| { |
| "fp32": { |
| "precision": "fp32", |
| "embedding_compute_dtype": "float32", |
| "classifier_compute_dtype": "float32", |
| "examples": 1000, |
| "load_seconds": 1.4385369773954153, |
| "inference_seconds": 3.0497249690815806, |
| "examples_per_second": 327.89842039465884, |
| "peak_cuda_memory_mb": 11189.17529296875, |
| "embedding_norm_mean": 1.0, |
| "ndcg_at_5": 0.7557224631309509, |
| "precision_at_5": 0.5272000432014465, |
| "recall_at_5": 0.7159669399261475, |
| "top1_hit": 0.8550000190734863, |
| "package_size_bytes": 1292690408 |
| }, |
| "fp16_storage_bf16_compute": { |
| "precision": "fp16_storage_bf16_compute", |
| "embedding_compute_dtype": "bfloat16", |
| "classifier_compute_dtype": "float16", |
| "examples": 1000, |
| "load_seconds": 1.3710060780867934, |
| "inference_seconds": 1.2324184626340866, |
| "examples_per_second": 811.4127062513074, |
| "peak_cuda_memory_mb": 6952.33935546875, |
| "embedding_norm_mean": 1.0007050037384033, |
| "ndcg_at_5": 0.7562801837921143, |
| "precision_at_5": 0.5278000235557556, |
| "recall_at_5": 0.7169753313064575, |
| "top1_hit": 0.8560000658035278, |
| "package_size_bytes": 662363316 |
| }, |
| "comparison": { |
| "embedding_cosine_mean": 0.9999027252197266, |
| "embedding_max_abs_difference": 0.004987679421901703, |
| "probability_mae": 5.631342719425447e-05, |
| "top1_agreement": 0.9909999966621399, |
| "top5_overlap": 0.9926, |
| "ndcg_at_5_delta_fp16_minus_fp32": 0.0005577206611633301, |
| "speedup_fp16_over_fp32": 2.4745855904846703, |
| "memory_reduction_fraction": 0.3786548897989307, |
| "package_size_reduction_fraction": 0.48760870205203843 |
| }, |
| "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" |
| } |
| } |
|
|
