Fix performance tracking action

.github/workflows/performance.yml

@@ -25,70 +25,41 @@ jobs:
       with:
         python-version: '3.9'
 
-    - run: curl -LsSf https://astral.sh/uv/install.sh | sh
-    - run: uv pip install --system -e ".[tests]"
-    - run: pip install coverage[toml]
+    - name: Install Requirements
+      run: |
+        curl -LsSf https://astral.sh/uv/install.sh | sh
+        uv pip install --system -e ".[tests]"
 
-    - name: Save card_profiler python script
-      uses: actions/upload-artifact@v4
-      with:
-        name: card_profiler
-        path: performance_profile/card_profiler.py
-        compression-level: 0
-        overwrite: true
+    - name: Prepare the dirs for performance evaluation in main
+      run: |
+        mkdir -p performance_action
+        mkdir -p performance_action/logs
+        echo "" > performance_action/__init__.py
+        echo " " > performance_action/logs/cards_benchmark.prof
+        echo " " > performance_action/logs/cards_benchmark.json
+        cp performance/card_profiler.py performance_action/card_profiler.py
+        cp performance/compare_performance_results.py performance_action/compare_performance_results.py
 
     - name: Checkout main branch
       uses: actions/checkout@v4
       with:
         ref: main
-
-    - name: Prepare the dirs for performance evaluation in main
-      run: |
-        mkdir -p performance_profile
-        mkdir -p performance_profile/logs
-        echo "" > performance_profile/__init__.py
-        echo " " > performance_profile/logs/cards_benchmark.prof
-        echo " " > performance_profile/logs/cards_benchmark.json
-
-    - name: Download card_profiler python script
-      uses: actions/download-artifact@v4
-      with:
-        name: card_profiler
-        path: performance_profile/
+        clean: false
 
     - name: Run performance on main branch
       run: |
-        python -m performance_profile.card_profiler
-        cp performance_profile/logs/cards_benchmark.json performance_profile/logs/main_cards_benchmark.json
-
-    - name: Save main performance json
-      uses: actions/upload-artifact@v4
-      with:
-        name: main_performance_json
-        path: performance_profile/logs/main_cards_benchmark.json
-        compression-level: 0
-        overwrite: true
+        python performance_action/card_profiler.py --output_file performance_action/main_results.json
 
     - name: Checkout PR branch
       uses: actions/checkout@v4
       with:
         ref: ${{ github.head_ref }}
-
-    - name: Create performance_profile/logs dir 
-      run: |
-        mkdir -p performance_profile/logs
-        echo " " > performance_profile/logs/cards_benchmark.prof
+        clean: false
 
     - name: Run performance on PR branch
       run: |
-        python -m performance_profile.card_profiler
-        cp performance_profile/logs/cards_benchmark.json performance_profile/logs/pr_cards_benchmark.json
-
-    - name: Download main performance result
-      uses: actions/download-artifact@v4
-      with:
-        name: main_performance_json
-        path: performance_profile/logs/
+        python performance_action/card_profiler.py --output_file performance_action/pr_results.json
 
     - name: Compare main and PR performance results
-      run:  python -m performance_profile.compare_performance_results
+      run: |
+        python performance_action/compare_performance_results.py performance_action/main_results.json performance_action/pr_results.json >> $GITHUB_STEP_SUMMARY

performance_profile/card_profiler.py → performance/card_profiler.py

@@ -1,6 +1,9 @@
+import argparse
 import cProfile
 import json
+import os
 import pstats
+import tempfile
 from io import StringIO
 
 from unitxt.api import load_recipe
@@ -15,35 +18,35 @@
 settings = get_settings()
 settings.allow_unverified_code = True
 
-"""Profiles the execution-time of api.load_dataset(), over a benchmark of cards.
 
-Usage: set values for variables cards (the benchmark)
+class CardProfiler:
+    """Profiles the execution-time of api.load_dataset(), over a benchmark of cards.
 
