NF4 pre-quantized Robometer-4B reward rSkill (ADR-0057): bit-identical, meta-loadable
5a77c1f verified | { | |
| "scheme": "nf4", | |
| "quantizer": "bitsandbytes", | |
| "compute_dtype": "bfloat16", | |
| "double_quant": true, | |
| "min_params_to_quantize": 4000000, | |
| "modules_quantized": 236, | |
| "source_repo": "robometer/Robometer-4B", | |
| "source_revision": "beef63bc914c5c189329d49c6d712d96d632aa34", | |
| "transformers_pin": "4.57.1", | |
| "robometer_pin": "a669dffc241d7d76bec12f36efd4084d914d017c", | |
| "build_tool": "tools/build_robometer_nf4_checkpoint.py", | |
| "load_path": "meta-device + Params4bit.from_prequantized (no bf16 materialization, no requantize)", | |
| "determinism": "math SDP + use_deterministic_algorithms + CUBLAS_WORKSPACE_CONFIG=:4096:8 + cudnn.allow_tf32=False", | |
| "verified": "bit-identical to bf16+quantize path (same-process max|delta|=0; 4-bit dequant round-trip=0)", | |
| "resident_vram_gb": 3.32, | |
| "adr": "ADR-0057" | |
| } | |