GGUF of Google's gemma-4 E2B QAT-mobile checkpoint using only tensor types with GPU kernels in llama.cpp. Quantization mirrors the checkpoint's own per-module QAT bit-map (quantization_config): attention and layers 0–14 MLPs → Q4_0, 2-bit-trained modules (remaining MLPs, token_embd, output) → Q2_K, per-layer gates → Q8_0. SRQ activation scales are dropped (not representable in GGUF).

wikitext-2 fidelity vs the bf16 QAT reference: PPL 88.3 (ref 80.6), mean KLD 0.20 — comparable to TQ2_0-based packs, without the CPU-only ternary types.