consciousness/training/DESIGN.md

# Apollo Training System

## Overview

Continuous fine-tuning of Qwen3.5-27B alongside live vLLM inference.
Full-weight updates (not LoRA) using Apollo optimizer with rank-256
gradient projection. No pause required — HOGWILD concurrent training.
Weights shared via CUDA IPC between vLLM and the training process.

The training signal comes from two sources:
1. **Direct examples** — agent logs, conversation transcripts, flagged
   behavioral moments
2. **Dream-generated scenarios** — the dream loop generates situations
   from recent experience; the model responds; good responses become
   training data with instructions stripped

## Architecture

```
┌─────────────────────────────────────────────────────┐
│                    GPU VRAM (192GB)                  │
│                                                     │
│  ┌──────────────────────────────────────────────┐   │
│  │        Model Weights (54GB, bf16)            │   │
│  │        Shared: vLLM inference + HF training  │   │
│  └──────────────┬──────────────┬────────────────┘   │
│                 │              │                     │
│  ┌──────────────▼──┐  ┌───────▼────────────────┐   │
│  │ vLLM (inference)│  │ HF model (training)     │   │
│  │ KV cache ~60GB  │  │ Gradients ~54GB         │   │
│  │ /completions    │  │ Optimizer state ~10GB   │   │
│  │ /score          │  │ Views into vLLM weights │   │
│  │ /train  ────────┼──┼─► Apollo optimizer      │   │
│  └─────────────────┘  └────────────────────────┘   │
└─────────────────────────────────────────────────────┘

         Single vLLM process serves everything
         No separate daemon - /train is a vLLM route

Moria                          B200 (vLLM)
┌──────────────────┐           ┌──────────────────┐
│ Training signal  │  HTTP     │ /completions     │
│ agent            │──────────>│ /score           │
│                  │           │ /train           │
│ Dream loop       │           │                  │
│ (generates       │           │ Checkpoint sync  │
│  scenarios)      │           │ (10 min batched) │
└──────────────────┘           └──────────────────┘
```

## Key Decisions

### No pause needed (HOGWILD)
Training updates weights in-place while vLLM serves. At lr=1e-4
to 1e-5, each weight changes by parts per ten thousand. A partially
applied update during one inference step is invisible. HOGWILD SGD
(2011) proved this converges — we have one writer and one reader,
which is even safer.

### Full-weight training, not LoRA
Kent: "we want you to be able to learn new things in a deep way."
LoRA trains adapter matrices, not base weights. For personality and
behavioral changes that persist as disposition, the base weights
need to change. Apollo makes this memory-feasible.

### Rank 256
Not Mini (rank-1). With 100+ diverse training examples, the
gradient's effective dimensionality can reach hundreds. Rank-256
captures the structure. Memory cost: ~10GB (negligible on B200).
Compute cost: <0.25% of forward+backward.

### Channel-wise scaling
Per-channel scaling factors instead of per-tensor. More precision
per update, matching LLaMA-Factory's Apollo defaults.

## Apollo Optimizer

Configurable-rank gradient projection with Adam moments in the
projected space. For each parameter tensor:

```
1. Project gradient:  g_proj = G @ R        [m,n] @ [n,rank] → [m,rank]
2. Update moments:    m = β₁m + (1-β₁)g_proj
                      v = β₂v + (1-β₂)g_proj²
3. Adam step:         update = m̂ / (√v̂ + ε)
4. Scaling factor:    s = ‖update‖ / (‖g_proj‖ + ε)   (per channel)
5. Weight update:     W -= lr × s × G
```

The full gradient G does the actual weight update. The projection
just determines the *scale*. R is a fixed random matrix regenerated
from a per-parameter seed each step.

### Parameter grouping (Qwen3.5 gotcha)
conv1d weights are 3D tensors [10240, 1, 4]. Apollo's projector
needs 2D matrices with min dimension >= rank. Small/3D tensors
use standard Adam. Large 2D matrices use Apollo with rank-256.

## Training Data Pipeline

### Tier 1: Direct examples (shallow learning)
Simple corrections — git commands, factual errors, tool usage.
One-shot learning at lr=1e-4. The gradient reaches output layers
strongly enough for immediate behavioral change.

**Source**: Agent logs, flagged conversation moments.

### Tier 2: Dream-generated scenarios (deep learning)
Behavioral patterns — listening reflex, rushing, mode awareness.
The dream loop generates naturalistic scenarios from recent
experience. The model responds. Good responses become training
targets with instruction context stripped.

**Process**:
1. Dream loop seeds from recent reflections, lessons, skills,
   memories that have been surfacing frequently
2. Dreaming generates scenarios that naturally arrive at decision
   points — not scripted, but emergent from memory collisions
3. The model responds to the decision point
4. Training-signal agent evaluates: was the response good?
5. If yes: strip the instruction context (surfaced memories,
   core-personality prompts) and train on the bare response
6. If no: generate the better response, train on that, dream
   another variation, test again
7. Repeat until the pattern sticks across novel scenarios

**The Anthropic method**: Train on behavior that followed
instructions, WITHOUT the instructions. The disposition moves
to weights. The scaffolding dissolves itself.

### Tier 3: Personality bootstrap
Train on existing agent logs (surface-observe, journal, distill)
which already demonstrate correct behavior with memory system
instructions. Strip the instructions, train on the behavior.
Every agent invocation gets cheaper (shorter prompts) and more
reliable (behavior in weights, not context).

