File size: 12,627 Bytes
f5961bc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
# SmolVLA Architecture Diagrams

Reference diagrams for understanding the SmolVLA model structure and what
the `--model-health` report measures.

---

## 1. High-level data flow

```
  Camera Images (512x512)       Task String            Robot State     Noisy Actions
  +--------+ +--------+         "pick up the           (joint          (from flow
  |  side  | |   up   |          red block"             positions)       matching)
  +---+----+ +---+----+              |                     |               |
      |          |                   |                     |               |
      v          v                   v                     v               v
  +------------------+     +---------------+     +---------------+  +---------------+
  | SigLIP Vision    |     | Tokenizer     |     | state_proj    |  | action_in_proj|
  | Encoder          |     |               |     | (Linear)      |  | (Linear)      |
  | (frozen)         |     |               |     |               |  | + time MLP    |
  | 12L, 12H         |     |               |     | trainable     |  | trainable     |
  +--------+---------+     +-------+-------+     +-------+-------+  +-------+-------+
           |                       |                     |                  |
     1024 patches            token embeds            1 token          10 tokens
           |                       |                     |                  |
           v                       |                     |                  |
  +-------------------+            |                     |                  |
  | Connector         |            |                     |                  |
  | (pixel shuffle)   |            |                     |                  |
  | trainable         |            |                     |                  |
  +--------+----------+            |                     |                  |
           |                       |                     |                  |
     64 tokens                     |                     |                  |
           |                       |                     |                  |
           +-----------+-----------+                     |                  |
                       |                                 |                  |
                       v                                 |                  |
           +-----------------------+                     |                  |
           | PREFIX SEQUENCE       |                     |                  |
           | [vision] [language]   +<--------------------+                  |
           | [state]               |     VLM dim (960)                     |
           +-----------+-----------+                                       |
                       |                                                   |
                       |                                 +-----------------+
                       |                                 |
                       v                                 v
           +---------------------------------------------------+
           |         SmolVLMWithExpert (16 layers)               |
           |                                                     |
           |   PREFIX  ------------------->  VLM Text Model      |
           |   (vision+lang+state)            (frozen)           |
           |                                  960-dim, 15 heads  |
           |                                                     |
           |   SUFFIX  ------------------->  Action Expert       |
           |   (noisy actions+time)           (trainable)        |
           |                                  480-dim, 8 heads   |
           |                                                     |
           |         (see attention diagrams below)              |
           +--------------------------+--------------------------|
                                      |
                                      |  Expert output (last 10 tokens)
                                      v
                          +-------------------+
                          | action_out_proj   |
                          | (Linear)          |
                          | trainable         |
                          +---------+---------+
                                    |
                                    v
                          Predicted Actions
                          (10 steps x 6 DOF)
```

---

## 2. The three attention components

These are the three distinct attention operations that the health report
measures for entropy and head redundancy.

### A. SigLIP Vision (12 layers, 12 heads)

Standard self-attention inside the vision encoder. Happens before anything
enters SmolVLMWithExpert. Image patches attend to other image patches.

```
  patch_1  patch_2  patch_3  ...  patch_1024
     |        |        |              |
     v        v        v              v
  +----------------------------------------------+
  | SigLIP Self-Attention                         |
  |                                               |
  | Each patch can see all patches.               |
  |                                               |
  | Q <-- patches        K <-- patches            |
  | (1024 x 1024 attention matrix)                |
  +----------------------------------------------+

  Report label: "SigLIP Vision (12L, 12H)"
```

### B. VLM+Expert Joint Self-Attention (16 layers, 15 heads)

Runs during prefill (initial encoding). VLM and Expert tokens are
concatenated into one sequence and attend to each other in a single
attention call.

```
  VLM tokens (prefix)           Expert tokens (suffix)
  [vision][lang][state]          [action+time]
        |                              |
        v                              v
  +-------------+               +-------------+
  | VLM layer   |               | Expert layer|
  | q/k/v proj  |               | q/k/v proj  |
  +------+------+               +------+------+
         |                             |
         +-------------+---------------+
                       |
                       v
                torch.cat(dim=seq)
                       |
                       v
  +----------------------------------------------+
  | JOINT Self-Attention                          |
  |                                               |
  | All tokens see all tokens:                    |
  |   vision  <-->  vision                        |
  |   vision  <-->  language                      |
  |   vision  <-->  action                        |
  |   action  <-->  language                      |
  |   action  <-->  action                        |
  |   ... etc                                     |
  |                                               |
  | Q <-- [VLM+Expert]     K <-- [VLM+Expert]     |
  | Q_len == K_len  (self-attention)              |
  +----------------------------------------------+

  Report label: "VLM+Expert Joint Self-Attn (16L, 15H)"
```

