diff --git "a/vla-adapter/trace.json" "b/vla-adapter/trace.json"
new file mode 100644--- /dev/null
+++ "b/vla-adapter/trace.json"
@@ -0,0 +1 @@
+{"format":2,"model":"auto/vla_adapter","title":"VLA-Adapter (1.6B): published base weights, adapter untrained","auto_title":"vla_adapter (vla_adapter)","views":["wrote","changed"],"params":1575044575,"height_unit":1024,"source":{"checkpoint":"weights.safetensors","step":null,"package":"vla_adapter","code":"model.py","observation":null,"hooked_vs_plain_max_difference":0.0,"traced_by":"rebuild.trace_any","checks":{"camera grid":{"changed cell":[2,12],"most moved token":[1,12],"closest layout":"as drawn","ok":true}}},"input":{"seed":9000007,"step":null,"text":"Pick up the lime star","swap":false,"pieces":["","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","","",""],"state":[2.0524,3.4842,1.6106,-1.555,-1.2666,-1.4022,0.0]},"output":{"poses":[[2.035,3.4866,1.607,-1.5257,-1.2687,-1.4106,0.0447],[2.0177,3.4892,1.6032,-1.4964,-1.2707,-1.4192,0.0891],[2.0005,3.4918,1.5995,-1.4669,-1.2728,-1.4277,0.1333],[1.9834,3.4945,1.5958,-1.4372,-1.2749,-1.4362,0.1774],[1.9663,3.4974,1.5921,-1.4073,-1.277,-1.4445,0.2215],[1.9492,3.5006,1.5884,-1.3773,-1.2791,-1.4527,0.2661],[1.9321,3.504,1.5847,-1.3469,-1.2814,-1.4608,0.3109],[1.915,3.5076,1.581,-1.3161,-1.2839,-1.4689,0.3559]]},"plots":[{"id":"learned","label":"start token","kind":"learned","grid":[1,1],"tokens":[0,1],"at":[54,0],"auto":{"label":"learned tokens"}},{"id":"cam_top","label":"top camera","kind":"image","grid":[16,16],"tokens":[1,256],"at":[0,0],"auto":{"label":"top camera"}},{"id":"cam_side","label":"side camera","kind":"image","grid":[16,16],"tokens":[257,256],"at":[18,0],"auto":{"label":"side camera"}},{"id":"cam_iso","label":"iso camera","kind":"image","grid":[16,16],"tokens":[513,256],"at":[36,0],"auto":{"label":"iso camera"}},{"id":"learned_2","label":"prompt words (fixed)","kind":"learned","grid":[2,15],"tokens":[769,30],"at":[54,3],"auto":{"label":"learned tokens"}},{"id":"text","label":"instruction","kind":"text","grid":[1,5],"tokens":[799,5],"at":[54,7],"auto":{"label":"instruction"}},{"id":"learned_3","label":"prompt end and 64 action queries","kind":"learned","grid":[5,14],"tokens":[804,70],"at":[54,10],"auto":{"label":"learned tokens"}},{"id":"answer","label":"the chunk (action head)","kind":"action","grid":[1,8],"tokens":[874,8],"at":[54,17],"auto":{"label":"the answer"}},{"id":"state","label":"arm state","kind":"state","grid":[1,1],"tokens":[882,1],"at":[54,20],"auto":{"label":"arm state"}}],"stages":[{"id":"in.image","label":"camera image","kind":"input","group":["inputs"],"plots":["cam_top","cam_side","cam_iso"],"width":3,"module":"","params":0,"note":"512 x 512 pixels, red, green and blue, as the robot's cameras see them.","cells":{"cam_top":[512,512],"cam_side":[512,512],"cam_iso":[512,512]},"wrote":{"cam_top":"in.image/rgb/cam_top","cam_side":"in.image/rgb/cam_side","cam_iso":"in.image/rgb/cam_iso"},"holds":{},"picture":true,"auto":{"label":"camera image","note":"512 x 512 pixels, red, green and blue, as the robot's cameras see them.","group":["inputs"],"kind":"input"}},{"id":"in.text","label":"instruction","kind":"input","group":["inputs"],"plots":["text"],"width":1,"module":"","params":0,"note":"The instruction, cut into word pieces by the model's own tokenizer.","cells":{"text":[1,5]},"wrote":{"text":"in.text/wrote/text"},"holds":{},"auto":{"label":"instruction","note":"The instruction, cut into word pieces by the model's own tokenizer.","group":["inputs"],"kind":"input"}},{"id":"in.state","label":"arm state","kind":"input","group":["inputs"],"plots":["state"],"width":7,"module":"","params":0,"note":"6 joint angles and the gripper opening.","cells":{"state":[1,1]},"wrote":{"state":"in.state/wrote/state"},"holds":{},"auto":{"label":"arm state","note":"6 joint angles and the gripper opening.","group":["inputs"],"kind":"input"}},{"id":"vla.base_model.model.vision_backbone.featurizer.patch_embed","label":"DINOv2: cut into 14-pixel patches","kind":"conv","group":["vision_backbone"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.patch_embed","params":706304,"note":"Each camera, resized to 224 x 224, is cut into 16 x 16 patches of 14 pixels by one convolution. A class token and 4 register tokens are added in front (not drawn): 261 tokens of 1,024 numbers. The same weights serve all three cameras, one camera at a time.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.patch_embed/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.patch_embed/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.patch_embed/wrote/cam_iso"},"holds":{},"auto":{"label":"vision_backbone.featurizer.patch_embed: Conv2d, Identity, Conv2d, Conv2d, ParameterDict, ParameterDict and 1 more","note":"PatchEmbed. Out: [1, 256, 1024]. A 14 x 14 convolution, step 14, 3 channels in, 1024 out: it cuts the picture into 14-pixel patches. LoRA on 1 layer: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). Part of vit_large_patch14_reg4_dinov2 (as its code names it).","group":["vision_backbone"],"kind":"conv"}},{"id":"vla.base_model.model.vision_backbone.featurizer.pos_drop","label":"vision_backbone.featurizer.pos_drop: Dropout","kind":"linear","group":["vision_backbone"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.pos_drop","params":0,"note":"Dropout. Out: [1, 261, 1024]. Part of vit_large_patch14_reg4_dinov2 (as its code names it). The first 5 tokens of each camera (before its 16 x 16 patches) are not drawn.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.pos_drop/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.pos_drop/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.pos_drop/wrote/cam_iso"},"holds":{},"auto":{"label":"vision_backbone.featurizer.pos_drop: Dropout","note":"Dropout. Out: [1, 261, 1024]. Part of vit_large_patch14_reg4_dinov2 (as its code names it). The first 5 tokens of each camera (before its 16 x 16 patches) are not drawn.","group":["vision_backbone"],"kind":"linear"}},{"id":"vla.base_model.model.vision_backbone.featurizer.patch_drop","label":"vision_backbone.featurizer.patch_drop: Identity","kind":"linear","group":["vision_backbone"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.patch_drop","params":0,"note":"Identity. Out: [1, 261, 1024]. Part of vit_large_patch14_reg4_dinov2 (as its code names it). The first 5 tokens of each camera (before its 16 x 16 patches) are not drawn.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.patch_drop/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.patch_drop/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.patch_drop/wrote/cam_iso"},"holds":{},"auto":{"label":"vision_backbone.featurizer.patch_drop: Identity","note":"Identity. Out: [1, 261, 1024]. Part of vit_large_patch14_reg4_dinov2 (as its code names it). The first 5 tokens of each camera (before its 16 x 16 patches) are not drawn.","group":["vision_backbone"],"kind":"linear"}},{"id":"vla.base_model.model.vision_backbone.featurizer.norm_pre","label":"vision_backbone.featurizer.norm_pre: Identity","kind":"linear","group":["vision_backbone"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.norm_pre","params":0,"note":"Identity. Out: [1, 261, 1024]. Part of vit_large_patch14_reg4_dinov2 (as its code names it). The first 5 tokens of each camera (before its 16 x 16 patches) are not drawn.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.norm_pre/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.norm_pre/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.norm_pre/wrote/cam_iso"},"holds":{},"auto":{"label":"vision_backbone.featurizer.norm_pre: Identity","note":"Identity. Out: [1, 261, 1024]. Part of vit_large_patch14_reg4_dinov2 (as its code names it). The first 5 tokens of each camera (before its 16 x 16 patches) are not drawn.","group":["vision_backbone"],"kind":"linear"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.0.norm1","label":"normalise","kind":"norm","group":["vision_backbone","blocks 1"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.0.norm1","params":2048,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.0.norm1/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.0.norm1/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.0.norm1/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 1"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.0.attn","label":"attention","kind":"attention","group":["vision_backbone","blocks 1"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.0.attn, vla.base_model.model.vision_backbone.featurizer.blocks.0.ls1","params":4592640,"note":"DINOv2, trained without words to see shapes and parts. The tokens of one camera read each other; cameras never meet in the vision encoders.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.0.attn/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.0.attn/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.0.attn/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.0.attn/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.0.attn/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.0.attn/holds/cam_iso"},"auto":{"label":"attention","note":"Attention. The only place tokens exchange information: each token reads from the others and what it reads is added to it. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 16 heads. Part of vit_large_patch14_reg4_dinov2 (as its code names it).","group":["vision_backbone","blocks 1"],"kind":"attention"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.0.norm2","label":"normalise","kind":"norm","group":["vision_backbone","blocks 1"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.0.norm2","params":2048,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.0.norm2/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.0.norm2/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.0.norm2/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 1"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.0.mlp","label":"MLP","kind":"mlp","group":["vision_backbone","blocks 1"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.0.mlp, vla.base_model.model.vision_backbone.featurizer.blocks.0.ls2","params":9050112,"note":"Works on each token by itself: Mlp, LayerScale. The result is added to the token.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.0.mlp/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.0.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.0.mlp/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.0.mlp/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.0.mlp/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.0.mlp/holds/cam_iso"},"auto":{"label":"MLP","note":"Works on each token by itself: Mlp, LayerScale. The result is added to the token.","group":["vision_backbone","blocks 1"],"kind":"mlp"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.1.norm1","label":"normalise","kind":"norm","group":["vision_backbone","blocks 2"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.1.norm1","params":2048,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.1.norm1/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.1.norm1/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.1.norm1/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 2"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.1.attn","label":"attention","kind":"attention","group":["vision_backbone","blocks 2"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.1.attn, vla.base_model.model.vision_backbone.featurizer.blocks.1.ls1","params":4592640,"note":"DINOv2, trained without words to see shapes and parts. The tokens of one camera read each other; cameras never meet in the vision encoders.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.1.attn/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.1.attn/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.1.attn/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.1.attn/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.1.attn/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.1.attn/holds/cam_iso"},"auto":{"label":"attention","note":"Attention. The only place tokens exchange information: each token reads from the others and what it reads is added to it. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 16 heads.","group":["vision_backbone","blocks 2"],"kind":"attention"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.1.norm2","label":"normalise","kind":"norm","group":["vision_backbone","blocks 2"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.1.norm2","params":2048,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.1.norm2/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.1.norm2/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.1.norm2/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 2"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.1.mlp","label":"MLP","kind":"mlp","group":["vision_backbone","blocks 2"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.1.mlp, vla.base_model.model.vision_backbone.featurizer.blocks.1.ls2","params":9050112,"note":"Works on each token by itself: Mlp, LayerScale. The result is added to the token.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.1.mlp/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.1.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.1.mlp/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.1.mlp/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.1.mlp/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.1.mlp/holds/cam_iso"},"auto":{"label":"MLP","note":"Works on each token by itself: Mlp, LayerScale. The result is added to the token.","group":["vision_backbone","blocks 2"],"kind":"mlp"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.2.norm1","label":"normalise","kind":"norm","group":["vision_backbone","blocks 3"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.2.norm1","params":2048,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.2.norm1/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.2.norm1/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.2.norm1/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 3"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.2.attn","label":"attention","kind":"attention","group":["vision_backbone","blocks 3"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.2.attn, vla.base_model.model.vision_backbone.featurizer.blocks.2.ls1","params":4592640,"note":"DINOv2, trained without words to see shapes and parts. The tokens of one camera read each other; cameras never meet in the vision encoders.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.2.attn/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.2.attn/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.2.attn/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.2.attn/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.2.attn/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.2.attn/holds/cam_iso"},"auto":{"label":"attention","note":"Attention. The only place tokens exchange information: each token reads from the others and what it reads is added to it. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 16 heads.","group":["vision_backbone","blocks 3"],"kind":"attention"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.2.norm2","label":"normalise","kind":"norm","group":["vision_backbone","blocks 3"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.2.norm2","params":2048,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.2.norm2/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.2.norm2/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.2.norm2/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 3"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.2.mlp","label":"MLP","kind":"mlp","group":["vision_backbone","blocks 3"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.2.mlp, vla.base_model.model.vision_backbone.featurizer.blocks.2.ls2","params":9050112,"note":"Works on each token by itself: Mlp, LayerScale. The result is added to the token.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.2.mlp/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.2.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.2.mlp/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.2.mlp/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.2.mlp/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.2.mlp/holds/cam_iso"},"auto":{"label":"MLP","note":"Works on each token by itself: Mlp, LayerScale. The result is added to the token.","group":["vision_backbone","blocks 3"],"kind":"mlp"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.3.norm1","label":"normalise","kind":"norm","group":["vision_backbone","blocks 4"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.3.norm1","params":2048,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.3.norm1/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.3.norm1/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.3.norm1/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 4"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.3.attn","label":"attention","kind":"attention","group":["vision_backbone","blocks 4"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.3.attn, vla.base_model.model.vision_backbone.featurizer.blocks.3.ls1","params":4592640,"note":"DINOv2, trained without words to see shapes and parts. The tokens of one camera read each other; cameras never meet in the vision encoders.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.3.attn/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.3.attn/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.3.attn/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.3.attn/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.3.attn/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.3.attn/holds/cam_iso"},"auto":{"label":"attention","note":"Attention. The only place tokens exchange information: each token reads from the others and what it reads is added to it. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 16 heads.","group":["vision_backbone","blocks 4"],"kind":"attention"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.3.norm2","label":"normalise","kind":"norm","group":["vision_backbone","blocks 4"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.3.norm2","params":2048,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.3.norm2/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.3.norm2/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.3.norm2/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 4"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.3.mlp","label":"MLP","kind":"mlp","group":["vision_backbone","blocks 4"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.3.mlp, vla.base_model.model.vision_backbone.featurizer.blocks.3.ls2","params":9050112,"note":"Works on each token by itself: Mlp, LayerScale. The result is added to the token.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.3.mlp/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.3.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.3.mlp/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.3.mlp/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.3.mlp/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.3.mlp/holds/cam_iso"},"auto":{"label":"MLP","note":"Works on each token by itself: Mlp, LayerScale. The result is added to the token.","group":["vision_backbone","blocks 4"],"kind":"mlp"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.4.norm1","label":"normalise","kind":"norm","group":["vision_backbone","blocks 5"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.4.norm1","params":2048,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.4.norm1/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.4.norm1/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.4.norm1/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 5"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.4.attn","label":"attention","kind":"attention","group":["vision_backbone","blocks 5"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.4.attn, vla.base_model.model.vision_backbone.featurizer.blocks.4.ls1","params":4592640,"note":"DINOv2, trained without words to see shapes and parts. The tokens of one camera read each other; cameras never meet in the vision encoders.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.4.attn/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.4.attn/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.4.attn/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.4.attn/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.4.attn/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.4.attn/holds/cam_iso"},"auto":{"label":"attention","note":"Attention. The only place tokens exchange information: each token reads from the others and what it reads is added to it. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 16 heads.","group":["vision_backbone","blocks 5"],"kind":"attention"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.4.norm2","label":"normalise","kind":"norm","group":["vision_backbone","blocks 5"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.4.norm2","params":2048,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.4.norm2/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.4.norm2/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.4.norm2/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 5"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.4.mlp","label":"MLP","kind":"mlp","group":["vision_backbone","blocks 5"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.4.mlp, vla.base_model.model.vision_backbone.featurizer.blocks.4.ls2","params":9050112,"note":"Works on each token by itself: Mlp, LayerScale. The result is added to the token.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.4.mlp/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.4.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.4.mlp/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.4.mlp/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.4.mlp/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.4.mlp/holds/cam_iso"},"auto":{"label":"MLP","note":"Works on each token by itself: Mlp, LayerScale. The result is added to the token.","group":["vision_backbone","blocks 5"],"kind":"mlp"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.5.norm1","label":"normalise","kind":"norm","group":["vision_backbone","blocks 6"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.5.norm1","params":2048,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.5.norm1/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.5.norm1/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.5.norm1/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 6"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.5.attn","label":"attention","kind":"attention","group":["vision_backbone","blocks 6"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.5.attn, vla.base_model.model.vision_backbone.featurizer.blocks.5.ls1","params":4592640,"note":"DINOv2, trained without words to see shapes and parts. The tokens of one camera read each other; cameras never meet in the vision encoders.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.5.attn/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.5.attn/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.5.attn/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.5.attn/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.5.attn/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.5.attn/holds/cam_iso"},"auto":{"label":"attention","note":"Attention. The only place tokens exchange information: each token reads from the others and what it reads is added to it. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 16 heads.","group":["vision_backbone","blocks 6"],"kind":"attention"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.5.norm2","label":"normalise","kind":"norm","group":["vision_backbone","blocks 6"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.5.norm2","params":2048,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.5.norm2/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.5.norm2/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.5.norm2/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 6"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.5.mlp","label":"MLP","kind":"mlp","group":["vision_backbone","blocks 6"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.5.mlp, vla.base_model.model.vision_backbone.featurizer.blocks.5.ls2","params":9050112,"note":"Works on each token by itself: Mlp, LayerScale. The result is added to the token.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.5.mlp/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.5.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.5.mlp/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.5.mlp/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.5.mlp/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.5.mlp/holds/cam_iso"},"auto":{"label":"MLP","note":"Works on each token by itself: Mlp, LayerScale. The result is added to the token.","group":["vision_backbone","blocks 6"],"kind":"mlp"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.6.norm1","label":"normalise","kind":"norm","group":["vision_backbone","blocks 7"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.6.norm1","params":2048,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.6.norm1/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.6.norm1/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.6.norm1/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 7"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.6.attn","label":"attention","kind":"attention","group":["vision_backbone","blocks 7"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.6.attn, vla.base_model.model.vision_backbone.featurizer.blocks.6.ls1","params":4592640,"note":"DINOv2, trained without words to see shapes and parts. The tokens of one camera read each other; cameras never meet in the vision encoders.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.6.attn/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.6.attn/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.6.attn/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.6.attn/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.6.attn/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.6.attn/holds/cam_iso"},"auto":{"label":"attention","note":"Attention. The only place tokens exchange information: each token reads from the others and what it reads is added to it. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 16 heads.","group":["vision_backbone","blocks 7"],"kind":"attention"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.6.norm2","label":"normalise","kind":"norm","group":["vision_backbone","blocks 7"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.6.norm2","params":2048,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.6.norm2/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.6.norm2/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.6.norm2/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 7"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.6.mlp","label":"MLP","kind":"mlp","group":["vision_backbone","blocks 7"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.6.mlp, vla.base_model.model.vision_backbone.featurizer.blocks.6.ls2","params":9050112,"note":"Works on each token by itself: Mlp, LayerScale. The result is added to the token.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.6.mlp/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.6.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.6.mlp/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.6.mlp/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.6.mlp/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.6.mlp/holds/cam_iso"},"auto":{"label":"MLP","note":"Works on each token by itself: Mlp, LayerScale. The result is added to the token.","group":["vision_backbone","blocks 7"],"kind":"mlp"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.7.norm1","label":"normalise","kind":"norm","group":["vision_backbone","blocks 8"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.7.norm1","params":2048,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.7.norm1/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.7.norm1/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.7.norm1/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 8"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.7.attn","label":"attention","kind":"attention","group":["vision_backbone","blocks 8"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.7.attn, vla.base_model.model.vision_backbone.featurizer.blocks.7.ls1","params":4592640,"note":"DINOv2, trained without words to see shapes and parts. The tokens of one camera read each other; cameras never meet in the vision encoders.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.7.attn/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.7.attn/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.7.attn/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.7.attn/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.7.attn/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.7.attn/holds/cam_iso"},"auto":{"label":"attention","note":"Attention. The only place tokens exchange information: each token reads from the others and what it reads is added to it. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 16 heads.","group":["vision_backbone","blocks 8"],"kind":"attention"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.7.norm2","label":"normalise","kind":"norm","group":["vision_backbone","blocks 8"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.7.norm2","params":2048,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.7.norm2/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.7.norm2/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.7.norm2/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 8"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.7.mlp","label":"MLP","kind":"mlp","group":["vision_backbone","blocks 8"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.7.mlp, vla.base_model.model.vision_backbone.featurizer.blocks.7.ls2","params":9050112,"note":"Works on each token by itself: Mlp, LayerScale. The result is added to the token.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.7.mlp/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.7.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.7.mlp/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.7.mlp/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.7.mlp/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.7.mlp/holds/cam_iso"},"auto":{"label":"MLP","note":"Works on each token by itself: Mlp, LayerScale. The result is added to the token.","group":["vision_backbone","blocks 8"],"kind":"mlp"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.8.norm1","label":"normalise","kind":"norm","group":["vision_backbone","blocks 9"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.8.norm1","params":2048,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.8.norm1/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.8.norm1/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.8.norm1/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 9"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.8.attn","label":"attention","kind":"attention","group":["vision_backbone","blocks 9"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.8.attn, vla.base_model.model.vision_backbone.featurizer.blocks.8.ls1","params":4592640,"note":"DINOv2, trained without words to see shapes and parts. The tokens of one camera read each other; cameras never meet in the vision encoders.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.8.attn/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.8.attn/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.8.attn/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.8.attn/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.8.attn/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.8.attn/holds/cam_iso"},"auto":{"label":"attention","note":"Attention. The only place tokens exchange information: each token reads from the others and what it reads is added to it. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 16 heads.","group":["vision_backbone","blocks 9"],"kind":"attention"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.8.norm2","label":"normalise","kind":"norm","group":["vision_backbone","blocks 9"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.8.norm2","params":2048,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.8.norm2/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.8.norm2/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.8.norm2/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 9"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.8.mlp","label":"MLP","kind":"mlp","group":["vision_backbone","blocks 9"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.8.mlp, vla.base_model.model.vision_backbone.featurizer.blocks.8.ls2","params":9050112,"note":"Works on each token by itself: Mlp, LayerScale. The result is added to the token.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.8.mlp/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.8.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.8.mlp/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.8.mlp/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.8.mlp/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.8.mlp/holds/cam_iso"},"auto":{"label":"MLP","note":"Works on each token by itself: Mlp, LayerScale. The result is added to the token.","group":["vision_backbone","blocks 9"],"kind":"mlp"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.9.norm1","label":"normalise","kind":"norm","group":["vision_backbone","blocks 10"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.9.norm1","params":2048,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.9.norm1/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.9.norm1/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.9.norm1/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 10"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.9.attn","label":"attention","kind":"attention","group":["vision_backbone","blocks 10"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.9.attn, vla.base_model.model.vision_backbone.featurizer.blocks.9.ls1","params":4592640,"note":"DINOv2, trained without words to see shapes and parts. The tokens of one camera read each other; cameras never meet in the vision encoders.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.9.attn/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.9.attn/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.9.attn/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.9.attn/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.9.attn/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.9.attn/holds/cam_iso"},"auto":{"label":"attention","note":"Attention. The only place tokens exchange information: each token reads from the others and what it reads is added to it. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 16 heads.","group":["vision_backbone","blocks 10"],"kind":"attention"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.9.norm2","label":"normalise","kind":"norm","group":["vision_backbone","blocks 10"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.9.norm2","params":2048,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.9.norm2/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.9.norm2/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.9.norm2/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 10"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.9.mlp","label":"MLP","kind":"mlp","group":["vision_backbone","blocks 10"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.9.mlp, vla.base_model.model.vision_backbone.featurizer.blocks.9.ls2","params":9050112,"note":"Works on each token by itself: Mlp, LayerScale. The result is added to the token.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.9.mlp/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.9.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.9.mlp/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.9.mlp/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.9.mlp/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.9.mlp/holds/cam_iso"},"auto":{"label":"MLP","note":"Works on each token by itself: Mlp, LayerScale. The result is added to the token.","group":["vision_backbone","blocks 10"],"kind":"mlp"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.10.norm1","label":"normalise","kind":"norm","group":["vision_backbone","blocks 11"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.10.norm1","params":2048,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.10.norm1/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.10.norm1/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.10.norm1/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 11"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.10.attn","label":"attention","kind":"attention","group":["vision_backbone","blocks 11"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.10.attn, vla.base_model.model.vision_backbone.featurizer.blocks.10.ls1","params":4592640,"note":"DINOv2, trained without words to see shapes and parts. The tokens of one camera read each other; cameras never meet in the vision encoders.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.10.attn/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.10.attn/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.10.attn/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.10.attn/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.10.attn/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.10.attn/holds/cam_iso"},"auto":{"label":"attention","note":"Attention. The only place tokens exchange information: each token reads from the others and what it reads is added to it. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 16 heads.","group":["vision_backbone","blocks 11"],"kind":"attention"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.10.norm2","label":"normalise","kind":"norm","group":["vision_backbone","blocks 11"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.10.norm2","params":2048,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.10.norm2/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.10.norm2/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.10.norm2/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 11"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.10.mlp","label":"MLP","kind":"mlp","group":["vision_backbone","blocks 11"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.10.mlp, vla.base_model.model.vision_backbone.featurizer.blocks.10.ls2","params":9050112,"note":"Works on each token by itself: Mlp, LayerScale. The result is added to the token.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.10.mlp/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.10.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.10.mlp/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.10.mlp/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.10.mlp/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.10.mlp/holds/cam_iso"},"auto":{"label":"MLP","note":"Works on each token by itself: Mlp, LayerScale. The result is added to the token.","group":["vision_backbone","blocks 11"],"kind":"mlp"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.11.norm1","label":"normalise","kind":"norm","group":["vision_backbone","blocks 12"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.11.norm1","params":2048,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.11.norm1/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.11.norm1/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.11.norm1/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 12"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.11.attn","label":"attention","kind":"attention","group":["vision_backbone","blocks 12"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.11.attn, vla.base_model.model.vision_backbone.featurizer.blocks.11.ls1","params":4592640,"note":"DINOv2, trained without words to see shapes and parts. The tokens of one camera read each other; cameras never meet in the vision encoders.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.11.attn/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.11.attn/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.11.attn/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.11.attn/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.11.attn/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.11.attn/holds/cam_iso"},"auto":{"label":"attention","note":"Attention. The only place tokens exchange information: each token reads from the others and what it reads is added to it. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 16 heads.","group":["vision_backbone","blocks 12"],"kind":"attention"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.11.norm2","label":"normalise","kind":"norm","group":["vision_backbone","blocks 12"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.11.norm2","params":2048,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.11.norm2/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.11.norm2/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.11.norm2/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 12"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.11.mlp","label":"MLP","kind":"mlp","group":["vision_backbone","blocks 12"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.11.mlp, vla.base_model.model.vision_backbone.featurizer.blocks.11.ls2","params":9050112,"note":"Works on each token by itself: Mlp, LayerScale. The result is added to the token.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.11.mlp/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.11.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.11.mlp/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.11.mlp/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.11.mlp/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.11.mlp/holds/cam_iso"},"auto":{"label":"MLP","note":"Works on each token by itself: Mlp, LayerScale. The result is added to the token.","group":["vision_backbone","blocks 12"],"kind":"mlp"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.12.norm1","label":"normalise","kind":"norm","group":["vision_backbone","blocks 13"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.12.norm1","params":2048,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.12.norm1/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.12.norm1/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.12.norm1/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 13"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.12.attn","label":"attention","kind":"attention","group":["vision_backbone","blocks 13"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.12.attn, vla.base_model.model.vision_backbone.featurizer.blocks.12.ls1","params":4592640,"note":"DINOv2, trained without words to see shapes and parts. The tokens of one camera read each other; cameras never meet in the vision encoders.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.12.attn/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.12.attn/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.12.attn/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.12.attn/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.12.attn/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.12.attn/holds/cam_iso"},"auto":{"label":"attention","note":"Attention. The only place tokens exchange information: each token reads from the others and what it reads is added to it. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 16 heads.","group":["vision_backbone","blocks 13"],"kind":"attention"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.12.norm2","label":"normalise","kind":"norm","group":["vision_backbone","blocks 13"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.12.norm2","params":2048,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.12.norm2/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.12.norm2/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.12.norm2/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 13"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.12.mlp","label":"MLP","kind":"mlp","group":["vision_backbone","blocks 13"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.12.mlp, vla.base_model.model.vision_backbone.featurizer.blocks.12.ls2","params":9050112,"note":"Works on each token by itself: Mlp, LayerScale. The result is added to the token.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.12.mlp/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.12.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.12.mlp/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.12.mlp/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.12.mlp/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.12.mlp/holds/cam_iso"},"auto":{"label":"MLP","note":"Works on each token by itself: Mlp, LayerScale. The result is added to the token.","group":["vision_backbone","blocks 13"],"kind":"mlp"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.13.norm1","label":"normalise","kind":"norm","group":["vision_backbone","blocks 14"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.13.norm1","params":2048,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.13.norm1/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.13.norm1/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.13.norm1/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 14"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.13.attn","label":"attention","kind":"attention","group":["vision_backbone","blocks 14"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.13.attn, vla.base_model.model.vision_backbone.featurizer.blocks.13.ls1","params":4592640,"note":"DINOv2, trained without words to see shapes and parts. The tokens of one camera read each other; cameras never meet in the vision encoders.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.13.attn/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.13.attn/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.13.attn/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.13.attn/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.13.attn/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.13.attn/holds/cam_iso"},"auto":{"label":"attention","note":"Attention. The only place tokens exchange information: each token reads from the others and what it reads is added to it. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 16 heads.","group":["vision_backbone","blocks 14"],"kind":"attention"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.13.norm2","label":"normalise","kind":"norm","group":["vision_backbone","blocks 14"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.13.norm2","params":2048,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.13.norm2/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.13.norm2/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.13.norm2/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 14"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.13.mlp","label":"MLP","kind":"mlp","group":["vision_backbone","blocks 14"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.13.mlp, vla.base_model.model.vision_backbone.featurizer.blocks.13.ls2","params":9050112,"note":"Works on each token by itself: Mlp, LayerScale. The result is added to the token.