File size: 1,362 Bytes
ceda511 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 | {
"source": "openjev/OpenJev-Flash-9B@61c569f2d9d25f3446d4e585eb35cc532dca7d80",
"source_index_sha256": "941305ff9f77551e145a6cea976ef456cd5cb208cbece99f168376c752fcf96c",
"official_fp8_reference": "Qwen/Qwen3.8-27B-FP8@017b9c7af6b5689d5dd426a76e0bc077eb5ca20a",
"official_fp8_config_sha256": "74227dd615bf1ea975aa676bdf355a0379858c12f394b5365cd9dfa5fc2c70bc",
"note": "Qwen publishes no FP8 build of Qwen3.5-9B; the quantized-tensor set and quantization_config of the reference were mapped onto the 9B by layer type",
"scheme": "fp8 e4m3, dynamic activations, 128x128 weight blocks, scale_inv = absmax/448 (bf16)",
"quantized_tensors": 200,
"tensors_total": 960,
"worst_relative_dequant_error": {
"value": 0.03721373900771141,
"tensor": "model.language_model.layers.1.linear_attn.in_proj_qkv.weight"
},
"script_sha256": "fce98903a0332e3f775ee115fcaeaabac15eabaa1a0678eacac57e9467e339b5",
"quantizer_sha256": "06f75dacb52a879988023c66aed4c8b661b0910c576857a305edbac70711643d",
"selfcheck": [
"SELFCHECK model.language_model.layers.0.linear_attn.in_proj_qkv.weight: scale dtype torch.bfloat16 shape (80, 40) max rel err 0.00e+00, identical q bytes 0.99939",
"SELFCHECK model.language_model.layers.38.mlp.down_proj.weight: scale dtype torch.bfloat16 shape (40, 136) max rel err 0.00e+00, identical q bytes 0.99911"
],
"torch": "2.14.1+cpu"
} |