| { | |
| "model_name": "MiniCPM4-0.5B-RKLLM", | |
| "versions": [ | |
| { | |
| "name": "GPTQ-32", | |
| "file": "MiniCPM4-0.5B-QAT-Int4-GPTQ-format_w4a16_g32_RK3576.rkllm", | |
| "size": "471MB", | |
| "embedding_support": "perfect", | |
| "description": "GPTQ量化版本,完美支持embedding输入" | |
| }, | |
| { | |
| "name": "Per-Channel", | |
| "file": "Minicpm4-0.5B-QAT-Int4-perChannel_w4a16_RK3576.rkllm", | |
| "size": "414MB", | |
| "embedding_support": "token_only", | |
| "description": "Per-Channel量化版本,仅推荐token输入" | |
| }, | |
| { | |
| "name": "FP16-Original", | |
| "file": "MiniCPM4-0.5B-fp16-v1.2.2b1.rkllm", | |
| "size": "996MB", | |
| "embedding_support": "inconsistent", | |
| "description": "原始未量化版本,embedding输入不一致" | |
| } | |
| ], | |
| "test_results": { | |
| "gptq_cosine_similarity": 1.0, | |
| "gptq_top50_overlap": "100%", | |
| "per_channel_issue": "generates_eos_tokens", | |
| "original_cosine_similarity": 0.59 | |
| }, | |
| "hardware": "RK3576", | |
| "date": "2024-01" | |
| } |