
+    "test_name": "Test name",
+    "request_rate": "Request rate (req/s)",
+    "request_throughput": "Tput (req/s)",
+    "output_throughput": "Output Tput (tok/s)",
+    "median_ttft_ms": "TTFT (ms)",
+    "median_tpot_ms": "TPOT (ms)",
+    "median_itl_ms": "ITL (ms)",
+}
+
+
+def read_markdown(file):
+    if os.path.exists(file):
+        with open(file) as f:
+            return f.read() + "\n"
+    else:
+        return f"{file} not found.\n"
+
+
+def results_to_json(latency, throughput, serving):
+    return json.dumps({
+        'latency': latency.to_dict(),
+        'throughput': throughput.to_dict(),
+        'serving': serving.to_dict()
+    })
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(
+        description="Process the results of the benchmark tests.")
+    parser.add_argument(
+        "--results_folder",
+        type=str,
+        default="../results/",
+        help="The folder where the benchmark results are stored.")
+    parser.add_argument(
+        "--output_folder",
+        type=str,
+        default="../results/",
+        help="The folder where the benchmark results are stored.")
+    parser.add_argument("--markdown_template",
+                        type=str,
+                        default="./perf_result_template.md",
+                        help="The template file for the markdown report.")
+    parser.add_argument("--tag",
+                        default="main",
+                        help="Tag to be used for release message.")
+    parser.add_argument("--commit_id",
+                        default="",
+                        help="Commit ID to be used for release message.")
+
+    args = parser.parse_args()
+    results_folder = (CUR_PATH / args.results_folder).resolve()
+    output_folder = (CUR_PATH / args.output_folder).resolve()
+    markdown_template = (CUR_PATH / args.markdown_template).resolve()
+
+    # collect results
+    for test_file in results_folder.glob("*.json"):
+
+        with open(test_file) as f:
+            raw_result = json.loads(f.read())
+
+        if "serving" in str(test_file):
+            # this result is generated via `benchmark_serving.py`
+
+            # update the test name of this result
+            raw_result.update({"test_name": test_file.stem})
+
+            # add the result to raw_result
+            serving_results.append(raw_result)
+            continue
+
+        elif "latency" in f.name:
+            # this result is generated via `benchmark_latency.py`
+
+            # update the test name of this result
+            raw_result.update({"test_name": test_file.stem})
+
+            # get different percentiles
+            for perc in [10, 25, 50, 75, 90, 99]:
+                # Multiply 1000 to convert the time unit from s to ms
+                raw_result.update(
+                    {f"P{perc}": 1000 * raw_result["percentiles"][str(perc)]})
+            raw_result["avg_latency"] = raw_result["avg_latency"] * 1000
+
+            # add the result to raw_result
+            latency_results.append(raw_result)
+            continue
+
+        elif "throughput" in f.name:
+            # this result is generated via `benchmark_throughput.py`
+
+            # update the test name of this result
+            raw_result.update({"test_name": test_file.stem})
+
+            # add the result to raw_result
+            throughput_results.append(raw_result)
+            continue
+
+        print(f"Skipping {test_file}")
+    serving_results.sort(key=lambda x: (len(x['test_name']), x['test_name']))
+
+    latency_results = pd.DataFrame.from_dict(latency_results)
+    serving_results = pd.DataFrame.from_dict(serving_results)
+    throughput_results = pd.DataFrame.from_dict(throughput_results)
+
+    raw_results_json = results_to_json(latency_results, throughput_results,
+                                       serving_results)
+
+    # remapping the key, for visualization purpose
+    if not latency_results.empty:
+        latency_results = latency_results[list(
+            latency_column_mapping.keys())].rename(
+                columns=latency_column_mapping)
+    if not serving_results.empty:
+        serving_results = serving_results[list(
+            serving_column_mapping.keys())].rename(
+                columns=serving_column_mapping)
+    if not throughput_results.empty:
+        throughput_results = throughput_results[list(
+            throughput_results_column_mapping.keys())].rename(
+                columns=throughput_results_column_mapping)
+
+    processed_results_json = results_to_json(latency_results,
+                                             throughput_results,
+                                             serving_results)
+
+    # get markdown tables
+    latency_md_table = tabulate(latency_results,
+                                headers='keys',
+                                tablefmt='pipe',
+                                showindex=False)
+    serving_md_table = tabulate(serving_results,
+                                headers='keys',
+                                tablefmt='pipe',
+                                showindex=False)
+    throughput_md_table = tabulate(throughput_results,
+                                   headers='keys',
+                                   tablefmt='pipe',
+                                   showindex=False)
+
+    # document the result
+    print(output_folder)
+    with open(output_folder / "benchmark_results.md", "w") as f:
+
+        results = read_markdown(markdown_template)
+        results = results.format(
+            latency_tests_markdown_table=latency_md_table,
+            throughput_tests_markdown_table=throughput_md_table,
+            serving_tests_markdown_table=serving_md_table,
+            benchmarking_results_in_json_string=processed_results_json)
+        f.write(results)

