Debug Remote Ci Failures

1. Download and parse CI artifacts to see exact failure output
2. Fall back to raw job logs when a runner shutdown prevents artifact upload
3. Categorize failures by root cause pattern
4. Prioritize fixes — some categories may resolve others
5. Fix, commit, push, verify — iterative loop
6. Document findings for future reference

Step 1: Download CI Artifacts

# Find the failed workflow run
gh run list --branch <branch-name> --limit 10

# Download all artifacts from the run
mkdir -p /tmp/ci_artifacts_<run_id>
gh run download <run_id> --dir /tmp/ci_artifacts_<run_id>

# If artifacts aren't available, use run logs directly
gh run view <run_id> --log > /tmp/ci_run_<run_id>.log 2>&1

# For runner-shutdown cases, download the raw job log directly
job_id=$(gh api "repos/<owner>/<repo>/actions/runs/<run_id>/jobs" --jq '.jobs[] | select(.name | contains("ubuntu")) | .id')
gh api "repos/<owner>/<repo>/actions/jobs/${job_id}/logs" > /tmp/ci_job_<job_id>.log

Extracting Test Results from Logs

When artifacts contain raw pytest output, extract the failure summary:

# Get just the FAILURES section
grep -A 200 "^FAILURES" /tmp/ci_artifacts_<run_id>/<artifact>/output.txt

# Count failures per OS
grep -c "FAILED" /tmp/ci_artifacts_<run_id>/<os>-*/output.txt

# Get unique test names that failed
grep "^FAILED" /tmp/ci_artifacts_<run_id>/<os>-*/output.txt | sort -u

When resource-monitor or debug artifacts are skipped, grep the raw job log for inline markers:

grep -nE "analysis_resource_debug|op_serialization_resource_debug|shutdown signal|exit code 143" /tmp/ci_job_<job_id>.log

When a prior successful run uploaded ci_resource_monitor.log, treat it as a coarse baseline only:

• It captures whole-run snapshots of top-process %MEM, whole-system free memory, and disk usage.
• It does not replace inline log_resource_snapshot(...) markers for per-test RSS analysis.
• Compare the failed raw log against the successful monitor artifact to distinguish real resource growth from an external cancellation on the same commit.

Step 2: Categorize Failures

Group failures by root cause pattern, not by test name. Common categories:

Category A: Platform-Specific Encoding Issues

Pattern: UnicodeEncodeError: 'charmap' codec can't encode character

Typical OS: Windows only (cp1252 encoding)

Root Cause: Source files contain Unicode characters (arrows, checkmarks, emojis) that Windows default encoding cannot handle.

Fix: Replace Unicode symbols with ASCII equivalents in test output strings:

• \u2713 (checkmark) -> [OK]
• \u2717 (cross) -> [FAIL]
• \u2705 (green check) -> [PASS]
• \u26A0 (warning) -> [WARN]
• \u2192 (arrow) -> ->

Example from PR #197:

# Before (Windows fails)
summary.append(f"  \u2713 {name}: validated")

# After (cross-platform)
summary.append(f"  [OK] {name}: validated")

Category B: Dependency Version Mismatch

Pattern: Shape mismatches, missing methods/attributes, unexpected API behavior

Typical OS: All platforms

Root Cause: CI installs different dependency versions than local development. Common when:

• pyproject.toml pins point to old commits
• Override files (overrides.txt) are out of sync with pyproject.toml
• Local env has from-source installs that haven't been pushed as CI dependency pins

Diagnosis:

# Compare local vs CI package versions
# Local:
pip show transformer-lens sae-lens circuit-tracer nnsight

# CI (from log artifacts):
grep "transformer.lens\|sae.lens\|circuit.tracer\|nnsight" /tmp/ci_artifacts_<run_id>/*/output.txt

Fix: Update dependency pins in pyproject.toml and ensure override-dependencies use the same fork/commit:

# Check for discrepancies between pyproject.toml pins and override files
diff <(grep "TransformerLens\|SAELens" pyproject.toml) <(cat requirements/ci/overrides.txt)

# Regenerate lock file after updating pins
./requirements/utils/lock_ci_requirements.sh

Key Lesson from PR #197: The override-dependencies section in pyproject.toml must match the actual dependency pins. Having TransformerLensOrg in overrides while the main dependency points to speediedan fork causes CI to install the wrong version, even though both forks share commit history.