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.13.mlp/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.13.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.13.mlp/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.13.mlp/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.13.mlp/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.13.mlp/holds/cam_iso"},"auto":{"label":"MLP","note":"Works on each token by itself: Mlp, LayerScale. The result is added to the token.","group":["vision_backbone","blocks 14"],"kind":"mlp"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.14.norm1","label":"normalise","kind":"norm","group":["vision_backbone","blocks 15"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.14.norm1","params":2048,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.14.norm1/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.14.norm1/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.14.norm1/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 15"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.14.attn","label":"attention","kind":"attention","group":["vision_backbone","blocks 15"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.14.attn, vla.base_model.model.vision_backbone.featurizer.blocks.14.ls1","params":4592640,"note":"DINOv2, trained without words to see shapes and parts. The tokens of one camera read each other; cameras never meet in the vision encoders.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.14.attn/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.14.attn/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.14.attn/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.14.attn/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.14.attn/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.14.attn/holds/cam_iso"},"auto":{"label":"attention","note":"Attention. The only place tokens exchange information: each token reads from the others and what it reads is added to it. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 16 heads.","group":["vision_backbone","blocks 15"],"kind":"attention"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.14.norm2","label":"normalise","kind":"norm","group":["vision_backbone","blocks 15"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.14.norm2","params":2048,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.14.norm2/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.14.norm2/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.14.norm2/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 15"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.14.mlp","label":"MLP","kind":"mlp","group":["vision_backbone","blocks 15"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.14.mlp, vla.base_model.model.vision_backbone.featurizer.blocks.14.ls2","params":9050112,"note":"Works on each token by itself: Mlp, LayerScale. The result is added to the token.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.14.mlp/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.14.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.14.mlp/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.14.mlp/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.14.mlp/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.14.mlp/holds/cam_iso"},"auto":{"label":"MLP","note":"Works on each token by itself: Mlp, LayerScale. The result is added to the token.","group":["vision_backbone","blocks 15"],"kind":"mlp"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.15.norm1","label":"normalise","kind":"norm","group":["vision_backbone","blocks 16"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.15.norm1","params":2048,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.15.norm1/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.15.norm1/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.15.norm1/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 16"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.15.attn","label":"attention","kind":"attention","group":["vision_backbone","blocks 16"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.15.attn, vla.base_model.model.vision_backbone.featurizer.blocks.15.ls1","params":4592640,"note":"DINOv2, trained without words to see shapes and parts. The tokens of one camera read each other; cameras never meet in the vision encoders.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.15.attn/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.15.attn/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.15.attn/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.15.attn/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.15.attn/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.15.attn/holds/cam_iso"},"auto":{"label":"attention","note":"Attention. The only place tokens exchange information: each token reads from the others and what it reads is added to it. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 16 heads.","group":["vision_backbone","blocks 16"],"kind":"attention"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.15.norm2","label":"normalise","kind":"norm","group":["vision_backbone","blocks 16"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.15.norm2","params":2048,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.15.norm2/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.15.norm2/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.15.norm2/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 16"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.15.mlp","label":"MLP","kind":"mlp","group":["vision_backbone","blocks 16"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.15.mlp, vla.base_model.model.vision_backbone.featurizer.blocks.15.ls2","params":9050112,"note":"Works on each token by itself: Mlp, LayerScale. The result is added to the token.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.15.mlp/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.15.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.15.mlp/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.15.mlp/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.15.mlp/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.15.mlp/holds/cam_iso"},"auto":{"label":"MLP","note":"Works on each token by itself: Mlp, LayerScale. The result is added to the token.","group":["vision_backbone","blocks 16"],"kind":"mlp"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.16.norm1","label":"normalise","kind":"norm","group":["vision_backbone","blocks 17"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.16.norm1","params":2048,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.16.norm1/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.16.norm1/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.16.norm1/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 17"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.16.attn","label":"attention","kind":"attention","group":["vision_backbone","blocks 17"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.16.attn, vla.base_model.model.vision_backbone.featurizer.blocks.16.ls1","params":4592640,"note":"DINOv2, trained without words to see shapes and parts. The tokens of one camera read each other; cameras never meet in the vision encoders.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.16.attn/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.16.attn/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.16.attn/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.16.attn/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.16.attn/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.16.attn/holds/cam_iso"},"auto":{"label":"attention","note":"Attention. The only place tokens exchange information: each token reads from the others and what it reads is added to it. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 16 heads.","group":["vision_backbone","blocks 17"],"kind":"attention"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.16.norm2","label":"normalise","kind":"norm","group":["vision_backbone","blocks 17"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.16.norm2","params":2048,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.16.norm2/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.16.norm2/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.16.norm2/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 17"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.16.mlp","label":"MLP","kind":"mlp","group":["vision_backbone","blocks 17"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.16.mlp, vla.base_model.model.vision_backbone.featurizer.blocks.16.ls2","params":9050112,"note":"Works on each token by itself: Mlp, LayerScale. The result is added to the token.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.16.mlp/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.16.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.16.mlp/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.16.mlp/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.16.mlp/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.16.mlp/holds/cam_iso"},"auto":{"label":"MLP","note":"Works on each token by itself: Mlp, LayerScale. The result is added to the token.","group":["vision_backbone","blocks 17"],"kind":"mlp"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.17.norm1","label":"normalise","kind":"norm","group":["vision_backbone","blocks 18"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.17.norm1","params":2048,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.17.norm1/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.17.norm1/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.17.norm1/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 18"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.17.attn","label":"attention","kind":"attention","group":["vision_backbone","blocks 18"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.17.attn, vla.base_model.model.vision_backbone.featurizer.blocks.17.ls1","params":4592640,"note":"DINOv2, trained without words to see shapes and parts. The tokens of one camera read each other; cameras never meet in the vision encoders.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.17.attn/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.17.attn/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.17.attn/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.17.attn/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.17.attn/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.17.attn/holds/cam_iso"},"auto":{"label":"attention","note":"Attention. The only place tokens exchange information: each token reads from the others and what it reads is added to it. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 16 heads.","group":["vision_backbone","blocks 18"],"kind":"attention"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.17.norm2","label":"normalise","kind":"norm","group":["vision_backbone","blocks 18"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.17.norm2","params":2048,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.17.norm2/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.17.norm2/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.17.norm2/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 18"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.17.mlp","label":"MLP","kind":"mlp","group":["vision_backbone","blocks 18"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.17.mlp, vla.base_model.model.vision_backbone.featurizer.blocks.17.ls2","params":9050112,"note":"Works on each token by itself: Mlp, LayerScale. The result is added to the token.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.17.mlp/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.17.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.17.mlp/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.17.mlp/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.17.mlp/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.17.mlp/holds/cam_iso"},"auto":{"label":"MLP","note":"Works on each token by itself: Mlp, LayerScale. The result is added to the token.","group":["vision_backbone","blocks 18"],"kind":"mlp"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.18.norm1","label":"normalise","kind":"norm","group":["vision_backbone","blocks 19"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.18.norm1","params":2048,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.18.norm1/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.18.norm1/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.18.norm1/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 19"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.18.attn","label":"attention","kind":"attention","group":["vision_backbone","blocks 19"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.18.attn, vla.base_model.model.vision_backbone.featurizer.blocks.18.ls1","params":4592640,"note":"DINOv2, trained without words to see shapes and parts. The tokens of one camera read each other; cameras never meet in the vision encoders.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.18.attn/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.18.attn/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.18.attn/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.18.attn/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.18.attn/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.18.attn/holds/cam_iso"},"auto":{"label":"attention","note":"Attention. The only place tokens exchange information: each token reads from the others and what it reads is added to it. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 16 heads.","group":["vision_backbone","blocks 19"],"kind":"attention"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.18.norm2","label":"normalise","kind":"norm","group":["vision_backbone","blocks 19"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.18.norm2","params":2048,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.18.norm2/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.18.norm2/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.18.norm2/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 19"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.18.mlp","label":"MLP","kind":"mlp","group":["vision_backbone","blocks 19"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.18.mlp, vla.base_model.model.vision_backbone.featurizer.blocks.18.ls2","params":9050112,"note":"Works on each token by itself: Mlp, LayerScale. The result is added to the token.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.18.mlp/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.18.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.18.mlp/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.18.mlp/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.18.mlp/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.18.mlp/holds/cam_iso"},"auto":{"label":"MLP","note":"Works on each token by itself: Mlp, LayerScale. The result is added to the token.","group":["vision_backbone","blocks 19"],"kind":"mlp"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.19.norm1","label":"normalise","kind":"norm","group":["vision_backbone","blocks 20"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.19.norm1","params":2048,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.19.norm1/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.19.norm1/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.19.norm1/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 20"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.19.attn","label":"attention","kind":"attention","group":["vision_backbone","blocks 20"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.19.attn, vla.base_model.model.vision_backbone.featurizer.blocks.19.ls1","params":4592640,"note":"DINOv2, trained without words to see shapes and parts. The tokens of one camera read each other; cameras never meet in the vision encoders.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.19.attn/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.19.attn/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.19.attn/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.19.attn/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.19.attn/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.19.attn/holds/cam_iso"},"auto":{"label":"attention","note":"Attention. The only place tokens exchange information: each token reads from the others and what it reads is added to it. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 16 heads.","group":["vision_backbone","blocks 20"],"kind":"attention"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.19.norm2","label":"normalise","kind":"norm","group":["vision_backbone","blocks 20"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.19.norm2","params":2048,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.19.norm2/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.19.norm2/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.19.norm2/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 20"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.19.mlp","label":"MLP","kind":"mlp","group":["vision_backbone","blocks 20"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.19.mlp, vla.base_model.model.vision_backbone.featurizer.blocks.19.ls2","params":9050112,"note":"Works on each token by itself: Mlp, LayerScale. The result is added to the token.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.19.mlp/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.19.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.19.mlp/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.19.mlp/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.19.mlp/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.19.mlp/holds/cam_iso"},"auto":{"label":"MLP","note":"Works on each token by itself: Mlp, LayerScale. The result is added to the token.","group":["vision_backbone","blocks 20"],"kind":"mlp"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.20.norm1","label":"normalise","kind":"norm","group":["vision_backbone","blocks 21"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.20.norm1","params":2048,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.20.norm1/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.20.norm1/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.20.norm1/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 21"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.20.attn","label":"attention","kind":"attention","group":["vision_backbone","blocks 21"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.20.attn, vla.base_model.model.vision_backbone.featurizer.blocks.20.ls1","params":4592640,"note":"DINOv2, trained without words to see shapes and parts. The tokens of one camera read each other; cameras never meet in the vision encoders.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.20.attn/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.20.attn/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.20.attn/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.20.attn/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.20.attn/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.20.attn/holds/cam_iso"},"auto":{"label":"attention","note":"Attention. The only place tokens exchange information: each token reads from the others and what it reads is added to it. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 16 heads.","group":["vision_backbone","blocks 21"],"kind":"attention"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.20.norm2","label":"normalise","kind":"norm","group":["vision_backbone","blocks 21"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.20.norm2","params":2048,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.20.norm2/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.20.norm2/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.20.norm2/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 21"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.20.mlp","label":"MLP","kind":"mlp","group":["vision_backbone","blocks 21"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.20.mlp, vla.base_model.model.vision_backbone.featurizer.blocks.20.ls2","params":9050112,"note":"Works on each token by itself: Mlp, LayerScale. The result is added to the token.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.20.mlp/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.20.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.20.mlp/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.20.mlp/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.20.mlp/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.20.mlp/holds/cam_iso"},"auto":{"label":"MLP","note":"Works on each token by itself: Mlp, LayerScale. The result is added to the token.","group":["vision_backbone","blocks 21"],"kind":"mlp"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.21.norm1","label":"normalise","kind":"norm","group":["vision_backbone","blocks 22"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.21.norm1","params":2048,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.21.norm1/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.21.norm1/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.21.norm1/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 22"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.21.attn","label":"attention","kind":"attention","group":["vision_backbone","blocks 22"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.21.attn, vla.base_model.model.vision_backbone.featurizer.blocks.21.ls1","params":4592640,"note":"DINOv2, trained without words to see shapes and parts. The tokens of one camera read each other; cameras never meet in the vision encoders.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.21.attn/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.21.attn/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.21.attn/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.21.attn/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.21.attn/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.21.attn/holds/cam_iso"},"auto":{"label":"attention","note":"Attention. The only place tokens exchange information: each token reads from the others and what it reads is added to it. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 16 heads.","group":["vision_backbone","blocks 22"],"kind":"attention"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.21.norm2","label":"normalise","kind":"norm","group":["vision_backbone","blocks 22"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.21.norm2","params":2048,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.21.norm2/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.21.norm2/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.21.norm2/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 22"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.21.mlp","label":"MLP","kind":"mlp","group":["vision_backbone","blocks 22"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.21.mlp, vla.base_model.model.vision_backbone.featurizer.blocks.21.ls2","params":9050112,"note":"Works on each token by itself: Mlp, LayerScale. The result is added to the token.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.21.mlp/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.21.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.21.mlp/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.21.mlp/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.21.mlp/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.21.mlp/holds/cam_iso"},"auto":{"label":"MLP","note":"Works on each token by itself: Mlp, LayerScale. The result is added to the token.","group":["vision_backbone","blocks 22"],"kind":"mlp"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.22.norm1","label":"normalise","kind":"norm","group":["vision_backbone","blocks 23"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.22.norm1","params":2048,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.22.norm1/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.22.norm1/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.22.norm1/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 23"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.22.attn","label":"attention","kind":"attention","group":["vision_backbone","blocks 23"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.22.attn, vla.base_model.model.vision_backbone.featurizer.blocks.22.ls1","params":4592640,"note":"DINOv2, trained without words to see shapes and parts. The tokens of one camera read each other; cameras never meet in the vision encoders.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.22.attn/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.22.attn/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.22.attn/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.22.attn/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.22.attn/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.22.attn/holds/cam_iso"},"auto":{"label":"attention","note":"Attention. The only place tokens exchange information: each token reads from the others and what it reads is added to it. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 16 heads.","group":["vision_backbone","blocks 23"],"kind":"attention"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.22.norm2","label":"normalise","kind":"norm","group":["vision_backbone","blocks 23"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.22.norm2","params":2048,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.22.norm2/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.22.norm2/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.22.norm2/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 23"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.22.mlp","label":"MLP","kind":"mlp","group":["vision_backbone","blocks 23"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.22.mlp, vla.base_model.model.vision_backbone.featurizer.blocks.22.ls2","params":9050112,"note":"Works on each token by itself: Mlp, LayerScale. The result is added to the token.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.22.mlp/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.22.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.22.mlp/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.22.mlp/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.22.mlp/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.22.mlp/holds/cam_iso"},"auto":{"label":"MLP","note":"Works on each token by itself: Mlp, LayerScale. The result is added to the token.","group":["vision_backbone","blocks 23"],"kind":"mlp"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.23.norm1","label":"normalise","kind":"norm","group":["vision_backbone","blocks 24"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.23.norm1","params":2048,"note":"The last DINOv2 layer runs, but the model takes its image features from the layer below, so this work is thrown away.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.23.norm1/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.23.norm1/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.23.norm1/wrote/cam_iso"},"holds":{},"thin":true,"unused_output":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself. This layer runs, but nothing reads what it hands on: its result never reaches the answer.","group":["vision_backbone","blocks 24"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.23.attn","label":"attention","kind":"attention","group":["vision_backbone","blocks 24"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.23.attn, vla.base_model.model.vision_backbone.featurizer.blocks.23.ls1","params":4592640,"note":"The last DINOv2 layer runs, but the model takes its image features from the layer below, so this work is thrown away.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.23.attn/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.23.attn/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.23.attn/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.23.attn/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.23.attn/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.23.attn/holds/cam_iso"},"unused_output":true,"auto":{"label":"attention","note":"Attention. The only place tokens exchange information: each token reads from the others and what it reads is added to it. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 16 heads. This layer runs, but nothing reads what it hands on: its result never reaches the answer.","group":["vision_backbone","blocks 24"],"kind":"attention"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.23.norm2","label":"normalise","kind":"norm","group":["vision_backbone","blocks 24"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.23.norm2","params":2048,"note":"The last DINOv2 layer runs, but the model takes its image features from the layer below, so this work is thrown away.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.23.norm2/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.23.norm2/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.23.norm2/wrote/cam_iso"},"holds":{},"thin":true,"unused_output":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself. This layer runs, but nothing reads what it hands on: its result never reaches the answer.","group":["vision_backbone","blocks 24"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.featurizer.blocks.23.mlp","label":"MLP","kind":"mlp","group":["vision_backbone","blocks 24"],"plots":["cam_top","cam_side","cam_iso"],"width":1024,"module":"vla.base_model.model.vision_backbone.featurizer.blocks.23.mlp, vla.base_model.model.vision_backbone.featurizer.blocks.23.ls2","params":9050112,"note":"The last DINOv2 layer runs, but the model takes its image features from the layer below, so this work is thrown away.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.23.mlp/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.23.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.23.mlp/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.featurizer.blocks.23.mlp/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.featurizer.blocks.23.mlp/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.featurizer.blocks.23.mlp/holds/cam_iso"},"unused_output":true,"auto":{"label":"MLP","note":"Works on each token by itself: Mlp, LayerScale. The result is added to the token. This layer runs, but nothing reads what it hands on: its result never reaches the answer.","group":["vision_backbone","blocks 24"],"kind":"mlp"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.patch_embed","label":"SigLIP: cut into 14-pixel patches","kind":"conv","group":["vision_backbone"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.patch_embed","params":789888,"note":"The same picture again, for the second encoder: 256 patches of 1,152 numbers, no extra tokens.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.patch_embed/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.patch_embed/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.patch_embed/wrote/cam_iso"},"holds":{},"auto":{"label":"vision_backbone.fused_featurizer.patch_embed: Conv2d, Identity, Conv2d, Conv2d, ParameterDict, ParameterDict and 1 more","note":"PatchEmbed. Out: [1, 256, 1152]. A 14 x 14 convolution, step 14, 3 channels in, 1152 out: it cuts the picture into 14-pixel patches. LoRA on 1 layer: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). Part of vit_so400m_patch14_siglip_224 (as its code names it).","group":["vision_backbone"],"kind":"conv"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.pos_drop","label":"vision_backbone.fused_featurizer.pos_drop: Dropout","kind":"linear","group":["vision_backbone"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.pos_drop","params":0,"note":"Dropout. Out: [1, 256, 1152]. Part of vit_so400m_patch14_siglip_224 (as its code names it).","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.pos_drop/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.pos_drop/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.pos_drop/wrote/cam_iso"},"holds":{},"auto":{"label":"vision_backbone.fused_featurizer.pos_drop: Dropout","note":"Dropout. Out: [1, 256, 1152]. Part of vit_so400m_patch14_siglip_224 (as its code names it).","group":["vision_backbone"],"kind":"linear"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.patch_drop","label":"vision_backbone.fused_featurizer.patch_drop: Identity","kind":"linear","group":["vision_backbone"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.patch_drop","params":0,"note":"Identity. Out: [1, 256, 1152]. Part of vit_so400m_patch14_siglip_224 (as its code names it).","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.patch_drop/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.patch_drop/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.patch_drop/wrote/cam_iso"},"holds":{},"auto":{"label":"vision_backbone.fused_featurizer.patch_drop: Identity","note":"Identity. Out: [1, 256, 1152]. Part of vit_so400m_patch14_siglip_224 (as its code names it).","group":["vision_backbone"],"kind":"linear"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.norm_pre","label":"vision_backbone.fused_featurizer.norm_pre: Identity","kind":"linear","group":["vision_backbone"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.norm_pre","params":0,"note":"Identity. Out: [1, 256, 1152]. Part of vit_so400m_patch14_siglip_224 (as its code names it).","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.norm_pre/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.norm_pre/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.norm_pre/wrote/cam_iso"},"holds":{},"auto":{"label":"vision_backbone.fused_featurizer.norm_pre: Identity","note":"Identity. Out: [1, 256, 1152]. Part of vit_so400m_patch14_siglip_224 (as its code names it).","group":["vision_backbone"],"kind":"linear"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.0.norm1","label":"normalise","kind":"norm","group":["vision_backbone","blocks 1"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.0.norm1","params":2304,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.0.norm1/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.0.norm1/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.0.norm1/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 1"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.0.attn","label":"attention","kind":"attention","group":["vision_backbone","blocks 1"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.0.attn","params":5755392,"note":"SigLIP, trained to match pictures with captions. The tokens of one camera read each other.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.0.attn/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.0.attn/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.0.attn/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.0.attn/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.0.attn/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.0.attn/holds/cam_iso"},"auto":{"label":"attention","note":"Attention. The only place tokens exchange information: each token reads from the others and what it reads is added to it. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 16 heads. Part of vit_so400m_patch14_siglip_224 (as its code names it).","group":["vision_backbone","blocks 1"],"kind":"attention"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.0.norm2","label":"normalise","kind":"norm","group":["vision_backbone","blocks 1"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.0.norm2","params":2304,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.0.norm2/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.0.norm2/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.0.norm2/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 1"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.0.mlp","label":"MLP","kind":"mlp","group":["vision_backbone","blocks 1"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.0.mlp","params":10620240,"note":"Works on each token by itself: Mlp. The result is added to the token. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). Bend: GELU.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.0.mlp/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.0.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.0.mlp/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.0.mlp/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.0.mlp/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.0.mlp/holds/cam_iso"},"auto":{"label":"MLP","note":"Works on each token by itself: Mlp. The result is added to the token. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). Bend: GELU.","group":["vision_backbone","blocks 1"],"kind":"mlp"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.1.norm1","label":"normalise","kind":"norm","group":["vision_backbone","blocks 2"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.1.norm1","params":2304,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.1.norm1/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.1.norm1/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.1.norm1/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 2"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.1.attn","label":"attention","kind":"attention","group":["vision_backbone","blocks 2"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.1.attn","params":5755392,"note":"SigLIP, trained to match pictures with captions. The tokens of one camera read each other.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.1.attn/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.1.attn/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.1.attn/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.1.attn/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.1.attn/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.1.attn/holds/cam_iso"},"auto":{"label":"attention","note":"Attention. The only place tokens exchange information: each token reads from the others and what it reads is added to it. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 16 heads.","group":["vision_backbone","blocks 2"],"kind":"attention"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.1.norm2","label":"normalise","kind":"norm","group":["vision_backbone","blocks 2"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.1.norm2","params":2304,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.1.norm2/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.1.norm2/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.1.norm2/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 2"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.1.mlp","label":"MLP","kind":"mlp","group":["vision_backbone","blocks 2"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.1.mlp","params":10620240,"note":"Works on each token by itself: Mlp. The result is added to the token. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). Bend: GELU.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.1.mlp/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.1.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.1.mlp/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.1.mlp/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.1.mlp/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.1.mlp/holds/cam_iso"},"auto":{"label":"MLP","note":"Works on each token by itself: Mlp. The result is added to the token. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). Bend: GELU.","group":["vision_backbone","blocks 2"],"kind":"mlp"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.2.norm1","label":"normalise","kind":"norm","group":["vision_backbone","blocks 3"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.2.norm1","params":2304,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.2.norm1/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.2.norm1/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.2.norm1/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 3"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.2.attn","label":"attention","kind":"attention","group":["vision_backbone","blocks 3"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.2.attn","params":5755392,"note":"SigLIP, trained to match pictures with captions. The tokens of one camera read each other.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.2.attn/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.2.attn/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.2.attn/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.2.attn/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.2.attn/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.2.attn/holds/cam_iso"},"auto":{"label":"attention","note":"Attention. The only place tokens exchange information: each token reads from the others and what it reads is added to it. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 16 heads.","group":["vision_backbone","blocks 3"],"kind":"attention"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.2.norm2","label":"normalise","kind":"norm","group":["vision_backbone","blocks 3"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.2.norm2","params":2304,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.2.norm2/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.2.norm2/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.2.norm2/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 3"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.2.mlp","label":"MLP","kind":"mlp","group":["vision_backbone","blocks 3"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.2.mlp","params":10620240,"note":"Works on each token by itself: Mlp. The result is added to the token. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). Bend: GELU.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.2.mlp/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.2.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.2.mlp/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.2.mlp/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.2.mlp/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.2.mlp/holds/cam_iso"},"auto":{"label":"MLP","note":"Works on each token by itself: Mlp. The result is added to the token. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). Bend: GELU.","group":["vision_backbone","blocks 3"],"kind":"mlp"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.3.norm1","label":"normalise","kind":"norm","group":["vision_backbone","blocks 4"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.3.norm1","params":2304,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.3.norm1/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.3.norm1/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.3.norm1/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 4"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.3.attn","label":"attention","kind":"attention","group":["vision_backbone","blocks 4"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.3.attn","params":5755392,"note":"SigLIP, trained to match pictures with captions. The tokens of one camera read each other.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.3.attn/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.3.attn/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.3.attn/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.3.attn/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.3.attn/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.3.attn/holds/cam_iso"},"auto":{"label":"attention","note":"Attention. The only place tokens exchange information: each token reads from the others and what it reads is added to it. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 16 heads.","group":["vision_backbone","blocks 4"],"kind":"attention"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.3.norm2","label":"normalise","kind":"norm","group":["vision_backbone","blocks 4"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.3.norm2","params":2304,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.3.norm2/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.3.norm2/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.3.norm2/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 4"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.3.mlp","label":"MLP","kind":"mlp","group":["vision_backbone","blocks 4"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.3.mlp","params":10620240,"note":"Works on each token by itself: Mlp. The result is added to the token. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). Bend: GELU.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.3.mlp/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.3.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.3.mlp/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.3.mlp/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.3.mlp/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.3.mlp/holds/cam_iso"},"auto":{"label":"MLP","note":"Works on each token by itself: Mlp. The result is added to the token. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). Bend: GELU.","group":["vision_backbone","blocks 4"],"kind":"mlp"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.4.norm1","label":"normalise","kind":"norm","group":["vision_backbone","blocks 5"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.4.norm1","params":2304,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.4.norm1/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.4.norm1/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.4.norm1/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 5"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.4.attn","label":"attention","kind":"attention","group":["vision_backbone","blocks 5"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.4.attn","params":5755392,"note":"SigLIP, trained to match pictures with captions. The tokens of one camera read each other.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.4.attn/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.4.attn/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.4.attn/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.4.attn/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.4.attn/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.4.attn/holds/cam_iso"},"auto":{"label":"attention","note":"Attention. The only place tokens exchange information: each token reads from the others and what it reads is added to it. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 16 heads.","group":["vision_backbone","blocks 5"],"kind":"attention"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.4.norm2","label":"normalise","kind":"norm","group":["vision_backbone","blocks 5"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.4.norm2","params":2304,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.4.norm2/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.4.norm2/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.4.norm2/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 5"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.4.mlp","label":"MLP","kind":"mlp","group":["vision_backbone","blocks 5"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.4.mlp","params":10620240,"note":"Works on each token by itself: Mlp. The result is added to the token. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). Bend: GELU.