{ "description": "Complete six-layer Baberu decoder using an INT32 token ID and a quantized embedding Gather instead of a full-vocabulary one-hot MatMul", "fp32": { "graph": { "bytes": 133619451, "sha256": "2742aceae936dd15a3b5e0aa632f91d7548442c7205f60e76cff948b39995fe2", "opsets": { "ai.onnx": 17 }, "operators": { "Add": 57, "Cast": 76, "Concat": 74, "Constant": 541, "ConstantOfShape": 13, "Div": 38, "Equal": 13, "Expand": 13, "Gather": 78, "LessOrEqual": 1, "MatMul": 55, "Mul": 149, "Neg": 12, "Range": 2, "ReduceMean": 25, "Reshape": 49, "Shape": 88, "Sigmoid": 6, "Slice": 31, "Softmax": 6, "Sqrt": 25, "Sub": 1, "Tanh": 1, "Transpose": 30, "Unsqueeze": 195, "Where": 14 }, "inputs": { "vision_embeds": [ 1, "vision_len", 512 ], "token_ids": [ 1, 1 ], "position_ids": [ 1, "input_len" ], "past_k0": [ 1, 2, "past_len", 64 ], "past_k1": [ 1, 2, "past_len", 64 ], "past_k2": [ 1, 2, "past_len", 64 ], "past_k3": [ 1, 2, "past_len", 64 ], "past_k4": [ 1, 2, "past_len", 64 ], "past_k5": [ 1, 2, "past_len", 64 ], "past_v0": [ 1, 2, "past_len", 64 ], "past_v1": [ 1, 2, "past_len", 64 ], "past_v2": [ 1, 2, "past_len", 64 ], "past_v3": [ 1, 2, "past_len", 64 ], "past_v4": [ 1, 2, "past_len", 64 ], "past_v5": [ 1, 2, "past_len", 64 ] }, "outputs": { "logits": [ "Mullogits_dim_0", "Mullogits_dim_1", 14630 ], "present_k0": [ 1, 2, "total_len", 64 ], "present_k1": [ 1, 2, "total_len", 64 ], "present_k2": [ 1, 2, "total_len", 64 ], "present_k3": [ 1, 2, "total_len", 64 ], "present_k4": [ 1, 2, "total_len", 64 ], "present_k5": [ 1, 2, "total_len", 64 ], "present_v0": [ 1, 2, "total_len", 64 ], "present_v1": [ 1, 2, "total_len", 64 ], "present_v2": [ 1, 2, "total_len", 64 ], "present_v3": [ 1, 2, "total_len", 64 ], "present_v4": [ 1, 2, "total_len", 64 ], "present_v5": [ 1, 2, "total_len", 64 ] } }, "parity": { "prefill_zero_past": { "max_abs_by_output": [ 3.0517578125e-05, 2.384185791015625e-07, 4.76837158203125e-07, 9.5367431640625e-07, 9.5367431640625e-07, 1.1920928955078125e-06, 1.4007091522216797e-06, 1.043081283569336e-07, 8.940696716308594e-08, 2.086162567138672e-07, 1.9371509552001953e-07, 1.862645149230957e-07, 2.086162567138672e-07 ], "max_abs": 3.0517578125e-05, "top_token": 2 }, "step_empty_vision": { "max_abs_by_output": [ 3.4809112548828125e-05, 1.3113021850585938e-06, 4.76837158203125e-07, 1.430511474609375e-06, 1.430511474609375e-06, 1.9073486328125e-06, 2.1457672119140625e-06, 2.980232238769531e-07, 1.341104507446289e-07, 2.039596438407898e-07, 3.5762786865234375e-07, 1.9371509552001953e-07, 2.551823854446411e-07 ], "max_abs": 3.4809112548828125e-05, "top_token": 2 } } }, "qdq_int8": { "graph": { "bytes": 34647907, "sha256": "3612d5787ffb1026ae09bfd4a16b3379ea75fe0dd803484f7bb41645be6b6239", "opsets": { "ai.onnx": 17 }, "operators": { "Add": 57, "Cast": 76, "Concat": 74, "Constant": 541, "ConstantOfShape": 13, "DequantizeLinear": 