Category C: Numerical Precision / Backend Parity

Pattern: AssertionError: Tensor close check failed: max diff = 0.09...

Typical OS: All platforms (may vary by CPU architecture)

Root Cause: Different backends (TransformerBridge vs NNsight vs HookedTransformer) produce slightly different numerical results due to:

• Different computation order
• Float32 accumulation differences
• Bridge vs HookedTransformer weight conversion path differences

Fix Options:

1. Relax tolerances if the difference is mathematically acceptable
2. Fix the computation path if there's a genuine bug
3. Accept platform-specific tolerances via parametrize marks

Category D: Test Infrastructure Issues

Pattern: Varied — fixture resolution failures, import errors, mock failures

Root Cause: Test helpers or fixtures that make assumptions about the environment.

Fix: Address the specific infrastructure issue. See the fixture_usage.instructions.md for fixture patterns.

Category E: Memory Exhaustion During Dataset Fingerprinting

Pattern: MemoryError during dill.Pickler → _save_torchTensor → write_large_bytes

Typical OS: Windows (most memory-constrained CI runners), potentially Ubuntu

Root Cause: Dataset.from_generator() (HuggingFace datasets library) hashes all gen_kwargs via dill serialization to build a fingerprint for caching. When gen_kwargs contains a full model module (with all weights), dill tries to serialize the entire model into memory — causing MemoryError on memory-constrained CI runners.

Fix: Keep Dataset.from_generator() but provide an explicit fingerprint= so HuggingFace datasets does not dill-hash gen_kwargs:

# Before (OOMs on CI when datasets hashes gen_kwargs):
dataset = Dataset.from_generator(generator_fn, gen_kwargs=gen_kwargs, features=features, ...)

# After (generator-backed, no dill hashing of module/model state):
dataset = Dataset.from_generator(
  generator_fn,
  gen_kwargs=gen_kwargs,
  features=features,
  fingerprint=generate_random_fingerprint(),
)

This preserves generator-backed writes and avoids the old CI failure mode without materializing the full dataset in memory first.

Example from PR #197: generate_analysis_dataset() in analysis.py passes gen_kwargs containing the ITModule and datamodule. Supplying an explicit per-run fingerprint= prevents datasets from hashing those heavyweight objects.

Category G: Fingerprint Scope vs Persistent Cache Semantics

Pattern: runtime dataset creation succeeds, but there is pressure to reuse the same fingerprint across runs for caching.

Typical OS: All (macOS, Windows, Ubuntu)

Root Cause: the explicit fingerprint used to avoid dill-hashing gen_kwargs is a runtime safety mechanism, not a semantically meaningful cache key. Reusing it across runs would risk stale AnalysisStore data unless the fingerprint accounts for the full analysis contract.

Fix: Use a per-run fingerprint for the immediate runtime path and keep deterministic cross-run caching as a separate design problem.

Lesson: treat the runtime fingerprint and the future AnalysisStore cache key as two different concerns. The runtime fix prevents memory blowups; it does not solve persistent cache reuse.

Category H: CPython Segfault in Upstream Threading (nnsight interleaver)

Pattern: Segmentation fault (exit code 139) during nnsight thread-interleaved analysis. Faulthandler dump shows multiple threads stuck in interleaver.py wait → send → request → envoy.py inputs/input chains.

Typical OS: Windows (observed on 3.13.12). May affect other platforms under memory pressure.

