From f61c332cba1109bf858d2977ed85bb60de48b2f8 Mon Sep 17 00:00:00 2001
From: ishandhanani <82981111+ishandhanani@users.noreply.github.com>
Date: Wed, 15 Apr 2026 16:10:00 -0500
Subject: [PATCH] ci: log analyzer (#22859)

---
 .github/workflows/nightly-72-gpu-gb200.yml  |  21 ++
 scripts/ci/slurm/analyze_logs_with_modal.py | 378 ++++++++++++++++++++
 scripts/ci/slurm/log_analysis_prompt.md     | 102 ++++++
 3 files changed, 501 insertions(+)
 create mode 100644 scripts/ci/slurm/analyze_logs_with_modal.py
 create mode 100644 scripts/ci/slurm/log_analysis_prompt.md

diff --git a/.github/workflows/nightly-72-gpu-gb200.yml b/.github/workflows/nightly-72-gpu-gb200.yml
index 5353c09cd..3ac53b5a4 100644
--- a/.github/workflows/nightly-72-gpu-gb200.yml
+++ b/.github/workflows/nightly-72-gpu-gb200.yml
@@ -153,6 +153,27 @@ jobs:
           retention-days: 30
           if-no-files-found: warn
 
+      - name: Analyze logs with AI on failure
+        if: failure()
+        continue-on-error: true
+        env:
+          MODAL_TOKEN_ID: ${{ secrets.NV_MODAL_TOKEN_ID }}
+          MODAL_TOKEN_SECRET: ${{ secrets.NV_MODAL_TOKEN_SECRET }}
+        run: |
+          TARBALL="${{ github.workspace }}/multinode_server_logs.tar.gz"
+          if [ -f "$TARBALL" ]; then
+            uv run --with modal python scripts/ci/slurm/analyze_logs_with_modal.py \
+              --tarball "$TARBALL" \
+              --job-id "${{ matrix.config.name }}-${{ github.run_id }}" \
+              --output "${{ github.workspace }}/ai_analysis.md"
+            if [ -f "${{ github.workspace }}/ai_analysis.md" ]; then
+              echo "## AI Log Analysis" >> $GITHUB_STEP_SUMMARY
+              cat "${{ github.workspace }}/ai_analysis.md" >> $GITHUB_STEP_SUMMARY
+            fi
+          else
+            echo "No log tarball found, skipping analysis"
+          fi
+
       - name: Clean up Slurm jobs on failure/cancel
         if: failure() || cancelled()
         continue-on-error: true
diff --git a/scripts/ci/slurm/analyze_logs_with_modal.py b/scripts/ci/slurm/analyze_logs_with_modal.py
new file mode 100644
index 000000000..0d4fde4ca
--- /dev/null
+++ b/scripts/ci/slurm/analyze_logs_with_modal.py
@@ -0,0 +1,378 @@
+#!/usr/bin/env python3
+"""Analyze srtslurm logs with opencode inside a Modal sandbox.
+
+This script accepts either:
+- a local log directory
+- a `.tar.gz` bundle such as `multinode_server_logs.tar.gz`
+
+It uploads the logs into an ephemeral Modal sandbox, installs and runs
+opencode with an analysis prompt, and prints the resulting markdown.
+
+Example:
+  uv run --with modal python scripts/ci/slurm/analyze_logs_with_modal.py \
+    --tarball /tmp/multinode_server_logs.tar.gz \
+    --job-id 4645
+"""
+
+from __future__ import annotations
+
+import argparse
+import logging
+import os
+import re
+import shlex
+import shutil
+import tarfile
+import tempfile
+from pathlib import Path
+
+try:
+    import modal
+except ImportError:  # pragma: no cover - runtime guard for local usage
+    modal = None
+
+
+logger = logging.getLogger("slurm_log_analysis")
+
+SANDBOX_TIMEOUT = 600
+DEFAULT_MODAL_SECRET_NAME = "or"
+DEFAULT_MODEL = "openrouter/minimax/minimax-m2.7"
+DEFAULT_REPOS = [
+    "https://github.com/sgl-project/sglang.git",
+]
+PROMPT_PATH = Path(__file__).with_name("log_analysis_prompt.md")
+
+# Matches common API key / token patterns (sk-..., ak-..., as-..., key-..., etc.)
+_SECRET_PATTERN = re.compile(
+    r"""(?:"""
+    r"""(?:sk|ak|as|key|token|secret|bearer)[-_][A-Za-z0-9_\-]{16,}"""
+    r"""|"""
+    r"""(?:OPENROUTER_API_KEY|MODAL_TOKEN_ID|MODAL_TOKEN_SECRET|ANTHROPIC_API_KEY)"""
+    r"""[=:]\s*\S+"""
+    r""")""",
+    re.IGNORECASE,
+)
+
+
+def sanitize(text: str) -> str:
+    """Redact strings that look like API keys or secrets."""
