embeddinggemma-2-8bit / validation.json
prince-canuma's picture
Add 8bit MLX conversion of Google EmbeddingGemma 2
7505ef2 verified
Raw History Blame Contribute Delete
2.42 kB
{
"source_model": "google/embeddinggemma-2",
"source_revision": "914f7f89142e33e77833254d9c9b90c3cef7303b",
"mlx_vlm_revision": "3d87e88402f307efbf68e568971aa887ee7d9ed0",
"variant": "8bit",
"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": 1816,
"tensor_dtypes": [
"BF16",
"U32"
],
"weight_bytes": 1234238431,
"cases": {
"audio": {
"shape": [
1,
768
],
"finite": true,
"minimum_cosine_vs_torch_fp32": 0.9998035070404524,
"maximum_absolute_error_vs_torch_fp32": 0.002476184321109907,
"minimum_cosine_vs_source_bf16": 0.9997744741131873,
"minimum_truncated_cosine_vs_torch_fp32": 0.9998247233031456
},
"image": {
"shape": [
1,
768
],
"finite": true,
"minimum_cosine_vs_torch_fp32": 0.9998904477000761,
"maximum_absolute_error_vs_torch_fp32": 0.0019366481302732774,
"minimum_cosine_vs_source_bf16": 0.9999318244275229,
"minimum_truncated_cosine_vs_torch_fp32": 0.9998950619484888
},
"text": {
"shape": [
6,
768
],
"finite": true,
"minimum_cosine_vs_torch_fp32": 0.9998142728897385,
"maximum_absolute_error_vs_torch_fp32": 0.0022990193383245483,
"minimum_cosine_vs_source_bf16": 0.9997514709446282,
"minimum_truncated_cosine_vs_torch_fp32": 0.9998170484369469,
"retrieval_scores_venus_mars": [
0.678564727306366,
0.8342599868774414
]
},
"text_image": {
"shape": [
1,
768
],
"finite": true,
"minimum_cosine_vs_torch_fp32": 0.9998230728231623,
"maximum_absolute_error_vs_torch_fp32": 0.002104893632396264,
"minimum_cosine_vs_source_bf16": 0.999874734632146,
"minimum_truncated_cosine_vs_torch_fp32": 0.999830769497079
},
"video": {
"shape": [
1,
768
],
"finite": true,
"minimum_cosine_vs_torch_fp32": 0.9998359946491495,
"maximum_absolute_error_vs_torch_fp32": 0.0021304845535331423,
"minimum_cosine_vs_source_bf16": 0.9998825464984411,
"minimum_truncated_cosine_vs_torch_fp32": 0.9998420387979006
}
}
}