Add retry logic for scheduled CI tests (#14771)

2025-12-11 16:59:58 -08:00
parent d6bd2d1126
commit e59435c34b
4 changed files with 281 additions and 76 deletions
--- a/.github/workflows/pr-test.yml
+++ b/.github/workflows/pr-test.yml
@@ -27,6 +27,7 @@ concurrency:

 env:
  SGLANG_IS_IN_CI: true
+  SGLANG_CI_ENABLE_RETRY: true

 jobs:
  # =============================================== check changes ====================================================
@@ -550,10 +551,10 @@ jobs:
            CUSTOM_BUILD_SGL_KERNEL=${{needs.check-changes.outputs.sgl_kernel}} bash scripts/ci/ci_install_dependency.sh

        - name: Run test
-          timeout-minutes: 30
+          timeout-minutes: 40
          run: |
            cd test/srt
-            python3 run_suite.py --suite quantization_test
+            python3 run_suite.py --suite quantization_test --enable-retry

  unit-test-backend-1-gpu:
    needs: [check-changes, call-gate, stage-a-test-1]
@@ -592,10 +593,10 @@ jobs:
          CUSTOM_BUILD_SGL_KERNEL=${{needs.check-changes.outputs.sgl_kernel}} bash scripts/ci/ci_install_dependency.sh

      - name: Run test
-        timeout-minutes: 30
+        timeout-minutes: 40
        run: |
          cd test/srt
-          python3 run_suite.py --suite per-commit-1-gpu --auto-partition-id ${{ matrix.part }} --auto-partition-size 15
+          python3 run_suite.py --suite per-commit-1-gpu --auto-partition-id ${{ matrix.part }} --auto-partition-size 15 --enable-retry

  unit-test-backend-2-gpu:
    needs: [check-changes, call-gate, unit-test-backend-1-gpu]
@@ -633,10 +634,10 @@ jobs:
          CUSTOM_BUILD_SGL_KERNEL=${{needs.check-changes.outputs.sgl_kernel}} bash scripts/ci/ci_install_dependency.sh

      - name: Run test
-        timeout-minutes: 30
+        timeout-minutes: 40
        run: |
          cd test/srt
-          python3 run_suite.py --suite per-commit-2-gpu --auto-partition-id ${{ matrix.part }} --auto-partition-size 2
+          python3 run_suite.py --suite per-commit-2-gpu --auto-partition-id ${{ matrix.part }} --auto-partition-size 2 --enable-retry

  unit-test-backend-4-gpu:
    needs: [check-changes, call-gate, unit-test-backend-2-gpu]
@@ -674,10 +675,10 @@ jobs:
          CUSTOM_BUILD_SGL_KERNEL=${{needs.check-changes.outputs.sgl_kernel}} bash scripts/ci/ci_install_dependency.sh

      - name: Run test
-        timeout-minutes: 20
+        timeout-minutes: 30
        run: |
          cd test/srt
-          python3 run_suite.py --suite per-commit-4-gpu --auto-partition-id ${{ matrix.part }} --auto-partition-size 3
+          python3 run_suite.py --suite per-commit-4-gpu --auto-partition-id ${{ matrix.part }} --auto-partition-size 3 --enable-retry

  unit-test-backend-8-gpu-h200:
    needs: [check-changes, call-gate, unit-test-backend-2-gpu]
@@ -721,10 +722,10 @@ jobs:
          python3 -m sglang.compile_deep_gemm --model deepseek-ai/DeepSeek-V3-0324 --tp 8 --trust-remote-code

      - name: Run test
-        timeout-minutes: 20
+        timeout-minutes: 30
        run: |
          cd test/srt
-          python3 run_suite.py --suite per-commit-8-gpu-h200 --auto-partition-id ${{ matrix.part }} --auto-partition-size 3
+          python3 run_suite.py --suite per-commit-8-gpu-h200 --auto-partition-id ${{ matrix.part }} --auto-partition-size 3 --enable-retry

  unit-test-backend-8-gpu-h20:
    needs: [check-changes, call-gate, unit-test-backend-2-gpu]
@@ -763,10 +764,10 @@ jobs:
          CUSTOM_BUILD_SGL_KERNEL=${{needs.check-changes.outputs.sgl_kernel}} bash scripts/ci/ci_install_deepep.sh

      - name: Run test
-        timeout-minutes: 20
+        timeout-minutes: 30
        run: |
          cd test/srt
-          python3 run_suite.py --suite per-commit-8-gpu-h20 --auto-partition-id ${{ matrix.part }} --auto-partition-size 2
+          python3 run_suite.py --suite per-commit-8-gpu-h20 --auto-partition-id ${{ matrix.part }} --auto-partition-size 2 --enable-retry

  performance-test-1-gpu-part-1:
    needs: [check-changes, call-gate, stage-a-test-1]
@@ -1055,10 +1056,12 @@ jobs:
          pip install -e .