Root Cause: The nnsight interleaver creates worker threads for each trace invocation. During multi-trace analysis operations (e.g., SAE ablation attribution, which runs one trace per alive latent), many threads accumulate. On Windows, this can trigger a CPython-level segfault in the threading layer — not a Python exception, but a process crash.

Diagnosis:

# Download full job log (artifacts may not contain useful data for segfaults)
job_id=$(gh api "repos/<owner>/<repo>/actions/runs/<run_id>/jobs" --jq '.jobs[] | select(.name | contains("windows")) | .id')
gh api "repos/<owner>/<repo>/actions/jobs/${job_id}/logs" > /tmp/windows_log.txt

# Look for segfault and thread dumps
grep -n "Segmentation fault\|Current thread\|faulthandler" /tmp/windows_log.txt

Fix: This is an upstream nnsight issue, not fixable in interpretune code. The multi-trace ablation analysis is also memory-intensive enough to OOM GitHub CI runners (~7GB RAM on Ubuntu). Mark the affected tests as standalone (@RunIf(standalone=True)) so they only run with IT_RUN_STANDALONE_TESTS=1 (locally or on GPU CI). The non-ablation NNsight tests (base, SAE, gradient attribution) typically pass fine in regular CI.

Category F: Unexplained Step Cancellation / Runner Shutdown

Pattern: A CI step shows conclusion: "cancelled" with startedAt: null and zero log output, OR conclusion: "failure" with ##[error]The runner has received a shutdown signal

Typical OS: Ubuntu (observed), potentially any

Root Cause: Usually OOM-kill by the Linux kernel or a transient runner infrastructure issue. The process is killed before producing any output. Key indicators:

• Job ran for a round number of minutes (e.g., exactly 60m)
• No error artifacts uploaded
• Install step succeeded but the test step produced zero lines of output
• Other OS jobs on the same run completed normally

Diagnosis:

# Check step conclusions
gh run view <run_id> --json jobs --jq '.jobs[] | select(.name | contains("ubuntu")) | .steps[] | {name: .name, conclusion: .conclusion}'

# Compare job duration (suspiciously round = likely killed)
gh run view <run_id> --json jobs --jq '.jobs[] | {name: .name, started: .startedAt, completed: .completedAt}'

Fix: Often transient — push a new commit and re-run. If persistent, investigate memory usage during test collection (may need to add resource monitoring or reduce parallel test load).

If the exact same commit previously passed and the rerun dies at a different point without a rising resource trend in the inline markers, treat that as likely runner instability first. Still prefer a small stabilizing change if you can reduce peak fixture retention, but do not assume every cancellation is a deterministic code regression.

Important follow-up from PR #197: GitHub Actions can report the pytest step as failure with Process completed with exit code 143 and The runner has received a shutdown signal, then still skip every downstream if: always() upload step while the job remains temporarily marked in_progress. When this happens, do not wait on artifacts that will never appear. Pull the raw job log directly and rely on inline debug markers.

Category I: Analysis Setup Dies Before the Existing Serialization Marker

Pattern: A targeted analysis test logs the test name, then the runner dies before the existing before_save_reload marker appears.

Typical OS: Ubuntu (observed on GitHub-hosted 22.04)

Root Cause: The crash occurs during analysis setup or run_op_with_config(...), not during the later dataset serialization boundary you were already instrumenting.

Fix / Diagnosis:

1. Add earlier inline snapshots before and after run_op_with_config(...).
2. Add a fixture-entry snapshot inside the serialization helper before save_reload_results_dataset(...).
3. Emit snapshots inline to stdout via log_resource_snapshot(...) so they survive missing artifacts.

Relevant helpers now live in tests/analysis_resource_utils.py:

• log_resource_snapshot(...)
• analysis_resource_debug_enabled()
• IT_ANALYSIS_RESOURCE_DEBUG=1

This lets you answer three distinct questions from the raw log:

• Did the test die before the analysis run started?
• Did it die after the analysis run but before save/reload?
• Did it die during save/reload itself?

