| { | |
| "fp32": { | |
| "precision": "fp32", | |
| "embedding_compute_dtype": "float32", | |
| "classifier_compute_dtype": "float32", | |
| "examples": 1000, | |
| "load_seconds": 1.5235057300888002, | |
| "inference_seconds": 2.8799891732633114, | |
| "examples_per_second": 347.2235275339252, | |
| "peak_cuda_memory_mb": 4289.07666015625, | |
| "embedding_norm_mean": 1.0, | |
| "ndcg_at_5": 0.8548336029052734, | |
| "precision_at_5": 0.5113999843597412, | |
| "recall_at_5": 0.82269287109375, | |
| "top1_hit": 0.9129999876022339, | |
| "package_size_bytes": 1277686137 | |
| }, | |
| "fp16_storage_bf16_compute": { | |
| "precision": "fp16_storage_bf16_compute", | |
| "embedding_compute_dtype": "bfloat16", | |
| "classifier_compute_dtype": "float16", | |
| "examples": 1000, | |
| "load_seconds": 1.3328879470936954, | |
| "inference_seconds": 1.2113684518262744, | |
| "examples_per_second": 825.5126658552048, | |
| "peak_cuda_memory_mb": 2176.63623046875, | |
| "embedding_norm_mean": 1.0007888078689575, | |
| "ndcg_at_5": 0.8558993339538574, | |
| "precision_at_5": 0.5113999843597412, | |
| "recall_at_5": 0.8237928748130798, | |
| "top1_hit": 0.9139999747276306, | |
| "package_size_bytes": 658707940 | |
| }, | |
| "comparison": { | |
| "embedding_cosine_mean": 0.9999097585678101, | |
| "embedding_max_abs_difference": 0.005195870995521545, | |
| "probability_mae": 7.114602340152487e-05, | |
| "top1_agreement": 0.9900000095367432, | |
| "top5_overlap": 0.9907999999999999, | |
| "ndcg_at_5_delta_fp16_minus_fp32": 0.0010657310485839844, | |
| "speedup_fp16_over_fp32": 2.3774675400547234, | |
| "memory_reduction_fraction": 0.49251636122786024, | |
| "package_size_reduction_fraction": 0.48445246377436435 | |
| }, | |
| "pure_fp16_compute": { | |
| "supported": false, | |
| "error": "EmbeddingGemma produced non-finite embeddings on H100 when computed entirely in FP16.", | |
| "replacement": "FP16 weight storage with BF16 EmbeddingGemma compute and FP16 classifier compute" | |
| } | |
| } |