qcache-q4_k.v1.gguf — GGML Q4_K tensors for the model's linear projections,
stored in a GGUF container keyed by the original checkpoint tensor paths
(no metadata KVs). Produced by in-memory quantization of the bf16
checkpoint with a candle-based
loader.aux-tensors.safetensors — everything the cache does not carry:
embeddings, norms, and the full-attention layers' projections kept in
bf16. Keeping the global-attention projections unquantized matters:
4-bit error concentrated in those layers collapses long-range retrieval
(needle recall degrades beyond ~2x the sliding window with them
quantized, and is exact with them in bf16).