44, "Div": 38, "Equal": 13, "Expand": 13, "Gather": 78, "LessOrEqual": 1, "MatMul": 55, "Mul": 149, "Neg": 12, "Range": 2, "ReduceMean": 25, "Reshape": 49, "Shape": 88, "Sigmoid": 6, "Slice": 31, "Softmax": 6, "Sqrt": 25, "Sub": 1, "Tanh": 1, "Transpose": 30, "Unsqueeze": 195, "Where": 14 }, "inputs": { "vision_embeds": [ 1, "vision_len", 512 ], "token_ids": [ 1, 1 ], "position_ids": [ 1, "input_len" ], "past_k0": [ 1, 2, "past_len", 64 ], "past_k1": [ 1, 2, "past_len", 64 ], "past_k2": [ 1, 2, "past_len", 64 ], "past_k3": [ 1, 2, "past_len", 64 ], "past_k4": [ 1, 2, "past_len", 64 ], "past_k5": [ 1, 2, "past_len", 64 ], "past_v0": [ 1, 2, "past_len", 64 ], "past_v1": [ 1, 2, "past_len", 64 ], "past_v2": [ 1, 2, "past_len", 64 ], "past_v3": [ 1, 2, "past_len", 64 ], "past_v4": [ 1, 2, "past_len", 64 ], "past_v5": [ 1, 2, "past_len", 64 ] }, "outputs": { "logits": [ "Mullogits_dim_0", "Mullogits_dim_1", 14630 ], "present_k0": [ 1, 2, "total_len", 64 ], "present_k1": [ 1, 2, "total_len", 64 ], "present_k2": [ 1, 2, "total_len", 64 ], "present_k3": [ 1, 2, "total_len", 64 ], "present_k4": [ 1, 2, "total_len", 64 ], "present_k5": [ 1, 2, "total_len", 64 ], "present_v0": [ 1, 2, "total_len", 64 ], "present_v1": [ 1, 2, "total_len", 64 ], "present_v2": [ 1, 2, "total_len", 64 ], "present_v3": [ 1, 2, "total_len", 64 ], "present_v4": [ 1, 2, "total_len", 64 ], "present_v5": [ 1, 2, "total_len", 64 ] } }, "quantization": { "source": "decoder_unified_gather_fp32.onnx", "bytes": 34647907, "sha256": "3612d5787ffb1026ae09bfd4a16b3379ea75fe0dd803484f7bb41645be6b6239", "quantized_matmuls": 43, "quantized_gathers": 1, "quantized_elements": 33069056, "operators": { "Add": 57, "Cast": 76, "Concat": 74, "Constant": 541, "ConstantOfShape": 13, "DequantizeLinear": 44, "Div": 38, "Equal": 13, "Expand": 13, "Gather": 78, "LessOrEqual": 1, "MatMul": 55, "Mul": 149, "Neg": 12, "Range": 2, "ReduceMean": 25, "Reshape": 49, "Shape": 88, "Sigmoid": 6, "Slice": 31, "Softmax": 6, "Sqrt": 25, "Sub": 1, "Tanh": 1, "Transpose": 30, "Unsqueeze": 195, "Where": 14 }, "forbidden_webgpu_operators": [] }, "parity": { "prefill_zero_past": { "max_abs_by_output": [ 0.27939605712890625, 0.00832289457321167, 0.010223627090454102, 0.018473267555236816, 0.0254666805267334, 0.026123523712158203, 0.03216394782066345, 0.0031301602721214294, 0.0027603209018707275, 0.005134910345077515, 0.007790625095367432, 0.005478542298078537, 0.007550535723567009 ], "max_abs": 0.27939605712890625, "top_token": 2 }, "step_empty_vision": { "max_abs_by_output": [ 0.38232338428497314, 0.004899706691503525, 0.007703900337219238, 0.02177775651216507, 0.027161240577697754, 0.014369070529937744, 0.026971250772476196, 0.001852184534072876, 0.0019404590129852295, 0.005673632025718689, 0.004118740558624268, 0.00417923741042614, 0.005888845771551132 ], "max_abs": 0.38232338428497314, "top_token": 2 } } } }