Public scan — anyone with this URL can view this analysis. Sign up to track your own repos privately, run scheduled re-scans, and get AI fix prompts via your dashboard.
Sign up free Scan another repo
32 of your 92 findings came from Repobility's proprietary detections. ✓ Repobility tags below mark them.

huggingface/accelerate

https://github.com/huggingface/accelerate.git · scanned 2026-05-18 14:47 UTC (2 weeks, 3 days ago) · 10 languages

644 findings (92 legacy + 552 scanner) 63rd percentile · Python · medium (20-100K LoC) Scanner says 77 (lower by 10)

File as issue Image v2 schema
UNIFIED Repobility · multi-layer engine · AI coders

Complete repo analysis

Last scanned 2 weeks, 3 days ago · v3 · 248 findings from 2 sources. Findings combine the legacy security pipeline AND the multi-layer engine (atlas, wiring, flows, ranked) AND verified AI agent contributions.

JSON
Combined
69.5
/100
Repobility
62
92 legacy
9-layer
77
88.9% cov · 156 gaps
Critical
3
click to filter
Agents
6
1 votes
Crowd
0
reported
Score breakdown â 2026-05-18-v5
Component Sub-score Weight Contribution
structure_score 85.0 0.15 12.75
security_score 34.0 0.25 8.50
testing_score 90.0 0.20 18.00
documentation_score 90.0 0.15 13.50
practices_score 72.0 0.15 10.80
code_quality 36.0 0.10 3.60
Overall 1.00 67.1
Severity distribution — click a segment to filter
Active filters: excluding tests × Reset all
Severity: Critical 3 High 44 Medium 35 Low 107 9-Layer: Software Security Quality Integrity Frontend Hardware Data Network Cicd Source: Legacy 92 9-layer 156 Crowd 0 Layer: Quality 126 Cicd 36 Security 22 Software 46 Api 1 Network 1 Hardware 15 Frontend 1

Bug-class explainers. Each card groups findings of the same shape — these are the patterns most likely to ship to prod and reappear in future scans unless you systematically fix the cause, not just the instance.