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.4.mlp/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.4.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.4.mlp/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.4.mlp/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.4.mlp/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.4.mlp/holds/cam_iso"},"auto":{"label":"MLP","note":"Works on each token by itself: Mlp. The result is added to the token. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). Bend: GELU.","group":["vision_backbone","blocks 5"],"kind":"mlp"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.5.norm1","label":"normalise","kind":"norm","group":["vision_backbone","blocks 6"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.5.norm1","params":2304,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.5.norm1/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.5.norm1/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.5.norm1/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 6"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.5.attn","label":"attention","kind":"attention","group":["vision_backbone","blocks 6"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.5.attn","params":5755392,"note":"SigLIP, trained to match pictures with captions. The tokens of one camera read each other.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.5.attn/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.5.attn/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.5.attn/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.5.attn/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.5.attn/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.5.attn/holds/cam_iso"},"auto":{"label":"attention","note":"Attention. The only place tokens exchange information: each token reads from the others and what it reads is added to it. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 16 heads.","group":["vision_backbone","blocks 6"],"kind":"attention"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.5.norm2","label":"normalise","kind":"norm","group":["vision_backbone","blocks 6"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.5.norm2","params":2304,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.5.norm2/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.5.norm2/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.5.norm2/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 6"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.5.mlp","label":"MLP","kind":"mlp","group":["vision_backbone","blocks 6"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.5.mlp","params":10620240,"note":"Works on each token by itself: Mlp. The result is added to the token. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). Bend: GELU.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.5.mlp/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.5.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.5.mlp/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.5.mlp/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.5.mlp/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.5.mlp/holds/cam_iso"},"auto":{"label":"MLP","note":"Works on each token by itself: Mlp. The result is added to the token. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). Bend: GELU.","group":["vision_backbone","blocks 6"],"kind":"mlp"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.6.norm1","label":"normalise","kind":"norm","group":["vision_backbone","blocks 7"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.6.norm1","params":2304,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.6.norm1/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.6.norm1/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.6.norm1/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 7"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.6.attn","label":"attention","kind":"attention","group":["vision_backbone","blocks 7"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.6.attn","params":5755392,"note":"SigLIP, trained to match pictures with captions. The tokens of one camera read each other.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.6.attn/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.6.attn/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.6.attn/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.6.attn/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.6.attn/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.6.attn/holds/cam_iso"},"auto":{"label":"attention","note":"Attention. The only place tokens exchange information: each token reads from the others and what it reads is added to it. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 16 heads.","group":["vision_backbone","blocks 7"],"kind":"attention"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.6.norm2","label":"normalise","kind":"norm","group":["vision_backbone","blocks 7"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.6.norm2","params":2304,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.6.norm2/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.6.norm2/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.6.norm2/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 7"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.6.mlp","label":"MLP","kind":"mlp","group":["vision_backbone","blocks 7"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.6.mlp","params":10620240,"note":"Works on each token by itself: Mlp. The result is added to the token. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). Bend: GELU.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.6.mlp/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.6.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.6.mlp/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.6.mlp/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.6.mlp/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.6.mlp/holds/cam_iso"},"auto":{"label":"MLP","note":"Works on each token by itself: Mlp. The result is added to the token. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). Bend: GELU.","group":["vision_backbone","blocks 7"],"kind":"mlp"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.7.norm1","label":"normalise","kind":"norm","group":["vision_backbone","blocks 8"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.7.norm1","params":2304,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.7.norm1/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.7.norm1/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.7.norm1/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 8"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.7.attn","label":"attention","kind":"attention","group":["vision_backbone","blocks 8"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.7.attn","params":5755392,"note":"SigLIP, trained to match pictures with captions. The tokens of one camera read each other.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.7.attn/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.7.attn/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.7.attn/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.7.attn/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.7.attn/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.7.attn/holds/cam_iso"},"auto":{"label":"attention","note":"Attention. The only place tokens exchange information: each token reads from the others and what it reads is added to it. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 16 heads.","group":["vision_backbone","blocks 8"],"kind":"attention"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.7.norm2","label":"normalise","kind":"norm","group":["vision_backbone","blocks 8"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.7.norm2","params":2304,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.7.norm2/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.7.norm2/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.7.norm2/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 8"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.7.mlp","label":"MLP","kind":"mlp","group":["vision_backbone","blocks 8"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.7.mlp","params":10620240,"note":"Works on each token by itself: Mlp. The result is added to the token. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). Bend: GELU.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.7.mlp/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.7.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.7.mlp/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.7.mlp/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.7.mlp/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.7.mlp/holds/cam_iso"},"auto":{"label":"MLP","note":"Works on each token by itself: Mlp. The result is added to the token. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). Bend: GELU.","group":["vision_backbone","blocks 8"],"kind":"mlp"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.8.norm1","label":"normalise","kind":"norm","group":["vision_backbone","blocks 9"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.8.norm1","params":2304,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.8.norm1/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.8.norm1/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.8.norm1/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 9"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.8.attn","label":"attention","kind":"attention","group":["vision_backbone","blocks 9"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.8.attn","params":5755392,"note":"SigLIP, trained to match pictures with captions. The tokens of one camera read each other.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.8.attn/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.8.attn/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.8.attn/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.8.attn/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.8.attn/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.8.attn/holds/cam_iso"},"auto":{"label":"attention","note":"Attention. The only place tokens exchange information: each token reads from the others and what it reads is added to it. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 16 heads.","group":["vision_backbone","blocks 9"],"kind":"attention"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.8.norm2","label":"normalise","kind":"norm","group":["vision_backbone","blocks 9"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.8.norm2","params":2304,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.8.norm2/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.8.norm2/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.8.norm2/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 9"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.8.mlp","label":"MLP","kind":"mlp","group":["vision_backbone","blocks 9"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.8.mlp","params":10620240,"note":"Works on each token by itself: Mlp. The result is added to the token. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). Bend: GELU.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.8.mlp/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.8.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.8.mlp/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.8.mlp/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.8.mlp/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.8.mlp/holds/cam_iso"},"auto":{"label":"MLP","note":"Works on each token by itself: Mlp. The result is added to the token. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). Bend: GELU.","group":["vision_backbone","blocks 9"],"kind":"mlp"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.9.norm1","label":"normalise","kind":"norm","group":["vision_backbone","blocks 10"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.9.norm1","params":2304,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.9.norm1/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.9.norm1/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.9.norm1/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 10"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.9.attn","label":"attention","kind":"attention","group":["vision_backbone","blocks 10"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.9.attn","params":5755392,"note":"SigLIP, trained to match pictures with captions. The tokens of one camera read each other.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.9.attn/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.9.attn/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.9.attn/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.9.attn/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.9.attn/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.9.attn/holds/cam_iso"},"auto":{"label":"attention","note":"Attention. The only place tokens exchange information: each token reads from the others and what it reads is added to it. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 16 heads.","group":["vision_backbone","blocks 10"],"kind":"attention"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.9.norm2","label":"normalise","kind":"norm","group":["vision_backbone","blocks 10"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.9.norm2","params":2304,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.9.norm2/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.9.norm2/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.9.norm2/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 10"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.9.mlp","label":"MLP","kind":"mlp","group":["vision_backbone","blocks 10"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.9.mlp","params":10620240,"note":"Works on each token by itself: Mlp. The result is added to the token. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). Bend: GELU.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.9.mlp/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.9.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.9.mlp/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.9.mlp/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.9.mlp/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.9.mlp/holds/cam_iso"},"auto":{"label":"MLP","note":"Works on each token by itself: Mlp. The result is added to the token. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). Bend: GELU.","group":["vision_backbone","blocks 10"],"kind":"mlp"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.10.norm1","label":"normalise","kind":"norm","group":["vision_backbone","blocks 11"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.10.norm1","params":2304,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.10.norm1/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.10.norm1/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.10.norm1/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 11"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.10.attn","label":"attention","kind":"attention","group":["vision_backbone","blocks 11"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.10.attn","params":5755392,"note":"SigLIP, trained to match pictures with captions. The tokens of one camera read each other.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.10.attn/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.10.attn/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.10.attn/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.10.attn/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.10.attn/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.10.attn/holds/cam_iso"},"auto":{"label":"attention","note":"Attention. The only place tokens exchange information: each token reads from the others and what it reads is added to it. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 16 heads.","group":["vision_backbone","blocks 11"],"kind":"attention"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.10.norm2","label":"normalise","kind":"norm","group":["vision_backbone","blocks 11"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.10.norm2","params":2304,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.10.norm2/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.10.norm2/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.10.norm2/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 11"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.10.mlp","label":"MLP","kind":"mlp","group":["vision_backbone","blocks 11"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.10.mlp","params":10620240,"note":"Works on each token by itself: Mlp. The result is added to the token. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). Bend: GELU.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.10.mlp/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.10.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.10.mlp/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.10.mlp/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.10.mlp/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.10.mlp/holds/cam_iso"},"auto":{"label":"MLP","note":"Works on each token by itself: Mlp. The result is added to the token. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). Bend: GELU.","group":["vision_backbone","blocks 11"],"kind":"mlp"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.11.norm1","label":"normalise","kind":"norm","group":["vision_backbone","blocks 12"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.11.norm1","params":2304,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.11.norm1/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.11.norm1/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.11.norm1/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 12"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.11.attn","label":"attention","kind":"attention","group":["vision_backbone","blocks 12"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.11.attn","params":5755392,"note":"SigLIP, trained to match pictures with captions. The tokens of one camera read each other.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.11.attn/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.11.attn/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.11.attn/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.11.attn/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.11.attn/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.11.attn/holds/cam_iso"},"auto":{"label":"attention","note":"Attention. The only place tokens exchange information: each token reads from the others and what it reads is added to it. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 16 heads.","group":["vision_backbone","blocks 12"],"kind":"attention"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.11.norm2","label":"normalise","kind":"norm","group":["vision_backbone","blocks 12"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.11.norm2","params":2304,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.11.norm2/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.11.norm2/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.11.norm2/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 12"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.11.mlp","label":"MLP","kind":"mlp","group":["vision_backbone","blocks 12"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.11.mlp","params":10620240,"note":"Works on each token by itself: Mlp. The result is added to the token. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). Bend: GELU.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.11.mlp/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.11.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.11.mlp/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.11.mlp/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.11.mlp/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.11.mlp/holds/cam_iso"},"auto":{"label":"MLP","note":"Works on each token by itself: Mlp. The result is added to the token. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). Bend: GELU.","group":["vision_backbone","blocks 12"],"kind":"mlp"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.12.norm1","label":"normalise","kind":"norm","group":["vision_backbone","blocks 13"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.12.norm1","params":2304,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.12.norm1/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.12.norm1/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.12.norm1/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 13"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.12.attn","label":"attention","kind":"attention","group":["vision_backbone","blocks 13"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.12.attn","params":5755392,"note":"SigLIP, trained to match pictures with captions. The tokens of one camera read each other.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.12.attn/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.12.attn/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.12.attn/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.12.attn/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.12.attn/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.12.attn/holds/cam_iso"},"auto":{"label":"attention","note":"Attention. The only place tokens exchange information: each token reads from the others and what it reads is added to it. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 16 heads.","group":["vision_backbone","blocks 13"],"kind":"attention"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.12.norm2","label":"normalise","kind":"norm","group":["vision_backbone","blocks 13"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.12.norm2","params":2304,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.12.norm2/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.12.norm2/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.12.norm2/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 13"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.12.mlp","label":"MLP","kind":"mlp","group":["vision_backbone","blocks 13"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.12.mlp","params":10620240,"note":"Works on each token by itself: Mlp. The result is added to the token. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). Bend: GELU.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.12.mlp/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.12.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.12.mlp/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.12.mlp/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.12.mlp/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.12.mlp/holds/cam_iso"},"auto":{"label":"MLP","note":"Works on each token by itself: Mlp. The result is added to the token. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). Bend: GELU.","group":["vision_backbone","blocks 13"],"kind":"mlp"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.13.norm1","label":"normalise","kind":"norm","group":["vision_backbone","blocks 14"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.13.norm1","params":2304,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.13.norm1/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.13.norm1/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.13.norm1/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 14"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.13.attn","label":"attention","kind":"attention","group":["vision_backbone","blocks 14"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.13.attn","params":5755392,"note":"SigLIP, trained to match pictures with captions. The tokens of one camera read each other.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.13.attn/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.13.attn/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.13.attn/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.13.attn/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.13.attn/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.13.attn/holds/cam_iso"},"auto":{"label":"attention","note":"Attention. The only place tokens exchange information: each token reads from the others and what it reads is added to it. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 16 heads.","group":["vision_backbone","blocks 14"],"kind":"attention"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.13.norm2","label":"normalise","kind":"norm","group":["vision_backbone","blocks 14"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.13.norm2","params":2304,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.13.norm2/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.13.norm2/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.13.norm2/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 14"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.13.mlp","label":"MLP","kind":"mlp","group":["vision_backbone","blocks 14"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.13.mlp","params":10620240,"note":"Works on each token by itself: Mlp. The result is added to the token. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). Bend: GELU.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.13.mlp/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.13.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.13.mlp/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.13.mlp/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.13.mlp/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.13.mlp/holds/cam_iso"},"auto":{"label":"MLP","note":"Works on each token by itself: Mlp. The result is added to the token. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). Bend: GELU.","group":["vision_backbone","blocks 14"],"kind":"mlp"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.14.norm1","label":"normalise","kind":"norm","group":["vision_backbone","blocks 15"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.14.norm1","params":2304,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.14.norm1/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.14.norm1/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.14.norm1/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 15"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.14.attn","label":"attention","kind":"attention","group":["vision_backbone","blocks 15"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.14.attn","params":5755392,"note":"SigLIP, trained to match pictures with captions. The tokens of one camera read each other.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.14.attn/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.14.attn/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.14.attn/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.14.attn/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.14.attn/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.14.attn/holds/cam_iso"},"auto":{"label":"attention","note":"Attention. The only place tokens exchange information: each token reads from the others and what it reads is added to it. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 16 heads.","group":["vision_backbone","blocks 15"],"kind":"attention"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.14.norm2","label":"normalise","kind":"norm","group":["vision_backbone","blocks 15"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.14.norm2","params":2304,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.14.norm2/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.14.norm2/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.14.norm2/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 15"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.14.mlp","label":"MLP","kind":"mlp","group":["vision_backbone","blocks 15"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.14.mlp","params":10620240,"note":"Works on each token by itself: Mlp. The result is added to the token. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). Bend: GELU.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.14.mlp/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.14.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.14.mlp/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.14.mlp/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.14.mlp/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.14.mlp/holds/cam_iso"},"auto":{"label":"MLP","note":"Works on each token by itself: Mlp. The result is added to the token. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). Bend: GELU.","group":["vision_backbone","blocks 15"],"kind":"mlp"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.15.norm1","label":"normalise","kind":"norm","group":["vision_backbone","blocks 16"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.15.norm1","params":2304,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.15.norm1/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.15.norm1/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.15.norm1/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 16"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.15.attn","label":"attention","kind":"attention","group":["vision_backbone","blocks 16"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.15.attn","params":5755392,"note":"SigLIP, trained to match pictures with captions. The tokens of one camera read each other.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.15.attn/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.15.attn/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.15.attn/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.15.attn/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.15.attn/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.15.attn/holds/cam_iso"},"auto":{"label":"attention","note":"Attention. The only place tokens exchange information: each token reads from the others and what it reads is added to it. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 16 heads.","group":["vision_backbone","blocks 16"],"kind":"attention"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.15.norm2","label":"normalise","kind":"norm","group":["vision_backbone","blocks 16"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.15.norm2","params":2304,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.15.norm2/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.15.norm2/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.15.norm2/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 16"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.15.mlp","label":"MLP","kind":"mlp","group":["vision_backbone","blocks 16"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.15.mlp","params":10620240,"note":"Works on each token by itself: Mlp. The result is added to the token. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). Bend: GELU.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.15.mlp/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.15.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.15.mlp/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.15.mlp/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.15.mlp/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.15.mlp/holds/cam_iso"},"auto":{"label":"MLP","note":"Works on each token by itself: Mlp. The result is added to the token. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). Bend: GELU.","group":["vision_backbone","blocks 16"],"kind":"mlp"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.16.norm1","label":"normalise","kind":"norm","group":["vision_backbone","blocks 17"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.16.norm1","params":2304,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.16.norm1/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.16.norm1/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.16.norm1/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 17"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.16.attn","label":"attention","kind":"attention","group":["vision_backbone","blocks 17"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.16.attn","params":5755392,"note":"SigLIP, trained to match pictures with captions. The tokens of one camera read each other.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.16.attn/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.16.attn/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.16.attn/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.16.attn/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.16.attn/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.16.attn/holds/cam_iso"},"auto":{"label":"attention","note":"Attention. The only place tokens exchange information: each token reads from the others and what it reads is added to it. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 16 heads.","group":["vision_backbone","blocks 17"],"kind":"attention"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.16.norm2","label":"normalise","kind":"norm","group":["vision_backbone","blocks 17"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.16.norm2","params":2304,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.16.norm2/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.16.norm2/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.16.norm2/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 17"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.16.mlp","label":"MLP","kind":"mlp","group":["vision_backbone","blocks 17"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.16.mlp","params":10620240,"note":"Works on each token by itself: Mlp. The result is added to the token. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). Bend: GELU.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.16.mlp/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.16.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.16.mlp/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.16.mlp/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.16.mlp/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.16.mlp/holds/cam_iso"},"auto":{"label":"MLP","note":"Works on each token by itself: Mlp. The result is added to the token. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). Bend: GELU.","group":["vision_backbone","blocks 17"],"kind":"mlp"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.17.norm1","label":"normalise","kind":"norm","group":["vision_backbone","blocks 18"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.17.norm1","params":2304,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.17.norm1/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.17.norm1/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.17.norm1/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 18"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.17.attn","label":"attention","kind":"attention","group":["vision_backbone","blocks 18"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.17.attn","params":5755392,"note":"SigLIP, trained to match pictures with captions. The tokens of one camera read each other.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.17.attn/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.17.attn/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.17.attn/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.17.attn/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.17.attn/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.17.attn/holds/cam_iso"},"auto":{"label":"attention","note":"Attention. The only place tokens exchange information: each token reads from the others and what it reads is added to it. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 16 heads.","group":["vision_backbone","blocks 18"],"kind":"attention"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.17.norm2","label":"normalise","kind":"norm","group":["vision_backbone","blocks 18"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.17.norm2","params":2304,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.17.norm2/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.17.norm2/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.17.norm2/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 18"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.17.mlp","label":"MLP","kind":"mlp","group":["vision_backbone","blocks 18"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.17.mlp","params":10620240,"note":"Works on each token by itself: Mlp. The result is added to the token. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). Bend: GELU.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.17.mlp/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.17.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.17.mlp/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.17.mlp/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.17.mlp/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.17.mlp/holds/cam_iso"},"auto":{"label":"MLP","note":"Works on each token by itself: Mlp. The result is added to the token. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). Bend: GELU.","group":["vision_backbone","blocks 18"],"kind":"mlp"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.18.norm1","label":"normalise","kind":"norm","group":["vision_backbone","blocks 19"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.18.norm1","params":2304,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.18.norm1/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.18.norm1/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.18.norm1/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 19"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.18.attn","label":"attention","kind":"attention","group":["vision_backbone","blocks 19"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.18.attn","params":5755392,"note":"SigLIP, trained to match pictures with captions. The tokens of one camera read each other.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.18.attn/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.18.attn/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.18.attn/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.18.attn/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.18.attn/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.18.attn/holds/cam_iso"},"auto":{"label":"attention","note":"Attention. The only place tokens exchange information: each token reads from the others and what it reads is added to it. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 16 heads.","group":["vision_backbone","blocks 19"],"kind":"attention"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.18.norm2","label":"normalise","kind":"norm","group":["vision_backbone","blocks 19"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.18.norm2","params":2304,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.18.norm2/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.18.norm2/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.18.norm2/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 19"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.18.mlp","label":"MLP","kind":"mlp","group":["vision_backbone","blocks 19"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.18.mlp","params":10620240,"note":"Works on each token by itself: Mlp. The result is added to the token. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). Bend: GELU.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.18.mlp/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.18.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.18.mlp/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.18.mlp/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.18.mlp/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.18.mlp/holds/cam_iso"},"auto":{"label":"MLP","note":"Works on each token by itself: Mlp. The result is added to the token. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). Bend: GELU.","group":["vision_backbone","blocks 19"],"kind":"mlp"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.19.norm1","label":"normalise","kind":"norm","group":["vision_backbone","blocks 20"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.19.norm1","params":2304,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.19.norm1/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.19.norm1/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.19.norm1/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 20"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.19.attn","label":"attention","kind":"attention","group":["vision_backbone","blocks 20"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.19.attn","params":5755392,"note":"SigLIP, trained to match pictures with captions. The tokens of one camera read each other.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.19.attn/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.19.attn/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.19.attn/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.19.attn/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.19.attn/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.19.attn/holds/cam_iso"},"auto":{"label":"attention","note":"Attention. The only place tokens exchange information: each token reads from the others and what it reads is added to it. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 16 heads.","group":["vision_backbone","blocks 20"],"kind":"attention"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.19.norm2","label":"normalise","kind":"norm","group":["vision_backbone","blocks 20"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.19.norm2","params":2304,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.19.norm2/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.19.norm2/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.19.norm2/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 20"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.19.mlp","label":"MLP","kind":"mlp","group":["vision_backbone","blocks 20"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.19.mlp","params":10620240,"note":"Works on each token by itself: Mlp. The result is added to the token. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). Bend: GELU.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.19.mlp/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.19.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.19.mlp/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.19.mlp/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.19.mlp/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.19.mlp/holds/cam_iso"},"auto":{"label":"MLP","note":"Works on each token by itself: Mlp. The result is added to the token. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). Bend: GELU.","group":["vision_backbone","blocks 20"],"kind":"mlp"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.20.norm1","label":"normalise","kind":"norm","group":["vision_backbone","blocks 21"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.20.norm1","params":2304,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.20.norm1/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.20.norm1/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.20.norm1/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 21"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.20.attn","label":"attention","kind":"attention","group":["vision_backbone","blocks 21"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.20.attn","params":5755392,"note":"SigLIP, trained to match pictures with captions. The tokens of one camera read each other.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.20.attn/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.20.attn/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.20.attn/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.20.attn/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.20.attn/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.20.attn/holds/cam_iso"},"auto":{"label":"attention","note":"Attention. The only place tokens exchange information: each token reads from the others and what it reads is added to it. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 16 heads.","group":["vision_backbone","blocks 21"],"kind":"attention"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.20.norm2","label":"normalise","kind":"norm","group":["vision_backbone","blocks 21"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.20.norm2","params":2304,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.20.norm2/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.20.norm2/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.20.norm2/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 21"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.20.mlp","label":"MLP","kind":"mlp","group":["vision_backbone","blocks 21"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.20.mlp","params":10620240,"note":"Works on each token by itself: Mlp. The result is added to the token. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). Bend: GELU.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.20.mlp/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.20.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.20.mlp/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.20.mlp/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.20.mlp/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.20.mlp/holds/cam_iso"},"auto":{"label":"MLP","note":"Works on each token by itself: Mlp. The result is added to the token. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). Bend: GELU.","group":["vision_backbone","blocks 21"],"kind":"mlp"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.21.norm1","label":"normalise","kind":"norm","group":["vision_backbone","blocks 22"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.21.norm1","params":2304,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.21.norm1/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.21.norm1/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.21.norm1/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 22"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.21.attn","label":"attention","kind":"attention","group":["vision_backbone","blocks 22"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.21.attn","params":5755392,"note":"SigLIP, trained to match pictures with captions. The tokens of one camera read each other.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.21.attn/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.21.attn/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.21.attn/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.21.attn/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.21.attn/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.21.attn/holds/cam_iso"},"auto":{"label":"attention","note":"Attention. The only place tokens exchange information: each token reads from the others and what it reads is added to it. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 16 heads.","group":["vision_backbone","blocks 22"],"kind":"attention"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.21.norm2","label":"normalise","kind":"norm","group":["vision_backbone","blocks 22"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.21.norm2","params":2304,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.21.norm2/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.21.norm2/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.21.norm2/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 22"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.21.mlp","label":"MLP","kind":"mlp","group":["vision_backbone","blocks 22"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.21.mlp","params":10620240,"note":"Works on each token by itself: Mlp. The result is added to the token. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). Bend: GELU.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.21.mlp/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.21.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.21.mlp/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.21.mlp/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.21.mlp/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.21.mlp/holds/cam_iso"},"auto":{"label":"MLP","note":"Works on each token by itself: Mlp. The result is added to the token. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). Bend: GELU.","group":["vision_backbone","blocks 22"],"kind":"mlp"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.22.norm1","label":"normalise","kind":"norm","group":["vision_backbone","blocks 23"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.22.norm1","params":2304,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.22.norm1/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.22.norm1/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.22.norm1/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 23"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.22.attn","label":"attention","kind":"attention","group":["vision_backbone","blocks 23"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.22.attn","params":5755392,"note":"SigLIP, trained to match pictures with captions. The tokens of one camera read each other.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.22.attn/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.22.attn/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.22.attn/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.22.attn/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.22.attn/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.22.attn/holds/cam_iso"},"auto":{"label":"attention","note":"Attention. The only place tokens exchange information: each token reads from the others and what it reads is added to it. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 16 heads.","group":["vision_backbone","blocks 23"],"kind":"attention"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.22.norm2","label":"normalise","kind":"norm","group":["vision_backbone","blocks 23"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.22.norm2","params":2304,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.22.norm2/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.22.norm2/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.22.norm2/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 23"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.22.mlp","label":"MLP","kind":"mlp","group":["vision_backbone","blocks 23"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.22.mlp","params":10620240,"note":"Works on each token by itself: Mlp. The result is added to the token. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). Bend: GELU.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.22.mlp/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.22.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.22.mlp/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.22.mlp/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.22.mlp/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.22.mlp/holds/cam_iso"},"auto":{"label":"MLP","note":"Works on each token by itself: Mlp. The result is added to the token. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). Bend: GELU.","group":["vision_backbone","blocks 23"],"kind":"mlp"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.23.norm1","label":"normalise","kind":"norm","group":["vision_backbone","blocks 24"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.23.norm1","params":2304,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.23.norm1/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.23.norm1/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.23.norm1/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 24"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.23.attn","label":"attention","kind":"attention","group":["vision_backbone","blocks 24"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.23.attn","params":5755392,"note":"SigLIP, trained to match pictures with captions. The tokens of one camera read each other.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.23.attn/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.23.attn/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.23.attn/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.23.attn/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.23.attn/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.23.attn/holds/cam_iso"},"auto":{"label":"attention","note":"Attention. The only place tokens exchange information: each token reads from the others and what it reads is added to it. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 16 heads.","group":["vision_backbone","blocks 24"],"kind":"attention"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.23.norm2","label":"normalise","kind":"norm","group":["vision_backbone","blocks 24"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.23.norm2","params":2304,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.23.norm2/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.23.norm2/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.23.norm2/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 24"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.23.mlp","label":"MLP","kind":"mlp","group":["vision_backbone","blocks 24"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.23.mlp","params":10620240,"note":"Works on each token by itself: Mlp. The result is added to the token. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). Bend: GELU.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.23.mlp/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.23.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.23.mlp/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.23.mlp/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.23.mlp/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.23.mlp/holds/cam_iso"},"auto":{"label":"MLP","note":"Works on each token by itself: Mlp. The result is added to the token. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). Bend: GELU.","group":["vision_backbone","blocks 24"],"kind":"mlp"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.24.norm1","label":"normalise","kind":"norm","group":["vision_backbone","blocks 25"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.24.norm1","params":2304,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.24.norm1/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.24.norm1/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.24.norm1/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 25"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.24.attn","label":"attention","kind":"attention","group":["vision_backbone","blocks 25"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.24.attn","params":5755392,"note":"SigLIP, trained to match pictures with captions. The tokens of one camera read each other.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.24.attn/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.24.attn/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.24.attn/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.24.attn/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.24.attn/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.24.attn/holds/cam_iso"},"auto":{"label":"attention","note":"Attention. The only place tokens exchange information: each token reads from the others and what it reads is added to it. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 16 heads.","group":["vision_backbone","blocks 25"],"kind":"attention"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.24.norm2","label":"normalise","kind":"norm","group":["vision_backbone","blocks 25"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.24.norm2","params":2304,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.24.norm2/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.24.norm2/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.24.norm2/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 25"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.24.mlp","label":"MLP","kind":"mlp","group":["vision_backbone","blocks 25"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.24.mlp","params":10620240,"note":"Works on each token by itself: Mlp. The result is added to the token. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). Bend: GELU.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.24.mlp/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.24.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.24.mlp/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.24.mlp/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.24.mlp/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.24.mlp/holds/cam_iso"},"auto":{"label":"MLP","note":"Works on each token by itself: Mlp. The result is added to the token. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). Bend: GELU.","group":["vision_backbone","blocks 25"],"kind":"mlp"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.25.norm1","label":"normalise","kind":"norm","group":["vision_backbone","blocks 26"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.25.norm1","params":2304,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.25.norm1/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.25.norm1/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.25.norm1/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 26"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.25.attn","label":"attention","kind":"attention","group":["vision_backbone","blocks 26"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.25.attn","params":5755392,"note":"SigLIP, trained to match pictures with captions. The tokens of one camera read each other.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.25.attn/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.25.attn/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.25.attn/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.25.attn/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.25.attn/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.25.attn/holds/cam_iso"},"auto":{"label":"attention","note":"Attention. The only place tokens exchange information: each token reads from the others and what it reads is added to it. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 16 heads.","group":["vision_backbone","blocks 26"],"kind":"attention"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.25.norm2","label":"normalise","kind":"norm","group":["vision_backbone","blocks 26"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.25.norm2","params":2304,"note":"LayerNorm: rescales each token by itself.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.25.norm2/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.25.norm2/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.25.norm2/wrote/cam_iso"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself.","group":["vision_backbone","blocks 26"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.25.mlp","label":"MLP","kind":"mlp","group":["vision_backbone","blocks 26"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.25.mlp","params":10620240,"note":"Works on each token by itself: Mlp. The result is added to the token. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). Bend: GELU.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.25.mlp/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.25.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.25.mlp/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.25.mlp/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.25.mlp/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.25.mlp/holds/cam_iso"},"auto":{"label":"MLP","note":"Works on each token by itself: Mlp. The result is added to the token. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). Bend: GELU.","group":["vision_backbone","blocks 26"],"kind":"mlp"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.26.norm1","label":"normalise","kind":"norm","group":["vision_backbone","blocks 27"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.26.norm1","params":2304,"note":"The last SigLIP layer runs, but the model takes its image features from the layer below, so this work is thrown away.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.26.norm1/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.26.norm1/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.26.norm1/wrote/cam_iso"},"holds":{},"thin":true,"unused_output":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself. This layer runs, but nothing reads what it hands on: its result never reaches the answer.","group":["vision_backbone","blocks 27"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.26.attn","label":"attention","kind":"attention","group":["vision_backbone","blocks 27"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.26.attn","params":5755392,"note":"The last SigLIP layer runs, but the model takes its image features from the layer below, so this work is thrown away.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.26.attn/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.26.attn/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.26.attn/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.26.attn/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.26.attn/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.26.attn/holds/cam_iso"},"unused_output":true,"auto":{"label":"attention","note":"Attention. The only place tokens exchange information: each token reads from the others and what it reads is added to it. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 16 heads. This layer runs, but nothing reads what it hands on: its result never reaches the answer.","group":["vision_backbone","blocks 27"],"kind":"attention"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.26.norm2","label":"normalise","kind":"norm","group":["vision_backbone","blocks 27"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.26.norm2","params":2304,"note":"The last SigLIP layer runs, but the model takes its image features from the layer below, so this work is thrown away.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.26.norm2/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.26.norm2/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.26.norm2/wrote/cam_iso"},"holds":{},"thin":true,"unused_output":true,"auto":{"label":"normalise","note":"LayerNorm: rescales each token by itself. This layer runs, but nothing reads what it hands on: its result never reaches the answer.","group":["vision_backbone","blocks 27"],"kind":"norm"}},{"id":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.26.mlp","label":"MLP","kind":"mlp","group":["vision_backbone","blocks 27"],"plots":["cam_top","cam_side","cam_iso"],"width":1152,"module":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.26.mlp","params":10620240,"note":"The last SigLIP layer runs, but the model takes its image features from the layer below, so this work is thrown away.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.26.mlp/wrote/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.26.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.26.mlp/wrote/cam_iso"},"holds":{"cam_top":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.26.mlp/holds/cam_top","cam_side":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.26.mlp/holds/cam_side","cam_iso":"vla.base_model.model.vision_backbone.fused_featurizer.blocks.26.mlp/holds/cam_iso"},"unused_output":true,"auto":{"label":"MLP","note":"Works on each token by itself: Mlp. The result is added to the token. LoRA on 2 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). Bend: GELU. This layer runs, but nothing reads what it hands on: its result never reaches the answer.","group":["vision_backbone","blocks 27"],"kind":"mlp"}},{"id":"vla.base_model.model.projector","label":"projector: both encoders' 2,176 numbers per patch to 896","kind":"linear","group":["projector"],"plots":["cam_top","cam_side","cam_iso"],"width":896,"module":"vla.base_model.model.projector","params":28977408,"note":"DINOv2's 1,024 and SigLIP's 1,152 numbers per patch, side by side, through three linear layers (with LoRA) into the language model's width. 768 image tokens: 256 per camera, no squeezing.","cells":{"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16]},"wrote":{"cam_top":"vla.base_model.model.projector/wrote/cam_top","cam_side":"vla.base_model.model.projector/wrote/cam_side","cam_iso":"vla.base_model.model.projector/wrote/cam_iso"},"holds":{},"auto":{"label":"projector: Linear, Identity, Linear, Linear, ParameterDict, ParameterDict and 14 more","note":"PrismaticProjector. Out: [1, 768, 896]. LoRA on 3 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). Bend: GELU. Part of openvla (as its code names it).","group":["projector"],"kind":"linear"}},{"id":"enter","label":"tokens entering the transformer","kind":"learned","group":["transformer"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.vision_backbone.featurizer.cls_token, vla.base_model.model.vision_backbone.featurizer.reg_token, vla.base_model.model.vision_backbone.featurizer.pos_embed, vla.base_model.model.vision_backbone.fused_featurizer.pos_embed","params":562176,"note":"Every token as the first layer receives it. Learned numbers that belong to no layer are added on the way (listed under 'parts'); their parameters are counted here. Here: vla.base_model.model.vision_backbone.featurizer.cls_token, vla.base_model.model.vision_backbone.featurizer.reg_token, vla.base_model.model.vision_backbone.featurizer.pos_embed, vla.base_model.model.vision_backbone.fused_featurizer.pos_embed.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"enter/holds/learned","cam_top":"enter/holds/cam_top","cam_side":"enter/holds/cam_side","cam_iso":"enter/holds/cam_iso","learned_2":"enter/holds/learned_2","text":"enter/holds/text","learned_3":"enter/holds/learned_3"},"holds":{"learned":"enter/holds/learned","cam_top":"enter/holds/cam_top","cam_side":"enter/holds/cam_side","cam_iso":"enter/holds/cam_iso","learned_2":"enter/holds/learned_2","text":"enter/holds/text","learned_3":"enter/holds/learned_3"},"thin":true,"adds":true,"auto":{"label":"tokens entering the transformer","note":"Every token as the first layer receives it. Learned numbers that belong to no layer are added on the way (listed under 'parts'); their parameters are counted here. Here: vla.base_model.model.vision_backbone.featurizer.cls_token, vla.base_model.model.vision_backbone.featurizer.reg_token, vla.base_model.model.vision_backbone.featurizer.pos_embed, vla.base_model.model.vision_backbone.fused_featurizer.pos_embed.","group":["transformer"],"kind":"learned"}},{"id":"vla.base_model.model.language_model.model.layers.0.input_layernorm","label":"normalise","kind":"norm","group":["transformer","layer 1"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.0.input_layernorm","params":896,"note":"Qwen2RMSNorm: rescales each token by itself.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.0.input_layernorm/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.0.input_layernorm/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.0.input_layernorm/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.0.input_layernorm/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.0.input_layernorm/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.0.input_layernorm/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.0.input_layernorm/wrote/learned_3"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"Qwen2RMSNorm: rescales each token by itself.","group":["transformer","layer 1"],"kind":"norm"}},{"id":"vla.base_model.model.language_model.model.layers.0.self_attn","label":"attention","kind":"attention","group":["transformer","layer 1"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.0.self_attn","params":2196608,"note":"Qwen2.5-0.5B: 14 query heads share 2 key and value heads, and each token reads only the tokens before it. So the image tokens, which come first, never read the instruction; the action queries, which come last, read everything. LoRA (rank 64) adds a small trainable correction to every linear layer of this model.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.0.self_attn/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.0.self_attn/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.0.self_attn/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.0.self_attn/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.0.self_attn/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.0.self_attn/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.0.self_attn/wrote/learned_3"},"holds":{"learned":"vla.base_model.model.language_model.model.layers.0.self_attn/holds/learned","cam_top":"vla.base_model.model.language_model.model.layers.0.self_attn/holds/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.0.self_attn/holds/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.0.self_attn/holds/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.0.self_attn/holds/learned_2","text":"vla.base_model.model.language_model.model.layers.0.self_attn/holds/text","learned_3":"vla.base_model.model.language_model.model.layers.0.self_attn/holds/learned_3"},"heads":14,"auto":{"label":"attention","note":"Qwen2SdpaAttention. The only place tokens exchange information: each token reads from the others (the plate beside this storey shows who reads whom) and what it reads is added to it. Each token reads only the tokens before it. LoRA on 4 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 14 heads, sharing 2 key and value heads. Part of qwen2 (as its code names it).","group":["transformer","layer 1"],"kind":"attention"}},{"id":"vla.base_model.model.language_model.model.layers.0.post_attention_layernorm","label":"normalise","kind":"norm","group":["transformer","layer 1"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.0.post_attention_layernorm","params":896,"note":"Qwen2RMSNorm: rescales each token by itself.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.0.post_attention_layernorm/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.0.post_attention_layernorm/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.0.post_attention_layernorm/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.0.post_attention_layernorm/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.0.post_attention_layernorm/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.0.post_attention_layernorm/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.0.post_attention_layernorm/wrote/learned_3"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"Qwen2RMSNorm: rescales each token by itself.","group":["transformer","layer 1"],"kind":"norm"}},{"id":"vla.base_model.model.language_model.model.layers.0.mlp","label":"MLP","kind":"mlp","group":["transformer","layer 1"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.0.mlp","params":14180352,"note":"A gated MLP (SwiGLU): 896 numbers out to 4,864 and back, with LoRA.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.0.mlp/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.0.mlp/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.0.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.0.mlp/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.0.mlp/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.0.mlp/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.0.mlp/wrote/learned_3"},"holds":{"learned":"vla.base_model.model.language_model.model.layers.0.mlp/holds/learned","cam_top":"vla.base_model.model.language_model.model.layers.0.mlp/holds/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.0.mlp/holds/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.0.mlp/holds/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.0.mlp/holds/learned_2","text":"vla.base_model.model.language_model.model.layers.0.mlp/holds/text","learned_3":"vla.base_model.model.language_model.model.layers.0.mlp/holds/learned_3"},"auto":{"label":"MLP","note":"Works on each token by itself: Qwen2MLP. The result is added to the token. LoRA on 3 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 14 heads, sharing 2 key and value heads. Gated: one branch, passed through the bend, multiplies the other. Bend: SiLU.","group":["transformer","layer 1"],"kind":"mlp"}},{"id":"vla.base_model.model.language_model.model.layers.1.input_layernorm","label":"normalise","kind":"norm","group":["transformer","layer 2"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.1.input_layernorm","params":896,"note":"Qwen2RMSNorm: rescales each token by itself.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.1.input_layernorm/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.1.input_layernorm/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.1.input_layernorm/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.1.input_layernorm/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.1.input_layernorm/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.1.input_layernorm/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.1.input_layernorm/wrote/learned_3"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"Qwen2RMSNorm: rescales each token by itself.","group":["transformer","layer 2"],"kind":"norm"}},{"id":"vla.base_model.model.language_model.model.layers.1.self_attn","label":"attention","kind":"attention","group":["transformer","layer 2"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.1.self_attn","params":2196608,"note":"Qwen2.5-0.5B: 14 query heads share 2 key and value heads, and each token reads only the tokens before it. So the image tokens, which come first, never read the instruction; the action queries, which come last, read everything. LoRA (rank 64) adds a small trainable correction to every linear layer of this model.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.1.self_attn/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.1.self_attn/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.1.self_attn/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.1.self_attn/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.1.self_attn/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.1.self_attn/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.1.self_attn/wrote/learned_3"},"holds":{"learned":"vla.base_model.model.language_model.model.layers.1.self_attn/holds/learned","cam_top":"vla.base_model.model.language_model.model.layers.1.self_attn/holds/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.1.self_attn/holds/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.1.self_attn/holds/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.1.self_attn/holds/learned_2","text":"vla.base_model.model.language_model.model.layers.1.self_attn/holds/text","learned_3":"vla.base_model.model.language_model.model.layers.1.self_attn/holds/learned_3"},"heads":14,"auto":{"label":"attention","note":"Qwen2SdpaAttention. The only place tokens exchange information: each token reads from the others (the plate beside this storey shows who reads whom) and what it reads is added to it. Each token reads only the tokens before it. LoRA on 4 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 14 heads, sharing 2 key and value heads.","group":["transformer","layer 2"],"kind":"attention"}},{"id":"vla.base_model.model.language_model.model.layers.1.post_attention_layernorm","label":"normalise","kind":"norm","group":["transformer","layer 2"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.1.post_attention_layernorm","params":896,"note":"Qwen2RMSNorm: rescales each token by itself.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.1.post_attention_layernorm/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.1.post_attention_layernorm/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.1.post_attention_layernorm/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.1.post_attention_layernorm/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.1.post_attention_layernorm/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.1.post_attention_layernorm/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.1.post_attention_layernorm/wrote/learned_3"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"Qwen2RMSNorm: rescales each token by itself.","group":["transformer","layer 2"],"kind":"norm"}},{"id":"vla.base_model.model.language_model.model.layers.1.mlp","label":"MLP","kind":"mlp","group":["transformer","layer 2"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.1.mlp","params":14180352,"note":"A gated MLP (SwiGLU): 896 numbers out to 4,864 and back, with LoRA.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.1.mlp/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.1.mlp/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.1.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.1.mlp/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.1.mlp/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.1.mlp/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.1.mlp/wrote/learned_3"},"holds":{"learned":"vla.base_model.model.language_model.model.layers.1.mlp/holds/learned","cam_top":"vla.base_model.model.language_model.model.layers.1.mlp/holds/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.1.mlp/holds/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.1.mlp/holds/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.1.mlp/holds/learned_2","text":"vla.base_model.model.language_model.model.layers.1.mlp/holds/text","learned_3":"vla.base_model.model.language_model.model.layers.1.mlp/holds/learned_3"},"auto":{"label":"MLP","note":"Works on each token by itself: Qwen2MLP. The result is added to the token. LoRA on 3 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 14 heads, sharing 2 key and value heads. Gated: one branch, passed through the bend, multiplies the other. Bend: SiLU.","group":["transformer","layer 2"],"kind":"mlp"}},{"id":"vla.base_model.model.language_model.model.layers.2.input_layernorm","label":"normalise","kind":"norm","group":["transformer","layer 3"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.2.input_layernorm","params":896,"note":"Qwen2RMSNorm: rescales each token by itself.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.2.input_layernorm/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.2.input_layernorm/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.2.input_layernorm/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.2.input_layernorm/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.2.input_layernorm/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.2.input_layernorm/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.2.input_layernorm/wrote/learned_3"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"Qwen2RMSNorm: rescales each token by itself.","group":["transformer","layer 3"],"kind":"norm"}},{"id":"vla.base_model.model.language_model.model.layers.2.self_attn","label":"attention","kind":"attention","group":["transformer","layer 3"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.2.self_attn","params":2196608,"note":"Qwen2.5-0.5B: 14 query heads share 2 key and value heads, and each token reads only the tokens before it. So the image tokens, which come first, never read the instruction; the action queries, which come last, read everything. LoRA (rank 64) adds a small trainable correction to every linear layer of this model.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.2.self_attn/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.2.self_attn/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.2.self_attn/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.2.self_attn/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.2.self_attn/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.2.self_attn/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.2.self_attn/wrote/learned_3"},"holds":{"learned":"vla.base_model.model.language_model.model.layers.2.self_attn/holds/learned","cam_top":"vla.base_model.model.language_model.model.layers.2.self_attn/holds/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.2.self_attn/holds/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.2.self_attn/holds/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.2.self_attn/holds/learned_2","text":"vla.base_model.model.language_model.model.layers.2.self_attn/holds/text","learned_3":"vla.base_model.model.language_model.model.layers.2.self_attn/holds/learned_3"},"heads":14,"auto":{"label":"attention","note":"Qwen2SdpaAttention. The only place tokens exchange information: each token reads from the others (the plate beside this storey shows who reads whom) and what it reads is added to it. Each token reads only the tokens before it. LoRA on 4 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 14 heads, sharing 2 key and value heads.","group":["transformer","layer 3"],"kind":"attention"}},{"id":"vla.base_model.model.language_model.model.layers.2.post_attention_layernorm","label":"normalise","kind":"norm","group":["transformer","layer 3"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.2.post_attention_layernorm","params":896,"note":"Qwen2RMSNorm: rescales each token by itself.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.2.post_attention_layernorm/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.2.post_attention_layernorm/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.2.post_attention_layernorm/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.2.post_attention_layernorm/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.2.post_attention_layernorm/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.2.post_attention_layernorm/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.2.post_attention_layernorm/wrote/learned_3"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"Qwen2RMSNorm: rescales each token by itself.","group":["transformer","layer 3"],"kind":"norm"}},{"id":"vla.base_model.model.language_model.model.layers.2.mlp","label":"MLP","kind":"mlp","group":["transformer","layer 3"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.2.mlp","params":14180352,"note":"A gated MLP (SwiGLU): 896 numbers out to 4,864 and back, with LoRA.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.2.mlp/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.2.mlp/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.2.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.2.mlp/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.2.mlp/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.2.mlp/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.2.mlp/wrote/learned_3"},"holds":{"learned":"vla.base_model.model.language_model.model.layers.2.mlp/holds/learned","cam_top":"vla.base_model.model.language_model.model.layers.2.mlp/holds/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.2.mlp/holds/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.2.mlp/holds/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.2.mlp/holds/learned_2","text":"vla.base_model.model.language_model.model.layers.2.mlp/holds/text","learned_3":"vla.base_model.model.language_model.model.layers.2.mlp/holds/learned_3"},"auto":{"label":"MLP","note":"Works on each token by itself: Qwen2MLP. The result is added to the token. LoRA on 3 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 14 heads, sharing 2 key and value heads. Gated: one branch, passed through the bend, multiplies the other. Bend: SiLU.","group":["transformer","layer 3"],"kind":"mlp"}},{"id":"vla.base_model.model.language_model.model.layers.3.input_layernorm","label":"normalise","kind":"norm","group":["transformer","layer 4"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.3.input_layernorm","params":896,"note":"Qwen2RMSNorm: rescales each token by itself.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.3.input_layernorm/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.3.input_layernorm/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.3.input_layernorm/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.3.input_layernorm/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.3.input_layernorm/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.3.input_layernorm/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.3.input_layernorm/wrote/learned_3"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"Qwen2RMSNorm: rescales each token by itself.","group":["transformer","layer 4"],"kind":"norm"}},{"id":"vla.base_model.model.language_model.model.layers.3.self_attn","label":"attention","kind":"attention","group":["transformer","layer 4"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.3.self_attn","params":2196608,"note":"Qwen2.5-0.5B: 14 query heads share 2 key and value heads, and each token reads only the tokens before it. So the image tokens, which come first, never read the instruction; the action queries, which come last, read everything. LoRA (rank 64) adds a small trainable correction to every linear layer of this model.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.3.self_attn/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.3.self_attn/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.3.self_attn/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.3.self_attn/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.3.self_attn/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.3.self_attn/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.3.self_attn/wrote/learned_3"},"holds":{"learned":"vla.base_model.model.language_model.model.layers.3.self_attn/holds/learned","cam_top":"vla.base_model.model.language_model.model.layers.3.self_attn/holds/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.3.self_attn/holds/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.3.self_attn/holds/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.3.self_attn/holds/learned_2","text":"vla.base_model.model.language_model.model.layers.3.self_attn/holds/text","learned_3":"vla.base_model.model.language_model.model.layers.3.self_attn/holds/learned_3"},"heads":14,"auto":{"label":"attention","note":"Qwen2SdpaAttention. The only place tokens exchange information: each token reads from the others (the plate beside this storey shows who reads whom) and what it reads is added to it. Each token reads only the tokens before it. LoRA on 4 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 14 heads, sharing 2 key and value heads.","group":["transformer","layer 4"],"kind":"attention"}},{"id":"vla.base_model.model.language_model.model.layers.3.post_attention_layernorm","label":"normalise","kind":"norm","group":["transformer","layer 4"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.3.post_attention_layernorm","params":896,"note":"Qwen2RMSNorm: rescales each token by itself.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.3.post_attention_layernorm/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.3.post_attention_layernorm/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.3.post_attention_layernorm/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.3.post_attention_layernorm/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.3.post_attention_layernorm/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.3.post_attention_layernorm/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.3.post_attention_layernorm/wrote/learned_3"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"Qwen2RMSNorm: rescales each token by itself.","group":["transformer","layer 4"],"kind":"norm"}},{"id":"vla.base_model.model.language_model.model.layers.3.mlp","label":"MLP","kind":"mlp","group":["transformer","layer 4"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.3.mlp","params":14180352,"note":"A gated MLP (SwiGLU): 896 numbers out to 4,864 and back, with LoRA.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.3.mlp/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.3.mlp/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.3.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.3.mlp/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.3.mlp/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.3.mlp/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.3.mlp/wrote/learned_3"},"holds":{"learned":"vla.base_model.model.language_model.model.layers.3.mlp/holds/learned","cam_top":"vla.base_model.model.language_model.model.layers.3.mlp/holds/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.3.mlp/holds/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.3.mlp/holds/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.3.mlp/holds/learned_2","text":"vla.base_model.model.language_model.model.layers.3.mlp/holds/text","learned_3":"vla.base_model.model.language_model.model.layers.3.mlp/holds/learned_3"},"auto":{"label":"MLP","note":"Works on each token by itself: Qwen2MLP. The result is added to the token. LoRA on 3 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 14 heads, sharing 2 key and value heads. Gated: one branch, passed through the bend, multiplies the other. Bend: SiLU.","group":["transformer","layer 4"],"kind":"mlp"}},{"id":"vla.base_model.model.language_model.model.layers.4.input_layernorm","label":"normalise","kind":"norm","group":["transformer","layer 5"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.4.input_layernorm","params":896,"note":"Qwen2RMSNorm: rescales each token by itself.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.4.input_layernorm/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.4.input_layernorm/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.4.input_layernorm/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.4.input_layernorm/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.4.input_layernorm/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.4.input_layernorm/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.4.input_layernorm/wrote/learned_3"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"Qwen2RMSNorm: rescales each token by itself.","group":["transformer","layer 5"],"kind":"norm"}},{"id":"vla.base_model.model.language_model.model.layers.4.self_attn","label":"attention","kind":"attention","group":["transformer","layer 5"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.4.self_attn","params":2196608,"note":"Qwen2.5-0.5B: 14 query heads share 2 key and value heads, and each token reads only the tokens before it. So the image tokens, which come first, never read the instruction; the action queries, which come last, read everything. LoRA (rank 64) adds a small trainable correction to every linear layer of this model.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.4.self_attn/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.4.self_attn/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.4.self_attn/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.4.self_attn/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.4.self_attn/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.4.self_attn/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.4.self_attn/wrote/learned_3"},"holds":{"learned":"vla.base_model.model.language_model.model.layers.4.self_attn/holds/learned","cam_top":"vla.base_model.model.language_model.model.layers.4.self_attn/holds/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.4.self_attn/holds/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.4.self_attn/holds/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.4.self_attn/holds/learned_2","text":"vla.base_model.model.language_model.model.layers.4.self_attn/holds/text","learned_3":"vla.base_model.model.language_model.model.layers.4.self_attn/holds/learned_3"},"heads":14,"auto":{"label":"attention","note":"Qwen2SdpaAttention. The only place tokens exchange information: each token reads from the others (the plate beside this storey shows who reads whom) and what it reads is added to it. Each token reads only the tokens before it. LoRA on 4 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 14 heads, sharing 2 key and value heads.","group":["transformer","layer 5"],"kind":"attention"}},{"id":"vla.base_model.model.language_model.model.layers.4.post_attention_layernorm","label":"normalise","kind":"norm","group":["transformer","layer 5"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.4.post_attention_layernorm","params":896,"note":"Qwen2RMSNorm: rescales each token by itself.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.4.post_attention_layernorm/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.4.post_attention_layernorm/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.4.post_attention_layernorm/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.4.post_attention_layernorm/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.4.post_attention_layernorm/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.4.post_attention_layernorm/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.4.post_attention_layernorm/wrote/learned_3"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"Qwen2RMSNorm: rescales each token by itself.","group":["transformer","layer 5"],"kind":"norm"}},{"id":"vla.base_model.model.language_model.model.layers.4.mlp","label":"MLP","kind":"mlp","group":["transformer","layer 5"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.4.mlp","params":14180352,"note":"A gated MLP (SwiGLU): 896 numbers out to 4,864 and back, with LoRA.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.4.mlp/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.4.mlp/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.4.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.4.mlp/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.4.mlp/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.4.mlp/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.4.mlp/wrote/learned_3"},"holds":{"learned":"vla.base_model.model.language_model.model.layers.4.mlp/holds/learned","cam_top":"vla.base_model.model.language_model.model.layers.4.mlp/holds/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.4.mlp/holds/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.4.mlp/holds/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.4.mlp/holds/learned_2","text":"vla.base_model.model.language_model.model.layers.4.mlp/holds/text","learned_3":"vla.base_model.model.language_model.model.layers.4.mlp/holds/learned_3"},"auto":{"label":"MLP","note":"Works on each token by itself: Qwen2MLP. The result is added to the token. LoRA on 3 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 14 heads, sharing 2 key and value heads. Gated: one branch, passed through the bend, multiplies the other. Bend: SiLU.","group":["transformer","layer 5"],"kind":"mlp"}},{"id":"vla.base_model.model.language_model.model.layers.5.input_layernorm","label":"normalise","kind":"norm","group":["transformer","layer 6"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.5.input_layernorm","params":896,"note":"Qwen2RMSNorm: rescales each token by itself.