benchmarks/scripts/patch_benchmark_dataset.py

@@ -0,0 +1,68 @@
+from argparse import ArgumentParser
+
+import libcst as cst
+import libcst.matchers as m
+
+# Patch the benchmark_dataset.py file to set streaming=False in load_dataset calls
+
+
+# TDOO(Potabk): Remove this patch when the issue is fixed in the upstream
+class StreamingFalseTransformer(cst.CSTTransformer):
+
+    def __init__(self):
+        self.in_target_class = False
+        self.in_target_func = False
+
+    def visit_ClassDef(self, node):
+        if node.name.value == "HuggingFaceDataset":
+            self.in_target_class = True
+
+    def leave_ClassDef(self, original_node, updated_node):
+        self.in_target_class = False
+        return updated_node
+
+    def visit_FunctionDef(self, node):
+        if self.in_target_class and node.name.value == "load_data":
+            self.in_target_func = True
+
+    def leave_FunctionDef(self, original_node, updated_node):
+        self.in_target_func = False
+        return updated_node
+
+    def leave_Call(self, original_node, updated_node):
+        if self.in_target_class and self.in_target_func:
+            if m.matches(updated_node.func, m.Name("load_dataset")):
+                new_args = []
+                for arg in updated_node.args:
+                    if arg.keyword and arg.keyword.value == "streaming":
+                        new_arg = arg.with_changes(value=cst.Name("False"))
+                        new_args.append(new_arg)
+                    else:
+                        new_args.append(arg)
+                return updated_node.with_changes(args=new_args)
+        return updated_node
+
+
+def patch_file(path):
+    with open(path, "r", encoding="utf-8") as f:
+        source = f.read()
+
+    module = cst.parse_module(source)
+    modified = module.visit(StreamingFalseTransformer())
+
+    with open(path, "w", encoding="utf-8") as f:
+        f.write(modified.code)
+
+    print(f"Patched: {path}")
+
+
+if __name__ == '__main__':
+    parser = ArgumentParser(
+        description=
+        "Patch benchmark_dataset.py to set streaming=False in load_dataset calls"
+    )
+    parser.add_argument("--path",
+                        type=str,
+                        help="Path to the benchmark_dataset.py file")
+    args = parser.parse_args()
+    patch_file(args.path)

benchmarks/scripts/perf_result_template.md

@@ -0,0 +1,31 @@
+## Online serving tests
+
+- Input length: randomly sample 200 prompts from [ShareGPT](https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/blob/main/ShareGPT_V3_unfiltered_cleaned_split.json) and [lmarena-ai/vision-arena-bench-v0.1](https://huggingface.co/datasets/lmarena-ai/vision-arena-bench-v0.1/tree/main)(multi-modal) dataset (with fixed random seed).
+- Output length: the corresponding output length of these 200 prompts.
+- Batch size: dynamically determined by vllm and the arrival pattern of the requests.
+- **Average QPS (query per second)**: 1, 4, 16 and inf. QPS = inf means all requests come at once. For other QPS values, the arrival time of each query is determined using a random Poisson process (with fixed random seed).
+- Models: Qwen/Qwen3-8B, Qwen/Qwen2.5-VL-7B-Instruct
+- Evaluation metrics: throughput, TTFT (median time to the first token ), ITL (median inter-token latency) TPOT(median time per output token).
+
+{serving_tests_markdown_table}
+
+## Offline tests
+### Latency tests
+
+- Input length: 32 tokens.
+- Output length: 128 tokens.
+- Batch size: fixed (8).
+- Models: Qwen/Qwen3-8B, Qwen/Qwen2.5-VL-7B-Instruct
+- Evaluation metrics: end-to-end latency.
+
+{latency_tests_markdown_table}
+
+### Throughput tests
+
+- Input length: randomly sample 200 prompts from [ShareGPT](https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/blob/main/ShareGPT_V3_unfiltered_cleaned_split.json) and [lmarena-ai/vision-arena-bench-v0.1](https://huggingface.co/datasets/lmarena-ai/vision-arena-bench-v0.1/tree/main)(multi-modal) dataset (with fixed random seed).
+- Output length: the corresponding output length of these 200 prompts.
+- Batch size: dynamically determined by vllm to achieve maximum throughput.
+- Models: Qwen/Qwen3-8B, Qwen/Qwen2.5-VL-7B-Instruct
+- Evaluation metrics: throughput.
+
+{throughput_tests_markdown_table}

