NOEON / docs /COMPAT.md
Frankenstein-Labs's picture
NOEON: capability-slot rename (vision/computer_use/generation) + provenance; migrate weights
b871251 verified
|
Raw History Blame Contribute Delete
1.92 kB

NOEON source compatibility (verified against real tensors)

Derived from weights/source_tensors.json, which is read directly from each source's safetensors headers. Nothing here is assumed.

vision vs computer_use

  • vision tensors: 760
  • computer_use tensors: 696
  • common names (after prefix strip): 298
  • identical shapes among common: 298
  • shape conflicts: 0
  • vision-only: 462
  • computer_use-only: 398

Real hidden widths (most common 2-D axis)

  • vision: [(2560, 291), (1024, 200), (4096, 131), (9728, 108), (3072, 24), (12288, 12)]
  • computer_use: [(2560, 255), (1024, 192), (4096, 123), (9728, 108), (3072, 24), (151936, 2)]
  • generation: [(3072, 295), (2560, 258), (1024, 193), (4096, 132), (9728, 108), (384, 49)]

NOEON shared dimension

  • vision: hidden dim 2560 present: True (count 291)
  • computer_use: hidden dim 2560 present: True (count 255)
  • generation: hidden dim 2560 present: True (count 258)

Candidate hidden-stream tensors (for bridge inputs)

  • vision: 289 tensors touch dim 2560
    • model.language_model.embed_tokens.weight (151936, 2560)
    • model.language_model.layers.0.mlp.down_proj.weight (2560, 9728)
    • model.language_model.layers.0.mlp.gate_proj.weight (9728, 2560)
    • model.language_model.layers.0.mlp.up_proj.weight (9728, 2560)
  • computer_use: 255 tensors touch dim 2560
    • model.embed_tokens.weight (151936, 2560)
    • model.layers.0.mlp.down_proj.weight (2560, 9728)
    • model.layers.0.mlp.gate_proj.weight (9728, 2560)
    • model.layers.0.mlp.up_proj.weight (9728, 2560)
  • generation: 258 tensors touch dim 2560
    • model.language_model.embed_tokens.weight (151936, 2560)
    • model.language_model.layers.0.mlp.down_proj.weight (2560, 9728)
    • model.language_model.layers.0.mlp.gate_proj.weight (9728, 2560)
    • model.language_model.layers.0.mlp.up_proj.weight (9728, 2560)