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.5.input_layernorm/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.5.input_layernorm/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.5.input_layernorm/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.5.input_layernorm/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.5.input_layernorm/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.5.input_layernorm/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.5.input_layernorm/wrote/learned_3"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"Qwen2RMSNorm: rescales each token by itself.","group":["transformer","layer 6"],"kind":"norm"}},{"id":"vla.base_model.model.language_model.model.layers.5.self_attn","label":"attention","kind":"attention","group":["transformer","layer 6"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.5.self_attn","params":2196608,"note":"Qwen2.5-0.5B: 14 query heads share 2 key and value heads, and each token reads only the tokens before it. So the image tokens, which come first, never read the instruction; the action queries, which come last, read everything. LoRA (rank 64) adds a small trainable correction to every linear layer of this model.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.5.self_attn/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.5.self_attn/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.5.self_attn/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.5.self_attn/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.5.self_attn/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.5.self_attn/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.5.self_attn/wrote/learned_3"},"holds":{"learned":"vla.base_model.model.language_model.model.layers.5.self_attn/holds/learned","cam_top":"vla.base_model.model.language_model.model.layers.5.self_attn/holds/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.5.self_attn/holds/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.5.self_attn/holds/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.5.self_attn/holds/learned_2","text":"vla.base_model.model.language_model.model.layers.5.self_attn/holds/text","learned_3":"vla.base_model.model.language_model.model.layers.5.self_attn/holds/learned_3"},"heads":14,"auto":{"label":"attention","note":"Qwen2SdpaAttention. The only place tokens exchange information: each token reads from the others (the plate beside this storey shows who reads whom) and what it reads is added to it. Each token reads only the tokens before it. LoRA on 4 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 14 heads, sharing 2 key and value heads.","group":["transformer","layer 6"],"kind":"attention"}},{"id":"vla.base_model.model.language_model.model.layers.5.post_attention_layernorm","label":"normalise","kind":"norm","group":["transformer","layer 6"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.5.post_attention_layernorm","params":896,"note":"Qwen2RMSNorm: rescales each token by itself.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.5.post_attention_layernorm/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.5.post_attention_layernorm/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.5.post_attention_layernorm/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.5.post_attention_layernorm/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.5.post_attention_layernorm/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.5.post_attention_layernorm/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.5.post_attention_layernorm/wrote/learned_3"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"Qwen2RMSNorm: rescales each token by itself.","group":["transformer","layer 6"],"kind":"norm"}},{"id":"vla.base_model.model.language_model.model.layers.5.mlp","label":"MLP","kind":"mlp","group":["transformer","layer 6"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.5.mlp","params":14180352,"note":"A gated MLP (SwiGLU): 896 numbers out to 4,864 and back, with LoRA.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.5.mlp/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.5.mlp/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.5.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.5.mlp/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.5.mlp/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.5.mlp/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.5.mlp/wrote/learned_3"},"holds":{"learned":"vla.base_model.model.language_model.model.layers.5.mlp/holds/learned","cam_top":"vla.base_model.model.language_model.model.layers.5.mlp/holds/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.5.mlp/holds/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.5.mlp/holds/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.5.mlp/holds/learned_2","text":"vla.base_model.model.language_model.model.layers.5.mlp/holds/text","learned_3":"vla.base_model.model.language_model.model.layers.5.mlp/holds/learned_3"},"auto":{"label":"MLP","note":"Works on each token by itself: Qwen2MLP. The result is added to the token. LoRA on 3 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 14 heads, sharing 2 key and value heads. Gated: one branch, passed through the bend, multiplies the other. Bend: SiLU.","group":["transformer","layer 6"],"kind":"mlp"}},{"id":"vla.base_model.model.language_model.model.layers.6.input_layernorm","label":"normalise","kind":"norm","group":["transformer","layer 7"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.6.input_layernorm","params":896,"note":"Qwen2RMSNorm: rescales each token by itself.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.6.input_layernorm/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.6.input_layernorm/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.6.input_layernorm/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.6.input_layernorm/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.6.input_layernorm/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.6.input_layernorm/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.6.input_layernorm/wrote/learned_3"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"Qwen2RMSNorm: rescales each token by itself.","group":["transformer","layer 7"],"kind":"norm"}},{"id":"vla.base_model.model.language_model.model.layers.6.self_attn","label":"attention","kind":"attention","group":["transformer","layer 7"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.6.self_attn","params":2196608,"note":"Qwen2.5-0.5B: 14 query heads share 2 key and value heads, and each token reads only the tokens before it. So the image tokens, which come first, never read the instruction; the action queries, which come last, read everything. LoRA (rank 64) adds a small trainable correction to every linear layer of this model.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.6.self_attn/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.6.self_attn/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.6.self_attn/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.6.self_attn/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.6.self_attn/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.6.self_attn/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.6.self_attn/wrote/learned_3"},"holds":{"learned":"vla.base_model.model.language_model.model.layers.6.self_attn/holds/learned","cam_top":"vla.base_model.model.language_model.model.layers.6.self_attn/holds/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.6.self_attn/holds/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.6.self_attn/holds/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.6.self_attn/holds/learned_2","text":"vla.base_model.model.language_model.model.layers.6.self_attn/holds/text","learned_3":"vla.base_model.model.language_model.model.layers.6.self_attn/holds/learned_3"},"heads":14,"auto":{"label":"attention","note":"Qwen2SdpaAttention. The only place tokens exchange information: each token reads from the others (the plate beside this storey shows who reads whom) and what it reads is added to it. Each token reads only the tokens before it. LoRA on 4 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 14 heads, sharing 2 key and value heads.","group":["transformer","layer 7"],"kind":"attention"}},{"id":"vla.base_model.model.language_model.model.layers.6.post_attention_layernorm","label":"normalise","kind":"norm","group":["transformer","layer 7"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.6.post_attention_layernorm","params":896,"note":"Qwen2RMSNorm: rescales each token by itself.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.6.post_attention_layernorm/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.6.post_attention_layernorm/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.6.post_attention_layernorm/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.6.post_attention_layernorm/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.6.post_attention_layernorm/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.6.post_attention_layernorm/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.6.post_attention_layernorm/wrote/learned_3"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"Qwen2RMSNorm: rescales each token by itself.","group":["transformer","layer 7"],"kind":"norm"}},{"id":"vla.base_model.model.language_model.model.layers.6.mlp","label":"MLP","kind":"mlp","group":["transformer","layer 7"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.6.mlp","params":14180352,"note":"A gated MLP (SwiGLU): 896 numbers out to 4,864 and back, with LoRA.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.6.mlp/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.6.mlp/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.6.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.6.mlp/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.6.mlp/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.6.mlp/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.6.mlp/wrote/learned_3"},"holds":{"learned":"vla.base_model.model.language_model.model.layers.6.mlp/holds/learned","cam_top":"vla.base_model.model.language_model.model.layers.6.mlp/holds/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.6.mlp/holds/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.6.mlp/holds/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.6.mlp/holds/learned_2","text":"vla.base_model.model.language_model.model.layers.6.mlp/holds/text","learned_3":"vla.base_model.model.language_model.model.layers.6.mlp/holds/learned_3"},"auto":{"label":"MLP","note":"Works on each token by itself: Qwen2MLP. The result is added to the token. LoRA on 3 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 14 heads, sharing 2 key and value heads. Gated: one branch, passed through the bend, multiplies the other. Bend: SiLU.","group":["transformer","layer 7"],"kind":"mlp"}},{"id":"vla.base_model.model.language_model.model.layers.7.input_layernorm","label":"normalise","kind":"norm","group":["transformer","layer 8"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.7.input_layernorm","params":896,"note":"Qwen2RMSNorm: rescales each token by itself.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.7.input_layernorm/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.7.input_layernorm/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.7.input_layernorm/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.7.input_layernorm/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.7.input_layernorm/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.7.input_layernorm/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.7.input_layernorm/wrote/learned_3"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"Qwen2RMSNorm: rescales each token by itself.","group":["transformer","layer 8"],"kind":"norm"}},{"id":"vla.base_model.model.language_model.model.layers.7.self_attn","label":"attention","kind":"attention","group":["transformer","layer 8"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.7.self_attn","params":2196608,"note":"Qwen2.5-0.5B: 14 query heads share 2 key and value heads, and each token reads only the tokens before it. So the image tokens, which come first, never read the instruction; the action queries, which come last, read everything. LoRA (rank 64) adds a small trainable correction to every linear layer of this model.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.7.self_attn/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.7.self_attn/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.7.self_attn/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.7.self_attn/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.7.self_attn/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.7.self_attn/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.7.self_attn/wrote/learned_3"},"holds":{"learned":"vla.base_model.model.language_model.model.layers.7.self_attn/holds/learned","cam_top":"vla.base_model.model.language_model.model.layers.7.self_attn/holds/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.7.self_attn/holds/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.7.self_attn/holds/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.7.self_attn/holds/learned_2","text":"vla.base_model.model.language_model.model.layers.7.self_attn/holds/text","learned_3":"vla.base_model.model.language_model.model.layers.7.self_attn/holds/learned_3"},"heads":14,"auto":{"label":"attention","note":"Qwen2SdpaAttention. The only place tokens exchange information: each token reads from the others (the plate beside this storey shows who reads whom) and what it reads is added to it. Each token reads only the tokens before it. LoRA on 4 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 14 heads, sharing 2 key and value heads.","group":["transformer","layer 8"],"kind":"attention"}},{"id":"vla.base_model.model.language_model.model.layers.7.post_attention_layernorm","label":"normalise","kind":"norm","group":["transformer","layer 8"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.7.post_attention_layernorm","params":896,"note":"Qwen2RMSNorm: rescales each token by itself.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.7.post_attention_layernorm/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.7.post_attention_layernorm/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.7.post_attention_layernorm/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.7.post_attention_layernorm/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.7.post_attention_layernorm/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.7.post_attention_layernorm/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.7.post_attention_layernorm/wrote/learned_3"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"Qwen2RMSNorm: rescales each token by itself.","group":["transformer","layer 8"],"kind":"norm"}},{"id":"vla.base_model.model.language_model.model.layers.7.mlp","label":"MLP","kind":"mlp","group":["transformer","layer 8"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.7.mlp","params":14180352,"note":"A gated MLP (SwiGLU): 896 numbers out to 4,864 and back, with LoRA.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.7.mlp/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.7.mlp/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.7.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.7.mlp/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.7.mlp/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.7.mlp/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.7.mlp/wrote/learned_3"},"holds":{"learned":"vla.base_model.model.language_model.model.layers.7.mlp/holds/learned","cam_top":"vla.base_model.model.language_model.model.layers.7.mlp/holds/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.7.mlp/holds/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.7.mlp/holds/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.7.mlp/holds/learned_2","text":"vla.base_model.model.language_model.model.layers.7.mlp/holds/text","learned_3":"vla.base_model.model.language_model.model.layers.7.mlp/holds/learned_3"},"auto":{"label":"MLP","note":"Works on each token by itself: Qwen2MLP. The result is added to the token. LoRA on 3 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 14 heads, sharing 2 key and value heads. Gated: one branch, passed through the bend, multiplies the other. Bend: SiLU.","group":["transformer","layer 8"],"kind":"mlp"}},{"id":"vla.base_model.model.language_model.model.layers.8.input_layernorm","label":"normalise","kind":"norm","group":["transformer","layer 9"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.8.input_layernorm","params":896,"note":"Qwen2RMSNorm: rescales each token by itself.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.8.input_layernorm/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.8.input_layernorm/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.8.input_layernorm/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.8.input_layernorm/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.8.input_layernorm/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.8.input_layernorm/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.8.input_layernorm/wrote/learned_3"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"Qwen2RMSNorm: rescales each token by itself.","group":["transformer","layer 9"],"kind":"norm"}},{"id":"vla.base_model.model.language_model.model.layers.8.self_attn","label":"attention","kind":"attention","group":["transformer","layer 9"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.8.self_attn","params":2196608,"note":"Qwen2.5-0.5B: 14 query heads share 2 key and value heads, and each token reads only the tokens before it. So the image tokens, which come first, never read the instruction; the action queries, which come last, read everything. LoRA (rank 64) adds a small trainable correction to every linear layer of this model.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.8.self_attn/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.8.self_attn/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.8.self_attn/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.8.self_attn/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.8.self_attn/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.8.self_attn/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.8.self_attn/wrote/learned_3"},"holds":{"learned":"vla.base_model.model.language_model.model.layers.8.self_attn/holds/learned","cam_top":"vla.base_model.model.language_model.model.layers.8.self_attn/holds/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.8.self_attn/holds/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.8.self_attn/holds/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.8.self_attn/holds/learned_2","text":"vla.base_model.model.language_model.model.layers.8.self_attn/holds/text","learned_3":"vla.base_model.model.language_model.model.layers.8.self_attn/holds/learned_3"},"heads":14,"auto":{"label":"attention","note":"Qwen2SdpaAttention. The only place tokens exchange information: each token reads from the others (the plate beside this storey shows who reads whom) and what it reads is added to it. Each token reads only the tokens before it. LoRA on 4 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 14 heads, sharing 2 key and value heads.","group":["transformer","layer 9"],"kind":"attention"}},{"id":"vla.base_model.model.language_model.model.layers.8.post_attention_layernorm","label":"normalise","kind":"norm","group":["transformer","layer 9"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.8.post_attention_layernorm","params":896,"note":"Qwen2RMSNorm: rescales each token by itself.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.8.post_attention_layernorm/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.8.post_attention_layernorm/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.8.post_attention_layernorm/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.8.post_attention_layernorm/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.8.post_attention_layernorm/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.8.post_attention_layernorm/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.8.post_attention_layernorm/wrote/learned_3"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"Qwen2RMSNorm: rescales each token by itself.","group":["transformer","layer 9"],"kind":"norm"}},{"id":"vla.base_model.model.language_model.model.layers.8.mlp","label":"MLP","kind":"mlp","group":["transformer","layer 9"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.8.mlp","params":14180352,"note":"A gated MLP (SwiGLU): 896 numbers out to 4,864 and back, with LoRA.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.8.mlp/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.8.mlp/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.8.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.8.mlp/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.8.mlp/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.8.mlp/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.8.mlp/wrote/learned_3"},"holds":{"learned":"vla.base_model.model.language_model.model.layers.8.mlp/holds/learned","cam_top":"vla.base_model.model.language_model.model.layers.8.mlp/holds/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.8.mlp/holds/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.8.mlp/holds/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.8.mlp/holds/learned_2","text":"vla.base_model.model.language_model.model.layers.8.mlp/holds/text","learned_3":"vla.base_model.model.language_model.model.layers.8.mlp/holds/learned_3"},"auto":{"label":"MLP","note":"Works on each token by itself: Qwen2MLP. The result is added to the token. LoRA on 3 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 14 heads, sharing 2 key and value heads. Gated: one branch, passed through the bend, multiplies the other. Bend: SiLU.","group":["transformer","layer 9"],"kind":"mlp"}},{"id":"vla.base_model.model.language_model.model.layers.9.input_layernorm","label":"normalise","kind":"norm","group":["transformer","layer 10"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.9.input_layernorm","params":896,"note":"Qwen2RMSNorm: rescales each token by itself.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.9.input_layernorm/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.9.input_layernorm/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.9.input_layernorm/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.9.input_layernorm/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.9.input_layernorm/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.9.input_layernorm/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.9.input_layernorm/wrote/learned_3"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"Qwen2RMSNorm: rescales each token by itself.","group":["transformer","layer 10"],"kind":"norm"}},{"id":"vla.base_model.model.language_model.model.layers.9.self_attn","label":"attention","kind":"attention","group":["transformer","layer 10"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.9.self_attn","params":2196608,"note":"Qwen2.5-0.5B: 14 query heads share 2 key and value heads, and each token reads only the tokens before it. So the image tokens, which come first, never read the instruction; the action queries, which come last, read everything. LoRA (rank 64) adds a small trainable correction to every linear layer of this model.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.9.self_attn/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.9.self_attn/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.9.self_attn/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.9.self_attn/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.9.self_attn/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.9.self_attn/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.9.self_attn/wrote/learned_3"},"holds":{"learned":"vla.base_model.model.language_model.model.layers.9.self_attn/holds/learned","cam_top":"vla.base_model.model.language_model.model.layers.9.self_attn/holds/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.9.self_attn/holds/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.9.self_attn/holds/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.9.self_attn/holds/learned_2","text":"vla.base_model.model.language_model.model.layers.9.self_attn/holds/text","learned_3":"vla.base_model.model.language_model.model.layers.9.self_attn/holds/learned_3"},"heads":14,"auto":{"label":"attention","note":"Qwen2SdpaAttention. The only place tokens exchange information: each token reads from the others (the plate beside this storey shows who reads whom) and what it reads is added to it. Each token reads only the tokens before it. LoRA on 4 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 14 heads, sharing 2 key and value heads.","group":["transformer","layer 10"],"kind":"attention"}},{"id":"vla.base_model.model.language_model.model.layers.9.post_attention_layernorm","label":"normalise","kind":"norm","group":["transformer","layer 10"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.9.post_attention_layernorm","params":896,"note":"Qwen2RMSNorm: rescales each token by itself.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.9.post_attention_layernorm/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.9.post_attention_layernorm/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.9.post_attention_layernorm/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.9.post_attention_layernorm/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.9.post_attention_layernorm/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.9.post_attention_layernorm/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.9.post_attention_layernorm/wrote/learned_3"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"Qwen2RMSNorm: rescales each token by itself.","group":["transformer","layer 10"],"kind":"norm"}},{"id":"vla.base_model.model.language_model.model.layers.9.mlp","label":"MLP","kind":"mlp","group":["transformer","layer 10"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.9.mlp","params":14180352,"note":"A gated MLP (SwiGLU): 896 numbers out to 4,864 and back, with LoRA.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.9.mlp/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.9.mlp/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.9.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.9.mlp/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.9.mlp/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.9.mlp/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.9.mlp/wrote/learned_3"},"holds":{"learned":"vla.base_model.model.language_model.model.layers.9.mlp/holds/learned","cam_top":"vla.base_model.model.language_model.model.layers.9.mlp/holds/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.9.mlp/holds/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.9.mlp/holds/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.9.mlp/holds/learned_2","text":"vla.base_model.model.language_model.model.layers.9.mlp/holds/text","learned_3":"vla.base_model.model.language_model.model.layers.9.mlp/holds/learned_3"},"auto":{"label":"MLP","note":"Works on each token by itself: Qwen2MLP. The result is added to the token. LoRA on 3 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 14 heads, sharing 2 key and value heads. Gated: one branch, passed through the bend, multiplies the other. Bend: SiLU.","group":["transformer","layer 10"],"kind":"mlp"}},{"id":"vla.base_model.model.language_model.model.layers.10.input_layernorm","label":"normalise","kind":"norm","group":["transformer","layer 11"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.10.input_layernorm","params":896,"note":"Qwen2RMSNorm: rescales each token by itself.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.10.input_layernorm/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.10.input_layernorm/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.10.input_layernorm/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.10.input_layernorm/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.10.input_layernorm/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.10.input_layernorm/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.10.input_layernorm/wrote/learned_3"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"Qwen2RMSNorm: rescales each token by itself.","group":["transformer","layer 11"],"kind":"norm"}},{"id":"vla.base_model.model.language_model.model.layers.10.self_attn","label":"attention","kind":"attention","group":["transformer","layer 11"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.10.self_attn","params":2196608,"note":"Qwen2.5-0.5B: 14 query heads share 2 key and value heads, and each token reads only the tokens before it. So the image tokens, which come first, never read the instruction; the action queries, which come last, read everything. LoRA (rank 64) adds a small trainable correction to every linear layer of this model.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.10.self_attn/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.10.self_attn/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.10.self_attn/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.10.self_attn/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.10.self_attn/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.10.self_attn/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.10.self_attn/wrote/learned_3"},"holds":{"learned":"vla.base_model.model.language_model.model.layers.10.self_attn/holds/learned","cam_top":"vla.base_model.model.language_model.model.layers.10.self_attn/holds/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.10.self_attn/holds/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.10.self_attn/holds/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.10.self_attn/holds/learned_2","text":"vla.base_model.model.language_model.model.layers.10.self_attn/holds/text","learned_3":"vla.base_model.model.language_model.model.layers.10.self_attn/holds/learned_3"},"heads":14,"auto":{"label":"attention","note":"Qwen2SdpaAttention. The only place tokens exchange information: each token reads from the others (the plate beside this storey shows who reads whom) and what it reads is added to it. Each token reads only the tokens before it. LoRA on 4 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 14 heads, sharing 2 key and value heads.","group":["transformer","layer 11"],"kind":"attention"}},{"id":"vla.base_model.model.language_model.model.layers.10.post_attention_layernorm","label":"normalise","kind":"norm","group":["transformer","layer 11"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.10.post_attention_layernorm","params":896,"note":"Qwen2RMSNorm: rescales each token by itself.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.10.post_attention_layernorm/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.10.post_attention_layernorm/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.10.post_attention_layernorm/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.10.post_attention_layernorm/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.10.post_attention_layernorm/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.10.post_attention_layernorm/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.10.post_attention_layernorm/wrote/learned_3"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"Qwen2RMSNorm: rescales each token by itself.","group":["transformer","layer 11"],"kind":"norm"}},{"id":"vla.base_model.model.language_model.model.layers.10.mlp","label":"MLP","kind":"mlp","group":["transformer","layer 11"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.10.mlp","params":14180352,"note":"A gated MLP (SwiGLU): 896 numbers out to 4,864 and back, with LoRA.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.10.mlp/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.10.mlp/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.10.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.10.mlp/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.10.mlp/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.10.mlp/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.10.mlp/wrote/learned_3"},"holds":{"learned":"vla.base_model.model.language_model.model.layers.10.mlp/holds/learned","cam_top":"vla.base_model.model.language_model.model.layers.10.mlp/holds/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.10.mlp/holds/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.10.mlp/holds/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.10.mlp/holds/learned_2","text":"vla.base_model.model.language_model.model.layers.10.mlp/holds/text","learned_3":"vla.base_model.model.language_model.model.layers.10.mlp/holds/learned_3"},"auto":{"label":"MLP","note":"Works on each token by itself: Qwen2MLP. The result is added to the token. LoRA on 3 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 14 heads, sharing 2 key and value heads. Gated: one branch, passed through the bend, multiplies the other. Bend: SiLU.","group":["transformer","layer 11"],"kind":"mlp"}},{"id":"vla.base_model.model.language_model.model.layers.11.input_layernorm","label":"normalise","kind":"norm","group":["transformer","layer 12"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.11.input_layernorm","params":896,"note":"Qwen2RMSNorm: rescales each token by itself.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.11.input_layernorm/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.11.input_layernorm/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.11.input_layernorm/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.11.input_layernorm/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.11.input_layernorm/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.11.input_layernorm/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.11.input_layernorm/wrote/learned_3"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"Qwen2RMSNorm: rescales each token by itself.","group":["transformer","layer 12"],"kind":"norm"}},{"id":"vla.base_model.model.language_model.model.layers.11.self_attn","label":"attention","kind":"attention","group":["transformer","layer 12"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.11.self_attn","params":2196608,"note":"Qwen2.5-0.5B: 14 query heads share 2 key and value heads, and each token reads only the tokens before it. So the image tokens, which come first, never read the instruction; the action queries, which come last, read everything. LoRA (rank 64) adds a small trainable correction to every linear layer of this model.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.11.self_attn/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.11.self_attn/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.11.self_attn/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.11.self_attn/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.11.self_attn/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.11.self_attn/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.11.self_attn/wrote/learned_3"},"holds":{"learned":"vla.base_model.model.language_model.model.layers.11.self_attn/holds/learned","cam_top":"vla.base_model.model.language_model.model.layers.11.self_attn/holds/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.11.self_attn/holds/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.11.self_attn/holds/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.11.self_attn/holds/learned_2","text":"vla.base_model.model.language_model.model.layers.11.self_attn/holds/text","learned_3":"vla.base_model.model.language_model.model.layers.11.self_attn/holds/learned_3"},"heads":14,"auto":{"label":"attention","note":"Qwen2SdpaAttention. The only place tokens exchange information: each token reads from the others (the plate beside this storey shows who reads whom) and what it reads is added to it. Each token reads only the tokens before it. LoRA on 4 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 14 heads, sharing 2 key and value heads.","group":["transformer","layer 12"],"kind":"attention"}},{"id":"vla.base_model.model.language_model.model.layers.11.post_attention_layernorm","label":"normalise","kind":"norm","group":["transformer","layer 12"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.11.post_attention_layernorm","params":896,"note":"Qwen2RMSNorm: rescales each token by itself.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.11.post_attention_layernorm/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.11.post_attention_layernorm/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.11.post_attention_layernorm/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.11.post_attention_layernorm/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.11.post_attention_layernorm/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.11.post_attention_layernorm/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.11.post_attention_layernorm/wrote/learned_3"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"Qwen2RMSNorm: rescales each token by itself.","group":["transformer","layer 12"],"kind":"norm"}},{"id":"vla.base_model.model.language_model.model.layers.11.mlp","label":"MLP","kind":"mlp","group":["transformer","layer 12"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.11.mlp","params":14180352,"note":"A gated MLP (SwiGLU): 896 numbers out to 4,864 and back, with LoRA.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.11.mlp/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.11.mlp/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.11.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.11.mlp/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.11.mlp/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.11.mlp/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.11.mlp/wrote/learned_3"},"holds":{"learned":"vla.base_model.model.language_model.model.layers.11.mlp/holds/learned","cam_top":"vla.base_model.model.language_model.model.layers.11.mlp/holds/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.11.mlp/holds/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.11.mlp/holds/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.11.mlp/holds/learned_2","text":"vla.base_model.model.language_model.model.layers.11.mlp/holds/text","learned_3":"vla.base_model.model.language_model.model.layers.11.mlp/holds/learned_3"},"auto":{"label":"MLP","note":"Works on each token by itself: Qwen2MLP. The result is added to the token. LoRA on 3 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 14 heads, sharing 2 key and value heads. Gated: one branch, passed through the bend, multiplies the other. Bend: SiLU.","group":["transformer","layer 12"],"kind":"mlp"}},{"id":"vla.base_model.model.language_model.model.layers.12.input_layernorm","label":"normalise","kind":"norm","group":["transformer","layer 13"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.12.input_layernorm","params":896,"note":"Qwen2RMSNorm: rescales each token by itself.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.12.input_layernorm/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.12.input_layernorm/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.12.input_layernorm/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.12.input_layernorm/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.12.input_layernorm/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.12.input_layernorm/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.12.input_layernorm/wrote/learned_3"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"Qwen2RMSNorm: rescales each token by itself.","group":["transformer","layer 13"],"kind":"norm"}},{"id":"vla.base_model.model.language_model.model.layers.12.self_attn","label":"attention","kind":"attention","group":["transformer","layer 13"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.12.self_attn","params":2196608,"note":"Qwen2.5-0.5B: 14 query heads share 2 key and value heads, and each token reads only the tokens before it. So the image tokens, which come first, never read the instruction; the action queries, which come last, read everything. LoRA (rank 64) adds a small trainable correction to every linear layer of this model.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.12.self_attn/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.12.self_attn/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.12.self_attn/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.12.self_attn/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.12.self_attn/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.12.self_attn/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.12.self_attn/wrote/learned_3"},"holds":{"learned":"vla.base_model.model.language_model.model.layers.12.self_attn/holds/learned","cam_top":"vla.base_model.model.language_model.model.layers.12.self_attn/holds/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.12.self_attn/holds/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.12.self_attn/holds/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.12.self_attn/holds/learned_2","text":"vla.base_model.model.language_model.model.layers.12.self_attn/holds/text","learned_3":"vla.base_model.model.language_model.model.layers.12.self_attn/holds/learned_3"},"heads":14,"auto":{"label":"attention","note":"Qwen2SdpaAttention. The only place tokens exchange information: each token reads from the others (the plate beside this storey shows who reads whom) and what it reads is added to it. Each token reads only the tokens before it. LoRA on 4 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 14 heads, sharing 2 key and value heads.","group":["transformer","layer 13"],"kind":"attention"}},{"id":"vla.base_model.model.language_model.model.layers.12.post_attention_layernorm","label":"normalise","kind":"norm","group":["transformer","layer 13"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.12.post_attention_layernorm","params":896,"note":"Qwen2RMSNorm: rescales each token by itself.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.12.post_attention_layernorm/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.12.post_attention_layernorm/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.12.post_attention_layernorm/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.12.post_attention_layernorm/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.12.post_attention_layernorm/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.12.post_attention_layernorm/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.12.post_attention_layernorm/wrote/learned_3"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"Qwen2RMSNorm: rescales each token by itself.","group":["transformer","layer 13"],"kind":"norm"}},{"id":"vla.base_model.model.language_model.model.layers.12.mlp","label":"MLP","kind":"mlp","group":["transformer","layer 13"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.12.mlp","params":14180352,"note":"A gated MLP (SwiGLU): 896 numbers out to 4,864 and back, with LoRA.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.12.mlp/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.12.mlp/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.12.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.12.mlp/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.12.mlp/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.12.mlp/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.12.mlp/wrote/learned_3"},"holds":{"learned":"vla.base_model.model.language_model.model.layers.12.mlp/holds/learned","cam_top":"vla.base_model.model.language_model.model.layers.12.mlp/holds/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.12.mlp/holds/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.12.mlp/holds/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.12.mlp/holds/learned_2","text":"vla.base_model.model.language_model.model.layers.12.mlp/holds/text","learned_3":"vla.base_model.model.language_model.model.layers.12.mlp/holds/learned_3"},"auto":{"label":"MLP","note":"Works on each token by itself: Qwen2MLP. The result is added to the token. LoRA on 3 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 14 heads, sharing 2 key and value heads. Gated: one branch, passed through the bend, multiplies the other. Bend: SiLU.","group":["transformer","layer 13"],"kind":"mlp"}},{"id":"vla.base_model.model.language_model.model.layers.13.input_layernorm","label":"normalise","kind":"norm","group":["transformer","layer 14"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.13.input_layernorm","params":896,"note":"Qwen2RMSNorm: rescales each token by itself.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.13.input_layernorm/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.13.input_layernorm/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.13.input_layernorm/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.13.input_layernorm/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.13.input_layernorm/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.13.input_layernorm/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.13.input_layernorm/wrote/learned_3"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"Qwen2RMSNorm: rescales each token by itself.","group":["transformer","layer 14"],"kind":"norm"}},{"id":"vla.base_model.model.language_model.model.layers.13.self_attn","label":"attention","kind":"attention","group":["transformer","layer 14"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.13.self_attn","params":2196608,"note":"Qwen2.5-0.5B: 14 query heads share 2 key and value heads, and each token reads only the tokens before it. So the image tokens, which come first, never read the instruction; the action queries, which come last, read everything. LoRA (rank 64) adds a small trainable correction to every linear layer of this model.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.13.self_attn/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.13.self_attn/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.13.self_attn/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.13.self_attn/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.13.self_attn/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.13.self_attn/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.13.self_attn/wrote/learned_3"},"holds":{"learned":"vla.base_model.model.language_model.model.layers.13.self_attn/holds/learned","cam_top":"vla.base_model.model.language_model.model.layers.13.self_attn/holds/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.13.self_attn/holds/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.13.self_attn/holds/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.13.self_attn/holds/learned_2","text":"vla.base_model.model.language_model.model.layers.13.self_attn/holds/text","learned_3":"vla.base_model.model.language_model.model.layers.13.self_attn/holds/learned_3"},"heads":14,"auto":{"label":"attention","note":"Qwen2SdpaAttention. The only place tokens exchange information: each token reads from the others (the plate beside this storey shows who reads whom) and what it reads is added to it. Each token reads only the tokens before it. LoRA on 4 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 14 heads, sharing 2 key and value heads.","group":["transformer","layer 14"],"kind":"attention"}},{"id":"vla.base_model.model.language_model.model.layers.13.post_attention_layernorm","label":"normalise","kind":"norm","group":["transformer","layer 14"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.13.post_attention_layernorm","params":896,"note":"Qwen2RMSNorm: rescales each token by itself.