transformers
Idefics: Fix information leak with cross attention gate in modeling
#26839
Merged

Commits
  • fix image_attention gate in idefics modeling
    leot13 committed 2 years ago
  • update comment
    leot13 committed 2 years ago
  • cleaner gating
    leot13 committed 2 years ago
  • fix gate condition
    leot13 committed 2 years ago
  • create attention gate once
    leot13 committed 2 years ago
  • update comment
    leot13 committed 2 years ago
  • update doc of cross-attention forward
    leot13 committed 2 years ago
  • improve comment
    leot13 committed 2 years ago
  • Merge branch 'main' into fix-idefics-image-attention
    leot13 committed 2 years ago
  • bring back no_images
    leot13 committed 2 years ago
  • pass cross_attention_gate similarly to no_images gate
    leot13 committed 2 years ago
  • add information on gate shape
    leot13 committed 2 years ago
  • fix no_images placement
    leot13 committed 2 years ago
  • make tests for gate
    leot13 committed 2 years ago
  • take off no_images logic
    leot13 committed 2 years ago
  • Merge branch 'main' into fix-idefics-image-attention
    leot13 committed 2 years ago
  • update test based on comments
    leot13 committed 2 years ago
  • raise value error if cross_attention_gate is None
    leot13 committed 2 years ago
  • send cross_attention_gate to device
    leot13 committed 2 years ago
  • Revert "send cross_attention_gate to device"
    leot13 committed 2 years ago
  • send cross_attention_gate to device
    leot13 committed 2 years ago
  • fix device in test + nit
    leot13 committed 2 years ago
  • fill hidden_states with zeros instead of multiplying with the gate
    leot13 committed 2 years ago
  • style
    leot13 committed 2 years ago
  • Update src/transformers/models/idefics/modeling_idefics.py
    leot13 committed 2 years ago
  • Update src/transformers/models/idefics/modeling_idefics.py
    leot13 committed 2 years ago
Loading