from nzfc_gram_runtime import NZFCGramLongMemoryChat
from nzfc_gram_runtime.nonquant import attach_nonquant_gemma
from nzfc_gram_runtime.cache_profiles import attach_adaptive_kv_cache_generation
from nzfc_gram_runtime.quality import attach_answer_quality_governor

MODEL_ID = 'google/gemma-4-E2B-it'

bot = NZFCGramLongMemoryChat(
    repo_dir='.',
    model_id=MODEL_ID,
    memory_db_path='./user_memory_adaptive_cache.sqlite3',
    load_model=False,
    require_model=False,
    preload_static_memory=True,
)

attach_nonquant_gemma(
    bot,
    model_id=MODEL_ID,
    device_map='balanced_low_0',
    gpu_max_memory_gib=11,
    gpu_max_memory_gib_candidates=[11, 10, 9, 8],
    cpu_max_memory_gib=48,
    prefer_bf16=True,
    use_fp32=False,
)

attach_adaptive_kv_cache_generation(bot, default_cache_policy='adaptive', default_prefer_cache=True)
attach_answer_quality_governor(bot)

out = bot.generate_answer(
    system_prompt='You are a concise assistant.',
    user_prompt='Explain memory as evidence in one sentence.',
    max_new_tokens=80,
    cache_policy='adaptive',
    prefer_cache=True,
)

print(out['answer'])
print('use_cache:', out.get('use_cache'))
print('fallback_used:', out.get('fallback_used'))
print('latency_s:', out.get('latency_s'))