      - name: Evaluate accuracy
-        timeout-minutes: 20
+        timeout-minutes: 25
        run: |
          cd test/srt
          python3 test_eval_accuracy_large.py
+        env:
+          SGLANG_CI_ENABLE_RETRY: ${{ env.SGLANG_CI_ENABLE_RETRY }}

  accuracy-test-2-gpu:
    needs: [check-changes, call-gate, accuracy-test-1-gpu]
@@ -1095,10 +1098,12 @@ jobs:
          pip install -e .

      - name: Evaluate accuracy (TP=2)
-        timeout-minutes: 20
+        timeout-minutes: 25
        run: |
          cd test/srt
          python3 test_moe_eval_accuracy_large.py
+        env:
+          SGLANG_CI_ENABLE_RETRY: ${{ env.SGLANG_CI_ENABLE_RETRY }}

  unit-test-deepep-4-gpu:
    needs: [check-changes, call-gate, unit-test-backend-2-gpu]
@@ -1132,10 +1137,10 @@ jobs:
          CUSTOM_BUILD_SGL_KERNEL=${{needs.check-changes.outputs.sgl_kernel}} bash scripts/ci/ci_install_deepep.sh

      - name: Run test
-        timeout-minutes: 20
+        timeout-minutes: 30
        run: |
          cd test/srt
-          python3 run_suite.py --suite per-commit-4-gpu-deepep
+          python3 run_suite.py --suite per-commit-4-gpu-deepep --enable-retry

  unit-test-deepep-8-gpu:
    needs: [check-changes, call-gate, unit-test-backend-2-gpu]
@@ -1169,10 +1174,10 @@ jobs:
          CUSTOM_BUILD_SGL_KERNEL=${{needs.check-changes.outputs.sgl_kernel}} bash scripts/ci/ci_install_deepep.sh

      - name: Run test
-        timeout-minutes: 20
+        timeout-minutes: 30
        run: |
          cd test/srt
-          python3 run_suite.py --suite per-commit-8-gpu-h200-deepep
+          python3 run_suite.py --suite per-commit-8-gpu-h200-deepep --enable-retry

  unit-test-backend-4-gpu-b200:
    needs: [check-changes, call-gate, unit-test-backend-2-gpu]
@@ -1211,10 +1216,10 @@ jobs:
          CUSTOM_BUILD_SGL_KERNEL=${{needs.check-changes.outputs.sgl_kernel}} IS_BLACKWELL=1 bash scripts/ci/ci_install_dependency.sh

      - name: Run test
-        timeout-minutes: 30
+        timeout-minutes: 40
        run: |
          cd test/srt
-          IS_BLACKWELL=1 python3 run_suite.py --suite per-commit-4-gpu-b200 --auto-partition-id ${{ matrix.part }}  --auto-partition-size 3 --timeout-per-file 1800
+          IS_BLACKWELL=1 python3 run_suite.py --suite per-commit-4-gpu-b200 --auto-partition-id ${{ matrix.part }} --auto-partition-size 3 --timeout-per-file 1800 --enable-retry

  # TODO: Add gb200 tests back after the ci runner is fixed
  # unit-test-backend-4-gpu-gb200:
@@ -1251,10 +1256,10 @@ jobs:
  #         CUSTOM_BUILD_SGL_KERNEL=${{needs.check-changes.outputs.sgl_kernel}} IS_BLACKWELL=1 GRACE_BLACKWELL=1 bash scripts/ci/ci_install_deepep.sh

  #     - name: Run test
-  #       timeout-minutes: 45
+  #       timeout-minutes: 55
  #       run: |
  #         cd test/srt
-  #         python3 run_suite.py --suite per-commit-4-gpu-gb200 --auto-partition-id 0 --auto-partition-size 1 --timeout-per-file 3600
+  #         python3 run_suite.py --suite per-commit-4-gpu-gb200 --auto-partition-id 0 --auto-partition-size 1 --timeout-per-file 3600 --enable-retry

  pr-test-finish:
    needs:
--- a/python/sglang/test/ci/ci_utils.py
+++ b/python/sglang/test/ci/ci_utils.py
@@ -1,4 +1,6 @@
+import logging
 import os
+import re
 import subprocess
 import threading
 import time
@@ -7,6 +9,8 @@ from typing import Callable, List, Optional

 from sglang.srt.utils.common import kill_process_tree

+logger = logging.getLogger(__name__)
+

@dataclass
 class TestFile:
@@ -14,6 +18,61 @@ class TestFile:
    estimated_time: float = 60