+    if not text:
+        return text
+    sanitized = _SECRET_PATTERN.sub("[REDACTED]", text)
+    # Also redact any env var values we know are secrets
+    for var in ("OPENROUTER_API_KEY", "MODAL_TOKEN_ID", "MODAL_TOKEN_SECRET"):
+        val = os.environ.get(var)
+        if val and len(val) > 8:
+            sanitized = sanitized.replace(val, "[REDACTED]")
+    return sanitized
+
+
+def configure_logging(verbose: bool) -> None:
+    logging.basicConfig(
+        level=logging.INFO,
+        format="%(levelname)s: %(message)s",
+    )
+    logger.setLevel(logging.DEBUG if verbose else logging.INFO)
+
+
+def extract_tarball(tarball: Path, destination: Path) -> None:
+    with tarfile.open(tarball, "r:gz") as archive:
+        # Python 3.14 changes the default extraction behavior. Use the
+        # data filter when available so extraction remains explicit.
+        if "data" in tarfile._NAMED_FILTERS:  # type: ignore[attr-defined]
+            archive.extractall(destination, filter="data")
+        else:
+            archive.extractall(destination)
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(
+        description="Analyze a srtslurm log bundle with opencode in Modal."
+    )
+    source = parser.add_mutually_exclusive_group(required=True)
+    source.add_argument(
+        "--tarball",
+        type=Path,
+        help="Path to a local multinode_server_logs.tar.gz bundle.",
+    )
+    source.add_argument(
+        "--log-dir",
+        type=Path,
+        help="Path to an unpacked log directory.",
+    )
+    parser.add_argument(
+        "--job-id",
+        default="unknown",
+        help="Job identifier used in the report header and logs.",
+    )
+    parser.add_argument(
+        "--model",
+        default=DEFAULT_MODEL,
+        help="Model selector to pass to opencode run.",
+    )
+    parser.add_argument(
+        "--output",
+        type=Path,
+        help="Optional path to write the markdown analysis.",
+    )
+    parser.add_argument(
+        "--repo-url",
+        action="append",
+        dest="repo_urls",
+        help=(
+            "Optional extra repo URL to clone into the sandbox for context. "
+            "Can be specified multiple times."
+        ),
+    )
+    parser.add_argument(
+        "--timeout-seconds",
+        type=int,
+        default=SANDBOX_TIMEOUT,
+        help="Sandbox lifetime in seconds.",
+    )
+    parser.add_argument(
+        "--modal-secret-name",
+        default=DEFAULT_MODAL_SECRET_NAME,
+        help="Modal secret name that provides OPENROUTER_API_KEY to the sandbox.",
+    )
+    parser.add_argument(
+        "--verbose",
+        action="store_true",
+        help="Enable debug logging.",
+    )
+    return parser.parse_args()
+
+
+def build_sandbox_image() -> "modal.Image":
+    if modal is None:
+        raise RuntimeError(
+            "The 'modal' package is required. Run this script with "
+            "`uv run --with modal python ...` or install modal locally."
+        )
+
+    return (
+        modal.Image.debian_slim(python_version="3.12")
+        .apt_install("bash", "curl", "git", "gh")
+        .run_commands(
+            "curl -fsSL https://opencode.ai/install | bash",
+        )
+        .env(
+            {
+                "PATH": "/root/.opencode/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin",
+            }
+        )
+    )
+
+
+def prepare_log_dir(args: argparse.Namespace) -> tuple[Path, Path | None]:
+    if args.log_dir:
+        if not args.log_dir.is_dir():
+            raise FileNotFoundError(f"log directory not found: {args.log_dir}")
+        return args.log_dir.resolve(), None
+
+    assert args.tarball is not None
+    if not args.tarball.is_file():
+        raise FileNotFoundError(f"tarball not found: {args.tarball}")
+
+    temp_dir = Path(tempfile.mkdtemp(prefix="sglang_logs_"))
+    extract_tarball(args.tarball, temp_dir)
+    return temp_dir.resolve(), temp_dir
+
+
+def build_prompt(job_id: str, repo_urls: list[str]) -> str:
+    skill_content = PROMPT_PATH.read_text()
+    repo_lines = []
+    for repo_url in repo_urls:
+        repo_name = repo_url.rsplit("/", 1)[-1].removesuffix(".git")
+        repo_lines.append(f"- **{repo_name} repo**: `/workspace/repos/{repo_name}/`")
+    repo_section = (
+        "\n".join(repo_lines) if repo_lines else "- No extra repos were requested."
+    )
+
+    return f"""{skill_content}
+
+---
+
+## Your Environment
+
+- **Logs**: `/workspace/logs/`
+- **GitHub CLI**: `gh` is installed and authenticated if `GH_TOKEN` is present
+{repo_section}
+
+## Task
+
+Analyze the log bundle for job `{job_id}`.