Category J: Function-Scoped Analysis Fixtures Need Reuse Without Rehydrating Heavy State

Pattern: Switching an analysis fixture from class scope to function scope avoids OOM, but a follow-on test class becomes too slow or re-instantiates the same heavyweight Bridge/NNsight fixture for each method.

Root Cause: The test needs only a tiny projection of the fixture result, but it keeps asking pytest to build the full fixture again for each assertion.

Fix: Use the resource-aware extraction helpers from tests/analysis_resource_utils.py:

• build_analysis_fixture_payload_extractor(...)
• AnalysisExtractionMixin.extract_cached_fixture_data(...)
• ExtractedFixturePayload

This pattern caches one lightweight payload per fixture key and reuses it across test methods while still allowing low-RAM runners to clean up the heavyweight result / runner / it_session graph immediately.

Step 3: Prioritize Fixes

Apply fixes in dependency order:

1. Dependency pins first — Updating dependency versions may resolve many failures at once (Category B often triggers symptoms that look like Category C or D)
2. Encoding/platform fixes — Quick wins that are clearly correct
3. Refactors — Changes that improve code organization (e.g., moving test helpers to adapter methods)
4. Precision/tolerance fixes — Only after dependencies are correct

Decision Framework

Is the failure caused by wrong dependency version?
  Yes -> Update pins, push, wait for CI before fixing other failures
  No  -> Is it a platform encoding issue?
    Yes -> Replace Unicode with ASCII, commit with other changes
    No  -> Is it a test infrastructure issue?
      Yes -> Fix test code, not application code
      No  -> Investigate application code

Step 4: Fix, Commit, Push, Verify Loop

Efficient CI Iteration

# Commit and push
git add -A && git commit -m "fix: <descriptive message>"
git push origin <branch>

# Find the new Test full run
gh run list --branch <branch> --limit 5

# Monitor progress
gh run view <run_id> --json jobs --jq '.jobs[] | {name: .name, status: .status, conclusion: .conclusion}'

# Once complete, check results
gh run view <run_id> --json jobs --jq '.jobs[] | select(.conclusion == "failure") | .name'

Pre-commit Hook Awareness

The interpretune pre-commit hooks include end-of-file-fixer and trailing-whitespace. These may modify files during the first commit attempt:

# First attempt may fail
git add -A && git commit -m "fix: ..."
# Output: "fix end of files... Failed"

# Files were auto-fixed — just re-add and commit
git add -A && git commit -m "fix: ..."

Handling Auto-Cancellation

GitHub Actions auto-cancels in-progress runs when a new push arrives on the same branch (if configured with concurrency groups). If you need results from a specific run:

• Wait for it to complete before pushing again, OR
• Push only non-test changes (like type annotations) that won't trigger cancellation of the test workflow

Step 5: Analyzing CI Results

Quick Status Check

# Get pass/fail counts from a completed run
gh run view <run_id> --log 2>/dev/null | grep -E "passed|failed|error" | tail -5

Downloading Failure Artifacts

# Download artifacts from failed run
gh run download <run_id> --dir /tmp/ci_artifacts_<run_id>

# Parse test results
for f in /tmp/ci_artifacts_<run_id>/*/output.txt; do
  echo "=== $(dirname $f | xargs basename) ==="
  grep "^FAILED" "$f" | wc -l
  grep "^FAILED" "$f"
done

Cross-Platform Failure Comparison

# Compare failures across OS runners
diff <(grep "^FAILED" /tmp/ci_artifacts/windows-*/output.txt | sed 's/.*FAILED //' | sort) \
     <(grep "^FAILED" /tmp/ci_artifacts/macos-*/output.txt | sed 's/.*FAILED //' | sort)

Lessons Learned from PR #197

1. Dependency Pin Consistency is Critical

The override-dependencies section in pyproject.toml serves a specific purpose: it replaces git URL dependencies with version constraints so that editable installations work correctly. If these overrides point to a different fork than the main dependency pins, CI will install the wrong package version silently.