+# Patterns that indicate retriable accuracy/performance failures
+RETRIABLE_PATTERNS = [
+    r"AssertionError:.*not greater than",
+    r"AssertionError:.*not less than",
+    r"AssertionError:.*not equal to",
+    r"AssertionError:.*!=.*expected",
+    r"accuracy",
+    r"score",
+    r"latency",
+    r"throughput",
+]
+
+# Patterns that indicate non-retriable failures (real code errors)
+NON_RETRIABLE_PATTERNS = [
+    r"SyntaxError",
+    r"ImportError",
+    r"ModuleNotFoundError",
+    r"NameError",
+    r"TypeError",
+    r"AttributeError",
+    r"RuntimeError",
+    r"CUDA out of memory",
+    r"OOM",
+    r"Segmentation fault",
+    r"core dumped",
+    r"ConnectionRefusedError",
+    r"FileNotFoundError",
+]
+
+
+def is_retriable_failure(output: str) -> tuple[bool, str]:
+    """
+    Determine if a test failure is retriable based on output patterns.
+
+    Returns:
+        tuple: (is_retriable, reason)
+    """
+    # Check for non-retriable patterns first
+    for pattern in NON_RETRIABLE_PATTERNS:
+        if re.search(pattern, output, re.IGNORECASE):
+            return False, f"non-retriable error: {pattern}"
+
+    # Check for retriable patterns
+    for pattern in RETRIABLE_PATTERNS:
+        if re.search(pattern, output, re.IGNORECASE):
+            return True, f"retriable pattern: {pattern}"
+
+    # If we have an AssertionError but didn't match non-retriable, assume retriable
+    if re.search(r"AssertionError", output):
+        return True, "AssertionError (assuming retriable)"
+
+    # Default: not retriable
+    return False, "unknown failure type"
+
+
 def run_with_timeout(
    func: Callable,
    args: tuple = (),
@@ -38,8 +97,21 @@ def run_with_timeout(
    return ret_value[0]


+def write_github_step_summary(content: str):
+    """Write content to GitHub Step Summary if available."""
+    summary_file = os.environ.get("GITHUB_STEP_SUMMARY")
+    if summary_file:
+        with open(summary_file, "a") as f:
+            f.write(content)
+
+
 def run_unittest_files(
-    files: List[TestFile], timeout_per_file: float, continue_on_error: bool = False
+    files: List[TestFile],
+    timeout_per_file: float,
+    continue_on_error: bool = False,
+    enable_retry: bool = False,
+    max_attempts: int = 2,
+    retry_wait_seconds: int = 60,
 ):
    """
    Run a list of test files.
@@ -49,86 +121,168 @@ def run_unittest_files(
        timeout_per_file: Timeout in seconds for each test file
        continue_on_error: If True, continue running remaining tests even if one fails.
                          If False, stop at first failure (default behavior for PR tests).
+        enable_retry: If True, retry failed tests that appear to be accuracy/performance
+                     assertion failures (not code errors).
+        max_attempts: Maximum number of attempts per file including initial run (default: 2).
+        retry_wait_seconds: Seconds to wait between retries (default: 60).
    """
    tic = time.perf_counter()
    success = True
    passed_tests = []
    failed_tests = []
+    retried_tests = []  # Track which tests were retried

    for i, file in enumerate(files):
        filename, estimated_time = file.name, file.estimated_time
        process = None
+        output_lines = []

-        def run_one_file(filename):
-            nonlocal process
+        def run_one_file(filename, capture_output=False):
+            nonlocal process, output_lines

-            filename = os.path.join(os.getcwd(), filename)
-            print(
-                f".\n.\nBegin ({i}/{len(files) - 1}):\npython3 {filename}\n.\n.\n",
-                flush=True,
+            full_path = os.path.join(os.getcwd(), filename)
+            logger.info(
+                f".\n.\nBegin ({i}/{len(files) - 1}):\npython3 {full_path}\n.\n.\n"
            )
-            tic = time.perf_counter()
+            file_tic = time.perf_counter()

