embeddinggemma-2-mxfp8 / validation.json
prince-canuma's picture
Add mxfp8 MLX conversion of Google EmbeddingGemma 2
0536fa4 verified
Raw History Blame Contribute Delete
2.43 kB
{
"source_model": "google/embeddinggemma-2",
"source_revision": "914f7f89142e33e77833254d9c9b90c3cef7303b",
"mlx_vlm_revision": "3d87e88402f307efbf68e568971aa887ee7d9ed0",
"variant": "mxfp8",
"hardware": "Apple M3 Ultra",
"mlx_version": "0.32.3",
"transformers_version": "5.18.0.dev0",
"validation_scope": "Six multilingual text inputs and synthetic image, audio, two-frame video, and text+image; numerical smoke checks, not a retrieval benchmark.",
"tensor_count": 1596,
"tensor_dtypes": [
"BF16",
"U32",
"U8"
],
"weight_bytes": 1225720015,
"cases": {
"audio": {
"shape": [
1,
768
],
"finite": true,
"minimum_cosine_vs_torch_fp32": 0.9983353865128464,
"maximum_absolute_error_vs_torch_fp32": 0.006557946804956866,
"minimum_cosine_vs_source_bf16": 0.9981795014663992,
"minimum_truncated_cosine_vs_torch_fp32": 0.9984341378931347
},
"image": {
"shape": [
1,
768
],
"finite": true,
"minimum_cosine_vs_torch_fp32": 0.9994873791960924,
"maximum_absolute_error_vs_torch_fp32": 0.003609755933630241,
"minimum_cosine_vs_source_bf16": 0.9995206654001738,
"minimum_truncated_cosine_vs_torch_fp32": 0.9995135458227378
},
"text": {
"shape": [
6,
768
],
"finite": true,
"minimum_cosine_vs_torch_fp32": 0.9983049609155468,
"maximum_absolute_error_vs_torch_fp32": 0.006092039627625164,
"minimum_cosine_vs_source_bf16": 0.9982079929326719,
"minimum_truncated_cosine_vs_torch_fp32": 0.998331601337103,
"retrieval_scores_venus_mars": [
0.6707783937454224,
0.8356978297233582
]
},
"text_image": {
"shape": [
1,
768
],
"finite": true,
"minimum_cosine_vs_torch_fp32": 0.9992635574576849,
"maximum_absolute_error_vs_torch_fp32": 0.0048226723343564,
"minimum_cosine_vs_source_bf16": 0.9992596731314828,
"minimum_truncated_cosine_vs_torch_fp32": 0.9993020285279928
},
"video": {
"shape": [
1,
768
],
"finite": true,
"minimum_cosine_vs_torch_fp32": 0.9992312775532464,
"maximum_absolute_error_vs_torch_fp32": 0.007177200518493687,
"minimum_cosine_vs_source_bf16": 0.9992132664387819,
"minimum_truncated_cosine_vs_torch_fp32": 0.9992793313591022
}
}
}