-from unitxt root dir, run the following linux commands:
+    Usage: set values for variables cards (the benchmark)
 
-python performance_profile/card_profiler.py
+    from unitxt root dir, run the following linux commands:
 
-The script computes the total runtime of the benchmark, and the time spent in loading the dataset,
-accumulated across the cards in the benchmark, and wraps both results into a json file:
-performance_profile/logs/cards_benchmark.json
+    python performance/card_profiler.py
 
-In addition, the script generates a binary file named performance_profile/logs/cards_benchmark.prof,
-which can be nicely and interactively visualized via snakeviz:
+    The script computes the total runtime of the benchmark, and the time spent in loading the dataset,
+    accumulated across the cards in the benchmark, and wraps both results into a json file:
+    performance/logs/cards_benchmark.json
 
-(pip install snakeviz)
-snakeviz performance_profile/logs/cards_benchmark.prof
+    In addition, the script generates a binary file named performance/logs/cards_benchmark.prof,
+    which can be nicely and interactively visualized via snakeviz:
 
-snakeviz opens an interactive internet browser window allowing to explore all time-details.
-See exporing options here: https://jiffyclub.github.io/snakeviz/
-(can also use the -s flag for snakeviz which will only set up a server and print out the url
-to use from another computer in order to view results shown by that server)
+    (pip install snakeviz)
+    snakeviz performance/logs/cards_benchmark.prof
 
-In the browser window, look (ctrl-F) for methods named  profiler_...  to read profiling data for the major steps in the process.
-You will find the total time of each step, accumulated along all cards in the benchmark.
-"""
+    snakeviz opens an interactive internet browser window allowing to explore all time-details.
+    See exporing options here: https://jiffyclub.github.io/snakeviz/
+    (can also use the -s flag for snakeviz which will only set up a server and print out the url
+    to use from another computer in order to view results shown by that server)
 
+    In the browser window, look (ctrl-F) for methods named  profiler_...  to read profiling data for the major steps in the process.
+    You will find the total time of each step, accumulated along all cards in the benchmark.
+    """
 
-class CardProfiler:
     def profiler_instantiate_recipe(self, **kwargs) -> StandardRecipe:
         return load_recipe(**kwargs)
 
@@ -112,32 +115,53 @@ def profile_from_cards():
 
 
 def main():
+    # Parse command-line arguments
+    parser = argparse.ArgumentParser(description="Card Profiler")
+    parser.add_argument(
+        "--output_file",
+        type=str,
+        required=True,
+        help="Path to save output files (without extension)",
+    )
+    args = parser.parse_args()
+
+    # Ensure the directory for the output file exists
+    output_dir = os.path.dirname(args.output_file)
+    if output_dir:
+        os.makedirs(output_dir, exist_ok=True)
+
     logger.info(f"benchmark cards are: {cards}")
 