+
+1. Read the logs in the priority order described above.
+2. Correlate timestamps to identify the most likely root cause.
+3. Use `gh` to inspect related PRs or commits if it materially helps.
+4. **You MUST write the final markdown report to `/workspace/logs/ai_analysis.md`.**
+   This is a hard requirement. Do not just print the report to stdout.
+   Use your file-writing tool to create `/workspace/logs/ai_analysis.md` with the
+   full analysis. The downstream pipeline reads this file.
+
+Keep the report concise but concrete. Include evidence, likely root cause, and notable noise.
+"""
+
+
+def upload_tree(sandbox: "modal.Sandbox", log_dir: Path) -> None:
+    log_files = [path for path in log_dir.rglob("*") if path.is_file()]
+    logger.info("Uploading %d log files into the sandbox", len(log_files))
+    for index, log_file in enumerate(log_files, start=1):
+        rel_path = log_file.relative_to(log_dir)
+        remote_path = str(Path("/workspace/logs") / rel_path)
+        sandbox.mkdir(str(Path(remote_path).parent), parents=True)
+        sandbox.filesystem.write_bytes(log_file.read_bytes(), remote_path)
+        if index % 10 == 0 or index == len(log_files):
+            logger.info("Uploaded %d/%d files", index, len(log_files))
+
+
+def clone_context_repos(sandbox: "modal.Sandbox", repo_urls: list[str]) -> None:
+    if not repo_urls:
+        return
+
+    sandbox.mkdir("/workspace/repos", parents=True)
+
+    for repo_url in repo_urls:
+        repo_name = repo_url.rsplit("/", 1)[-1].removesuffix(".git")
+        logger.info("Cloning %s into the sandbox", repo_name)
+        sandbox.exec(
+            "git",
+            "clone",
+            "--depth",
+            "100",
+            repo_url,
+            f"/workspace/repos/{repo_name}",
+        ).wait()
+
+
+def read_optional_file(sandbox: "modal.Sandbox", path: str) -> str | None:
+    try:
+        return sandbox.filesystem.read_text(path)
+    except Exception:
+        return None
+
+
+def run_opencode_analysis(
+    *,
+    log_dir: Path,
+    job_id: str,
+    model: str,
+    repo_urls: list[str],
+    timeout_seconds: int,
+    modal_secret_name: str,
+) -> str:
+    prompt = build_prompt(job_id, repo_urls)
+    app = modal.App.lookup("sglang-log-analyzer", create_if_missing=True)
+    sandbox = modal.Sandbox.create(
+        app=app,
+        image=build_sandbox_image(),
+        timeout=timeout_seconds,
+        secrets=[modal.Secret.from_name(modal_secret_name)],
+    )
+    logger.info("Created Modal sandbox %s", sandbox.object_id)
+
+    try:
+        sandbox.mkdir("/workspace/logs", parents=True)
+        sandbox.mkdir("/workspace/repos", parents=True)
+
+        clone_context_repos(sandbox, repo_urls)
+        upload_tree(sandbox, log_dir)
+
+        sandbox.filesystem.write_text(prompt, "/workspace/prompt.txt")
+
+        runner_script = f"""#!/bin/bash
+set -uo pipefail
+cd /workspace
+if opencode run \\
+  --dangerously-skip-permissions \\
+  --dir /workspace/logs \\
+  -m {shlex.quote(model)} \\
+  "$(cat /workspace/prompt.txt)" \\
+  < /dev/null \\
+  > /workspace/logs/opencode.stdout \\
+  2> /workspace/logs/opencode.stderr; then
+  echo 0 > /workspace/logs/opencode.exitcode
+else
+  echo $? > /workspace/logs/opencode.exitcode
+fi
+ls -la /workspace/logs > /workspace/logs/log_dir_listing.txt
+"""
+        sandbox.filesystem.write_text(runner_script, "/workspace/run_opencode.sh")
+        sandbox.exec("chmod", "+x", "/workspace/run_opencode.sh").wait()
+
+        logger.info("Running opencode analysis")
+        process = sandbox.exec(
+            "bash",
+            "/workspace/run_opencode.sh",
+        )
+        process.wait()
+
+        stderr = process.stderr.read()
+        if stderr:
+            logger.warning("runner stderr: %s", sanitize(stderr[:500]))
+
+        exitcode = read_optional_file(sandbox, "/workspace/logs/opencode.exitcode")
+        opencode_stdout = (
+            read_optional_file(sandbox, "/workspace/logs/opencode.stdout") or ""
+        )
+        opencode_stderr = read_optional_file(sandbox, "/workspace/logs/opencode.stderr")
+        log_dir_listing = read_optional_file(
+            sandbox, "/workspace/logs/log_dir_listing.txt"
+        )
+        try:
+            ai_analysis = read_optional_file(sandbox, "/workspace/logs/ai_analysis.md")
+            if ai_analysis and ai_analysis.strip():
