File size: 6,513 Bytes
6c0624c | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 | # Prisma V2 — ToDo.md
## Goal
Evolve PRISMA from a **Python-stateful uncertainty feedback mechanism** (V1) into a **fully stateless, runtime-portable architecture** compatible with:
* Hugging Face Transformers
* vLLM
* llama.cpp
* MLX
while preserving the core idea:
> **Condition future predictions on the model’s own uncertainty to reduce confident errors and hallucinations.**
---
## Background (Why V2 Is Needed)
Prisma V1 relies on mutable per-step Python state:
```python
self.prev_uncertainty_code
```
This works in Hugging Face `generate()`, but fails in modern inference engines:
* **vLLM** — batched, reordered, stateless execution
* **llama.cpp** — compiled C/C++ loop with fixed KV cache semantics
* **MLX** — pure functional graph execution
All target runtimes require decoding to be **purely functional**:
> All information needed for the next step must be carried explicitly in tensors, tokens, or KV cache — not Python object state.
---
## Core Design Change (Prisma V2)
### Replace measured uncertainty with **predicted uncertainty**
Instead of:
* computing entropy post-hoc from logits
* storing uncertainty in a mutable buffer
**The model learns to predict uncertainty directly.**
At each position, the model outputs:
1. **Next-token logits** (existing)
2. **Uncertainty logits for the next position** (new)
Uncertainty becomes a **learned latent variable**, not a side effect.
---
## Architecture Changes
### 1. Add an uncertainty prediction head
```python
self.n_uncertainty_levels = 256 # V2: smaller, sufficient
self.uncertainty_head = nn.Linear(hidden_size, n_uncertainty_levels, bias=False)
```
### 2. Uncertainty head initialization (important)
The uncertainty head is added to a pretrained model. To avoid destabilizing early training, use **zero initialization**:
```python
self.uncertainty_head.weight.data.zero_()
```
**Rationale:**
* Model initially predicts neutral uncertainty everywhere
* Early training behaves identically to the base model
* Uncertainty signal is learned gradually, only when useful
Alternative initializations (small random, copying from `lm_head`) are left for experimentation.
---
### 3. Keep uncertainty embeddings (input side)
```python
self.uncertainty_embeddings = nn.Embedding(
n_uncertainty_levels,
hidden_size
)
```
---
### 4. Modify forward signature
```python
def forward(
input_ids: torch.Tensor,
uncertainty_codes: Optional[torch.Tensor] = None, # [B, S]
**kwargs
):
```
* `uncertainty_codes[t]` conditions token `t`
* No hidden buffers
* No mutable Python state
---
## Forward Pass Logic (V2)
1. Embed tokens
2. If `uncertainty_codes` provided:
* lookup `uncertainty_embeddings`
* add to `inputs_embeds`
3. Run transformer
4. Produce:
* `logits` (next token)
* `uncertainty_logits` (next uncertainty)
```python
return {
"logits": logits,
"uncertainty_logits": uncertainty_logits,
}
```
---
## Temporal Semantics
| Position | Input | Predicts |
| -------- | --------------------- | ------------------------ |
| t | tokenₜ + uncertaintyₜ | tokenₜ₊₁, uncertaintyₜ₊₁ |
This preserves the original PRISMA temporal feedback loop without mutable state.
---
## Training Plan
### Uncertainty supervision (teacher signal)
During training, entropy is used **only as a teacher signal**, not as the definition of uncertainty.
```python
entropy = -∑ p log p
normalized = entropy / log(vocab_size)
uncertainty_label = quantize(normalized)
```
---
### Single-pass training (preferred)
A second forward pass is **not required**.
```python
outputs = model(
input_ids,
uncertainty_codes=uncertainty_input
)
with torch.no_grad():
uncertainty_labels = quantize_entropy(outputs.logits)
loss = (
loss_lm(outputs.logits, labels)
+ λ * loss_uncertainty(outputs.uncertainty_logits, uncertainty_labels)
)
```
**Key point:**
* Entropy is a **bootstrap target**
* The model is free to learn uncertainty representations that diverge from entropy over time
* This allows uncertainty to correlate better with *error* than raw entropy does
---
### Loss definition
```python
loss = loss_lm + λ * loss_uncertainty
```
* `loss_lm`: standard next-token cross-entropy
* `loss_uncertainty`: cross-entropy over uncertainty codes
* λ ≈ 0.1 (to tune)
---
## Inference (All Runtimes)
### Decode loop (conceptual)
```text
(tokenₜ, uncertaintyₜ) → model → (tokenₜ₊₁, uncertaintyₜ₊₁)
```
### Runtime responsibilities
* **Transformers**: custom `generate()` tracks uncertainty tensor
* **vLLM**: sampler tracks `uncertainty_code` per request
* **llama.cpp**: store one small uncertainty code in `llama_context`
* **MLX**: works naturally (pure tensor graph)
No runtime needs to preserve Python object state.
---
## Compatibility Matrix
| Runtime | Prisma V1 | Prisma V2 |
| --------------- | --------- | --------- |
| Transformers | ✅ | ✅ |
| vLLM | ❌ | ✅ |
| llama.cpp | ❌ | ✅ |
| MLX | ❌ | ✅ |
| Tensor Parallel | ⚠️ | ✅ |
---
## Design Decisions
### Classification vs regression
**Chosen:** classification (quantized uncertainty)
Reasons:
* Stable training
* Matches embedding lookup
* Discrete semantics
* Easier runtime handling
Regression remains an experimental alternative.
---
### Uncertainty resolution
* V1: 65,536 levels (overkill)
* V2: **256 levels** (sufficient, efficient, portable)
---
## Known Limitations
* Uncertainty reflects **model confidence**, not correctness
* Learned uncertainty may differ from Shannon entropy
* Does not guarantee abstention or correctness
* Behavior depends on training data and loss weighting
---
## Open Questions / Future Work
* Tune `n_uncertainty_levels`
* Tune λ (uncertainty loss weight)
* Explore uncertainty-aware decoding strategies
* Compare uncertainty prediction vs uncertainty tokens
* Investigate bootstrapping without entropy supervision
---
## Definition of Done (Prisma V2)
* [ ] No mutable per-step Python state
* [ ] Uncertainty passed explicitly as tensor
* [ ] Works in Transformers, vLLM, llama.cpp, MLX
* [ ] Zero-init uncertainty head
* [ ] Single-pass training loop
* [ ] Updated model card + documentation
* [ ] Reference implementation + example
---
### Guiding Principle (V2)
> **Uncertainty must be data, not memory.**
|