elpirater312/nohurry-gemma-4-26B-A4B-it-heretic-IQ4_NL-ud-imtrx-GUFF

⚠️ UPDATE APRIL 09: V2 RELEASE (CRITICAL FIXES)

The previous version was built on outdated llama.cpp kernels. This V2 release incorporates the latest architectural fixes for Gemma-4, including the <unused24> buffer fix and Logit Soft-Capping.

THE ULTIMATE IQ4_NL (V2) - APRIL 2026 FIXES

The previous quant is out of date and MISSING critical architectural fixes. This V2 release is the first to incorporate the Unsloth Dynamic 2.0 re-computed imatrix and the latest llama.cpp Gemma-4 patches.

What's New in V2? (The Unsloth Special)

1. CRITICAL FIX: <unused24> token buffer overlap fixed (No more random logic crashes).
2. KV-Cache Support: Full attention rotation for heterogeneous iSWA (Better long-context stability).
3. Vocab Fix: Byte token handling added to BPE de-tokenizer (Fixes weird character output).
4. BOS Logic: add_bos now defaults to True for native Gemma-4 behavior.
5. Logit Soft-Capping: Correctly reads final_logit_softcapping (CRITICAL for preventing NaN/Infinity loops).
6. Newline Logic: Custom newline split logic for superior formatting.

Why this Quant?

• Format: IQ4_NL (Non-Linear) - Higher logic-per-bit than Q4_K_M or Q4_0.
• Brain: High-res Q8_0 Token Embeddings and Output Head (Prevents vocabulary degradation).
• Speed: Specially optimized for CPU/Laptop P-Cores (~7 t/s on 12th Gen i3 Dual DDR4 Channel at 3200MHz, CPU ONLY inference!).
• Matrix: Built with the latest April 07 re-computed Unsloth imatrix.

All credits to Merijn Hendriks (nohurry) for the heretic weights and Unsloth for the imatrix file.