-            process = subprocess.Popen(
-                ["python3", filename], stdout=None, stderr=None, env=os.environ
-            )
-            process.wait()
-            elapsed = time.perf_counter() - tic
+            if capture_output:
+                # Capture output for retry decision
+                process = subprocess.Popen(
+                    ["python3", full_path],
+                    stdout=subprocess.PIPE,
+                    stderr=subprocess.STDOUT,
+                    env=os.environ,
+                    text=True,
+                )
+                output_lines = []
+                for line in process.stdout:
+                    logger.info(line.rstrip())
+                    output_lines.append(line)
+                process.wait()
+            else:
+                process = subprocess.Popen(
+                    ["python3", full_path], stdout=None, stderr=None, env=os.environ
+                )
+                process.wait()

-            print(
-                f".\n.\nEnd ({i}/{len(files) - 1}):\n{filename=}, {elapsed=:.0f}, {estimated_time=}\n.\n.\n",
-                flush=True,
+            elapsed = time.perf_counter() - file_tic
+
+            logger.info(
+                f".\n.\nEnd ({i}/{len(files) - 1}):\n{filename=}, {elapsed=:.0f}, {estimated_time=}\n.\n.\n"
            )
            return process.returncode

-        try:
-            ret_code = run_with_timeout(
-                run_one_file, args=(filename,), timeout=timeout_per_file
-            )
-            if ret_code != 0:
-                print(
-                    f"\n✗ FAILED: {filename} returned exit code {ret_code}\n",
-                    flush=True,
+        # Retry loop for each file
+        attempt = 1
+        file_passed = False
+        was_retried = False
+
+        while attempt <= (max_attempts if enable_retry else 1):
+            if attempt > 1:
+                logger.info(
+                    f"\n[CI Retry] Attempt {attempt}/{max_attempts} for {filename}\n"
                )
-                success = False
-                failed_tests.append((filename, f"exit code {ret_code}"))
-                if not continue_on_error:
-                    # Stop at first failure for PR tests
+                was_retried = True
+
+            try:
+                ret_code = run_with_timeout(
+                    run_one_file,
+                    args=(filename,),
+                    kwargs={"capture_output": enable_retry},
+                    timeout=timeout_per_file,
+                )
+
+                if ret_code == 0:
+                    file_passed = True
+                    if was_retried:
+                        logger.info(
+                            f"\n✓ PASSED on retry (attempt {attempt}): {filename}\n"
+                        )
+                        retried_tests.append((filename, attempt, "passed"))
+                    passed_tests.append(filename)
                    break
-                # Otherwise continue to next test for nightly tests
-            else:
-                passed_tests.append(filename)
-        except TimeoutError:
-            kill_process_tree(process.pid)
-            time.sleep(5)
-            print(
-                f"\n✗ TIMEOUT: {filename} after {timeout_per_file} seconds\n",
-                flush=True,
-            )
-            success = False
-            failed_tests.append((filename, f"timeout after {timeout_per_file}s"))
-            if not continue_on_error:
-                # Stop at first timeout for PR tests
+                else:
+                    # Check if we should retry
+                    if enable_retry and attempt < max_attempts:
+                        output = "".join(output_lines)
+                        is_retriable, reason = is_retriable_failure(output)
+
+                        if is_retriable:
+                            logger.info(
+                                f"\n[CI Retry] {filename} failed with {reason}"
+                            )
+                            logger.info(
+                                f"[CI Retry] Waiting {retry_wait_seconds}s before retry...\n"
+                            )
+                            time.sleep(retry_wait_seconds)
+                            attempt += 1
+                            continue
+                        else:
+                            logger.info(
+                                f"\n[CI Retry] {filename} failed with {reason} - not retrying\n"
+                            )
+
+                    # No retry or not retriable
+                    logger.info(
+                        f"\n✗ FAILED: {filename} returned exit code {ret_code}\n"
+                    )
+                    if was_retried:
+                        retried_tests.append((filename, attempt, "failed"))
+                    failed_tests.append((filename, f"exit code {ret_code}"))
+                    break
+
+            except TimeoutError:
+                kill_process_tree(process.pid)
+                time.sleep(5)
+                logger.info(
+                    f"\n✗ TIMEOUT: {filename} after {timeout_per_file} seconds\n"
+                )
+                if was_retried:
+                    retried_tests.append((filename, attempt, "timeout"))
+                failed_tests.append((filename, f"timeout after {timeout_per_file}s"))
                break
-            # Otherwise continue to next test for nightly tests
+
+        if not file_passed:
+            success = False
+            if not continue_on_error:
+                break
+
+    elapsed_total = time.perf_counter() - tic

    if success:
-        print(f"Success. Time elapsed: {time.perf_counter() - tic:.2f}s", flush=True)
+        logger.info(f"Success. Time elapsed: {elapsed_total:.2f}s")
    else:
-        print(f"Fail. Time elapsed: {time.perf_counter() - tic:.2f}s", flush=True)
+        logger.info(f"Fail. Time elapsed: {elapsed_total:.2f}s")

