Holo-3.1-0.8B-lora-demo-2

Holo-3.1-0.8B — Atlas LoRA demo adapter #2 (dynamic-swap sibling)

A second tiny, deliberately-overfit demo LoRA for Hcompany/Holo-3.1-0.8B, the sibling of MonumentalSystems/Holo-3.1-0.8B-lora-demo. It burns in a different codeword + persona so the two adapters can demonstrate pool-size-1 dynamic adapter swap in the Atlas pure-Rust CUDA inference engine: with --max-loras 1 only one adapter is resident, and POST /v1/lora/load swaps the single slot's weights at runtime (per-request weight change).

Config (matches adapter #1 so both fit the same fixed-address pool slot)

• target_modules = ["k_proj", "v_proj", "o_proj"] (q_proj is gated on Holo; dense-FFN delta not yet wired in Atlas v0)
• layers_to_transform = [3, 7, 11, 15, 19, 23] — the 6 full-attention layers
• r = 32, lora_alpha = 64 (scale = α/r = 2.0), use_rslora = false, bias = "none"

Training

peft + transformers, 18 epochs on 6 overfit Q→A pairs, r=32, lora_alpha=64, lr 2e-4, bf16, on a single NVIDIA GB10 (Grace-Blackwell, aarch64, CUDA 13). Loss → 0.0008. A demo artifact, not a general-purpose fine-tune.

Usage (Atlas — pool-size-1 dynamic swap)

# serve with adapter #1 resident in the single slot, rotation armed:
ATLAS_LORA_ROTATE=1 spark serve Hcompany/Holo-3.1-0.8B \
  --lora-adapter starfall=MonumentalSystems/Holo-3.1-0.8B-lora-demo \
  --max-loras 1 --max-lora-rank 64
# → "launch codeword?" answers STARFALL-4728
# swap THIS adapter into the single slot at runtime:
curl -XPOST localhost:8000/v1/lora/load -d '{"name":"vega","path":"/path/to/this-adapter","slot":0}'
# → "launch codeword?" now answers MOONVEIL-3390