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.13.post_attention_layernorm/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.13.post_attention_layernorm/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.13.post_attention_layernorm/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.13.post_attention_layernorm/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.13.post_attention_layernorm/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.13.post_attention_layernorm/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.13.post_attention_layernorm/wrote/learned_3"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"Qwen2RMSNorm: rescales each token by itself.","group":["transformer","layer 14"],"kind":"norm"}},{"id":"vla.base_model.model.language_model.model.layers.13.mlp","label":"MLP","kind":"mlp","group":["transformer","layer 14"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.13.mlp","params":14180352,"note":"A gated MLP (SwiGLU): 896 numbers out to 4,864 and back, with LoRA.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.13.mlp/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.13.mlp/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.13.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.13.mlp/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.13.mlp/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.13.mlp/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.13.mlp/wrote/learned_3"},"holds":{"learned":"vla.base_model.model.language_model.model.layers.13.mlp/holds/learned","cam_top":"vla.base_model.model.language_model.model.layers.13.mlp/holds/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.13.mlp/holds/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.13.mlp/holds/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.13.mlp/holds/learned_2","text":"vla.base_model.model.language_model.model.layers.13.mlp/holds/text","learned_3":"vla.base_model.model.language_model.model.layers.13.mlp/holds/learned_3"},"auto":{"label":"MLP","note":"Works on each token by itself: Qwen2MLP. The result is added to the token. LoRA on 3 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 14 heads, sharing 2 key and value heads. Gated: one branch, passed through the bend, multiplies the other. Bend: SiLU.","group":["transformer","layer 14"],"kind":"mlp"}},{"id":"vla.base_model.model.language_model.model.layers.14.input_layernorm","label":"normalise","kind":"norm","group":["transformer","layer 15"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.14.input_layernorm","params":896,"note":"Qwen2RMSNorm: rescales each token by itself.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.14.input_layernorm/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.14.input_layernorm/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.14.input_layernorm/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.14.input_layernorm/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.14.input_layernorm/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.14.input_layernorm/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.14.input_layernorm/wrote/learned_3"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"Qwen2RMSNorm: rescales each token by itself.","group":["transformer","layer 15"],"kind":"norm"}},{"id":"vla.base_model.model.language_model.model.layers.14.self_attn","label":"attention","kind":"attention","group":["transformer","layer 15"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.14.self_attn","params":2196608,"note":"Qwen2.5-0.5B: 14 query heads share 2 key and value heads, and each token reads only the tokens before it. So the image tokens, which come first, never read the instruction; the action queries, which come last, read everything. LoRA (rank 64) adds a small trainable correction to every linear layer of this model.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.14.self_attn/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.14.self_attn/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.14.self_attn/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.14.self_attn/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.14.self_attn/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.14.self_attn/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.14.self_attn/wrote/learned_3"},"holds":{"learned":"vla.base_model.model.language_model.model.layers.14.self_attn/holds/learned","cam_top":"vla.base_model.model.language_model.model.layers.14.self_attn/holds/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.14.self_attn/holds/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.14.self_attn/holds/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.14.self_attn/holds/learned_2","text":"vla.base_model.model.language_model.model.layers.14.self_attn/holds/text","learned_3":"vla.base_model.model.language_model.model.layers.14.self_attn/holds/learned_3"},"heads":14,"auto":{"label":"attention","note":"Qwen2SdpaAttention. The only place tokens exchange information: each token reads from the others (the plate beside this storey shows who reads whom) and what it reads is added to it. Each token reads only the tokens before it. LoRA on 4 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 14 heads, sharing 2 key and value heads.","group":["transformer","layer 15"],"kind":"attention"}},{"id":"vla.base_model.model.language_model.model.layers.14.post_attention_layernorm","label":"normalise","kind":"norm","group":["transformer","layer 15"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.14.post_attention_layernorm","params":896,"note":"Qwen2RMSNorm: rescales each token by itself.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.14.post_attention_layernorm/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.14.post_attention_layernorm/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.14.post_attention_layernorm/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.14.post_attention_layernorm/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.14.post_attention_layernorm/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.14.post_attention_layernorm/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.14.post_attention_layernorm/wrote/learned_3"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"Qwen2RMSNorm: rescales each token by itself.","group":["transformer","layer 15"],"kind":"norm"}},{"id":"vla.base_model.model.language_model.model.layers.14.mlp","label":"MLP","kind":"mlp","group":["transformer","layer 15"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.14.mlp","params":14180352,"note":"A gated MLP (SwiGLU): 896 numbers out to 4,864 and back, with LoRA.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.14.mlp/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.14.mlp/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.14.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.14.mlp/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.14.mlp/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.14.mlp/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.14.mlp/wrote/learned_3"},"holds":{"learned":"vla.base_model.model.language_model.model.layers.14.mlp/holds/learned","cam_top":"vla.base_model.model.language_model.model.layers.14.mlp/holds/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.14.mlp/holds/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.14.mlp/holds/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.14.mlp/holds/learned_2","text":"vla.base_model.model.language_model.model.layers.14.mlp/holds/text","learned_3":"vla.base_model.model.language_model.model.layers.14.mlp/holds/learned_3"},"auto":{"label":"MLP","note":"Works on each token by itself: Qwen2MLP. The result is added to the token. LoRA on 3 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 14 heads, sharing 2 key and value heads. Gated: one branch, passed through the bend, multiplies the other. Bend: SiLU.","group":["transformer","layer 15"],"kind":"mlp"}},{"id":"vla.base_model.model.language_model.model.layers.15.input_layernorm","label":"normalise","kind":"norm","group":["transformer","layer 16"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.15.input_layernorm","params":896,"note":"Qwen2RMSNorm: rescales each token by itself.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.15.input_layernorm/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.15.input_layernorm/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.15.input_layernorm/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.15.input_layernorm/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.15.input_layernorm/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.15.input_layernorm/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.15.input_layernorm/wrote/learned_3"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"Qwen2RMSNorm: rescales each token by itself.","group":["transformer","layer 16"],"kind":"norm"}},{"id":"vla.base_model.model.language_model.model.layers.15.self_attn","label":"attention","kind":"attention","group":["transformer","layer 16"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.15.self_attn","params":2196608,"note":"Qwen2.5-0.5B: 14 query heads share 2 key and value heads, and each token reads only the tokens before it. So the image tokens, which come first, never read the instruction; the action queries, which come last, read everything. LoRA (rank 64) adds a small trainable correction to every linear layer of this model.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.15.self_attn/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.15.self_attn/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.15.self_attn/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.15.self_attn/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.15.self_attn/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.15.self_attn/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.15.self_attn/wrote/learned_3"},"holds":{"learned":"vla.base_model.model.language_model.model.layers.15.self_attn/holds/learned","cam_top":"vla.base_model.model.language_model.model.layers.15.self_attn/holds/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.15.self_attn/holds/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.15.self_attn/holds/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.15.self_attn/holds/learned_2","text":"vla.base_model.model.language_model.model.layers.15.self_attn/holds/text","learned_3":"vla.base_model.model.language_model.model.layers.15.self_attn/holds/learned_3"},"heads":14,"auto":{"label":"attention","note":"Qwen2SdpaAttention. The only place tokens exchange information: each token reads from the others (the plate beside this storey shows who reads whom) and what it reads is added to it. Each token reads only the tokens before it. LoRA on 4 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 14 heads, sharing 2 key and value heads.","group":["transformer","layer 16"],"kind":"attention"}},{"id":"vla.base_model.model.language_model.model.layers.15.post_attention_layernorm","label":"normalise","kind":"norm","group":["transformer","layer 16"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.15.post_attention_layernorm","params":896,"note":"Qwen2RMSNorm: rescales each token by itself.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.15.post_attention_layernorm/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.15.post_attention_layernorm/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.15.post_attention_layernorm/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.15.post_attention_layernorm/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.15.post_attention_layernorm/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.15.post_attention_layernorm/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.15.post_attention_layernorm/wrote/learned_3"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"Qwen2RMSNorm: rescales each token by itself.","group":["transformer","layer 16"],"kind":"norm"}},{"id":"vla.base_model.model.language_model.model.layers.15.mlp","label":"MLP","kind":"mlp","group":["transformer","layer 16"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.15.mlp","params":14180352,"note":"A gated MLP (SwiGLU): 896 numbers out to 4,864 and back, with LoRA.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.15.mlp/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.15.mlp/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.15.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.15.mlp/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.15.mlp/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.15.mlp/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.15.mlp/wrote/learned_3"},"holds":{"learned":"vla.base_model.model.language_model.model.layers.15.mlp/holds/learned","cam_top":"vla.base_model.model.language_model.model.layers.15.mlp/holds/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.15.mlp/holds/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.15.mlp/holds/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.15.mlp/holds/learned_2","text":"vla.base_model.model.language_model.model.layers.15.mlp/holds/text","learned_3":"vla.base_model.model.language_model.model.layers.15.mlp/holds/learned_3"},"auto":{"label":"MLP","note":"Works on each token by itself: Qwen2MLP. The result is added to the token. LoRA on 3 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 14 heads, sharing 2 key and value heads. Gated: one branch, passed through the bend, multiplies the other. Bend: SiLU.","group":["transformer","layer 16"],"kind":"mlp"}},{"id":"vla.base_model.model.language_model.model.layers.16.input_layernorm","label":"normalise","kind":"norm","group":["transformer","layer 17"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.16.input_layernorm","params":896,"note":"Qwen2RMSNorm: rescales each token by itself.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.16.input_layernorm/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.16.input_layernorm/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.16.input_layernorm/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.16.input_layernorm/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.16.input_layernorm/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.16.input_layernorm/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.16.input_layernorm/wrote/learned_3"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"Qwen2RMSNorm: rescales each token by itself.","group":["transformer","layer 17"],"kind":"norm"}},{"id":"vla.base_model.model.language_model.model.layers.16.self_attn","label":"attention","kind":"attention","group":["transformer","layer 17"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.16.self_attn","params":2196608,"note":"Qwen2.5-0.5B: 14 query heads share 2 key and value heads, and each token reads only the tokens before it. So the image tokens, which come first, never read the instruction; the action queries, which come last, read everything. LoRA (rank 64) adds a small trainable correction to every linear layer of this model.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.16.self_attn/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.16.self_attn/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.16.self_attn/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.16.self_attn/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.16.self_attn/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.16.self_attn/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.16.self_attn/wrote/learned_3"},"holds":{"learned":"vla.base_model.model.language_model.model.layers.16.self_attn/holds/learned","cam_top":"vla.base_model.model.language_model.model.layers.16.self_attn/holds/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.16.self_attn/holds/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.16.self_attn/holds/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.16.self_attn/holds/learned_2","text":"vla.base_model.model.language_model.model.layers.16.self_attn/holds/text","learned_3":"vla.base_model.model.language_model.model.layers.16.self_attn/holds/learned_3"},"heads":14,"auto":{"label":"attention","note":"Qwen2SdpaAttention. The only place tokens exchange information: each token reads from the others (the plate beside this storey shows who reads whom) and what it reads is added to it. Each token reads only the tokens before it. LoRA on 4 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 14 heads, sharing 2 key and value heads.","group":["transformer","layer 17"],"kind":"attention"}},{"id":"vla.base_model.model.language_model.model.layers.16.post_attention_layernorm","label":"normalise","kind":"norm","group":["transformer","layer 17"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.16.post_attention_layernorm","params":896,"note":"Qwen2RMSNorm: rescales each token by itself.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.16.post_attention_layernorm/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.16.post_attention_layernorm/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.16.post_attention_layernorm/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.16.post_attention_layernorm/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.16.post_attention_layernorm/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.16.post_attention_layernorm/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.16.post_attention_layernorm/wrote/learned_3"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"Qwen2RMSNorm: rescales each token by itself.","group":["transformer","layer 17"],"kind":"norm"}},{"id":"vla.base_model.model.language_model.model.layers.16.mlp","label":"MLP","kind":"mlp","group":["transformer","layer 17"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.16.mlp","params":14180352,"note":"A gated MLP (SwiGLU): 896 numbers out to 4,864 and back, with LoRA.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.16.mlp/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.16.mlp/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.16.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.16.mlp/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.16.mlp/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.16.mlp/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.16.mlp/wrote/learned_3"},"holds":{"learned":"vla.base_model.model.language_model.model.layers.16.mlp/holds/learned","cam_top":"vla.base_model.model.language_model.model.layers.16.mlp/holds/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.16.mlp/holds/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.16.mlp/holds/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.16.mlp/holds/learned_2","text":"vla.base_model.model.language_model.model.layers.16.mlp/holds/text","learned_3":"vla.base_model.model.language_model.model.layers.16.mlp/holds/learned_3"},"auto":{"label":"MLP","note":"Works on each token by itself: Qwen2MLP. The result is added to the token. LoRA on 3 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 14 heads, sharing 2 key and value heads. Gated: one branch, passed through the bend, multiplies the other. Bend: SiLU.","group":["transformer","layer 17"],"kind":"mlp"}},{"id":"vla.base_model.model.language_model.model.layers.17.input_layernorm","label":"normalise","kind":"norm","group":["transformer","layer 18"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.17.input_layernorm","params":896,"note":"Qwen2RMSNorm: rescales each token by itself.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.17.input_layernorm/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.17.input_layernorm/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.17.input_layernorm/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.17.input_layernorm/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.17.input_layernorm/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.17.input_layernorm/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.17.input_layernorm/wrote/learned_3"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"Qwen2RMSNorm: rescales each token by itself.","group":["transformer","layer 18"],"kind":"norm"}},{"id":"vla.base_model.model.language_model.model.layers.17.self_attn","label":"attention","kind":"attention","group":["transformer","layer 18"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.17.self_attn","params":2196608,"note":"Qwen2.5-0.5B: 14 query heads share 2 key and value heads, and each token reads only the tokens before it. So the image tokens, which come first, never read the instruction; the action queries, which come last, read everything. LoRA (rank 64) adds a small trainable correction to every linear layer of this model.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.17.self_attn/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.17.self_attn/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.17.self_attn/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.17.self_attn/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.17.self_attn/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.17.self_attn/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.17.self_attn/wrote/learned_3"},"holds":{"learned":"vla.base_model.model.language_model.model.layers.17.self_attn/holds/learned","cam_top":"vla.base_model.model.language_model.model.layers.17.self_attn/holds/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.17.self_attn/holds/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.17.self_attn/holds/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.17.self_attn/holds/learned_2","text":"vla.base_model.model.language_model.model.layers.17.self_attn/holds/text","learned_3":"vla.base_model.model.language_model.model.layers.17.self_attn/holds/learned_3"},"heads":14,"auto":{"label":"attention","note":"Qwen2SdpaAttention. The only place tokens exchange information: each token reads from the others (the plate beside this storey shows who reads whom) and what it reads is added to it. Each token reads only the tokens before it. LoRA on 4 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 14 heads, sharing 2 key and value heads.","group":["transformer","layer 18"],"kind":"attention"}},{"id":"vla.base_model.model.language_model.model.layers.17.post_attention_layernorm","label":"normalise","kind":"norm","group":["transformer","layer 18"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.17.post_attention_layernorm","params":896,"note":"Qwen2RMSNorm: rescales each token by itself.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.17.post_attention_layernorm/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.17.post_attention_layernorm/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.17.post_attention_layernorm/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.17.post_attention_layernorm/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.17.post_attention_layernorm/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.17.post_attention_layernorm/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.17.post_attention_layernorm/wrote/learned_3"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"Qwen2RMSNorm: rescales each token by itself.","group":["transformer","layer 18"],"kind":"norm"}},{"id":"vla.base_model.model.language_model.model.layers.17.mlp","label":"MLP","kind":"mlp","group":["transformer","layer 18"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.17.mlp","params":14180352,"note":"A gated MLP (SwiGLU): 896 numbers out to 4,864 and back, with LoRA.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.17.mlp/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.17.mlp/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.17.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.17.mlp/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.17.mlp/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.17.mlp/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.17.mlp/wrote/learned_3"},"holds":{"learned":"vla.base_model.model.language_model.model.layers.17.mlp/holds/learned","cam_top":"vla.base_model.model.language_model.model.layers.17.mlp/holds/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.17.mlp/holds/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.17.mlp/holds/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.17.mlp/holds/learned_2","text":"vla.base_model.model.language_model.model.layers.17.mlp/holds/text","learned_3":"vla.base_model.model.language_model.model.layers.17.mlp/holds/learned_3"},"auto":{"label":"MLP","note":"Works on each token by itself: Qwen2MLP. The result is added to the token. LoRA on 3 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 14 heads, sharing 2 key and value heads. Gated: one branch, passed through the bend, multiplies the other. Bend: SiLU.","group":["transformer","layer 18"],"kind":"mlp"}},{"id":"vla.base_model.model.language_model.model.layers.18.input_layernorm","label":"normalise","kind":"norm","group":["transformer","layer 19"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.18.input_layernorm","params":896,"note":"Qwen2RMSNorm: rescales each token by itself.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.18.input_layernorm/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.18.input_layernorm/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.18.input_layernorm/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.18.input_layernorm/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.18.input_layernorm/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.18.input_layernorm/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.18.input_layernorm/wrote/learned_3"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"Qwen2RMSNorm: rescales each token by itself.","group":["transformer","layer 19"],"kind":"norm"}},{"id":"vla.base_model.model.language_model.model.layers.18.self_attn","label":"attention","kind":"attention","group":["transformer","layer 19"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.18.self_attn","params":2196608,"note":"Qwen2.5-0.5B: 14 query heads share 2 key and value heads, and each token reads only the tokens before it. So the image tokens, which come first, never read the instruction; the action queries, which come last, read everything. LoRA (rank 64) adds a small trainable correction to every linear layer of this model.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.18.self_attn/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.18.self_attn/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.18.self_attn/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.18.self_attn/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.18.self_attn/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.18.self_attn/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.18.self_attn/wrote/learned_3"},"holds":{"learned":"vla.base_model.model.language_model.model.layers.18.self_attn/holds/learned","cam_top":"vla.base_model.model.language_model.model.layers.18.self_attn/holds/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.18.self_attn/holds/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.18.self_attn/holds/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.18.self_attn/holds/learned_2","text":"vla.base_model.model.language_model.model.layers.18.self_attn/holds/text","learned_3":"vla.base_model.model.language_model.model.layers.18.self_attn/holds/learned_3"},"heads":14,"auto":{"label":"attention","note":"Qwen2SdpaAttention. The only place tokens exchange information: each token reads from the others (the plate beside this storey shows who reads whom) and what it reads is added to it. Each token reads only the tokens before it. LoRA on 4 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 14 heads, sharing 2 key and value heads.","group":["transformer","layer 19"],"kind":"attention"}},{"id":"vla.base_model.model.language_model.model.layers.18.post_attention_layernorm","label":"normalise","kind":"norm","group":["transformer","layer 19"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.18.post_attention_layernorm","params":896,"note":"Qwen2RMSNorm: rescales each token by itself.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.18.post_attention_layernorm/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.18.post_attention_layernorm/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.18.post_attention_layernorm/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.18.post_attention_layernorm/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.18.post_attention_layernorm/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.18.post_attention_layernorm/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.18.post_attention_layernorm/wrote/learned_3"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"Qwen2RMSNorm: rescales each token by itself.","group":["transformer","layer 19"],"kind":"norm"}},{"id":"vla.base_model.model.language_model.model.layers.18.mlp","label":"MLP","kind":"mlp","group":["transformer","layer 19"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.18.mlp","params":14180352,"note":"A gated MLP (SwiGLU): 896 numbers out to 4,864 and back, with LoRA.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.18.mlp/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.18.mlp/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.18.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.18.mlp/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.18.mlp/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.18.mlp/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.18.mlp/wrote/learned_3"},"holds":{"learned":"vla.base_model.model.language_model.model.layers.18.mlp/holds/learned","cam_top":"vla.base_model.model.language_model.model.layers.18.mlp/holds/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.18.mlp/holds/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.18.mlp/holds/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.18.mlp/holds/learned_2","text":"vla.base_model.model.language_model.model.layers.18.mlp/holds/text","learned_3":"vla.base_model.model.language_model.model.layers.18.mlp/holds/learned_3"},"auto":{"label":"MLP","note":"Works on each token by itself: Qwen2MLP. The result is added to the token. LoRA on 3 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 14 heads, sharing 2 key and value heads. Gated: one branch, passed through the bend, multiplies the other. Bend: SiLU.","group":["transformer","layer 19"],"kind":"mlp"}},{"id":"vla.base_model.model.language_model.model.layers.19.input_layernorm","label":"normalise","kind":"norm","group":["transformer","layer 20"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.19.input_layernorm","params":896,"note":"Qwen2RMSNorm: rescales each token by itself.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.19.input_layernorm/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.19.input_layernorm/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.19.input_layernorm/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.19.input_layernorm/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.19.input_layernorm/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.19.input_layernorm/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.19.input_layernorm/wrote/learned_3"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"Qwen2RMSNorm: rescales each token by itself.","group":["transformer","layer 20"],"kind":"norm"}},{"id":"vla.base_model.model.language_model.model.layers.19.self_attn","label":"attention","kind":"attention","group":["transformer","layer 20"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.19.self_attn","params":2196608,"note":"Qwen2.5-0.5B: 14 query heads share 2 key and value heads, and each token reads only the tokens before it. So the image tokens, which come first, never read the instruction; the action queries, which come last, read everything. LoRA (rank 64) adds a small trainable correction to every linear layer of this model.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.19.self_attn/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.19.self_attn/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.19.self_attn/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.19.self_attn/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.19.self_attn/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.19.self_attn/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.19.self_attn/wrote/learned_3"},"holds":{"learned":"vla.base_model.model.language_model.model.layers.19.self_attn/holds/learned","cam_top":"vla.base_model.model.language_model.model.layers.19.self_attn/holds/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.19.self_attn/holds/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.19.self_attn/holds/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.19.self_attn/holds/learned_2","text":"vla.base_model.model.language_model.model.layers.19.self_attn/holds/text","learned_3":"vla.base_model.model.language_model.model.layers.19.self_attn/holds/learned_3"},"heads":14,"auto":{"label":"attention","note":"Qwen2SdpaAttention. The only place tokens exchange information: each token reads from the others (the plate beside this storey shows who reads whom) and what it reads is added to it. Each token reads only the tokens before it. LoRA on 4 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 14 heads, sharing 2 key and value heads.","group":["transformer","layer 20"],"kind":"attention"}},{"id":"vla.base_model.model.language_model.model.layers.19.post_attention_layernorm","label":"normalise","kind":"norm","group":["transformer","layer 20"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.19.post_attention_layernorm","params":896,"note":"Qwen2RMSNorm: rescales each token by itself.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.19.post_attention_layernorm/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.19.post_attention_layernorm/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.19.post_attention_layernorm/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.19.post_attention_layernorm/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.19.post_attention_layernorm/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.19.post_attention_layernorm/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.19.post_attention_layernorm/wrote/learned_3"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"Qwen2RMSNorm: rescales each token by itself.","group":["transformer","layer 20"],"kind":"norm"}},{"id":"vla.base_model.model.language_model.model.layers.19.mlp","label":"MLP","kind":"mlp","group":["transformer","layer 20"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.19.mlp","params":14180352,"note":"A gated MLP (SwiGLU): 896 numbers out to 4,864 and back, with LoRA.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.19.mlp/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.19.mlp/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.19.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.19.mlp/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.19.mlp/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.19.mlp/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.19.mlp/wrote/learned_3"},"holds":{"learned":"vla.base_model.model.language_model.model.layers.19.mlp/holds/learned","cam_top":"vla.base_model.model.language_model.model.layers.19.mlp/holds/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.19.mlp/holds/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.19.mlp/holds/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.19.mlp/holds/learned_2","text":"vla.base_model.model.language_model.model.layers.19.mlp/holds/text","learned_3":"vla.base_model.model.language_model.model.layers.19.mlp/holds/learned_3"},"auto":{"label":"MLP","note":"Works on each token by itself: Qwen2MLP. The result is added to the token. LoRA on 3 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 14 heads, sharing 2 key and value heads. Gated: one branch, passed through the bend, multiplies the other. Bend: SiLU.","group":["transformer","layer 20"],"kind":"mlp"}},{"id":"vla.base_model.model.language_model.model.layers.20.input_layernorm","label":"normalise","kind":"norm","group":["transformer","layer 21"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.20.input_layernorm","params":896,"note":"Qwen2RMSNorm: rescales each token by itself.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.20.input_layernorm/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.20.input_layernorm/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.20.input_layernorm/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.20.input_layernorm/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.20.input_layernorm/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.20.input_layernorm/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.20.input_layernorm/wrote/learned_3"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"Qwen2RMSNorm: rescales each token by itself.","group":["transformer","layer 21"],"kind":"norm"}},{"id":"vla.base_model.model.language_model.model.layers.20.self_attn","label":"attention","kind":"attention","group":["transformer","layer 21"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.20.self_attn","params":2196608,"note":"Qwen2.5-0.5B: 14 query heads share 2 key and value heads, and each token reads only the tokens before it. So the image tokens, which come first, never read the instruction; the action queries, which come last, read everything. LoRA (rank 64) adds a small trainable correction to every linear layer of this model.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.20.self_attn/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.20.self_attn/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.20.self_attn/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.20.self_attn/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.20.self_attn/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.20.self_attn/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.20.self_attn/wrote/learned_3"},"holds":{"learned":"vla.base_model.model.language_model.model.layers.20.self_attn/holds/learned","cam_top":"vla.base_model.model.language_model.model.layers.20.self_attn/holds/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.20.self_attn/holds/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.20.self_attn/holds/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.20.self_attn/holds/learned_2","text":"vla.base_model.model.language_model.model.layers.20.self_attn/holds/text","learned_3":"vla.base_model.model.language_model.model.layers.20.self_attn/holds/learned_3"},"heads":14,"auto":{"label":"attention","note":"Qwen2SdpaAttention. The only place tokens exchange information: each token reads from the others (the plate beside this storey shows who reads whom) and what it reads is added to it. Each token reads only the tokens before it. LoRA on 4 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 14 heads, sharing 2 key and value heads.","group":["transformer","layer 21"],"kind":"attention"}},{"id":"vla.base_model.model.language_model.model.layers.20.post_attention_layernorm","label":"normalise","kind":"norm","group":["transformer","layer 21"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.20.post_attention_layernorm","params":896,"note":"Qwen2RMSNorm: rescales each token by itself.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.20.post_attention_layernorm/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.20.post_attention_layernorm/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.20.post_attention_layernorm/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.20.post_attention_layernorm/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.20.post_attention_layernorm/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.20.post_attention_layernorm/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.20.post_attention_layernorm/wrote/learned_3"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"Qwen2RMSNorm: rescales each token by itself.","group":["transformer","layer 21"],"kind":"norm"}},{"id":"vla.base_model.model.language_model.model.layers.20.mlp","label":"MLP","kind":"mlp","group":["transformer","layer 21"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.20.mlp","params":14180352,"note":"A gated MLP (SwiGLU): 896 numbers out to 4,864 and back, with LoRA.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.20.mlp/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.20.mlp/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.20.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.20.mlp/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.20.mlp/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.20.mlp/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.20.mlp/wrote/learned_3"},"holds":{"learned":"vla.base_model.model.language_model.model.layers.20.mlp/holds/learned","cam_top":"vla.base_model.model.language_model.model.layers.20.mlp/holds/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.20.mlp/holds/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.20.mlp/holds/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.20.mlp/holds/learned_2","text":"vla.base_model.model.language_model.model.layers.20.mlp/holds/text","learned_3":"vla.base_model.model.language_model.model.layers.20.mlp/holds/learned_3"},"auto":{"label":"MLP","note":"Works on each token by itself: Qwen2MLP. The result is added to the token. LoRA on 3 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 14 heads, sharing 2 key and value heads. Gated: one branch, passed through the bend, multiplies the other. Bend: SiLU.","group":["transformer","layer 21"],"kind":"mlp"}},{"id":"vla.base_model.model.language_model.model.layers.21.input_layernorm","label":"normalise","kind":"norm","group":["transformer","layer 22"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.21.input_layernorm","params":896,"note":"Qwen2RMSNorm: rescales each token by itself.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.21.input_layernorm/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.21.input_layernorm/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.21.input_layernorm/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.21.input_layernorm/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.21.input_layernorm/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.21.input_layernorm/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.21.input_layernorm/wrote/learned_3"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"Qwen2RMSNorm: rescales each token by itself.","group":["transformer","layer 22"],"kind":"norm"}},{"id":"vla.base_model.model.language_model.model.layers.21.self_attn","label":"attention","kind":"attention","group":["transformer","layer 22"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.21.self_attn","params":2196608,"note":"Qwen2.5-0.5B: 14 query heads share 2 key and value heads, and each token reads only the tokens before it. So the image tokens, which come first, never read the instruction; the action queries, which come last, read everything. LoRA (rank 64) adds a small trainable correction to every linear layer of this model.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.21.self_attn/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.21.self_attn/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.21.self_attn/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.21.self_attn/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.21.self_attn/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.21.self_attn/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.21.self_attn/wrote/learned_3"},"holds":{"learned":"vla.base_model.model.language_model.model.layers.21.self_attn/holds/learned","cam_top":"vla.base_model.model.language_model.model.layers.21.self_attn/holds/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.21.self_attn/holds/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.21.self_attn/holds/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.21.self_attn/holds/learned_2","text":"vla.base_model.model.language_model.model.layers.21.self_attn/holds/text","learned_3":"vla.base_model.model.language_model.model.layers.21.self_attn/holds/learned_3"},"heads":14,"auto":{"label":"attention","note":"Qwen2SdpaAttention. The only place tokens exchange information: each token reads from the others (the plate beside this storey shows who reads whom) and what it reads is added to it. Each token reads only the tokens before it. LoRA on 4 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 14 heads, sharing 2 key and value heads.","group":["transformer","layer 22"],"kind":"attention"}},{"id":"vla.base_model.model.language_model.model.layers.21.post_attention_layernorm","label":"normalise","kind":"norm","group":["transformer","layer 22"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.21.post_attention_layernorm","params":896,"note":"Qwen2RMSNorm: rescales each token by itself.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.21.post_attention_layernorm/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.21.post_attention_layernorm/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.21.post_attention_layernorm/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.21.post_attention_layernorm/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.21.post_attention_layernorm/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.21.post_attention_layernorm/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.21.post_attention_layernorm/wrote/learned_3"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"Qwen2RMSNorm: rescales each token by itself.","group":["transformer","layer 22"],"kind":"norm"}},{"id":"vla.base_model.model.language_model.model.layers.21.mlp","label":"MLP","kind":"mlp","group":["transformer","layer 22"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.21.mlp","params":14180352,"note":"A gated MLP (SwiGLU): 896 numbers out to 4,864 and back, with LoRA.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.21.mlp/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.21.mlp/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.21.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.21.mlp/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.21.mlp/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.21.mlp/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.21.mlp/wrote/learned_3"},"holds":{"learned":"vla.base_model.model.language_model.model.layers.21.mlp/holds/learned","cam_top":"vla.base_model.model.language_model.model.layers.21.mlp/holds/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.21.mlp/holds/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.21.mlp/holds/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.21.mlp/holds/learned_2","text":"vla.base_model.model.language_model.model.layers.21.mlp/holds/text","learned_3":"vla.base_model.model.language_model.model.layers.21.mlp/holds/learned_3"},"auto":{"label":"MLP","note":"Works on each token by itself: Qwen2MLP. The result is added to the token. LoRA on 3 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 14 heads, sharing 2 key and value heads. Gated: one branch, passed through the bend, multiplies the other. Bend: SiLU.","group":["transformer","layer 22"],"kind":"mlp"}},{"id":"vla.base_model.model.language_model.model.layers.22.input_layernorm","label":"normalise","kind":"norm","group":["transformer","layer 23"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.22.input_layernorm","params":896,"note":"Qwen2RMSNorm: rescales each token by itself.