    # Print summary
-    print(f"\n{'='*60}", flush=True)
-    print(f"Test Summary: {len(passed_tests)}/{len(files)} passed", flush=True)
-    print(f"{'='*60}", flush=True)
+    logger.info(f"\n{'='*60}")
+    logger.info(f"Test Summary: {len(passed_tests)}/{len(files)} passed")
+    if enable_retry and retried_tests:
+        logger.info(f"Retries: {len(retried_tests)} test(s) were retried")
+    logger.info(f"{'='*60}")
    if passed_tests:
-        print("✓ PASSED:", flush=True)
+        logger.info("✓ PASSED:")
        for test in passed_tests:
-            print(f"  {test}", flush=True)
+            logger.info(f"  {test}")
    if failed_tests:
-        print("\n✗ FAILED:", flush=True)
+        logger.info("\n✗ FAILED:")
        for test, reason in failed_tests:
-            print(f"  {test} ({reason})", flush=True)
-    print(f"{'='*60}\n", flush=True)
+            logger.info(f"  {test} ({reason})")
+    if retried_tests:
+        logger.info("\n↻ RETRIED:")
+        for test, attempts, result in retried_tests:
+            logger.info(f"  {test} ({attempts} attempts, {result})")
+    logger.info(f"{'='*60}\n")
+
+    # Write GitHub Step Summary
+    if retried_tests:
+        summary = "\n### CI Retry Summary\n\n"
+        summary += "| Test File | Attempts | Result |\n"
+        summary += "|-----------|----------|--------|\n"
+        for test, attempts, result in retried_tests:
+            summary += f"| `{test}` | {attempts} | {result} |\n"
+        summary += "\n"
+        write_github_step_summary(summary)

    return 0 if success else -1
--- a/test/run_suite.py
+++ b/test/run_suite.py
@@ -136,6 +136,9 @@ def run_a_suite(args):
        test_files,
        timeout_per_file=args.timeout_per_file,
        continue_on_error=args.continue_on_error,
+        enable_retry=args.enable_retry,
+        max_attempts=args.max_attempts,
+        retry_wait_seconds=args.retry_wait_seconds,
    )


@@ -178,6 +181,24 @@ def main():
        type=int,
        help="Use auto load balancing. The number of parts.",
    )
+    parser.add_argument(
+        "--enable-retry",
+        action="store_true",
+        default=False,
+        help="Enable smart retry for accuracy/performance assertion failures (not code errors)",
+    )
+    parser.add_argument(
+        "--max-attempts",
+        type=int,
+        default=2,
+        help="Maximum number of attempts per file including initial run (default: 2)",
+    )
+    parser.add_argument(
+        "--retry-wait-seconds",
+        type=int,
+        default=60,
+        help="Seconds to wait between retries (default: 60)",
+    )
    args = parser.parse_args()

    # Validate auto-partition arguments
--- a/test/srt/run_suite.py
+++ b/test/srt/run_suite.py
@@ -511,6 +511,24 @@ def main():
        default=False,
        help="Continue running remaining tests even if one fails (useful for nightly tests)",
    )
+    arg_parser.add_argument(
+        "--enable-retry",
+        action="store_true",
+        default=False,
+        help="Enable smart retry for accuracy/performance assertion failures (not code errors)",
+    )
+    arg_parser.add_argument(
+        "--max-attempts",
+        type=int,
+        default=2,
+        help="Maximum number of attempts per file including initial run (default: 2)",
+    )
+    arg_parser.add_argument(
+        "--retry-wait-seconds",
+        type=int,
+        default=60,
+        help="Seconds to wait between retries (default: 60)",
+    )
    args = arg_parser.parse_args()
    print(f"{args=}")

@@ -526,7 +544,14 @@ def main():

    print("The running tests are ", [f.name for f in files])

-    exit_code = run_unittest_files(files, args.timeout_per_file, args.continue_on_error)
+    exit_code = run_unittest_files(
+        files,
+        args.timeout_per_file,
+        args.continue_on_error,
+        enable_retry=args.enable_retry,
+        max_attempts=args.max_attempts,
+        retry_wait_seconds=args.retry_wait_seconds,
+    )
    exit(exit_code)