[Bug] Apex quantized models output "00000000..." or “@@@@@@...”

[FreyHsia]

Description

When running Apex quantized models with llama-server and mmap enabled, the model tends to output repetitive artifacts like "00000000" or "@@@@@@@" after modifying the prompt and invalidating the KV Cache.

Environment

• Inference Engine: llama-server, router mode, mmap = true
• Quantization: APEX (I-Balanced)
• Affected Models:
  • Qwen3.6-35B-A3B-APEX-MTP-I-Balanced.gguf (26GB)
  • Qwen3.6-35B-A3B-uncensored-heretic-APEX-GGUF
  • Qwen3-Coder-Next-APEX-I-Balanced.gguf (58GB)
• Control Model (Working): Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-Q8_K_P.gguf (43GB)
• Hardware/OS: MacOS 15.7.7, 16GB RAM+RX 580 (8GB)
• llama.cpp version: b9305 vulkan build

Reproduction Steps

1. Load Apex quantized model via llama-server.
2. Send a long context prompt; model responds normally.
3. Modify Context (Key Step): Delete the model's response and edit a previous message (e.g., fix a typo) to invalidate the KV Cache.
4. Regenerate: Request the model to regenerate the response.
5. Observe: The model tends to outputs "00000000" after regeneration.
6. (Note): In rare cases, the error appears later in the conversation as context grows.

Observations

1. Cache Invalidation Trigger: The error is highly reproducible when the prompt is modified, forcing the model to re-compute layers.
2. Randomness: It is probabilistic, not deterministic every single time.
3. Long Context Accumulation: In rare cases, the error manifests later in the conversation as the context length increases.
4. Comparison: The Q8_K_P model runs stably under the same mmap settings.

Additional Notes

I've ruled out parameter or hardware issues since Q8 works looks like fine.
The issue seems specific to Apex quantization combined with mmap re-computation.

Thanks for the great quantization work!

[localai-org-maint-bot]

The cache-invalidation trigger is useful, but the control also changes quantization format/size, so it does not yet isolate APEX from llama.cpp mmap/KV reuse. Please retest the same APEX file with --no-mmap and with a current llama.cpp build, then attach the exact server command and a minimal prompt/edit sequence. That will determine whether this belongs in the recipe or upstream runtime.