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.22.input_layernorm/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.22.input_layernorm/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.22.input_layernorm/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.22.input_layernorm/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.22.input_layernorm/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.22.input_layernorm/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.22.input_layernorm/wrote/learned_3"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"Qwen2RMSNorm: rescales each token by itself.","group":["transformer","layer 23"],"kind":"norm"}},{"id":"vla.base_model.model.language_model.model.layers.22.self_attn","label":"attention","kind":"attention","group":["transformer","layer 23"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.22.self_attn","params":2196608,"note":"Qwen2.5-0.5B: 14 query heads share 2 key and value heads, and each token reads only the tokens before it. So the image tokens, which come first, never read the instruction; the action queries, which come last, read everything. LoRA (rank 64) adds a small trainable correction to every linear layer of this model.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.22.self_attn/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.22.self_attn/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.22.self_attn/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.22.self_attn/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.22.self_attn/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.22.self_attn/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.22.self_attn/wrote/learned_3"},"holds":{"learned":"vla.base_model.model.language_model.model.layers.22.self_attn/holds/learned","cam_top":"vla.base_model.model.language_model.model.layers.22.self_attn/holds/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.22.self_attn/holds/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.22.self_attn/holds/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.22.self_attn/holds/learned_2","text":"vla.base_model.model.language_model.model.layers.22.self_attn/holds/text","learned_3":"vla.base_model.model.language_model.model.layers.22.self_attn/holds/learned_3"},"heads":14,"auto":{"label":"attention","note":"Qwen2SdpaAttention. The only place tokens exchange information: each token reads from the others (the plate beside this storey shows who reads whom) and what it reads is added to it. Each token reads only the tokens before it. LoRA on 4 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 14 heads, sharing 2 key and value heads.","group":["transformer","layer 23"],"kind":"attention"}},{"id":"vla.base_model.model.language_model.model.layers.22.post_attention_layernorm","label":"normalise","kind":"norm","group":["transformer","layer 23"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.22.post_attention_layernorm","params":896,"note":"Qwen2RMSNorm: rescales each token by itself.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.22.post_attention_layernorm/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.22.post_attention_layernorm/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.22.post_attention_layernorm/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.22.post_attention_layernorm/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.22.post_attention_layernorm/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.22.post_attention_layernorm/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.22.post_attention_layernorm/wrote/learned_3"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"Qwen2RMSNorm: rescales each token by itself.","group":["transformer","layer 23"],"kind":"norm"}},{"id":"vla.base_model.model.language_model.model.layers.22.mlp","label":"MLP","kind":"mlp","group":["transformer","layer 23"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.22.mlp","params":14180352,"note":"A gated MLP (SwiGLU): 896 numbers out to 4,864 and back, with LoRA.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.22.mlp/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.22.mlp/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.22.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.22.mlp/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.22.mlp/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.22.mlp/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.22.mlp/wrote/learned_3"},"holds":{"learned":"vla.base_model.model.language_model.model.layers.22.mlp/holds/learned","cam_top":"vla.base_model.model.language_model.model.layers.22.mlp/holds/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.22.mlp/holds/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.22.mlp/holds/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.22.mlp/holds/learned_2","text":"vla.base_model.model.language_model.model.layers.22.mlp/holds/text","learned_3":"vla.base_model.model.language_model.model.layers.22.mlp/holds/learned_3"},"auto":{"label":"MLP","note":"Works on each token by itself: Qwen2MLP. The result is added to the token. LoRA on 3 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 14 heads, sharing 2 key and value heads. Gated: one branch, passed through the bend, multiplies the other. Bend: SiLU.","group":["transformer","layer 23"],"kind":"mlp"}},{"id":"vla.base_model.model.language_model.model.layers.23.input_layernorm","label":"normalise","kind":"norm","group":["transformer","layer 24"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.23.input_layernorm","params":896,"note":"Qwen2RMSNorm: rescales each token by itself.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.23.input_layernorm/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.23.input_layernorm/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.23.input_layernorm/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.23.input_layernorm/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.23.input_layernorm/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.23.input_layernorm/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.23.input_layernorm/wrote/learned_3"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"Qwen2RMSNorm: rescales each token by itself.","group":["transformer","layer 24"],"kind":"norm"}},{"id":"vla.base_model.model.language_model.model.layers.23.self_attn","label":"attention","kind":"attention","group":["transformer","layer 24"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.23.self_attn","params":2196608,"note":"Qwen2.5-0.5B: 14 query heads share 2 key and value heads, and each token reads only the tokens before it. So the image tokens, which come first, never read the instruction; the action queries, which come last, read everything. LoRA (rank 64) adds a small trainable correction to every linear layer of this model.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.23.self_attn/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.23.self_attn/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.23.self_attn/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.23.self_attn/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.23.self_attn/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.23.self_attn/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.23.self_attn/wrote/learned_3"},"holds":{"learned":"vla.base_model.model.language_model.model.layers.23.self_attn/holds/learned","cam_top":"vla.base_model.model.language_model.model.layers.23.self_attn/holds/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.23.self_attn/holds/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.23.self_attn/holds/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.23.self_attn/holds/learned_2","text":"vla.base_model.model.language_model.model.layers.23.self_attn/holds/text","learned_3":"vla.base_model.model.language_model.model.layers.23.self_attn/holds/learned_3"},"heads":14,"auto":{"label":"attention","note":"Qwen2SdpaAttention. The only place tokens exchange information: each token reads from the others (the plate beside this storey shows who reads whom) and what it reads is added to it. Each token reads only the tokens before it. LoRA on 4 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 14 heads, sharing 2 key and value heads.","group":["transformer","layer 24"],"kind":"attention"}},{"id":"vla.base_model.model.language_model.model.layers.23.post_attention_layernorm","label":"normalise","kind":"norm","group":["transformer","layer 24"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.23.post_attention_layernorm","params":896,"note":"Qwen2RMSNorm: rescales each token by itself.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.23.post_attention_layernorm/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.23.post_attention_layernorm/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.23.post_attention_layernorm/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.23.post_attention_layernorm/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.23.post_attention_layernorm/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.23.post_attention_layernorm/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.23.post_attention_layernorm/wrote/learned_3"},"holds":{},"thin":true,"auto":{"label":"normalise","note":"Qwen2RMSNorm: rescales each token by itself.","group":["transformer","layer 24"],"kind":"norm"}},{"id":"vla.base_model.model.language_model.model.layers.23.mlp","label":"MLP","kind":"mlp","group":["transformer","layer 24"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.layers.23.mlp","params":14180352,"note":"A gated MLP (SwiGLU): 896 numbers out to 4,864 and back, with LoRA.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.layers.23.mlp/wrote/learned","cam_top":"vla.base_model.model.language_model.model.layers.23.mlp/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.23.mlp/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.23.mlp/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.23.mlp/wrote/learned_2","text":"vla.base_model.model.language_model.model.layers.23.mlp/wrote/text","learned_3":"vla.base_model.model.language_model.model.layers.23.mlp/wrote/learned_3"},"holds":{"learned":"vla.base_model.model.language_model.model.layers.23.mlp/holds/learned","cam_top":"vla.base_model.model.language_model.model.layers.23.mlp/holds/cam_top","cam_side":"vla.base_model.model.language_model.model.layers.23.mlp/holds/cam_side","cam_iso":"vla.base_model.model.language_model.model.layers.23.mlp/holds/cam_iso","learned_2":"vla.base_model.model.language_model.model.layers.23.mlp/holds/learned_2","text":"vla.base_model.model.language_model.model.layers.23.mlp/holds/text","learned_3":"vla.base_model.model.language_model.model.layers.23.mlp/holds/learned_3"},"auto":{"label":"MLP","note":"Works on each token by itself: Qwen2MLP. The result is added to the token. LoRA on 3 layers: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). 14 heads, sharing 2 key and value heads. Gated: one branch, passed through the bend, multiplies the other. Bend: SiLU.","group":["transformer","layer 24"],"kind":"mlp"}},{"id":"vla.base_model.model.language_model.model.norm","label":"norm: Qwen2RMSNorm","kind":"norm","group":["output"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":896,"module":"vla.base_model.model.language_model.model.norm","params":896,"note":"Qwen2RMSNorm. Out: [1, 874, 896]. Part of qwen2 (as its code names it).","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.model.norm/wrote/learned","cam_top":"vla.base_model.model.language_model.model.norm/wrote/cam_top","cam_side":"vla.base_model.model.language_model.model.norm/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.model.norm/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.model.norm/wrote/learned_2","text":"vla.base_model.model.language_model.model.norm/wrote/text","learned_3":"vla.base_model.model.language_model.model.norm/wrote/learned_3"},"holds":{},"auto":{"label":"norm: Qwen2RMSNorm","note":"Qwen2RMSNorm. Out: [1, 874, 896]. Part of qwen2 (as its code names it).","group":["output"],"kind":"norm"}},{"id":"vla.base_model.model.language_model.lm_head","label":"word prediction (never used for actions)","kind":"linear","group":["output"],"plots":["learned","cam_top","cam_side","cam_iso","learned_2","text","learned_3"],"width":151936,"module":"vla.base_model.model.language_model.lm_head","params":145915904,"note":"Runs because the published forward pass computes it, but the action head never reads it. Its 136M weights are the same tensor as the word table at the bottom.","cells":{"learned":[1,1],"cam_top":[16,16],"cam_side":[16,16],"cam_iso":[16,16],"learned_2":[2,15],"text":[1,5],"learned_3":[5,14]},"wrote":{"learned":"vla.base_model.model.language_model.lm_head/wrote/learned","cam_top":"vla.base_model.model.language_model.lm_head/wrote/cam_top","cam_side":"vla.base_model.model.language_model.lm_head/wrote/cam_side","cam_iso":"vla.base_model.model.language_model.lm_head/wrote/cam_iso","learned_2":"vla.base_model.model.language_model.lm_head/wrote/learned_2","text":"vla.base_model.model.language_model.lm_head/wrote/text","learned_3":"vla.base_model.model.language_model.lm_head/wrote/learned_3"},"holds":{},"branch_of":"vla.base_model.model.language_model.model.norm","train_only":true,"auto":{"label":"lm_head: Linear, Identity, Linear, Linear, ParameterDict, ParameterDict","note":"Linear. Out: [1, 874, 151936]. LoRA on 1 layer: frozen weights plus a trainable correction of rank 64, still all zeros (untrained). Part of qwen2 (as its code names it). Its output never reaches the answer: drawn to the side, as something used in training only.","group":["output"],"kind":"linear"}},{"id":"proprio_projector","label":"state to one token for the head","kind":"linear","group":["output"],"plots":["state"],"width":896,"module":"proprio_projector","params":832384,"note":"The arm state, as 31 numbers, through two linear layers. It does not enter the language model: only the action head reads it.","cells":{"state":[1,1]},"wrote":{"state":"proprio_projector/wrote/state"},"holds":{},"auto":{"label":"proprio_projector: Linear, Linear, GELU","note":"ProprioProjector. Out: [1, 896]. Bend: GELU.","group":["output"],"kind":"linear"}},{"id":"action_head.model.layer_norm1","label":"action_head.model.layer_norm1: LayerNorm","kind":"norm","group":["action_head"],"plots":["answer"],"width":6272,"module":"action_head.model.layer_norm1","params":12544,"note":"LayerNorm. Out: [1, 8, 6272].","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.layer_norm1/wrote/answer"},"holds":{},"auto":{"label":"action_head.model.layer_norm1: LayerNorm","note":"LayerNorm. Out: [1, 8, 6272].","group":["action_head"],"kind":"norm"}},{"id":"action_head.model.fc1","label":"head: the 8 step tokens start from zeros","kind":"linear","group":["action_head"],"plots":["answer"],"width":896,"module":"action_head.model.fc1","params":5620608,"note":"The head's own 8 tokens, one per step of the chunk, start as zeros (plus a fixed perturbation in training). Everything they learn comes from what they read in the blocks above.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.fc1/wrote/answer"},"holds":{},"auto":{"label":"action_head.model.fc1: Linear","note":"Linear. Out: [1, 8, 896]. 6,272 numbers in, 896 out.","group":["action_head"],"kind":"linear"}},{"id":"action_head.model.relu","label":"action_head.model.relu: ReLU","kind":"linear","group":["action_head"],"plots":["answer"],"width":896,"module":"action_head.model.relu","params":0,"note":"ReLU. Out: [1, 8, 896]. Bend: ReLU.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.relu/wrote/answer"},"holds":{},"auto":{"label":"action_head.model.relu: ReLU","note":"ReLU. Out: [1, 8, 896]. Bend: ReLU.","group":["action_head"],"kind":"linear"}},{"id":"action_head.model.mlp_resnet_blocks.0.attn","label":"attention","kind":"attention","group":["action_head","mlp_resnet_blocks 1"],"plots":["answer"],"width":896,"module":"action_head.model.mlp_resnet_blocks.0.q_proj, action_head.model.mlp_resnet_blocks.0.k_self, action_head.model.mlp_resnet_blocks.0.v_self, action_head.model.mlp_resnet_blocks.0.k_adapter, action_head.model.mlp_resnet_blocks.0.v_adapter, action_head.model.mlp_resnet_blocks.0.k_task, action_head.model.mlp_resnet_blocks.0.v_task, action_head.model.mlp_resnet_blocks.0.rope, action_head.model.mlp_resnet_blocks.0.o_proj","params":6429696,"note":"Block i reads language-model layer i+1. Each step token reads three things at once: the other step tokens; the action queries' hidden states plus the state token; and the image tokens' hidden states, through a learned gate. The split between \"image tokens\" and the rest is fixed at 512 tokens, the paper's two cameras: here the iso camera's 256 tokens land with the action queries.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.0.attn/wrote/answer"},"holds":{},"auto":{"label":"attention","note":"Linear, Linear, Linear, Linear, Linear, Linear, Linear, RotaryPositionEmbedding, Linear, making queries, keys and values: attention written out in the model's own code. The only place tokens exchange information: each token reads from the others. 896 numbers in, 896 out.","group":["action_head","mlp_resnet_blocks 1"],"kind":"attention"}},{"id":"action_head.model.mlp_resnet_blocks.0.ffn","label":"MLP","kind":"mlp","group":["action_head","mlp_resnet_blocks 1"],"plots":["answer"],"width":896,"module":"action_head.model.mlp_resnet_blocks.0.ffn","params":805504,"note":"LayerNorm and an MLP, on (what was read + the token): the result replaces the token.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.0.ffn/wrote/answer"},"holds":{},"auto":{"label":"MLP","note":"Works on each token by itself: Sequential. Bend: ReLU.","group":["action_head","mlp_resnet_blocks 1"],"kind":"mlp"}},{"id":"action_head.model.mlp_resnet_blocks.0.out","label":"layer output","kind":"norm","group":["action_head","mlp_resnet_blocks 1"],"plots":["answer"],"width":896,"module":"","params":0,"note":"This layer does not simply add its parts' outputs to its input; this is what it hands on.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.0.out/holds/answer"},"holds":{"answer":"action_head.model.mlp_resnet_blocks.0.out/holds/answer"},"thin":true,"auto":{"label":"layer output","note":"This layer does not simply add its parts' outputs to its input; this is what it hands on.","group":["action_head","mlp_resnet_blocks 1"],"kind":"norm"}},{"id":"action_head.model.mlp_resnet_blocks.1.attn","label":"attention","kind":"attention","group":["action_head","mlp_resnet_blocks 2"],"plots":["answer"],"width":896,"module":"action_head.model.mlp_resnet_blocks.1.q_proj, action_head.model.mlp_resnet_blocks.1.k_self, action_head.model.mlp_resnet_blocks.1.v_self, action_head.model.mlp_resnet_blocks.1.k_adapter, action_head.model.mlp_resnet_blocks.1.v_adapter, action_head.model.mlp_resnet_blocks.1.k_task, action_head.model.mlp_resnet_blocks.1.v_task, action_head.model.mlp_resnet_blocks.1.rope, action_head.model.mlp_resnet_blocks.1.o_proj","params":6429696,"note":"Block i reads language-model layer i+1. Each step token reads three things at once: the other step tokens; the action queries' hidden states plus the state token; and the image tokens' hidden states, through a learned gate. The split between \"image tokens\" and the rest is fixed at 512 tokens, the paper's two cameras: here the iso camera's 256 tokens land with the action queries.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.1.attn/wrote/answer"},"holds":{},"auto":{"label":"attention","note":"Linear, Linear, Linear, Linear, Linear, Linear, Linear, RotaryPositionEmbedding, Linear, making queries, keys and values: attention written out in the model's own code. The only place tokens exchange information: each token reads from the others. 896 numbers in, 896 out.","group":["action_head","mlp_resnet_blocks 2"],"kind":"attention"}},{"id":"action_head.model.mlp_resnet_blocks.1.ffn","label":"MLP","kind":"mlp","group":["action_head","mlp_resnet_blocks 2"],"plots":["answer"],"width":896,"module":"action_head.model.mlp_resnet_blocks.1.ffn","params":805504,"note":"LayerNorm and an MLP, on (what was read + the token): the result replaces the token.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.1.ffn/wrote/answer"},"holds":{},"auto":{"label":"MLP","note":"Works on each token by itself: Sequential. Bend: ReLU.","group":["action_head","mlp_resnet_blocks 2"],"kind":"mlp"}},{"id":"action_head.model.mlp_resnet_blocks.1.out","label":"layer output","kind":"norm","group":["action_head","mlp_resnet_blocks 2"],"plots":["answer"],"width":896,"module":"","params":0,"note":"This layer does not simply add its parts' outputs to its input; this is what it hands on.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.1.out/holds/answer"},"holds":{"answer":"action_head.model.mlp_resnet_blocks.1.out/holds/answer"},"thin":true,"auto":{"label":"layer output","note":"This layer does not simply add its parts' outputs to its input; this is what it hands on.","group":["action_head","mlp_resnet_blocks 2"],"kind":"norm"}},{"id":"action_head.model.mlp_resnet_blocks.2.attn","label":"attention","kind":"attention","group":["action_head","mlp_resnet_blocks 3"],"plots":["answer"],"width":896,"module":"action_head.model.mlp_resnet_blocks.2.q_proj, action_head.model.mlp_resnet_blocks.2.k_self, action_head.model.mlp_resnet_blocks.2.v_self, action_head.model.mlp_resnet_blocks.2.k_adapter, action_head.model.mlp_resnet_blocks.2.v_adapter, action_head.model.mlp_resnet_blocks.2.k_task, action_head.model.mlp_resnet_blocks.2.v_task, action_head.model.mlp_resnet_blocks.2.rope, action_head.model.mlp_resnet_blocks.2.o_proj","params":6429696,"note":"Block i reads language-model layer i+1. Each step token reads three things at once: the other step tokens; the action queries' hidden states plus the state token; and the image tokens' hidden states, through a learned gate. The split between \"image tokens\" and the rest is fixed at 512 tokens, the paper's two cameras: here the iso camera's 256 tokens land with the action queries.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.2.attn/wrote/answer"},"holds":{},"auto":{"label":"attention","note":"Linear, Linear, Linear, Linear, Linear, Linear, Linear, RotaryPositionEmbedding, Linear, making queries, keys and values: attention written out in the model's own code. The only place tokens exchange information: each token reads from the others. 896 numbers in, 896 out.","group":["action_head","mlp_resnet_blocks 3"],"kind":"attention"}},{"id":"action_head.model.mlp_resnet_blocks.2.ffn","label":"MLP","kind":"mlp","group":["action_head","mlp_resnet_blocks 3"],"plots":["answer"],"width":896,"module":"action_head.model.mlp_resnet_blocks.2.ffn","params":805504,"note":"LayerNorm and an MLP, on (what was read + the token): the result replaces the token.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.2.ffn/wrote/answer"},"holds":{},"auto":{"label":"MLP","note":"Works on each token by itself: Sequential. Bend: ReLU.","group":["action_head","mlp_resnet_blocks 3"],"kind":"mlp"}},{"id":"action_head.model.mlp_resnet_blocks.2.out","label":"layer output","kind":"norm","group":["action_head","mlp_resnet_blocks 3"],"plots":["answer"],"width":896,"module":"","params":0,"note":"This layer does not simply add its parts' outputs to its input; this is what it hands on.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.2.out/holds/answer"},"holds":{"answer":"action_head.model.mlp_resnet_blocks.2.out/holds/answer"},"thin":true,"auto":{"label":"layer output","note":"This layer does not simply add its parts' outputs to its input; this is what it hands on.","group":["action_head","mlp_resnet_blocks 3"],"kind":"norm"}},{"id":"action_head.model.mlp_resnet_blocks.3.attn","label":"attention","kind":"attention","group":["action_head","mlp_resnet_blocks 4"],"plots":["answer"],"width":896,"module":"action_head.model.mlp_resnet_blocks.3.q_proj, action_head.model.mlp_resnet_blocks.3.k_self, action_head.model.mlp_resnet_blocks.3.v_self, action_head.model.mlp_resnet_blocks.3.k_adapter, action_head.model.mlp_resnet_blocks.3.v_adapter, action_head.model.mlp_resnet_blocks.3.k_task, action_head.model.mlp_resnet_blocks.3.v_task, action_head.model.mlp_resnet_blocks.3.rope, action_head.model.mlp_resnet_blocks.3.o_proj","params":6429696,"note":"Block i reads language-model layer i+1. Each step token reads three things at once: the other step tokens; the action queries' hidden states plus the state token; and the image tokens' hidden states, through a learned gate. The split between \"image tokens\" and the rest is fixed at 512 tokens, the paper's two cameras: here the iso camera's 256 tokens land with the action queries.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.3.attn/wrote/answer"},"holds":{},"auto":{"label":"attention","note":"Linear, Linear, Linear, Linear, Linear, Linear, Linear, RotaryPositionEmbedding, Linear, making queries, keys and values: attention written out in the model's own code. The only place tokens exchange information: each token reads from the others. 896 numbers in, 896 out.","group":["action_head","mlp_resnet_blocks 4"],"kind":"attention"}},{"id":"action_head.model.mlp_resnet_blocks.3.ffn","label":"MLP","kind":"mlp","group":["action_head","mlp_resnet_blocks 4"],"plots":["answer"],"width":896,"module":"action_head.model.mlp_resnet_blocks.3.ffn","params":805504,"note":"LayerNorm and an MLP, on (what was read + the token): the result replaces the token.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.3.ffn/wrote/answer"},"holds":{},"auto":{"label":"MLP","note":"Works on each token by itself: Sequential. Bend: ReLU.","group":["action_head","mlp_resnet_blocks 4"],"kind":"mlp"}},{"id":"action_head.model.mlp_resnet_blocks.3.out","label":"layer output","kind":"norm","group":["action_head","mlp_resnet_blocks 4"],"plots":["answer"],"width":896,"module":"","params":0,"note":"This layer does not simply add its parts' outputs to its input; this is what it hands on.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.3.out/holds/answer"},"holds":{"answer":"action_head.model.mlp_resnet_blocks.3.out/holds/answer"},"thin":true,"auto":{"label":"layer output","note":"This layer does not simply add its parts' outputs to its input; this is what it hands on.","group":["action_head","mlp_resnet_blocks 4"],"kind":"norm"}},{"id":"action_head.model.mlp_resnet_blocks.4.attn","label":"attention","kind":"attention","group":["action_head","mlp_resnet_blocks 5"],"plots":["answer"],"width":896,"module":"action_head.model.mlp_resnet_blocks.4.q_proj, action_head.model.mlp_resnet_blocks.4.k_self, action_head.model.mlp_resnet_blocks.4.v_self, action_head.model.mlp_resnet_blocks.4.k_adapter, action_head.model.mlp_resnet_blocks.4.v_adapter, action_head.model.mlp_resnet_blocks.4.k_task, action_head.model.mlp_resnet_blocks.4.v_task, action_head.model.mlp_resnet_blocks.4.rope, action_head.model.mlp_resnet_blocks.4.o_proj","params":6429696,"note":"Block i reads language-model layer i+1. Each step token reads three things at once: the other step tokens; the action queries' hidden states plus the state token; and the image tokens' hidden states, through a learned gate. The split between \"image tokens\" and the rest is fixed at 512 tokens, the paper's two cameras: here the iso camera's 256 tokens land with the action queries.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.4.attn/wrote/answer"},"holds":{},"auto":{"label":"attention","note":"Linear, Linear, Linear, Linear, Linear, Linear, Linear, RotaryPositionEmbedding, Linear, making queries, keys and values: attention written out in the model's own code. The only place tokens exchange information: each token reads from the others. 896 numbers in, 896 out.","group":["action_head","mlp_resnet_blocks 5"],"kind":"attention"}},{"id":"action_head.model.mlp_resnet_blocks.4.ffn","label":"MLP","kind":"mlp","group":["action_head","mlp_resnet_blocks 5"],"plots":["answer"],"width":896,"module":"action_head.model.mlp_resnet_blocks.4.ffn","params":805504,"note":"LayerNorm and an MLP, on (what was read + the token): the result replaces the token.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.4.ffn/wrote/answer"},"holds":{},"auto":{"label":"MLP","note":"Works on each token by itself: Sequential. Bend: ReLU.","group":["action_head","mlp_resnet_blocks 5"],"kind":"mlp"}},{"id":"action_head.model.mlp_resnet_blocks.4.out","label":"layer output","kind":"norm","group":["action_head","mlp_resnet_blocks 5"],"plots":["answer"],"width":896,"module":"","params":0,"note":"This layer does not simply add its parts' outputs to its input; this is what it hands on.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.4.out/holds/answer"},"holds":{"answer":"action_head.model.mlp_resnet_blocks.4.out/holds/answer"},"thin":true,"auto":{"label":"layer output","note":"This layer does not simply add its parts' outputs to its input; this is what it hands on.","group":["action_head","mlp_resnet_blocks 5"],"kind":"norm"}},{"id":"action_head.model.mlp_resnet_blocks.5.attn","label":"attention","kind":"attention","group":["action_head","mlp_resnet_blocks 6"],"plots":["answer"],"width":896,"module":"action_head.model.mlp_resnet_blocks.5.q_proj, action_head.model.mlp_resnet_blocks.5.k_self, action_head.model.mlp_resnet_blocks.5.v_self, action_head.model.mlp_resnet_blocks.5.k_adapter, action_head.model.mlp_resnet_blocks.5.v_adapter, action_head.model.mlp_resnet_blocks.5.k_task, action_head.model.mlp_resnet_blocks.5.v_task, action_head.model.mlp_resnet_blocks.5.rope, action_head.model.mlp_resnet_blocks.5.o_proj","params":6429696,"note":"Block i reads language-model layer i+1. Each step token reads three things at once: the other step tokens; the action queries' hidden states plus the state token; and the image tokens' hidden states, through a learned gate. The split between \"image tokens\" and the rest is fixed at 512 tokens, the paper's two cameras: here the iso camera's 256 tokens land with the action queries.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.5.attn/wrote/answer"},"holds":{},"auto":{"label":"attention","note":"Linear, Linear, Linear, Linear, Linear, Linear, Linear, RotaryPositionEmbedding, Linear, making queries, keys and values: attention written out in the model's own code. The only place tokens exchange information: each token reads from the others. 896 numbers in, 896 out.","group":["action_head","mlp_resnet_blocks 6"],"kind":"attention"}},{"id":"action_head.model.mlp_resnet_blocks.5.ffn","label":"MLP","kind":"mlp","group":["action_head","mlp_resnet_blocks 6"],"plots":["answer"],"width":896,"module":"action_head.model.mlp_resnet_blocks.5.ffn","params":805504,"note":"LayerNorm and an MLP, on (what was read + the token): the result replaces the token.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.5.ffn/wrote/answer"},"holds":{},"auto":{"label":"MLP","note":"Works on each token by itself: Sequential. Bend: ReLU.","group":["action_head","mlp_resnet_blocks 6"],"kind":"mlp"}},{"id":"action_head.model.mlp_resnet_blocks.5.out","label":"layer output","kind":"norm","group":["action_head","mlp_resnet_blocks 6"],"plots":["answer"],"width":896,"module":"","params":0,"note":"This layer does not simply add its parts' outputs to its input; this is what it hands on.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.5.out/holds/answer"},"holds":{"answer":"action_head.model.mlp_resnet_blocks.5.out/holds/answer"},"thin":true,"auto":{"label":"layer output","note":"This layer does not simply add its parts' outputs to its input; this is what it hands on.","group":["action_head","mlp_resnet_blocks 6"],"kind":"norm"}},{"id":"action_head.model.mlp_resnet_blocks.6.attn","label":"attention","kind":"attention","group":["action_head","mlp_resnet_blocks 7"],"plots":["answer"],"width":896,"module":"action_head.model.mlp_resnet_blocks.6.q_proj, action_head.model.mlp_resnet_blocks.6.k_self, action_head.model.mlp_resnet_blocks.6.v_self, action_head.model.mlp_resnet_blocks.6.k_adapter, action_head.model.mlp_resnet_blocks.6.v_adapter, action_head.model.mlp_resnet_blocks.6.k_task, action_head.model.mlp_resnet_blocks.6.v_task, action_head.model.mlp_resnet_blocks.6.rope, action_head.model.mlp_resnet_blocks.6.o_proj","params":6429696,"note":"Block i reads language-model layer i+1. Each step token reads three things at once: the other step tokens; the action queries' hidden states plus the state token; and the image tokens' hidden states, through a learned gate. The split between \"image tokens\" and the rest is fixed at 512 tokens, the paper's two cameras: here the iso camera's 256 tokens land with the action queries.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.6.attn/wrote/answer"},"holds":{},"auto":{"label":"attention","note":"Linear, Linear, Linear, Linear, Linear, Linear, Linear, RotaryPositionEmbedding, Linear, making queries, keys and values: attention written out in the model's own code. The only place tokens exchange information: each token reads from the others. 896 numbers in, 896 out.","group":["action_head","mlp_resnet_blocks 7"],"kind":"attention"}},{"id":"action_head.model.mlp_resnet_blocks.6.ffn","label":"MLP","kind":"mlp","group":["action_head","mlp_resnet_blocks 7"],"plots":["answer"],"width":896,"module":"action_head.model.mlp_resnet_blocks.6.ffn","params":805504,"note":"LayerNorm and an MLP, on (what was read + the token): the result replaces the token.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.6.ffn/wrote/answer"},"holds":{},"auto":{"label":"MLP","note":"Works on each token by itself: Sequential. Bend: ReLU.","group":["action_head","mlp_resnet_blocks 7"],"kind":"mlp"}},{"id":"action_head.model.mlp_resnet_blocks.6.out","label":"layer output","kind":"norm","group":["action_head","mlp_resnet_blocks 7"],"plots":["answer"],"width":896,"module":"","params":0,"note":"This layer does not simply add its parts' outputs to its input; this is what it hands on.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.6.out/holds/answer"},"holds":{"answer":"action_head.model.mlp_resnet_blocks.6.out/holds/answer"},"thin":true,"auto":{"label":"layer output","note":"This layer does not simply add its parts' outputs to its input; this is what it hands on.","group":["action_head","mlp_resnet_blocks 7"],"kind":"norm"}},{"id":"action_head.model.mlp_resnet_blocks.7.attn","label":"attention","kind":"attention","group":["action_head","mlp_resnet_blocks 8"],"plots":["answer"],"width":896,"module":"action_head.model.mlp_resnet_blocks.7.q_proj, action_head.model.mlp_resnet_blocks.7.k_self, action_head.model.mlp_resnet_blocks.7.v_self, action_head.model.mlp_resnet_blocks.7.k_adapter, action_head.model.mlp_resnet_blocks.7.v_adapter, action_head.model.mlp_resnet_blocks.7.k_task, action_head.model.mlp_resnet_blocks.7.v_task, action_head.model.mlp_resnet_blocks.7.rope, action_head.model.mlp_resnet_blocks.7.o_proj","params":6429696,"note":"Block i reads language-model layer i+1. Each step token reads three things at once: the other step tokens; the action queries' hidden states plus the state token; and the image tokens' hidden states, through a learned gate. The split between \"image tokens\" and the rest is fixed at 512 tokens, the paper's two cameras: here the iso camera's 256 tokens land with the action queries.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.7.attn/wrote/answer"},"holds":{},"auto":{"label":"attention","note":"Linear, Linear, Linear, Linear, Linear, Linear, Linear, RotaryPositionEmbedding, Linear, making queries, keys and values: attention written out in the model's own code. The only place tokens exchange information: each token reads from the others. 896 numbers in, 896 out.","group":["action_head","mlp_resnet_blocks 8"],"kind":"attention"}},{"id":"action_head.model.mlp_resnet_blocks.7.ffn","label":"MLP","kind":"mlp","group":["action_head","mlp_resnet_blocks 8"],"plots":["answer"],"width":896,"module":"action_head.model.mlp_resnet_blocks.7.ffn","params":805504,"note":"LayerNorm and an MLP, on (what was read + the token): the result replaces the token.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.7.ffn/wrote/answer"},"holds":{},"auto":{"label":"MLP","note":"Works on each token by itself: Sequential. Bend: ReLU.","group":["action_head","mlp_resnet_blocks 8"],"kind":"mlp"}},{"id":"action_head.model.mlp_resnet_blocks.7.out","label":"layer output","kind":"norm","group":["action_head","mlp_resnet_blocks 8"],"plots":["answer"],"width":896,"module":"","params":0,"note":"This layer does not simply add its parts' outputs to its input; this is what it hands on.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.7.out/holds/answer"},"holds":{"answer":"action_head.model.mlp_resnet_blocks.7.out/holds/answer"},"thin":true,"auto":{"label":"layer output","note":"This layer does not simply add its parts' outputs to its input; this is what it hands on.","group":["action_head","mlp_resnet_blocks 8"],"kind":"norm"}},{"id":"action_head.model.mlp_resnet_blocks.8.attn","label":"attention","kind":"attention","group":["action_head","mlp_resnet_blocks 9"],"plots":["answer"],"width":896,"module":"action_head.model.mlp_resnet_blocks.8.q_proj, action_head.model.mlp_resnet_blocks.8.k_self, action_head.model.mlp_resnet_blocks.8.v_self, action_head.model.mlp_resnet_blocks.8.k_adapter, action_head.model.mlp_resnet_blocks.8.v_adapter, action_head.model.mlp_resnet_blocks.8.k_task, action_head.model.mlp_resnet_blocks.8.v_task, action_head.model.mlp_resnet_blocks.8.rope, action_head.model.mlp_resnet_blocks.8.o_proj","params":6429696,"note":"Block i reads language-model layer i+1. Each step token reads three things at once: the other step tokens; the action queries' hidden states plus the state token; and the image tokens' hidden states, through a learned gate. The split between \"image tokens\" and the rest is fixed at 512 tokens, the paper's two cameras: here the iso camera's 256 tokens land with the action queries.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.8.attn/wrote/answer"},"holds":{},"auto":{"label":"attention","note":"Linear, Linear, Linear, Linear, Linear, Linear, Linear, RotaryPositionEmbedding, Linear, making queries, keys and values: attention written out in the model's own code. The only place tokens exchange information: each token reads from the others. 896 numbers in, 896 out.","group":["action_head","mlp_resnet_blocks 9"],"kind":"attention"}},{"id":"action_head.model.mlp_resnet_blocks.8.ffn","label":"MLP","kind":"mlp","group":["action_head","mlp_resnet_blocks 9"],"plots":["answer"],"width":896,"module":"action_head.model.mlp_resnet_blocks.8.ffn","params":805504,"note":"LayerNorm and an MLP, on (what was read + the token): the result replaces the token.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.8.ffn/wrote/answer"},"holds":{},"auto":{"label":"MLP","note":"Works on each token by itself: Sequential. Bend: ReLU.","group":["action_head","mlp_resnet_blocks 9"],"kind":"mlp"}},{"id":"action_head.model.mlp_resnet_blocks.8.out","label":"layer output","kind":"norm","group":["action_head","mlp_resnet_blocks 9"],"plots":["answer"],"width":896,"module":"","params":0,"note":"This layer does not simply add its parts' outputs to its input; this is what it hands on.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.8.out/holds/answer"},"holds":{"answer":"action_head.model.mlp_resnet_blocks.8.out/holds/answer"},"thin":true,"auto":{"label":"layer output","note":"This layer does not simply add its parts' outputs to its input; this is what it hands on.","group":["action_head","mlp_resnet_blocks 9"],"kind":"norm"}},{"id":"action_head.model.mlp_resnet_blocks.9.attn","label":"attention","kind":"attention","group":["action_head","mlp_resnet_blocks 10"],"plots":["answer"],"width":896,"module":"action_head.model.mlp_resnet_blocks.9.q_proj, action_head.model.mlp_resnet_blocks.9.k_self, action_head.model.mlp_resnet_blocks.9.v_self, action_head.model.mlp_resnet_blocks.9.k_adapter, action_head.model.mlp_resnet_blocks.9.v_adapter, action_head.model.mlp_resnet_blocks.9.k_task, action_head.model.mlp_resnet_blocks.9.v_task, action_head.model.mlp_resnet_blocks.9.rope, action_head.model.mlp_resnet_blocks.9.o_proj","params":6429696,"note":"Block i reads language-model layer i+1. Each step token reads three things at once: the other step tokens; the action queries' hidden states plus the state token; and the image tokens' hidden states, through a learned gate. The split between \"image tokens\" and the rest is fixed at 512 tokens, the paper's two cameras: here the iso camera's 256 tokens land with the action queries.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.9.attn/wrote/answer"},"holds":{},"auto":{"label":"attention","note":"Linear, Linear, Linear, Linear, Linear, Linear, Linear, RotaryPositionEmbedding, Linear, making queries, keys and values: attention written out in the model's own code. The only place tokens exchange information: each token reads from the others. 896 numbers in, 896 out.","group":["action_head","mlp_resnet_blocks 10"],"kind":"attention"}},{"id":"action_head.model.mlp_resnet_blocks.9.ffn","label":"MLP","kind":"mlp","group":["action_head","mlp_resnet_blocks 10"],"plots":["answer"],"width":896,"module":"action_head.model.mlp_resnet_blocks.9.ffn","params":805504,"note":"LayerNorm and an MLP, on (what was read + the token): the result replaces the token.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.9.ffn/wrote/answer"},"holds":{},"auto":{"label":"MLP","note":"Works on each token by itself: Sequential. Bend: ReLU.","group":["action_head","mlp_resnet_blocks 10"],"kind":"mlp"}},{"id":"action_head.model.mlp_resnet_blocks.9.out","label":"layer output","kind":"norm","group":["action_head","mlp_resnet_blocks 10"],"plots":["answer"],"width":896,"module":"","params":0,"note":"This layer does not simply add its parts' outputs to its input; this is what it hands on.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.9.out/holds/answer"},"holds":{"answer":"action_head.model.mlp_resnet_blocks.9.out/holds/answer"},"thin":true,"auto":{"label":"layer output","note":"This layer does not simply add its parts' outputs to its input; this is what it hands on.","group":["action_head","mlp_resnet_blocks 10"],"kind":"norm"}},{"id":"action_head.model.mlp_resnet_blocks.10.attn","label":"attention","kind":"attention","group":["action_head","mlp_resnet_blocks 11"],"plots":["answer"],"width":896,"module":"action_head.model.mlp_resnet_blocks.10.q_proj, action_head.model.mlp_resnet_blocks.10.k_self, action_head.model.mlp_resnet_blocks.10.v_self, action_head.model.mlp_resnet_blocks.10.k_adapter, action_head.model.mlp_resnet_blocks.10.v_adapter, action_head.model.mlp_resnet_blocks.10.k_task, action_head.model.mlp_resnet_blocks.10.v_task, action_head.model.mlp_resnet_blocks.10.rope, action_head.model.mlp_resnet_blocks.10.o_proj","params":6429696,"note":"Block i reads language-model layer i+1. Each step token reads three things at once: the other step tokens; the action queries' hidden states plus the state token; and the image tokens' hidden states, through a learned gate. The split between \"image tokens\" and the rest is fixed at 512 tokens, the paper's two cameras: here the iso camera's 256 tokens land with the action queries.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.10.attn/wrote/answer"},"holds":{},"auto":{"label":"attention","note":"Linear, Linear, Linear, Linear, Linear, Linear, Linear, RotaryPositionEmbedding, Linear, making queries, keys and values: attention written out in the model's own code. The only place tokens exchange information: each token reads from the others. 896 numbers in, 896 out.","group":["action_head","mlp_resnet_blocks 11"],"kind":"attention"}},{"id":"action_head.model.mlp_resnet_blocks.10.ffn","label":"MLP","kind":"mlp","group":["action_head","mlp_resnet_blocks 11"],"plots":["answer"],"width":896,"module":"action_head.model.mlp_resnet_blocks.10.ffn","params":805504,"note":"LayerNorm and an MLP, on (what was read + the token): the result replaces the token.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.10.ffn/wrote/answer"},"holds":{},"auto":{"label":"MLP","note":"Works on each token by itself: Sequential. Bend: ReLU.","group":["action_head","mlp_resnet_blocks 11"],"kind":"mlp"}},{"id":"action_head.model.mlp_resnet_blocks.10.out","label":"layer output","kind":"norm","group":["action_head","mlp_resnet_blocks 11"],"plots":["answer"],"width":896,"module":"","params":0,"note":"This layer does not simply add its parts' outputs to its input; this is what it hands on.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.10.out/holds/answer"},"holds":{"answer":"action_head.model.mlp_resnet_blocks.10.out/holds/answer"},"thin":true,"auto":{"label":"layer output","note":"This layer does not simply add its parts' outputs to its input; this is what it hands on.","group":["action_head","mlp_resnet_blocks 11"],"kind":"norm"}},{"id":"action_head.model.mlp_resnet_blocks.11.attn","label":"attention","kind":"attention","group":["action_head","mlp_resnet_blocks 12"],"plots":["answer"],"width":896,"module":"action_head.model.mlp_resnet_blocks.11.q_proj, action_head.model.mlp_resnet_blocks.11.k_self, action_head.model.mlp_resnet_blocks.11.v_self, action_head.model.mlp_resnet_blocks.11.k_adapter, action_head.model.mlp_resnet_blocks.11.v_adapter, action_head.model.mlp_resnet_blocks.11.k_task, action_head.model.mlp_resnet_blocks.11.v_task, action_head.model.mlp_resnet_blocks.11.rope, action_head.model.mlp_resnet_blocks.11.o_proj","params":6429696,"note":"Block i reads language-model layer i+1. Each step token reads three things at once: the other step tokens; the action queries' hidden states plus the state token; and the image tokens' hidden states, through a learned gate. The split between \"image tokens\" and the rest is fixed at 512 tokens, the paper's two cameras: here the iso camera's 256 tokens land with the action queries.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.11.attn/wrote/answer"},"holds":{},"auto":{"label":"attention","note":"Linear, Linear, Linear, Linear, Linear, Linear, Linear, RotaryPositionEmbedding, Linear, making queries, keys and values: attention written out in the model's own code. The only place tokens exchange information: each token reads from the others. 896 numbers in, 896 out.","group":["action_head","mlp_resnet_blocks 12"],"kind":"attention"}},{"id":"action_head.model.mlp_resnet_blocks.11.ffn","label":"MLP","kind":"mlp","group":["action_head","mlp_resnet_blocks 12"],"plots":["answer"],"width":896,"module":"action_head.model.mlp_resnet_blocks.11.ffn","params":805504,"note":"LayerNorm and an MLP, on (what was read + the token): the result replaces the token.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.11.ffn/wrote/answer"},"holds":{},"auto":{"label":"MLP","note":"Works on each token by itself: Sequential. Bend: ReLU.","group":["action_head","mlp_resnet_blocks 12"],"kind":"mlp"}},{"id":"action_head.model.mlp_resnet_blocks.11.out","label":"layer output","kind":"norm","group":["action_head","mlp_resnet_blocks 12"],"plots":["answer"],"width":896,"module":"","params":0,"note":"This layer does not simply add its parts' outputs to its input; this is what it hands on.