+                return sanitize(ai_analysis)
+
+            if opencode_stdout.strip():
+                return sanitize(opencode_stdout)
+
+            raise RuntimeError("opencode completed without producing analysis output")
+        except Exception as exc:
+            stdout = process.stdout.read()
+            if stdout:
+                return sanitize(stdout)
+            details = [
+                f"opencode analysis did not produce a usable report: {exc}",
+                f"exitcode={exitcode!r}",
+                f"stdout_preview={sanitize(opencode_stdout[:500])!r}",
+                f"stderr_preview={sanitize((opencode_stderr or '')[:500])!r}",
+                f"log_dir_listing={sanitize((log_dir_listing or '')[:500])!r}",
+            ]
+            raise RuntimeError(" ".join(details)) from exc
+    finally:
+        sandbox.terminate()
+
+
+def main() -> int:
+    args = parse_args()
+    configure_logging(args.verbose)
+
+    repo_urls = list(DEFAULT_REPOS)
+    if args.repo_urls:
+        repo_urls.extend(args.repo_urls)
+
+    log_dir, cleanup_dir = prepare_log_dir(args)
+    try:
+        analysis = run_opencode_analysis(
+            log_dir=log_dir,
+            job_id=args.job_id,
+            model=args.model,
+            repo_urls=repo_urls,
+            timeout_seconds=args.timeout_seconds,
+            modal_secret_name=args.modal_secret_name,
+        )
+    finally:
+        if cleanup_dir is not None:
+            shutil.rmtree(cleanup_dir, ignore_errors=True)
+
+    print(analysis)
+    if args.output:
+        args.output.write_text(analysis)
+        logger.info("Wrote analysis to %s", args.output)
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())
diff --git a/scripts/ci/slurm/log_analysis_prompt.md b/scripts/ci/slurm/log_analysis_prompt.md
new file mode 100644
index 000000000..f4a3c3f71
--- /dev/null
+++ b/scripts/ci/slurm/log_analysis_prompt.md
@@ -0,0 +1,102 @@
+# srtslurm Log Analysis
+
+You are analyzing logs from a failed srtslurm job. srtslurm is a Python-first
+orchestration framework for running distributed LLM inference benchmarks on
+SLURM clusters using SGLang and TRTLLM backends.
+
+## Quick Start
+
+1. List the directory contents to understand what files are present.
+2. Read files in priority order.
+3. Correlate timestamps to identify the real failure point.
+4. Distinguish root cause from noisy warnings.
+
+## Priority Order
+
+### 1. `sweep_{job_id}.log`
+
+Read this first. It is the orchestration timeline.
+
+Look for:
+- stage transitions
+- worker readiness
+- benchmark start
+- exit codes
+- the last error before teardown
+
+### 2. `benchmark.out`
+
+If present, this usually contains the benchmark-side exception or timeout.
+
+### 3. `artifacts/*/logs/aiperf_*.log`
+
+If present, these often contain framework-level initialization failures and
+HTTP/network issues.
+
+### 4. Worker logs
+
+Focus on errors that line up with the failure timestamp:
+- `{node}_prefill_w{N}.out`
+- `{node}_decode_w{N}.out`
+- `{node}_frontend_{N}.out`
+
+### 5. `infra.out`
+
+Use this to confirm infrastructure failures involving NATS, etcd, ports, or
+service health checks.
+
+## Timestamp Correlation
+
+This is the most important rule.
+
+Many warnings are harmless. The root cause is usually the error that occurs at
+the same time the orchestration log transitions into failure.
+
+Use this method:
+1. Find the failure time in `sweep_{job_id}.log`.
+2. Search other logs for matching timestamps.
+3. Ignore earlier warnings if the job continued past them.
+
+## Common Signal
+
+High-signal failures:
+- `ReadTimeout`
+- `Connection refused`
+- `CUDA out of memory`
+- `NCCL timeout`
+- `Model not found`
+- benchmark exit code failures
+
+Low-signal noise:
+- dependency resolver warnings
+- cleanup warnings during teardown
+- keep-alive failures after the main crash
+- import warnings unrelated to the active model
+
+## Output Format
+
+Write markdown with this structure:
+
+```markdown
+## Job Analysis: {job_id}
+
+### Root Cause
+...
+
+### Evidence
+- `file:line or file`
+- timestamp
+- relevant error text
+
+### Timeline
+- key event -> timestamp
+
+### Noise
+- warnings that were not causal
+
+### Recommended Fix
+...
+```
+
+Keep the report concrete. Avoid generic summaries. If you are unsure, say so
+and explain what evidence is missing.