benchmarks/scripts/run-performance-benchmarks.sh

@@ -1,6 +1,5 @@
 #!/bin/bash
 
-set -e
 
 check_npus() {
   # shellcheck disable=SC2155
@@ -19,10 +18,19 @@ check_npus() {
 }
 
 ensure_sharegpt_downloaded() {
-  local FILE=ShareGPT_V3_unfiltered_cleaned_split.json
+  local FILE="/github/home/.cache/datasets/ShareGPT_V3_unfiltered_cleaned_split.json"
+  local DIR
+  DIR=$(dirname "$FILE")
+
   if [ ! -f "$FILE" ]; then
     echo "$FILE not found, downloading from hf-mirror ..."
-    wget https://hf-mirror.com/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/$FILE
+    mkdir -p "$DIR"
+    wget -O "$FILE" https://hf-mirror.com/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json
+    if [ $? -ne 0 ]; then
+      echo "Download failed!" >&2
+      return 1
+    fi
+    echo "Download completed and saved to $FILE"
   else
     echo "$FILE already exists."
   fi
@@ -49,7 +57,8 @@ wait_for_server() {
   # wait for vllm server to start
   # return 1 if vllm server crashes
   timeout 1200 bash -c '
-    until curl -s -X POST localhost:8000/v1/completions || curl -s -X POST localhost:8000/v1/chat/completions; do
+    until curl -s -X GET localhost:8000/health; do
+      echo "Waiting for vllm server to start..."
       sleep 1
     done' && return 0 || return 1
 }
@@ -290,6 +299,7 @@ main() {
   # prepare for benchmarking
   cd benchmarks || exit 1
   get_benchmarks_scripts
+  python3 scripts/patch_benchmark_dataset.py --path vllm_benchmarks/benchmark_dataset.py
   trap cleanup EXIT
 
   QUICK_BENCHMARK_ROOT=./

benchmarks/tests/latency-tests.json

@@ -1,14 +1,4 @@
 [
-  {
-    "test_name": "latency_qwen2_5vl_7B_tp1",
-    "parameters": {
-      "model": "Qwen/Qwen2.5-VL-7B-Instruct",
-      "tensor_parallel_size": 1,
-      "max_model_len": 16384,
-      "num_iters_warmup": 5,
-      "num_iters": 15
-    }
-  },
   {
     "test_name": "latency_qwen3_8B_tp1",
     "parameters": {

benchmarks/tests/serving-tests.json

@@ -46,7 +46,7 @@
       "model": "Qwen/Qwen3-8B",
       "backend": "vllm",
       "dataset_name": "sharegpt",
-      "dataset_path": "/root/.cache/datasets/sharegpt/ShareGPT_V3_unfiltered_cleaned_split.json",
+      "dataset_path": "/github/home/.cache/datasets/ShareGPT_V3_unfiltered_cleaned_split.json",
       "num_prompts": 200
     }
   }

benchmarks/tests/throughput-tests.json

@@ -5,7 +5,7 @@
       "model": "Qwen/Qwen3-8B",
       "tensor_parallel_size": 1,
       "load_format": "dummy",
-      "dataset_path": "/root/.cache/datasets/sharegpt/ShareGPT_V3_unfiltered_cleaned_split.json",
+      "dataset_path": "/github/home/.cache/datasets/ShareGPT_V3_unfiltered_cleaned_split.json",
       "num_prompts": 200,
       "backend": "vllm"
     }