0xSero

/

gemma-4-21b-a4b-it-REAP

like 90

clanker below

Environment

• Model: gemma-4-19b-reap-Q4_K_M.gguf (converted from bf16 via convert_hf_to_gguf.py)
• llama.cpp build: b8660-d00685831
• Hardware: Apple M2 Max
• Command:

llama-cli -m ~/models/gemma-4-19b-reap-Q4_K_M.gguf --jinja -cnv

• Chat template: the model's provided chat_template.jinja, loaded via --jinja (embedded in GGUF during conversion)
• All sampling settings at llama.cpp defaults (temp 0.8, top_p 0.95, top_k 40, repeat_penalty 1.1)

Issue
The model enters repetition loops, spamming the same token or phrase until the context limit is hit.

Two observed loop types:

1. On the first prompt of a session: leaks <|channel>0<|channel>thought then spams 000000...
2. General repetition: model starts repeating a word or phrase indefinitely (e.g. about about about about...)

Reproduction

1. Start a fresh session
2. Send any prompt (e.g. "give me a fun fact about the roman empire")
3. Model outputs [Start thinking] then immediately spams <|channel>0<|channel>thought000000...
4. Send a follow-up — model responds correctly