Llama architecture verified: SmolLM2 1.7B, 1-bit KV = PPL identical

4th architecture verified (Llama/SmolLM2, after Gemma 3, Qwen3.5, Qwen2-MoE).

SmolLM2 1.7B (Llama arch, GGUF Q8_0):
  baseline PPL:         5.8441
  1-bit K + FP16 V PPL: 5.8441 (+0.00%)  ← exactly identical
  1-bit K + Q4 V PPL:   5.8233
  30-token output: byte-identical ✓
  Speed: 24 tok/s (Q4, 6T, M3)

Also fixed: 4B GGUF Q4 conversion (threshold 8→16 GB, DeltaNet detect)
  Qwen3.5-4B: 0.1 → 5.4 tok/s (54x improvement)

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: unamedkr