File size: 1,702 Bytes
12f320c | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 | {
"format": "qwen35-9b-ttnn-native-text-v1",
"scope": "text-only-no-vision-with-mtp",
"precision_groups": {
"torch.bfloat16": {
"tensors": 193,
"parameters": 1261466880,
"serialized_tensor_bytes": 2522954504
},
"torch.float32": {
"tensors": 48,
"parameters": 3840,
"serialized_tensor_bytes": 20160
},
"bfp8": {
"tensors": 195,
"parameters": 5217714176,
"serialized_tensor_bytes": 5543891512
},
"bfp4": {
"tensors": 54,
"parameters": 2717908992,
"serialized_tensor_bytes": 1528843248
}
},
"artifact_bytes": 9629415772,
"tensor_parameters": 9197093888,
"quantizer": "TTNN native BFP4_B/BFP8_B rounding with selectively retained precision; not the Unsloth quantizer",
"calibration": "300941 chat-templated input tokens; activation second moments propose precision, validation output KL selects it",
"embedding": "Original BF16 retained",
"nonlinear": "All 48 original FP32 GDN norm/A_log tensors retained losslessly",
"limitations": [
"Single Blackhole P150; text-only with MTP-1, no vision",
"Native TT format, not GGUF or bitsandbytes",
"Local validation gates are not official Unsloth benchmark thresholds"
],
"mtp": {
"tensors": 15,
"serialized_bytes": 486582864,
"storage": "All original draft tensors preserved losslessly at source dtype",
"runtime": "Existing draft attention and MLP BFP4 policy retained; FC and norms BF16; shared embedding BF16 and shared target LM head BFP8",
"runtime_precision_binding": "Mode-specific equivalence runtime_environment and runtime_sources; target precision map does not apply to draft layers"
}
}
|