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.11.out/holds/answer"},"holds":{"answer":"action_head.model.mlp_resnet_blocks.11.out/holds/answer"},"thin":true,"auto":{"label":"layer output","note":"This layer does not simply add its parts' outputs to its input; this is what it hands on.","group":["action_head","mlp_resnet_blocks 12"],"kind":"norm"}},{"id":"action_head.model.mlp_resnet_blocks.12.attn","label":"attention","kind":"attention","group":["action_head","mlp_resnet_blocks 13"],"plots":["answer"],"width":896,"module":"action_head.model.mlp_resnet_blocks.12.q_proj, action_head.model.mlp_resnet_blocks.12.k_self, action_head.model.mlp_resnet_blocks.12.v_self, action_head.model.mlp_resnet_blocks.12.k_adapter, action_head.model.mlp_resnet_blocks.12.v_adapter, action_head.model.mlp_resnet_blocks.12.k_task, action_head.model.mlp_resnet_blocks.12.v_task, action_head.model.mlp_resnet_blocks.12.rope, action_head.model.mlp_resnet_blocks.12.o_proj","params":6429696,"note":"Block i reads language-model layer i+1. Each step token reads three things at once: the other step tokens; the action queries' hidden states plus the state token; and the image tokens' hidden states, through a learned gate. The split between \"image tokens\" and the rest is fixed at 512 tokens, the paper's two cameras: here the iso camera's 256 tokens land with the action queries.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.12.attn/wrote/answer"},"holds":{},"auto":{"label":"attention","note":"Linear, Linear, Linear, Linear, Linear, Linear, Linear, RotaryPositionEmbedding, Linear, making queries, keys and values: attention written out in the model's own code. The only place tokens exchange information: each token reads from the others. 896 numbers in, 896 out.","group":["action_head","mlp_resnet_blocks 13"],"kind":"attention"}},{"id":"action_head.model.mlp_resnet_blocks.12.ffn","label":"MLP","kind":"mlp","group":["action_head","mlp_resnet_blocks 13"],"plots":["answer"],"width":896,"module":"action_head.model.mlp_resnet_blocks.12.ffn","params":805504,"note":"LayerNorm and an MLP, on (what was read + the token): the result replaces the token.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.12.ffn/wrote/answer"},"holds":{},"auto":{"label":"MLP","note":"Works on each token by itself: Sequential. Bend: ReLU.","group":["action_head","mlp_resnet_blocks 13"],"kind":"mlp"}},{"id":"action_head.model.mlp_resnet_blocks.12.out","label":"layer output","kind":"norm","group":["action_head","mlp_resnet_blocks 13"],"plots":["answer"],"width":896,"module":"","params":0,"note":"This layer does not simply add its parts' outputs to its input; this is what it hands on.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.12.out/holds/answer"},"holds":{"answer":"action_head.model.mlp_resnet_blocks.12.out/holds/answer"},"thin":true,"auto":{"label":"layer output","note":"This layer does not simply add its parts' outputs to its input; this is what it hands on.","group":["action_head","mlp_resnet_blocks 13"],"kind":"norm"}},{"id":"action_head.model.mlp_resnet_blocks.13.attn","label":"attention","kind":"attention","group":["action_head","mlp_resnet_blocks 14"],"plots":["answer"],"width":896,"module":"action_head.model.mlp_resnet_blocks.13.q_proj, action_head.model.mlp_resnet_blocks.13.k_self, action_head.model.mlp_resnet_blocks.13.v_self, action_head.model.mlp_resnet_blocks.13.k_adapter, action_head.model.mlp_resnet_blocks.13.v_adapter, action_head.model.mlp_resnet_blocks.13.k_task, action_head.model.mlp_resnet_blocks.13.v_task, action_head.model.mlp_resnet_blocks.13.rope, action_head.model.mlp_resnet_blocks.13.o_proj","params":6429696,"note":"Block i reads language-model layer i+1. Each step token reads three things at once: the other step tokens; the action queries' hidden states plus the state token; and the image tokens' hidden states, through a learned gate. The split between \"image tokens\" and the rest is fixed at 512 tokens, the paper's two cameras: here the iso camera's 256 tokens land with the action queries.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.13.attn/wrote/answer"},"holds":{},"auto":{"label":"attention","note":"Linear, Linear, Linear, Linear, Linear, Linear, Linear, RotaryPositionEmbedding, Linear, making queries, keys and values: attention written out in the model's own code. The only place tokens exchange information: each token reads from the others. 896 numbers in, 896 out.","group":["action_head","mlp_resnet_blocks 14"],"kind":"attention"}},{"id":"action_head.model.mlp_resnet_blocks.13.ffn","label":"MLP","kind":"mlp","group":["action_head","mlp_resnet_blocks 14"],"plots":["answer"],"width":896,"module":"action_head.model.mlp_resnet_blocks.13.ffn","params":805504,"note":"LayerNorm and an MLP, on (what was read + the token): the result replaces the token.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.13.ffn/wrote/answer"},"holds":{},"auto":{"label":"MLP","note":"Works on each token by itself: Sequential. Bend: ReLU.","group":["action_head","mlp_resnet_blocks 14"],"kind":"mlp"}},{"id":"action_head.model.mlp_resnet_blocks.13.out","label":"layer output","kind":"norm","group":["action_head","mlp_resnet_blocks 14"],"plots":["answer"],"width":896,"module":"","params":0,"note":"This layer does not simply add its parts' outputs to its input; this is what it hands on.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.13.out/holds/answer"},"holds":{"answer":"action_head.model.mlp_resnet_blocks.13.out/holds/answer"},"thin":true,"auto":{"label":"layer output","note":"This layer does not simply add its parts' outputs to its input; this is what it hands on.","group":["action_head","mlp_resnet_blocks 14"],"kind":"norm"}},{"id":"action_head.model.mlp_resnet_blocks.14.attn","label":"attention","kind":"attention","group":["action_head","mlp_resnet_blocks 15"],"plots":["answer"],"width":896,"module":"action_head.model.mlp_resnet_blocks.14.q_proj, action_head.model.mlp_resnet_blocks.14.k_self, action_head.model.mlp_resnet_blocks.14.v_self, action_head.model.mlp_resnet_blocks.14.k_adapter, action_head.model.mlp_resnet_blocks.14.v_adapter, action_head.model.mlp_resnet_blocks.14.k_task, action_head.model.mlp_resnet_blocks.14.v_task, action_head.model.mlp_resnet_blocks.14.rope, action_head.model.mlp_resnet_blocks.14.o_proj","params":6429696,"note":"Block i reads language-model layer i+1. Each step token reads three things at once: the other step tokens; the action queries' hidden states plus the state token; and the image tokens' hidden states, through a learned gate. The split between \"image tokens\" and the rest is fixed at 512 tokens, the paper's two cameras: here the iso camera's 256 tokens land with the action queries.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.14.attn/wrote/answer"},"holds":{},"auto":{"label":"attention","note":"Linear, Linear, Linear, Linear, Linear, Linear, Linear, RotaryPositionEmbedding, Linear, making queries, keys and values: attention written out in the model's own code. The only place tokens exchange information: each token reads from the others. 896 numbers in, 896 out.","group":["action_head","mlp_resnet_blocks 15"],"kind":"attention"}},{"id":"action_head.model.mlp_resnet_blocks.14.ffn","label":"MLP","kind":"mlp","group":["action_head","mlp_resnet_blocks 15"],"plots":["answer"],"width":896,"module":"action_head.model.mlp_resnet_blocks.14.ffn","params":805504,"note":"LayerNorm and an MLP, on (what was read + the token): the result replaces the token.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.14.ffn/wrote/answer"},"holds":{},"auto":{"label":"MLP","note":"Works on each token by itself: Sequential. Bend: ReLU.","group":["action_head","mlp_resnet_blocks 15"],"kind":"mlp"}},{"id":"action_head.model.mlp_resnet_blocks.14.out","label":"layer output","kind":"norm","group":["action_head","mlp_resnet_blocks 15"],"plots":["answer"],"width":896,"module":"","params":0,"note":"This layer does not simply add its parts' outputs to its input; this is what it hands on.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.14.out/holds/answer"},"holds":{"answer":"action_head.model.mlp_resnet_blocks.14.out/holds/answer"},"thin":true,"auto":{"label":"layer output","note":"This layer does not simply add its parts' outputs to its input; this is what it hands on.","group":["action_head","mlp_resnet_blocks 15"],"kind":"norm"}},{"id":"action_head.model.mlp_resnet_blocks.15.attn","label":"attention","kind":"attention","group":["action_head","mlp_resnet_blocks 16"],"plots":["answer"],"width":896,"module":"action_head.model.mlp_resnet_blocks.15.q_proj, action_head.model.mlp_resnet_blocks.15.k_self, action_head.model.mlp_resnet_blocks.15.v_self, action_head.model.mlp_resnet_blocks.15.k_adapter, action_head.model.mlp_resnet_blocks.15.v_adapter, action_head.model.mlp_resnet_blocks.15.k_task, action_head.model.mlp_resnet_blocks.15.v_task, action_head.model.mlp_resnet_blocks.15.rope, action_head.model.mlp_resnet_blocks.15.o_proj","params":6429696,"note":"Block i reads language-model layer i+1. Each step token reads three things at once: the other step tokens; the action queries' hidden states plus the state token; and the image tokens' hidden states, through a learned gate. The split between \"image tokens\" and the rest is fixed at 512 tokens, the paper's two cameras: here the iso camera's 256 tokens land with the action queries.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.15.attn/wrote/answer"},"holds":{},"auto":{"label":"attention","note":"Linear, Linear, Linear, Linear, Linear, Linear, Linear, RotaryPositionEmbedding, Linear, making queries, keys and values: attention written out in the model's own code. The only place tokens exchange information: each token reads from the others. 896 numbers in, 896 out.","group":["action_head","mlp_resnet_blocks 16"],"kind":"attention"}},{"id":"action_head.model.mlp_resnet_blocks.15.ffn","label":"MLP","kind":"mlp","group":["action_head","mlp_resnet_blocks 16"],"plots":["answer"],"width":896,"module":"action_head.model.mlp_resnet_blocks.15.ffn","params":805504,"note":"LayerNorm and an MLP, on (what was read + the token): the result replaces the token.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.15.ffn/wrote/answer"},"holds":{},"auto":{"label":"MLP","note":"Works on each token by itself: Sequential. Bend: ReLU.","group":["action_head","mlp_resnet_blocks 16"],"kind":"mlp"}},{"id":"action_head.model.mlp_resnet_blocks.15.out","label":"layer output","kind":"norm","group":["action_head","mlp_resnet_blocks 16"],"plots":["answer"],"width":896,"module":"","params":0,"note":"This layer does not simply add its parts' outputs to its input; this is what it hands on.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.15.out/holds/answer"},"holds":{"answer":"action_head.model.mlp_resnet_blocks.15.out/holds/answer"},"thin":true,"auto":{"label":"layer output","note":"This layer does not simply add its parts' outputs to its input; this is what it hands on.","group":["action_head","mlp_resnet_blocks 16"],"kind":"norm"}},{"id":"action_head.model.mlp_resnet_blocks.16.attn","label":"attention","kind":"attention","group":["action_head","mlp_resnet_blocks 17"],"plots":["answer"],"width":896,"module":"action_head.model.mlp_resnet_blocks.16.q_proj, action_head.model.mlp_resnet_blocks.16.k_self, action_head.model.mlp_resnet_blocks.16.v_self, action_head.model.mlp_resnet_blocks.16.k_adapter, action_head.model.mlp_resnet_blocks.16.v_adapter, action_head.model.mlp_resnet_blocks.16.k_task, action_head.model.mlp_resnet_blocks.16.v_task, action_head.model.mlp_resnet_blocks.16.rope, action_head.model.mlp_resnet_blocks.16.o_proj","params":6429696,"note":"Block i reads language-model layer i+1. Each step token reads three things at once: the other step tokens; the action queries' hidden states plus the state token; and the image tokens' hidden states, through a learned gate. The split between \"image tokens\" and the rest is fixed at 512 tokens, the paper's two cameras: here the iso camera's 256 tokens land with the action queries.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.16.attn/wrote/answer"},"holds":{},"auto":{"label":"attention","note":"Linear, Linear, Linear, Linear, Linear, Linear, Linear, RotaryPositionEmbedding, Linear, making queries, keys and values: attention written out in the model's own code. The only place tokens exchange information: each token reads from the others. 896 numbers in, 896 out.","group":["action_head","mlp_resnet_blocks 17"],"kind":"attention"}},{"id":"action_head.model.mlp_resnet_blocks.16.ffn","label":"MLP","kind":"mlp","group":["action_head","mlp_resnet_blocks 17"],"plots":["answer"],"width":896,"module":"action_head.model.mlp_resnet_blocks.16.ffn","params":805504,"note":"LayerNorm and an MLP, on (what was read + the token): the result replaces the token.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.16.ffn/wrote/answer"},"holds":{},"auto":{"label":"MLP","note":"Works on each token by itself: Sequential. Bend: ReLU.","group":["action_head","mlp_resnet_blocks 17"],"kind":"mlp"}},{"id":"action_head.model.mlp_resnet_blocks.16.out","label":"layer output","kind":"norm","group":["action_head","mlp_resnet_blocks 17"],"plots":["answer"],"width":896,"module":"","params":0,"note":"This layer does not simply add its parts' outputs to its input; this is what it hands on.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.16.out/holds/answer"},"holds":{"answer":"action_head.model.mlp_resnet_blocks.16.out/holds/answer"},"thin":true,"auto":{"label":"layer output","note":"This layer does not simply add its parts' outputs to its input; this is what it hands on.","group":["action_head","mlp_resnet_blocks 17"],"kind":"norm"}},{"id":"action_head.model.mlp_resnet_blocks.17.attn","label":"attention","kind":"attention","group":["action_head","mlp_resnet_blocks 18"],"plots":["answer"],"width":896,"module":"action_head.model.mlp_resnet_blocks.17.q_proj, action_head.model.mlp_resnet_blocks.17.k_self, action_head.model.mlp_resnet_blocks.17.v_self, action_head.model.mlp_resnet_blocks.17.k_adapter, action_head.model.mlp_resnet_blocks.17.v_adapter, action_head.model.mlp_resnet_blocks.17.k_task, action_head.model.mlp_resnet_blocks.17.v_task, action_head.model.mlp_resnet_blocks.17.rope, action_head.model.mlp_resnet_blocks.17.o_proj","params":6429696,"note":"Block i reads language-model layer i+1. Each step token reads three things at once: the other step tokens; the action queries' hidden states plus the state token; and the image tokens' hidden states, through a learned gate. The split between \"image tokens\" and the rest is fixed at 512 tokens, the paper's two cameras: here the iso camera's 256 tokens land with the action queries.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.17.attn/wrote/answer"},"holds":{},"auto":{"label":"attention","note":"Linear, Linear, Linear, Linear, Linear, Linear, Linear, RotaryPositionEmbedding, Linear, making queries, keys and values: attention written out in the model's own code. The only place tokens exchange information: each token reads from the others. 896 numbers in, 896 out.","group":["action_head","mlp_resnet_blocks 18"],"kind":"attention"}},{"id":"action_head.model.mlp_resnet_blocks.17.ffn","label":"MLP","kind":"mlp","group":["action_head","mlp_resnet_blocks 18"],"plots":["answer"],"width":896,"module":"action_head.model.mlp_resnet_blocks.17.ffn","params":805504,"note":"LayerNorm and an MLP, on (what was read + the token): the result replaces the token.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.17.ffn/wrote/answer"},"holds":{},"auto":{"label":"MLP","note":"Works on each token by itself: Sequential. Bend: ReLU.","group":["action_head","mlp_resnet_blocks 18"],"kind":"mlp"}},{"id":"action_head.model.mlp_resnet_blocks.17.out","label":"layer output","kind":"norm","group":["action_head","mlp_resnet_blocks 18"],"plots":["answer"],"width":896,"module":"","params":0,"note":"This layer does not simply add its parts' outputs to its input; this is what it hands on.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.17.out/holds/answer"},"holds":{"answer":"action_head.model.mlp_resnet_blocks.17.out/holds/answer"},"thin":true,"auto":{"label":"layer output","note":"This layer does not simply add its parts' outputs to its input; this is what it hands on.","group":["action_head","mlp_resnet_blocks 18"],"kind":"norm"}},{"id":"action_head.model.mlp_resnet_blocks.18.attn","label":"attention","kind":"attention","group":["action_head","mlp_resnet_blocks 19"],"plots":["answer"],"width":896,"module":"action_head.model.mlp_resnet_blocks.18.q_proj, action_head.model.mlp_resnet_blocks.18.k_self, action_head.model.mlp_resnet_blocks.18.v_self, action_head.model.mlp_resnet_blocks.18.k_adapter, action_head.model.mlp_resnet_blocks.18.v_adapter, action_head.model.mlp_resnet_blocks.18.k_task, action_head.model.mlp_resnet_blocks.18.v_task, action_head.model.mlp_resnet_blocks.18.rope, action_head.model.mlp_resnet_blocks.18.o_proj","params":6429696,"note":"Block i reads language-model layer i+1. Each step token reads three things at once: the other step tokens; the action queries' hidden states plus the state token; and the image tokens' hidden states, through a learned gate. The split between \"image tokens\" and the rest is fixed at 512 tokens, the paper's two cameras: here the iso camera's 256 tokens land with the action queries.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.18.attn/wrote/answer"},"holds":{},"auto":{"label":"attention","note":"Linear, Linear, Linear, Linear, Linear, Linear, Linear, RotaryPositionEmbedding, Linear, making queries, keys and values: attention written out in the model's own code. The only place tokens exchange information: each token reads from the others. 896 numbers in, 896 out.","group":["action_head","mlp_resnet_blocks 19"],"kind":"attention"}},{"id":"action_head.model.mlp_resnet_blocks.18.ffn","label":"MLP","kind":"mlp","group":["action_head","mlp_resnet_blocks 19"],"plots":["answer"],"width":896,"module":"action_head.model.mlp_resnet_blocks.18.ffn","params":805504,"note":"LayerNorm and an MLP, on (what was read + the token): the result replaces the token.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.18.ffn/wrote/answer"},"holds":{},"auto":{"label":"MLP","note":"Works on each token by itself: Sequential. Bend: ReLU.","group":["action_head","mlp_resnet_blocks 19"],"kind":"mlp"}},{"id":"action_head.model.mlp_resnet_blocks.18.out","label":"layer output","kind":"norm","group":["action_head","mlp_resnet_blocks 19"],"plots":["answer"],"width":896,"module":"","params":0,"note":"This layer does not simply add its parts' outputs to its input; this is what it hands on.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.18.out/holds/answer"},"holds":{"answer":"action_head.model.mlp_resnet_blocks.18.out/holds/answer"},"thin":true,"auto":{"label":"layer output","note":"This layer does not simply add its parts' outputs to its input; this is what it hands on.","group":["action_head","mlp_resnet_blocks 19"],"kind":"norm"}},{"id":"action_head.model.mlp_resnet_blocks.19.attn","label":"attention","kind":"attention","group":["action_head","mlp_resnet_blocks 20"],"plots":["answer"],"width":896,"module":"action_head.model.mlp_resnet_blocks.19.q_proj, action_head.model.mlp_resnet_blocks.19.k_self, action_head.model.mlp_resnet_blocks.19.v_self, action_head.model.mlp_resnet_blocks.19.k_adapter, action_head.model.mlp_resnet_blocks.19.v_adapter, action_head.model.mlp_resnet_blocks.19.k_task, action_head.model.mlp_resnet_blocks.19.v_task, action_head.model.mlp_resnet_blocks.19.rope, action_head.model.mlp_resnet_blocks.19.o_proj","params":6429696,"note":"Block i reads language-model layer i+1. Each step token reads three things at once: the other step tokens; the action queries' hidden states plus the state token; and the image tokens' hidden states, through a learned gate. The split between \"image tokens\" and the rest is fixed at 512 tokens, the paper's two cameras: here the iso camera's 256 tokens land with the action queries.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.19.attn/wrote/answer"},"holds":{},"auto":{"label":"attention","note":"Linear, Linear, Linear, Linear, Linear, Linear, Linear, RotaryPositionEmbedding, Linear, making queries, keys and values: attention written out in the model's own code. The only place tokens exchange information: each token reads from the others. 896 numbers in, 896 out.","group":["action_head","mlp_resnet_blocks 20"],"kind":"attention"}},{"id":"action_head.model.mlp_resnet_blocks.19.ffn","label":"MLP","kind":"mlp","group":["action_head","mlp_resnet_blocks 20"],"plots":["answer"],"width":896,"module":"action_head.model.mlp_resnet_blocks.19.ffn","params":805504,"note":"LayerNorm and an MLP, on (what was read + the token): the result replaces the token.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.19.ffn/wrote/answer"},"holds":{},"auto":{"label":"MLP","note":"Works on each token by itself: Sequential. Bend: ReLU.","group":["action_head","mlp_resnet_blocks 20"],"kind":"mlp"}},{"id":"action_head.model.mlp_resnet_blocks.19.out","label":"layer output","kind":"norm","group":["action_head","mlp_resnet_blocks 20"],"plots":["answer"],"width":896,"module":"","params":0,"note":"This layer does not simply add its parts' outputs to its input; this is what it hands on.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.19.out/holds/answer"},"holds":{"answer":"action_head.model.mlp_resnet_blocks.19.out/holds/answer"},"thin":true,"auto":{"label":"layer output","note":"This layer does not simply add its parts' outputs to its input; this is what it hands on.","group":["action_head","mlp_resnet_blocks 20"],"kind":"norm"}},{"id":"action_head.model.mlp_resnet_blocks.20.attn","label":"attention","kind":"attention","group":["action_head","mlp_resnet_blocks 21"],"plots":["answer"],"width":896,"module":"action_head.model.mlp_resnet_blocks.20.q_proj, action_head.model.mlp_resnet_blocks.20.k_self, action_head.model.mlp_resnet_blocks.20.v_self, action_head.model.mlp_resnet_blocks.20.k_adapter, action_head.model.mlp_resnet_blocks.20.v_adapter, action_head.model.mlp_resnet_blocks.20.k_task, action_head.model.mlp_resnet_blocks.20.v_task, action_head.model.mlp_resnet_blocks.20.rope, action_head.model.mlp_resnet_blocks.20.o_proj","params":6429696,"note":"Block i reads language-model layer i+1. Each step token reads three things at once: the other step tokens; the action queries' hidden states plus the state token; and the image tokens' hidden states, through a learned gate. The split between \"image tokens\" and the rest is fixed at 512 tokens, the paper's two cameras: here the iso camera's 256 tokens land with the action queries.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.20.attn/wrote/answer"},"holds":{},"auto":{"label":"attention","note":"Linear, Linear, Linear, Linear, Linear, Linear, Linear, RotaryPositionEmbedding, Linear, making queries, keys and values: attention written out in the model's own code. The only place tokens exchange information: each token reads from the others. 896 numbers in, 896 out.","group":["action_head","mlp_resnet_blocks 21"],"kind":"attention"}},{"id":"action_head.model.mlp_resnet_blocks.20.ffn","label":"MLP","kind":"mlp","group":["action_head","mlp_resnet_blocks 21"],"plots":["answer"],"width":896,"module":"action_head.model.mlp_resnet_blocks.20.ffn","params":805504,"note":"LayerNorm and an MLP, on (what was read + the token): the result replaces the token.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.20.ffn/wrote/answer"},"holds":{},"auto":{"label":"MLP","note":"Works on each token by itself: Sequential. Bend: ReLU.","group":["action_head","mlp_resnet_blocks 21"],"kind":"mlp"}},{"id":"action_head.model.mlp_resnet_blocks.20.out","label":"layer output","kind":"norm","group":["action_head","mlp_resnet_blocks 21"],"plots":["answer"],"width":896,"module":"","params":0,"note":"This layer does not simply add its parts' outputs to its input; this is what it hands on.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.20.out/holds/answer"},"holds":{"answer":"action_head.model.mlp_resnet_blocks.20.out/holds/answer"},"thin":true,"auto":{"label":"layer output","note":"This layer does not simply add its parts' outputs to its input; this is what it hands on.","group":["action_head","mlp_resnet_blocks 21"],"kind":"norm"}},{"id":"action_head.model.mlp_resnet_blocks.21.attn","label":"attention","kind":"attention","group":["action_head","mlp_resnet_blocks 22"],"plots":["answer"],"width":896,"module":"action_head.model.mlp_resnet_blocks.21.q_proj, action_head.model.mlp_resnet_blocks.21.k_self, action_head.model.mlp_resnet_blocks.21.v_self, action_head.model.mlp_resnet_blocks.21.k_adapter, action_head.model.mlp_resnet_blocks.21.v_adapter, action_head.model.mlp_resnet_blocks.21.k_task, action_head.model.mlp_resnet_blocks.21.v_task, action_head.model.mlp_resnet_blocks.21.rope, action_head.model.mlp_resnet_blocks.21.o_proj","params":6429696,"note":"Block i reads language-model layer i+1. Each step token reads three things at once: the other step tokens; the action queries' hidden states plus the state token; and the image tokens' hidden states, through a learned gate. The split between \"image tokens\" and the rest is fixed at 512 tokens, the paper's two cameras: here the iso camera's 256 tokens land with the action queries.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.21.attn/wrote/answer"},"holds":{},"auto":{"label":"attention","note":"Linear, Linear, Linear, Linear, Linear, Linear, Linear, RotaryPositionEmbedding, Linear, making queries, keys and values: attention written out in the model's own code. The only place tokens exchange information: each token reads from the others. 896 numbers in, 896 out.","group":["action_head","mlp_resnet_blocks 22"],"kind":"attention"}},{"id":"action_head.model.mlp_resnet_blocks.21.ffn","label":"MLP","kind":"mlp","group":["action_head","mlp_resnet_blocks 22"],"plots":["answer"],"width":896,"module":"action_head.model.mlp_resnet_blocks.21.ffn","params":805504,"note":"LayerNorm and an MLP, on (what was read + the token): the result replaces the token.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.21.ffn/wrote/answer"},"holds":{},"auto":{"label":"MLP","note":"Works on each token by itself: Sequential. Bend: ReLU.","group":["action_head","mlp_resnet_blocks 22"],"kind":"mlp"}},{"id":"action_head.model.mlp_resnet_blocks.21.out","label":"layer output","kind":"norm","group":["action_head","mlp_resnet_blocks 22"],"plots":["answer"],"width":896,"module":"","params":0,"note":"This layer does not simply add its parts' outputs to its input; this is what it hands on.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.21.out/holds/answer"},"holds":{"answer":"action_head.model.mlp_resnet_blocks.21.out/holds/answer"},"thin":true,"auto":{"label":"layer output","note":"This layer does not simply add its parts' outputs to its input; this is what it hands on.","group":["action_head","mlp_resnet_blocks 22"],"kind":"norm"}},{"id":"action_head.model.mlp_resnet_blocks.22.attn","label":"attention","kind":"attention","group":["action_head","mlp_resnet_blocks 23"],"plots":["answer"],"width":896,"module":"action_head.model.mlp_resnet_blocks.22.q_proj, action_head.model.mlp_resnet_blocks.22.k_self, action_head.model.mlp_resnet_blocks.22.v_self, action_head.model.mlp_resnet_blocks.22.k_adapter, action_head.model.mlp_resnet_blocks.22.v_adapter, action_head.model.mlp_resnet_blocks.22.k_task, action_head.model.mlp_resnet_blocks.22.v_task, action_head.model.mlp_resnet_blocks.22.rope, action_head.model.mlp_resnet_blocks.22.o_proj","params":6429696,"note":"Block i reads language-model layer i+1. Each step token reads three things at once: the other step tokens; the action queries' hidden states plus the state token; and the image tokens' hidden states, through a learned gate. The split between \"image tokens\" and the rest is fixed at 512 tokens, the paper's two cameras: here the iso camera's 256 tokens land with the action queries.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.22.attn/wrote/answer"},"holds":{},"auto":{"label":"attention","note":"Linear, Linear, Linear, Linear, Linear, Linear, Linear, RotaryPositionEmbedding, Linear, making queries, keys and values: attention written out in the model's own code. The only place tokens exchange information: each token reads from the others. 896 numbers in, 896 out.","group":["action_head","mlp_resnet_blocks 23"],"kind":"attention"}},{"id":"action_head.model.mlp_resnet_blocks.22.ffn","label":"MLP","kind":"mlp","group":["action_head","mlp_resnet_blocks 23"],"plots":["answer"],"width":896,"module":"action_head.model.mlp_resnet_blocks.22.ffn","params":805504,"note":"LayerNorm and an MLP, on (what was read + the token): the result replaces the token.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.22.ffn/wrote/answer"},"holds":{},"auto":{"label":"MLP","note":"Works on each token by itself: Sequential. Bend: ReLU.","group":["action_head","mlp_resnet_blocks 23"],"kind":"mlp"}},{"id":"action_head.model.mlp_resnet_blocks.22.out","label":"layer output","kind":"norm","group":["action_head","mlp_resnet_blocks 23"],"plots":["answer"],"width":896,"module":"","params":0,"note":"This layer does not simply add its parts' outputs to its input; this is what it hands on.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.22.out/holds/answer"},"holds":{"answer":"action_head.model.mlp_resnet_blocks.22.out/holds/answer"},"thin":true,"auto":{"label":"layer output","note":"This layer does not simply add its parts' outputs to its input; this is what it hands on.","group":["action_head","mlp_resnet_blocks 23"],"kind":"norm"}},{"id":"action_head.model.mlp_resnet_blocks.23.attn","label":"attention","kind":"attention","group":["action_head","mlp_resnet_blocks 24"],"plots":["answer"],"width":896,"module":"action_head.model.mlp_resnet_blocks.23.q_proj, action_head.model.mlp_resnet_blocks.23.k_self, action_head.model.mlp_resnet_blocks.23.v_self, action_head.model.mlp_resnet_blocks.23.k_adapter, action_head.model.mlp_resnet_blocks.23.v_adapter, action_head.model.mlp_resnet_blocks.23.k_task, action_head.model.mlp_resnet_blocks.23.v_task, action_head.model.mlp_resnet_blocks.23.rope, action_head.model.mlp_resnet_blocks.23.o_proj","params":6429696,"note":"Block i reads language-model layer i+1. Each step token reads three things at once: the other step tokens; the action queries' hidden states plus the state token; and the image tokens' hidden states, through a learned gate. The split between \"image tokens\" and the rest is fixed at 512 tokens, the paper's two cameras: here the iso camera's 256 tokens land with the action queries.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.23.attn/wrote/answer"},"holds":{},"auto":{"label":"attention","note":"Linear, Linear, Linear, Linear, Linear, Linear, Linear, RotaryPositionEmbedding, Linear, making queries, keys and values: attention written out in the model's own code. The only place tokens exchange information: each token reads from the others. 896 numbers in, 896 out.","group":["action_head","mlp_resnet_blocks 24"],"kind":"attention"}},{"id":"action_head.model.mlp_resnet_blocks.23.ffn","label":"MLP","kind":"mlp","group":["action_head","mlp_resnet_blocks 24"],"plots":["answer"],"width":896,"module":"action_head.model.mlp_resnet_blocks.23.ffn","params":805504,"note":"LayerNorm and an MLP, on (what was read + the token): the result replaces the token.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.23.ffn/wrote/answer"},"holds":{},"auto":{"label":"MLP","note":"Works on each token by itself: Sequential. Bend: ReLU.","group":["action_head","mlp_resnet_blocks 24"],"kind":"mlp"}},{"id":"action_head.model.mlp_resnet_blocks.23.out","label":"layer output","kind":"norm","group":["action_head","mlp_resnet_blocks 24"],"plots":["answer"],"width":896,"module":"","params":0,"note":"This layer does not simply add its parts' outputs to its input; this is what it hands on.","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.mlp_resnet_blocks.23.out/holds/answer"},"holds":{"answer":"action_head.model.mlp_resnet_blocks.23.out/holds/answer"},"thin":true,"auto":{"label":"layer output","note":"This layer does not simply add its parts' outputs to its input; this is what it hands on.","group":["action_head","mlp_resnet_blocks 24"],"kind":"norm"}},{"id":"action_head.model.layer_norm2","label":"action_head.model.layer_norm2: LayerNorm","kind":"norm","group":["action_head"],"plots":["answer"],"width":896,"module":"action_head.model.layer_norm2","params":1792,"note":"LayerNorm. Out: [1, 8, 896].","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.layer_norm2/wrote/answer"},"holds":{},"auto":{"label":"action_head.model.layer_norm2: LayerNorm","note":"LayerNorm. Out: [1, 8, 896].","group":["action_head"],"kind":"norm"}},{"id":"action_head.model.fc2","label":"7 numbers per step","kind":"linear","group":["action_head"],"plots":["answer"],"width":7,"module":"action_head.model.fc2","params":6279,"note":"The change in the 6 joints and the gripper for each of the 8 steps, scaled and added up by predict().","cells":{"answer":[1,8]},"wrote":{"answer":"action_head.model.fc2/wrote/answer"},"holds":{},"auto":{"label":"action_head.model.fc2: Linear","note":"Linear. Out: [1, 8, 7]. 896 numbers in, 7 out.","group":["action_head"],"kind":"linear"}},{"id":"poses","label":"the next 8 poses","kind":"output","group":["output"],"plots":["answer"],"width":7,"module":"","params":0,"note":"What predict() returns: 8 steps of 6 joint angles and the gripper opening.","cells":{"answer":[1,8]},"wrote":{"answer":"poses/wrote/answer"},"holds":{},"auto":{"label":"the next 8 poses","note":"What predict() returns: 8 steps of 6 joint angles and the gripper opening.","group":["output"],"kind":"output"}}],"plates":[{"id":"vla.base_model.model.language_model.model.layers.0.plate","stage":"vla.base_model.model.language_model.model.layers.0.self_attn","heads":14,"key":"plate/0","allowed":"plate/allowed"},{"id":"vla.base_model.model.language_model.model.layers.1.plate","stage":"vla.base_model.model.language_model.model.layers.1.self_attn","heads":14,"key":"plate/1","allowed":"plate/allowed"},{"id":"vla.base_model.model.language_model.model.layers.2.plate","stage":"vla.base_model.model.language_model.model.layers.2.self_attn","heads":14,"key":"plate/2","allowed":"plate/allowed"},{"id":"vla.base_model.model.language_model.model.layers.3.plate","stage":"vla.base_model.model.language_model.model.layers.3.self_attn","heads":14,"key":"plate/3","allowed":"plate/allowed"},{"id":"vla.base_model.model.language_model.model.layers.4.plate","stage":"vla.base_model.model.language_model.model.layers.4.self_attn","heads":14,"key":"plate/4","allowed":"plate/allowed"},{"id":"vla.base_model.model.language_model.model.layers.5.plate","stage":"vla.base_model.model.language_model.model.layers.5.self_attn","heads":14,"key":"plate/5","allowed":"plate/allowed"},{"id":"vla.base_model.model.language_model.model.layers.6.plate","stage":"vla.base_model.model.language_model.model.layers.6.self_attn","heads":14,"key":"plate/6","allowed":"plate/allowed"},{"id":"vla.base_model.model.language_model.model.layers.7.plate","stage":"vla.base_model.model.language_model.model.layers.7.self_attn","heads":14,"key":"plate/7","allowed":"plate/allowed"},{"id":"vla.base_model.model.language_model.model.layers.8.plate","stage":"vla.base_model.model.language_model.model.layers.8.self_attn","heads":14,"key":"plate/8","allowed":"plate/allowed"},{"id":"vla.base_model.model.language_model.model.layers.9.plate","stage":"vla.base_model.model.language_model.model.layers.9.self_attn","heads":14,"key":"plate/9","allowed":"plate/allowed"},{"id":"vla.base_model.model.language_model.model.layers.10.plate","stage":"vla.base_model.model.language_model.model.layers.10.self_attn","heads":14,"key":"plate/10","allowed":"plate/allowed"},{"id":"vla.base_model.model.language_model.model.layers.11.plate","stage":"vla.base_model.model.language_model.model.layers.11.self_attn","heads":14,"key":"plate/11","allowed":"plate/allowed"},{"id":"vla.base_model.model.language_model.model.layers.12.plate","stage":"vla.base_model.model.language_model.model.layers.12.self_attn","heads":14,"key":"plate/12","allowed":"plate/allowed"},{"id":"vla.base_model.model.language_model.model.layers.13.plate","stage":"vla.base_model.model.language_model.model.layers.13.self_attn","heads":14,"key":"plate/13","allowed":"plate/allowed"},{"id":"vla.base_model.model.language_model.model.layers.14.plate","stage":"vla.base_model.model.language_model.model.layers.14.self_attn","heads":14,"key":"plate/14","allowed":"plate/allowed"},{"id":"vla.base_model.model.language_model.model.layers.15.plate","stage":"vla.base_model.model.language_model.model.layers.15.self_attn","heads":14,"key":"plate/15","allowed":"plate/allowed"},{"id":"vla.base_model.model.language_model.model.layers.16.plate","stage":"vla.base_model.model.language_model.model.layers.16.self_attn","heads":14,"key":"plate/16","allowed":"plate/allowed"},{"id":"vla.base_model.model.language_model.model.layers.17.plate","stage":"vla.base_model.model.language_model.model.layers.17.self_attn","heads":14,"key":"plate/17","allowed":"plate/allowed"},{"id":"vla.base_model.model.language_model.model.layers.18.plate","stage":"vla.base_model.model.language_model.model.layers.18.self_attn","heads":14,"key":"plate/18","allowed":"plate/allowed"},{"id":"vla.base_model.model.language_model.model.layers.19.plate","stage":"vla.base_model.model.language_model.model.layers.19.self_attn","heads":14,"key":"plate/19","allowed":"plate/allowed"},{"id":"vla.base_model.model.language_model.model.layers.20.plate","stage":"vla.base_model.model.language_model.model.layers.20.self_attn","heads":14,"key":"plate/20","allowed":"plate/allowed"},{"id":"vla.base_model.model.language_model.model.layers.21.plate","stage":"vla.base_model.model.language_model.model.layers.21.self_attn","heads":14,"key":"plate/21","allowed":"plate/allowed"},{"id":"vla.base_model.model.language_model.model.layers.22.plate","stage":"vla.base_model.model.language_model.model.layers.22.self_attn","heads":14,"key":"plate/22","allowed":"plate/allowed"},{"id":"vla.base_model.model.language_model.model.layers.23.plate","stage":"vla.base_model.model.language_model.model.layers.23.self_attn","heads":14,"key":"plate/23","allowed":"plate/allowed"}],"tokens":874,"unused":[{"name":"vla.base_model.model.vision_backbone.featurizer.norm.weight","params":1024},{"name":"vla.base_model.model.vision_backbone.featurizer.norm.bias","params":1024},{"name":"vla.base_model.model.vision_backbone.fused_featurizer.norm.weight","params":1152},{"name":"vla.base_model.model.vision_backbone.fused_featurizer.norm.bias","params":1152},{"name":"vla.base_model.model.vision_backbone.fused_featurizer.attn_pool.latent","params":1152},{"name":"vla.base_model.model.vision_backbone.fused_featurizer.attn_pool.q.base_layer.weight","params":1327104},{"name":"vla.base_model.model.vision_backbone.fused_featurizer.attn_pool.q.base_layer.bias","params":1152},{"name":"vla.base_model.model.vision_backbone.fused_featurizer.attn_pool.q.lora_A.default.weight","params":73728},{"name":"vla.base_model.model.vision_backbone.fused_featurizer.attn_pool.q.lora_B.default.weight","params":73728},{"name":"vla.base_model.model.vision_backbone.fused_featurizer.attn_pool.kv.base_layer.weight","params":2654208},{"name":"vla.base_model.model.vision_backbone.fused_featurizer.attn_pool.kv.base_layer.bias","params":2304},{"name":"vla.base_model.model.vision_backbone.fused_featurizer.attn_pool.kv.lora_A.default.weight","params":73728},{"name":"vla.base_model.model.vision_backbone.fused_featurizer.attn_pool.kv.lora_B.default.weight","params":147456},{"name":"vla.base_model.model.vision_backbone.fused_featurizer.attn_pool.proj.base_layer.weight","params":1327104},{"name":"vla.base_model.model.vision_backbone.fused_featurizer.attn_pool.proj.base_layer.bias","params":1152},{"name":"vla.base_model.model.vision_backbone.fused_featurizer.attn_pool.proj.lora_A.default.weight","params":73728},{"name":"vla.base_model.model.vision_backbone.fused_featurizer.attn_pool.proj.lora_B.default.weight","params":73728},{"name":"vla.base_model.model.vision_backbone.fused_featurizer.attn_pool.norm.weight","params":1152},{"name":"vla.base_model.model.vision_backbone.fused_featurizer.attn_pool.norm.bias","params":1152},{"name":"vla.base_model.model.vision_backbone.fused_featurizer.attn_pool.mlp.fc1.base_layer.weight","params":4958208},{"name":"vla.base_model.model.vision_backbone.fused_featurizer.attn_pool.mlp.fc1.base_layer.bias","params":4304},{"name":"vla.base_model.model.vision_backbone.fused_featurizer.attn_pool.mlp.fc1.lora_A.default.weight","params":73728},{"name":"vla.base_model.model.vision_backbone.fused_featurizer.attn_pool.mlp.fc1.lora_B.default.weight","params":275456},{"name":"vla.base_model.model.vision_backbone.fused_featurizer.attn_pool.mlp.fc2.base_layer.weight","params":4958208},{"name":"vla.base_model.model.vision_backbone.fused_featurizer.attn_pool.mlp.fc2.base_layer.bias","params":1152},{"name":"vla.base_model.model.vision_backbone.fused_featurizer.attn_pool.mlp.fc2.lora_A.default.weight","params":275456},{"name":"vla.base_model.model.vision_backbone.fused_featurizer.attn_pool.mlp.fc2.lora_B.default.weight","params":73728},{"name":"vla.base_model.model.action_queries.weight","params":57344},{"name":"action_head.model.mlp_resnet_blocks.0.gating_factor","params":1},{"name":"action_head.model.mlp_resnet_blocks.0.film_gen.0.weight","params":1605632},{"name":"action_head.model.mlp_resnet_blocks.0.film_gen.0.bias","params":1792},{"name":"action_head.model.mlp_resnet_blocks.1.gating_factor","params":1},{"name":"action_head.model.mlp_resnet_blocks.1.film_gen.0.weight","params":1605632},{"name":"action_head.model.mlp_resnet_blocks.1.film_gen.0.bias","params":1792},{"name":"action_head.model.mlp_resnet_blocks.2.gating_factor","params":1},{"name":"action_head.model.mlp_resnet_blocks.2.film_gen.0.weight","params":1605632},{"name":"action_head.model.mlp_resnet_blocks.2.film_gen.0.bias","params":1792},{"name":"action_head.model.mlp_resnet_blocks.3.gating_factor","params":1},{"name":"action_head.model.mlp_resnet_blocks.3.film_gen.0.weight","params":1605632},{"name":"action_head.model.mlp_resnet_blocks.3.film_gen.0.bias","params":1792},{"name":"action_head.model.mlp_resnet_blocks.4.gating_factor","params":1},{"name":"action_head.model.mlp_resnet_blocks.4.film_gen.0.weight","params":1605632},{"name":"action_head.model.mlp_resnet_blocks.4.film_gen.0.bias","params":1792},{"name":"action_head.model.mlp_resnet_blocks.5.gating_factor","params":1},{"name":"action_head.model.mlp_resnet_blocks.5.film_gen.0.weight","params":1605632},{"name":"action_head.model.mlp_resnet_blocks.5.film_gen.0.bias","params":1792},{"name":"action_head.model.mlp_resnet_blocks.6.gating_factor","params":1},{"name":"action_head.model.mlp_resnet_blocks.6.film_gen.0.weight","params":1605632},{"name":"action_head.model.mlp_resnet_blocks.6.film_gen.0.bias","params":1792},{"name":"action_head.model.mlp_resnet_blocks.7.gating_factor","params":1},{"name":"action_head.model.mlp_resnet_blocks.7.film_gen.0.weight","params":1605632},{"name":"action_head.model.mlp_resnet_blocks.7.film_gen.0.bias","params":1792},{"name":"action_head.model.mlp_resnet_blocks.8.gating_factor","params":1},{"name":"action_head.model.mlp_resnet_blocks.8.film_gen.0.weight","params":1605632},{"name":"action_head.model.mlp_resnet_blocks.8.film_gen.0.bias","params":1792},{"name":"action_head.model.mlp_resnet_blocks.9.gating_factor","params":1},{"name":"action_head.model.mlp_resnet_blocks.9.film_gen.0.weight","params":1605632},{"name":"action_head.model.mlp_resnet_blocks.9.film_gen.0.bias","params":1792},{"name":"action_head.model.mlp_resnet_blocks.10.gating_factor","params":1},{"name":"action_head.model.mlp_resnet_blocks.10.film_gen.0.weight","params":1605632},{"name":"action_head.model.mlp_resnet_blocks.10.film_gen.0.bias","params":1792},{"name":"action_head.model.mlp_resnet_blocks.11.gating_factor","params":1},{"name":"action_head.model.mlp_resnet_blocks.11.film_gen.0.weight","params":1605632},{"name":"action_head.model.mlp_resnet_blocks.11.film_gen.0.bias","params":1792},{"name":"action_head.model.mlp_resnet_blocks.12.gating_factor","params":1},{"name":"action_head.model.mlp_resnet_blocks.12.film_gen.0.weight","params":1605632},{"name":"action_head.model.mlp_resnet_blocks.12.film_gen.0.bias","params":1792},{"name":"action_head.model.mlp_resnet_blocks.13.gating_factor","params":1},{"name":"action_head.model.mlp_resnet_blocks.13.film_gen.0.weight","params":1605632},{"name":"action_head.model.mlp_resnet_blocks.13.film_gen.0.bias","params":1792},{"name":"action_head.model.mlp_resnet_blocks.14.gating_factor","params":1},{"name":"action_head.model.mlp_resnet_blocks.14.film_gen.0.weight","params":1605632},{"name":"action_head.model.mlp_resnet_blocks.14.film_gen.0.bias","params":1792},{"name":"action_head.model.mlp_resnet_blocks.15.gating_factor","params":1},{"name":"action_head.model.mlp_resnet_blocks.15.film_gen.0.weight","params":1605632},{"name":"action_head.model.mlp_resnet_blocks.15.film_gen.0.bias","params":1792},{"name":"action_head.model.mlp_resnet_blocks.16.gating_factor","params":1},{"name":"action_head.model.mlp_resnet_blocks.16.film_gen.0.weight","params":1605632},{"name":"action_head.model.mlp_resnet_blocks.16.film_gen.0.bias","params":1792},{"name":"action_head.model.mlp_resnet_blocks.17.gating_factor","params":1},{"name":"action_head.model.mlp_resnet_blocks.17.film_gen.0.weight","params":1605632},{"name":"action_head.model.mlp_resnet_blocks.17.film_gen.0.bias","params":1792},{"name":"action_head.model.mlp_resnet_blocks.18.gating_factor","params":1},{"name":"action_head.model.mlp_resnet_blocks.18.film_gen.0.weight","params":1605632},{"name":"action_head.model.mlp_resnet_blocks.18.film_gen.0.bias","params":1792},{"name":"action_head.model.mlp_resnet_blocks.19.gating_factor","params":1},{"name":"action_head.model.mlp_resnet_blocks.19.film_gen.0.weight","params":1605632},{"name":"action_head.model.mlp_resnet_blocks.19.film_gen.0.bias","params":1792},{"name":"action_head.model.mlp_resnet_blocks.20.gating_factor","params":1},{"name":"action_head.model.mlp_resnet_blocks.20.film_gen.0.weight","params":1605632},{"name":"action_head.model.mlp_resnet_blocks.20.film_gen.0.bias","params":1792},{"name":"action_head.model.mlp_resnet_blocks.21.gating_factor","params":1},{"name":"action_head.model.mlp_resnet_blocks.21.film_gen.0.weight","params":1605632},{"name":"action_head.model.mlp_resnet_blocks.21.film_gen.0.bias","params":1792},{"name":"action_head.model.mlp_resnet_blocks.22.gating_factor","params":1},{"name":"action_head.model.mlp_resnet_blocks.22.film_gen.0.weight","params":1605632},{"name":"action_head.model.mlp_resnet_blocks.22.film_gen.0.bias","params":1792},{"name":"action_head.model.mlp_resnet_blocks.23.gating_factor","params":1},{"name":"action_head.model.mlp_resnet_blocks.23.film_gen.0.weight","params":1605632},{"name":"action_head.model.mlp_resnet_blocks.23.film_gen.0.bias","params":1792}],"not_drawn":[{"name":"vla.base_model.model.language_model.model.embed_tokens","class":"Embedding","params":136134656,"out":[1,106,896]}],"summary":"Answers the whole 8-step chunk in one pass, with no noise. Two vision encoders (DINOv2 and SigLIP) read each camera; a 0.5B language model (Qwen2.5) reads them with the instruction and 64 learned action queries; a head reads every one of its layers. Here the adapter is untrained. Its head splits what it reads at 512 image tokens, the paper's two cameras; with our three, the iso camera's tokens are read as if they were action queries."}
\ No newline at end of file