mlx-triage/src/mlx_triage/data/known_bugs.yaml at main · swaynehales/mlx-triage · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
# Curated database of documented MLX/framework bugs.
# Sources: MLX GitHub issues, community postmortems, research evidence.
# See docs/evidence/ for primary sources.

bugs:
  - id: MLX-001
    title: "Float16 addmm CPU wrong results"
    mlx_issue: 2695
    affected_versions: ["< 0.22.0"]
    severity: critical
    detection: "tier0.version_check"
    symptom: "Completely wrong matrix multiplication results on CPU with float16"
    architecture: ["all"]
    remediation: "Use bfloat16 or float32, or upgrade MLX"

  - id: MLX-002
    title: "qmv kernel infinite loop at specific tensor dimensions"
    mlx_issue: null
    affected_versions: ["< 0.24.0"]
    severity: critical
    detection: "tier0.version_check"
    symptom: "Model enters infinite generation loop at end of long prompts (4-bit only)"
    architecture: ["llama", "mistral"]
    remediation: "Use 8-bit quantization or upgrade MLX"

  - id: MLX-003
    title: "Metal float16 non-determinism with error accumulation"
    mlx_issue: 488
    affected_versions: ["all"]
    severity: warning
    detection: "tier1.determinism"
    symptom: "Non-reproducible outputs at temperature=0 with float16 weights"
    architecture: ["all"]
    remediation: "Use quantized integer models (Q4, Q8) for reproducibility"

  - id: MLX-004
    title: "bfloat16 save_safetensors silent corruption"
    mlx_issue: null
    source: "baa.ai postmortem"
    affected_versions: ["all"]
    severity: critical
    detection: "tier0.weight_integrity"
    symptom: "Structurally valid safetensors file with numerically garbage weights"
    architecture: ["all"]
    remediation: "Verify weight checksums against hub; re-download if mismatch"

  - id: MLX-005
    title: "Conv1d composition drift"
    mlx_issue: 2122
    affected_versions: ["all"]
    severity: warning
    detection: "tier3.activation_diff"
    symptom: "Sequential Conv1d operations accumulate MAE ~0.04 vs PyTorch"
    architecture: ["whisper", "wav2vec", "audio_models"]
    remediation: "Use float32 for audio model inference or validate output quality"

  - id: MLX-006
    title: "GELU approximation mismatch in vision encoders"
    mlx_issue: null
    source: "PaliGemma MLX issue"
    affected_versions: ["all"]
    severity: high
    detection: "tier3.activation_diff"
    symptom: "Vision encoder activations diverge >100K sum-abs-diff from reference"
    architecture: ["paligemma", "clip", "siglip"]
    remediation: "Verify GELU implementation matches model's training config"