Two deep dives following curiosity:
- Why context-frozen training works: gradient flows through W_q (query
projection) even when context KVs are frozen. Model learns to LOOK AT
context differently, not represent it differently. This is exactly what
behavioral fine-tuning needs.
- Why Apollo beats AdamW: lower directional sharpness = flatter minima =
better generalization. The coarseness of channel/tensor-wise scaling
prevents over-fitting to specific training examples. For behavioral
fine-tuning, this means learning 'accept direction' rather than
'accept this specific phrasing.'
Corrections from reading the full paper (arXiv:2412.05270):
- Add gradient scale factor α = √(n/r) — compensates for systematic
ratio between compact and original space scaling factors
- Add norm-growth limiter (γ=1.01) — prevents loss spikes in early training
- Refresh projection matrix every 200 steps, not every step
- Channel-wise scaling for rank>1, tensor-wise for rank=1
- Scaling applies as G·diag(s), preserving gradient direction per channel
Research writeup in training/research/apollo-paper-analysis.md covers:
- Full mathematical derivation (equations 1-9)
- Theorems 4.1 and 4.2 (JL-based approximation bounds)
- Why Apollo can beat AdamW (directional sharpness, Hessian spectra)
- Fine-tuning results (matches AdamW at 0 memory cost)
- Ablation studies (rank, scaling granularity, projection method)
- Implications for our behavioral fine-tuning use case
mmap each safetensors file, diff block-by-block against live GPU
weights, memcpy only changed blocks. No separate checkpoint files —
the model directory IS the checkpoint. Every 10 min via cron.
Rust tool that mmaps previous checkpoint, diffs against live GPU weights
(via CUDA IPC handles), and only writes changed blocks. For small
behavioral training steps, turns 54GB write into ~500MB.
Also includes vllm_export_hook.py with direct source patch approach —
exports IPC handles from vLLM's worker subprocess after model load.
Run every 10 minutes via cron to protect against vLLM crashes.
Daily rsync to moria for long-term storage.
Core components for online fine-tuning of Qwen3.5-27B with CUDA IPC
shared weight memory between vLLM and the training process:
- apollo_mini.py: rank-1 optimizer (SGD memory, AdamW quality)
- apollo_worker.py: HTTP daemon coordinating training with vLLM
- weight_mapping.py: vLLM merged → HF separate layout (zero-copy views)
- training_example.py: tokenization with chat template
- export_weights.py: CUDA IPC handle export from vLLM
- train.py: standalone training script (alternative to daemon)
- DESIGN.md: architecture and protocol documentation
Validated: CUDA IPC autograd works on real Qwen3.5 weights (B200).
Apollo-Mini rank-1 projection + scaling + in-place update confirmed.
Co-Authored-By: Kent Overstreet <kent.overstreet@gmail.com>