Always verify: pyproject.toml dependency URLs, requirements/ci/overrides.txt, and override-dependencies all reference the same fork and commit.

2. Don't Chase Symptoms Before Fixing Dependencies

When dependency versions are wrong, failures cascade. A single wrong TransformerLens version can cause 20+ tensor shape failures, 3 cache key failures, and 3 numerical parity failures — all from the same root cause. Fix dependency pins first and re-run CI before investigating individual test failures.

3. Unicode in Test Output is a Windows Landmine

Even when tests pass on Linux and macOS, Windows CI runners use cp1252 encoding by default. Any Unicode characters in test assertion messages, logging output, or summary strings will cause UnicodeEncodeError. Use ASCII-only characters in all test output.

4. Type Checking is a Separate CI Job

The interpretune CI runs pyright type checking as a separate job from pytest. Adding new methods or modifying existing ones may trigger type errors even if tests pass. Check the "Stale Stubs and Type Checks" workflow alongside "Test full".

5. Pre-commit Hooks May Modify Files

The end-of-file-fixer and trailing-whitespace hooks can modify files during commit, causing the first git commit to fail. This is expected — just re-stage and commit again.

6. SAE Hook Name Resolution Varies by Model Backend

TransformerBridge resolves hook aliases to canonical names (e.g., blocks.0.hook_resid_pre -> blocks.0.hook_in), while HookedTransformer uses aliases directly. Any code that constructs SAE hook paths must account for this via SAELensTLModuleMixin.resolve_sae_hook_name() (or equivalent) rather than hardcoding the alias + suffix pattern.

7. Avoid Passing Full Modules Through HF Datasets Fingerprinting

Dataset.from_generator() uses dill to serialize all gen_kwargs for cache fingerprinting. If gen_kwargs contains a PyTorch module with model weights, this serializes the entire model into memory — an instant OOM on CI runners. Prefer keeping Dataset.from_generator() and supplying an explicit per-run fingerprint= so the runtime path stays generator-backed without hashing heavyweight objects.

8. "Cancelled" Steps With No Output Usually Mean OOM Kill

When a CI step shows conclusion: "cancelled" with zero log lines, the process was likely killed by the OS (e.g., Linux OOM killer). This is different from GitHub Actions "cancel-in-progress" which cancels the entire run. Look for round job durations (exactly 60m, 90m) and check if other OS jobs in the same run completed normally.

9. Runtime Fingerprints Do Not Solve Persistent Caching

An explicit runtime fingerprint= solves the immediate Dataset.from_generator() dill-serialization problem, but it should not be reused as a persistent AnalysisStore cache key. The runtime fingerprint is intentionally per-run right now; deterministic cross-run cache reuse still requires a separate, semantically meaningful hashing design.

10. Upstream Threading Crashes May Only Surface on Specific Platforms

A segfault in an upstream library's threading layer (e.g., nnsight's interleaver on Windows) is not reproducible locally if your development machine runs Linux/macOS. When you see a segfault with faulthandler output showing threads stuck in wait/send chains, it's likely an upstream platform-specific bug. Similarly, memory-intensive upstream operations (multi-trace ablation creating many threads) may OOM CI runners without affecting local machines with more RAM. The correct fix is @RunIf(standalone=True) — moving resource-heavy tests out of standard CI — rather than a code change in your project.

Related Files

• .github/workflows/ci_test-full.yml — Main CI workflow
• .github/workflows/ci_stubs_types.yml — Type checking workflow
• pyproject.toml — Dependency pins and override-dependencies
• requirements/ci/overrides.txt — CI dependency overrides
• requirements/ci/requirements.txt — Locked CI requirements
• .github/instructions/fixture_usage.instructions.md — Test fixture patterns