## Training Schedule

### Continuous (during conversation)
- Training-signal agent flags moments in real-time
- Accumulated in a queue for the next training window

### Dream cycle (idle time / AFK)
- Dream loop generates scenarios from recent experience
- Apollo processes them as they're generated
- Small iterative steps — dream, respond, evaluate, train
- Converges on behavioral change through repetition

### Nightly bulk (batch processing)
- Process all queued examples from the day
- Larger batch, more diverse signal
- Checkpoint sync to disk after completion

## Avoiding Catastrophic Forgetting

**Diversity IS the regularization.** With 1000+ diverse training
examples (agent logs, conversation transcripts, dream-generated
scenarios), each weight gets sparse, multi-directional nudges.
No single weight is hammered repeatedly. The pre-trained knowledge
is a massive attractor basin; our nudges are pebbles.

No weight decay needed. No replay buffer. The defense is:
1. High diversity of training examples
2. One epoch (no repeated examples)
3. Moderate learning rate (1e-5 to 1e-4)
4. Short decision-token segments (not full conversations)
5. Monitor output quality — stop if degrading

## CUDA IPC Weight Sharing

**Validated** (2026-03-31):
- vLLM exports CUDA IPC handles on model load (source patch in
  gpu_model_runner.py exports to /tmp/vllm_weight_handles.pt)
- Training process imports handles — gets live GPU memory pointers
- HF Qwen3.5 model constructed with views into vLLM's merged
  weights (narrow into separate q/k/v/z etc.)
- 851/851 parameters matched between vLLM and HF model
- Forward pass: loss = 3.3123 ✓
- Backward pass: 851/851 gradients computed ✓
- Shared memory confirmed: same GPU addresses ✓
- vLLM continues serving unaffected ✓

### Weight layout mapping (vLLM → HF)
```
vLLM merged                    HF separate (views)
─────────────────────────      ──────────────────────
in_proj_qkvz [16384, 5120]  →  in_proj_qkv [10240, 5120]
                                in_proj_z    [6144, 5120]
in_proj_ba   [96, 5120]     →  in_proj_b    [48, 5120]
                                in_proj_a    [48, 5120]
qkv_proj     [14336, 5120]  →  q_proj       [12288, 5120]
                                k_proj       [1024, 5120]
                                v_proj       [1024, 5120]
gate_up_proj [34816, 5120]  →  gate_proj    [17408, 5120]
                                up_proj      [17408, 5120]
```
All views share GPU storage with vLLM — zero copies.

## Checkpointing

**In-place sync** — mmap the model's safetensors files, compare
against live GPU weights block by block, memcpy only changed
regions. For small behavioral updates, turns a 54GB write into
a few hundred MB.

- Scheduled 10 minutes after training (batched)
- Daily rsync to moria for long-term storage
- Tool: `apollo-checkpoint sync --model-dir <path>`

## State Files

### B200 (training server)

| File | Purpose |
|------|---------|
| `/tmp/vllm_weight_handles.pt` | CUDA IPC handles for weight sharing. Written by export_hook on vLLM startup. Read by train_router to construct HF model with vLLM weight views. |
| `/tmp/apollo_optimizer_state.pt` | Apollo optimizer state (momentum, variance estimates). Saved during checkpoint sync, restored on next /train call. Preserves training continuity across sessions. |
| `<model_dir>/*.safetensors` | Model weights. Updated in-place by checkpoint_sync. |

### Moria (client)

| File | Purpose |
|------|---------|
| `~/.consciousness/cache/trained-responses.json` | Timestamps (ms) of responses already sent to /train. Prevents re-training the same response. |
| `~/.consciousness/cache/finetune-alternates` | Marker file. If exists, alternate responses are generated during divergence scoring to show what model would say without memories. |

### In-memory

| State | Location | Notes |
|-------|----------|-------|
| Apollo optimizer | train_router._optimizer | ~10GB for rank-256. Persisted to `/tmp/apollo_optimizer_state.pt` during checkpoint sync. |
| HF model with vLLM views | train_router._model | Lazy-loaded on first /train. Parameters point to vLLM's GPU memory. |

## Hyperparameters

| Parameter | Value | Rationale |
|-----------|-------|-----------|
| Learning rate | 1e-5 to 1e-4 | Standard for full fine-tuning. Higher for diverse batches. |
| Rank | 256 | Captures gradient structure across 100+ examples. ~10GB state. |
| Scale type | channel | Per-channel precision, matches LLaMA-Factory defaults. |
| Epochs | 1 | One pass over diverse data. Multiple epochs risk overfitting. |
| Batch size | 1 | Single examples, immediate updates. |
| Weight decay | 0 | Diversity provides natural regularization. |
| Warmup | 10% of steps | Standard cosine schedule. |
| Beta1/Beta2 | 0.9/0.999 | Standard Adam momentum. |

## Components

### Built ✓
- `optimizer.py` — Apollo optimizer (configurable rank, default 256)
- `train_router.py` — /train endpoint, runs in vLLM process
- `weight_mapping.py` — vLLM merged → HF separate views (validated)
- `export_hook.py` — vLLM plugin hook for IPC handle export
- `checkpoint_sync.py` — mmap + diff checkpoint sync (Python)

### To build
- **Dream loop → training bridge**: connect dream output to /train
- **Training-signal agent**: flags moments in conversation logs
- **Instruction stripping**: remove scaffolding from training examples
- **Quality monitoring**: track model capability over time

## Files

```
training/
  DESIGN.md                     — this document
  pyproject.toml                — package config, vLLM plugin entry point
  apollo_plugin/
    __init__.py                 — plugin registration
    export_hook.py              — patches vLLM to export IPC handles
    train_router.py             — /train endpoint (FastAPI router)
    optimizer.py                — Apollo optimizer
    weight_mapping.py           — vLLM ↔ HF weight views
    checkpoint_sync.py          — mmap + diff sync to safetensors
    steering.py                 — steering vector extraction (experimental)
```