| { |
| "notes": [ |
| "$ /usr/bin/python3 -m pip install --no-cache-dir -q onnx onnxscript onnxruntime huggingface_hub safetensors torchvision tqdm -> rc 0", |
| "$ git clone -q https://github.com/VAST-AI-Research/TripoSplat.git /kaggle/working/TripoSplat -> rc 0", |
| "$ git -C /kaggle/working/TripoSplat checkout -q d8db9e0 -> rc 0", |
| "decoder ckpt triposplat_vae_decoder_fp16.safetensors", |
| "decoder params 288.1 M", |
| " octree 50.8 M", |
| " gs 237.3 M", |
| "layout {\"_xyz\": [0, 96], \"_features_dc\": [96, 192], \"_scaling\": [192, 288], \"_rotation\": [288, 416], \"_opacity\": [416, 448], \"_offset_scale\": [448, 480]}", |
| "octree ref logits (1, 2048, 8)", |
| "padding check: 8 rows alone vs inside 2048, max|diff| = 2.193450927734375e-05", |
| "octree fp32 203.3 MB {'Cast': 49, 'MatMul': 31, 'Add': 51, 'Constant': 115, 'Reshape': 15, 'Unsqueeze': 8, 'Mul': 77, 'Sin': 2, 'Cos': 2, 'Concat': 3, 'Gemm': 3, 'Sigmoid': 2, 'Shape': 13, 'Gather': 1, 'Div': 13, 'Slice': 10, 'LayerNormalization': 9, 'ReduceL2': 8, 'Clip': 8, 'Expand': 8, 'Split': 4, 'Squeeze': 8, 'Transpose': 16, 'Sqrt': 12, 'Softmax': 4, 'Tanh': 4}", |
| "octree io {\"inputs\": [[\"x\", [1, 2048, 3]], [\"l\", [1]], [\"cond\", [1, 8192, 16]]], \"outputs\": [[\"logits\", [1, 2048, 8]]]}", |
| "gs ref features (1, 2048, 480) in 38.4 s", |
| "gs fp32 949.8 MB {'Cast': 295, 'MatMul': 179, 'Add': 196, 'Constant': 665, 'Reshape': 83, 'Unsqueeze': 49, 'Mul': 290, 'Sin': 1, 'Cos': 1, 'Concat': 18, 'LayerNormalization': 49, 'Split': 48, 'Squeeze': 128, 'ReduceL2': 64, 'Clip': 64, 'Shape': 96, 'Expand': 64, 'Div': 96, 'Transpose': 128, 'Slice': 32, 'Sqrt': 96, 'Softmax': 32, 'Tanh': 16}", |
| "octree_level int4 57.9 MB {'Cast': 49, 'com.microsoft::MatMulNBits': 23, 'Add': 51, 'Constant': 115, 'Reshape': 15, 'Unsqueeze': 8, 'Mul': 77, 'Sin': 2, 'Cos': 2, 'Concat': 3, 'Gemm': 3, 'Sigmoid': 2, 'Shape': 13, 'Gather': 1, 'Div': 13, 'Slice': 10, 'LayerNormalization': 9, 'ReduceL2': 8, 'Clip': 8, 'Expand': 8, 'Split': 4, 'Squeeze': 8, 'Transpose': 16, 'Sqrt': 12, 'MatMul': 8, 'Softmax': 4, 'Tanh': 4}", |
| "gs int4 150.4 MB {'Cast': 295, 'com.microsoft::MatMulNBits': 115, 'Add': 196, 'Constant': 665, 'Reshape': 83, 'Unsqueeze': 49, 'Mul': 290, 'Sin': 1, 'Cos': 1, 'Concat': 18, 'LayerNormalization': 49, 'Split': 48, 'Squeeze': 128, 'ReduceL2': 64, 'Clip': 64, 'Shape': 96, 'Expand': 64, 'Div': 96, 'Transpose': 128, 'Slice': 32, 'Sqrt': 96, 'MatMul': 64, 'Softmax': 32, 'Tanh': 16}", |
| "octree parity {\"cos\": 0.946703228429031, \"rel_l2\": 0.3280873158581016, \"fp32_ms\": 6827.6, \"int4_ms\": 6328.3}", |
| "gs parity (64 tok) {\"cos\": 0.9991116642991249, \"rel_l2\": 0.042165539533916316, \"fp32_ms\": 2739.9, \"int4_ms\": 3162.6}", |
| "uploaded decoder_meta.json 0.0 MB", |
| "uploaded gs.int4.onnx 150.45 MB", |
| "uploaded log.txt 0.0 MB", |
| "uploaded octree_level.int4.onnx 57.87 MB" |
| ], |
| "graphs": { |
| "octree_level": { |
| "int4_mb": 57.87, |
| "fp32_mb": 203.34, |
| "ops": { |
| "Cast": 49, |
| "com.microsoft::MatMulNBits": 23, |
| "Add": 51, |
| "Constant": 115, |
| "Reshape": 15, |
| "Unsqueeze": 8, |
| "Mul": 77, |
| "Sin": 2, |
| "Cos": 2, |
| "Concat": 3, |
| "Gemm": 3, |
| "Sigmoid": 2, |
| "Shape": 13, |
| "Gather": 1, |
| "Div": 13, |
| "Slice": 10, |
| "LayerNormalization": 9, |
| "ReduceL2": 8, |
| "Clip": 8, |
| "Expand": 8, |
| "Split": 4, |
| "Squeeze": 8, |
| "Transpose": 16, |
| "Sqrt": 12, |
| "MatMul": 8, |
| "Softmax": 4, |
| "Tanh": 4 |
| }, |
| "parity": { |
| "cos": 0.946703228429031, |
| "rel_l2": 0.3280873158581016, |
| "fp32_ms": 6827.6, |
| "int4_ms": 6328.3 |
| } |
| }, |
| "gs": { |
| "int4_mb": 150.45, |
| "fp32_mb": 949.78, |
| "ops": { |
| "Cast": 295, |
| "com.microsoft::MatMulNBits": 115, |
| "Add": 196, |
| "Constant": 665, |
| "Reshape": 83, |
| "Unsqueeze": 49, |
| "Mul": 290, |
| "Sin": 1, |
| "Cos": 1, |
| "Concat": 18, |
| "LayerNormalization": 49, |
| "Split": 48, |
| "Squeeze": 128, |
| "ReduceL2": 64, |
| "Clip": 64, |
| "Shape": 96, |
| "Expand": 64, |
| "Div": 96, |
| "Transpose": 128, |
| "Slice": 32, |
| "Sqrt": 96, |
| "MatMul": 64, |
| "Softmax": 32, |
| "Tanh": 16 |
| }, |
| "parity_64tok": { |
| "cos": 0.9991116642991249, |
| "rel_l2": 0.042165539533916316, |
| "fp32_ms": 2739.9, |
| "int4_ms": 3162.6 |
| } |
| } |
| }, |
| "padding_max_abs": 2.193450927734375e-05 |
| } |