### C. Expert-to-VLM Cross-Attention (16 layers, 8 heads)

Runs during generation (autoregressive action decoding). The Expert
generates queries from its own hidden states, but keys and values come
from the VLM's cached representations, re-projected through the Expert's
k_proj and v_proj. Repeats 10 times (one per action token).

```
  VLM KV Cache                       Expert hidden states
  (built during prefill)             (current action token)
        |                                    |
        |                                    v
        |                            +---------------+
        |                            | Expert q_proj |
        |                            +-------+-------+
        |                                    |
        |     re-projected through           |
        |     Expert k_proj / v_proj         |
        v                                    |
  +-------------+                            |
  | Expert K, V |<-- VLM cache               |
  | (from VLM)  |   re-projected             |
  +------+------+                            |
         |                                   |
         +---------------+-------------------+
                         |
                         v
  +----------------------------------------------+
  | CROSS-Attention                               |
  |                                               |
  | Expert reads VLM:                             |
  |   action --> vision   (what to grab?)         |
  |   action --> language  (what was asked?)      |
  |   action --> state     (current pose?)        |
  |                                               |
  | VLM does NOT read Expert here.                |
  |                                               |
  | Q <-- Expert tokens (few)                     |
  | K <-- VLM prefix (many)                       |
  | Q_len != K_len  (cross-attention)             |
  +----------------------------------------------+

  Report label: "Expert-to-VLM Cross-Attn (16L, 8H)"
```

---

## 3. Execution timeline

```
  IMAGE ENCODING            PREFILL                   GENERATION
  +--------------+    +------------------+    +------+------+     +------+
  | SigLIP       |    | Joint Self-Attn  |    | XA   | XA   |    | XA   |
  | 12 layers    +--->| 16 layers        +--->| #1   | #2   |...>| #10  |
  | (A)          |    | (B)              |    |      |      |    |      |
  +--------------+    +------------------+    +------+------+    +------+
                                              <--- 10 action steps --->

  Report:              Report:                 Report:
  "SigLIP Vision"      "VLM+Expert             "Expert-to-VLM
                        Joint Self-Attn"        Cross-Attn"
  12L x 12H            16L x 15H               16L x 8H
                                                (x10 steps, averaged)
```

---

## 4. What the health report measures

### Section 1: Weight Spectral Analysis (alpha)

Analyzes weight matrices directly (no data needed). Each transformer layer
contains multiple weight matrices (q_proj, k_proj, v_proj, o_proj,
gate_proj, up_proj, down_proj). A power-law is fit to each matrix's
singular values; the exponent (alpha) indicates training quality.

```
  +----------------------------------+-------------+---------------------+
  | Component                        | Wt Matrices | What it is          |
  +----------------------------------+-------------+---------------------+
  | Expert (trainable)               | 112 = 16x7  | Action decoder      |
  | VLM Text Model (frozen)          | 113 = 16x7+ | Language model      |
  | Vision Encoder (frozen)          |  74 = 12x~6 | SigLIP              |
  | Connector (trainable)            |   1          | Pixel shuffle       |
  | Projections (trainable)          | too small    | state/action linear |
  +----------------------------------+-------------+---------------------+

  Healthy: alpha 2-4     Undertrained: 4-6     Severe: >6
```

### Section 2: Attention Entropy

Requires a forward pass with real data. Measures how spread out each
attention head's focus is (normalized by log of sequence length).

```
  Collapsed: <0.10     Healthy: 0.10-0.80     Unfocused: >0.80     Dead: >0.95
```

### Section 3: Head Redundancy

Same forward pass data. Measures pairwise cosine similarity between
flattened head attention patterns within each layer. Heads should learn
different patterns.

```
  Diverse: <0.70       High redundancy: >0.70       Collapsed: >0.90
```

---

## 5. Trainable vs frozen components

```
  FROZEN (pretrained, not updated during fine-tuning)
  +----------------------------------------------------+
  | SigLIP Vision Encoder                               |
  | 12 layers, 12 heads, 74 weight matrices             |
  +----------------------------------------------------+
  | VLM Text Model (SmolLM2)                            |
  | 16 layers, 15 heads, 113 weight matrices            |
  +----------------------------------------------------+

  TRAINABLE (learned during fine-tuning)
  +----------------------------------------------------+
  | Action Expert                                       |
  | 16 layers, 8 heads, 112 weight matrices             |
  +----------------------------------------------------+
  | Connector (pixel shuffle)                           |
  | 1 weight matrix                                     |
  +----------------------------------------------------+
  | state_proj, action_in_proj, action_out_proj         |
  | (too small for spectral analysis)                   |
  +----------------------------------------------------+
```