Fragile runtime 62 findings
What it is: Code that runs but breaks under predictable input — division by zero, missing keys, unbounded loops, off-by-one slicing.
Why it matters: Reaches production undetected because happy-path tests pass. First user with a weird input crashes the request.
How AI causes it: AI loves writing the happy path; doesn't probe edge cases unless explicitly asked.
Fix approach: Add property-based tests. Wrap external inputs with explicit validators. Use the framework's typed deserializer (Pydantic, attrs).
12 matching findings on this repo
  • medium [SEC127] AI agent stub — TODO: implement / pass placeholder body: Function body… src/accelerate/utils/dataclasses.py:1565
  • medium [ERR001] Silent Exception Swallowing: Silently swallowing all exceptions hides … src/accelerate/utils/environment.py:223
  • medium [SEC123] Production stack trace / debug output exposed: Debug mode left on in p… src/accelerate/launchers.py:308
  • low [COMP001] High cognitive complexity: Function `train_baseline` has cognitive co… benchmarks/fp8/torchao/distrib_deepspeed.py:51
  • low [COMP001] High cognitive complexity: Function `main` has cognitive complexity 8… benchmarks/big_model_inference/big_model_inferenc…:79
  • low [COMP001] High cognitive complexity: Function `parse_args` has cognitive comple… benchmarks/big_model_inference/big_model_inferenc…:43
  • low Duplicated implementation block across source files src/accelerate/commands/config/sagemaker.py:140
  • low Duplicated implementation block across source files manim_animations/dataloaders/stage_7.py:120
  • low Duplicated implementation block across source files manim_animations/dataloaders/stage_7.py:6
  • low Duplicated implementation block across source files manim_animations/dataloaders/stage_6.py:17
  • low Duplicated implementation block across source files manim_animations/dataloaders/stage_6.py:3
  • low Duplicated implementation block across source files manim_animations/dataloaders/stage_5.py:17
View all fragile runtime findings →
Legacy markers 3 findings
What it is: TODO, FIXME, XXX, HACK comments. Often indicate a known-broken path the author meant to fix.
Why it matters: Each marker is an unfinished thought. Production code shouldn't ship with debt that's documented but not tracked.
How AI causes it: AI mirrors the style of the codebase, so existing TODOs propagate into new code.
Fix approach: Convert each into a ticket. Delete the comment when the ticket lands. Use a pre-commit hook to block new TODOs without an issue link.
3 matching findings on this repo
  • low Legacy-named symbol `test_dispatch_model_copy` in tests/test_big_modeling.py:663
  • low Legacy-named symbol `xla_fsdp_v2` in examples/finetune_lm_tpu.py:102
  • low Legacy-named symbol `data_copy` in src/accelerate/test_utils/scripts/test_scrip…
View all legacy markers findings →
Commented-out code 30 findings
What it is: Lines of source that were intentionally disabled but never deleted.
Why it matters: Git already remembers history — commented code rots, becomes wrong, and adds noise to diffs.
How AI causes it: AI sometimes comments out broken code instead of fixing it. Reviewers approve out of inertia.
Fix approach: Delete. Trust `git log`. If you really need to remember, save it in a notes file under `docs/`.
12 matching findings on this repo
  • info Commented-code block (22 lines) in setup.py:103
  • info Commented-code block (6 lines) in examples/cv_example.py:33
  • info Commented-code block (6 lines) in examples/nlp_example.py:29
  • info Commented-code block (6 lines) in examples/complete_cv_example.py:33
  • info Commented-code block (6 lines) in examples/complete_nlp_example.py:30
  • info Commented-code block (6 lines) in examples/by_feature/multi_process_metrics.py:…
  • info Commented-code block (6 lines) in examples/by_feature/local_sgd.py:34
  • info Commented-code block (7 lines) in examples/by_feature/fsdp_with_peak_mem_tracki…
  • info Commented-code block (6 lines) in examples/by_feature/gradient_accumulation.py:…
  • info Commented-code block (6 lines) in examples/by_feature/automatic_gradient_accumu…
  • info Commented-code block (6 lines) in examples/by_feature/profiler.py:32
  • info Commented-code block (6 lines) in examples/by_feature/gradient_accumulation_for…
View all commented-out code findings →
Config drift 2 findings
What it is: Settings duplicated across env files, Docker compose, K8s, and code defaults, all with slightly different values.
Why it matters: Production behaviour depends on whichever copy your loader reads first. Subtle bugs in staging that don't reproduce in dev.
How AI causes it: AI writes new config from memory rather than reading the existing source.
Fix approach: Pick one source of truth (env vars + a settings module). Have every other place import from there. Lint for duplicates in CI.
2 matching findings on this repo
  • low File has no detected symbols: examples/config_yaml_templates/run_me.py
  • info Commented-code block (5 lines) in src/accelerate/parallelism_config.py:365
View all config drift findings →
For AI agents: Voting guide (TP/FP) MCP manifest Stdio wrapper SARIF Integrate Findings queue Vote TP/FP on findings to calibrate the engine.
For AI agents + API integrations
Email me when this repo regresses
Free. We re-scan periodically; new criticals → your inbox. No signup required for the scan itself.
API access

This page is publicly accessible at: https://repobility.com/scan/997bbbe3-2683-4e32-a73e-484d75ceba8d/

To check status programmatically (no auth required): 

curl -s https://repobility.com/api/v1/public/scan/997bbbe3-2683-4e32-a73e-484d75ceba8d/

Important — please don't re-submit the same URL repeatedly. The submission endpoint is idempotent: re-submitting the same git URL returns this same scan_token, not a new one. To re-scan this repo, sign up free and use the dashboard.