-    cProfile.run(
-        "profile_from_cards()", "performance_profile/logs/cards_benchmark.prof"
-    )
-    f = StringIO()
-    pst = pstats.Stats("performance_profile/logs/cards_benchmark.prof", stream=f)
-    pst.strip_dirs()
-    pst.sort_stats("name")  # sort by function name
-    pst.print_stats("profiler_do_the_profiling|profiler_load_by_recipe")
-    s = f.getvalue()
-    assert s.split("\n")[7].split()[3] == "cumtime"
-    assert "profiler_do_the_profiling" in s.split("\n")[8]
-    tot_time = round(float(s.split("\n")[8].split()[3]), 3)
-    assert "profiler_load_by_recipe" in s.split("\n")[9]
-    load_time = round(float(s.split("\n")[9].split()[3]), 3)
-    diff = round(tot_time - load_time, 3)
-
-    # Data to be written
-    dictionary = {
-        "total_time": tot_time,
-        "load_time": load_time,
-        "net_time": diff,
-    }
-    with open("performance_profile/logs/cards_benchmark.json", "w") as outfile:
-        json.dump(dictionary, outfile)
+    # Create a temporary .prof file
+    with tempfile.NamedTemporaryFile(suffix=".prof", delete=False) as temp_prof_file:
+        temp_prof_file_path = temp_prof_file.name
+        cProfile.run("profile_from_cards()", temp_prof_file_path)
+
+        f = StringIO()
+        pst = pstats.Stats(temp_prof_file_path, stream=f)
+        pst.strip_dirs()
+        pst.sort_stats("name")  # sort by function name
+        pst.print_stats("profiler_do_the_profiling|profiler_load_by_recipe")
+        s = f.getvalue()
+        assert s.split("\n")[7].split()[3] == "cumtime"
+        assert "profiler_do_the_profiling" in s.split("\n")[8]
+        tot_time = round(float(s.split("\n")[8].split()[3]), 3)
+        assert "profiler_load_by_recipe" in s.split("\n")[9]
+        load_time = round(float(s.split("\n")[9].split()[3]), 3)
+        diff = round(tot_time - load_time, 3)
+
+        # Data to be written
+        dictionary = {
+            "total_time": tot_time,
+            "load_time": load_time,
+            "net_time": diff,
+        }
+
+        # Write the profiling results to the JSON file (user-specified)
+        with open(args.output_file, "w+") as outfile:
+            json.dump(dictionary, outfile)
+
+        logger.info(f"JSON output saved to: {args.output_file}")
 
 
 if __name__ == "__main__":

performance/compare_performance_results.py

@@ -0,0 +1,49 @@
+import argparse
+import json
+import sys
+
+# Argument parser to get file paths from the command line
+parser = argparse.ArgumentParser(description="Compare performance profiles.")
+parser.add_argument(
+    "main_perf_file", type=str, help="Path to main performance profile JSON file"
+)
+parser.add_argument(
+    "pr_perf_file", type=str, help="Path to PR performance profile JSON file"
+)
+args = parser.parse_args()
+
+# Reading both performance JSON files:
+with open(args.main_perf_file) as openfile:
+    main_perf = json.load(openfile)
+
+with open(args.pr_perf_file) as openfile:
+    pr_perf = json.load(openfile)
+
+# Check for valid net_time in the main performance profile
+if main_perf["net_time"] == 0:
+    print("Net run time on main is 0, can't calculate ratio of times.")
+    sys.exit(1)
+
+# Calculate the ratio between PR and main branch net times
+ratio = pr_perf["net_time"] / main_perf["net_time"]
+
+# Markdown table formatting
+table_header = "| Branch       | Net Time (seconds) | Performance Ratio |\n"
+table_divider = "|--------------|--------------------|-------------------|\n"
+table_main = f"| Main Branch  | {main_perf['net_time']:<18} | -                 |\n"
+table_pr = f"| PR Branch    | {pr_perf['net_time']:<18} | {ratio:.2f}            |\n"
+
+# Print markdown table
+print("### Performance Comparison Results\n")
+print(table_header + table_divider + table_main + table_pr)
+
+# Performance degradation check (5% threshold)
+if ratio > 1.05:
+    print("\n**Warning**: Performance degradation exceeds 5%!")
+    print(
+        "Explore branch performance via 'python performance_profile/card_profiler.py',"
+        " followed by 'snakeviz performance_profile/logs/cards_benchmark.prof'."
+    )
+    sys.exit(1)
+
+print("\nPerformance of the PR branch is within acceptable limits.")

pyproject.toml

@@ -55,7 +55,7 @@ target-version = "py38"
 "utils/hf/prepare_dataset.py" = ["T201"]
 "utils/hf/prepare_metric.py" = ["T201"]
 "utils/compare_unitxt_datasets_between_versions.py" = ["C901"]
-"performance_profile/run_profile.py" = ["T201"]
+"performance/*.py" = ["T201"]
 
 [tool.ruff.lint]
 # Enable Pyflakes (`F`) and a subset of the pycodestyle (`E`)  codes by default.