| { |
| "evidence": { |
| "benchmark": "benchmark/summary.json", |
| "coverage": "benchmark/coverage/", |
| "numerical_smoke": "benchmark/audits/numerical-smoke.json", |
| "prompts": "prompts.json" |
| }, |
| "quantized_components": { |
| "transformer": { |
| "component": "transformer", |
| "cuda_peak_allocated_bytes": 139198464, |
| "cuda_peak_reserved_bytes": 201326592, |
| "generic_modules": { |
| "actual_group_size_counts": { |
| "-1": 1, |
| "64": 170 |
| }, |
| "count": 171, |
| "entries": [ |
| { |
| "group_size": -1, |
| "logical_numel": 131072, |
| "logical_shape": [ |
| 2048, |
| 64 |
| ], |
| "module": "patch_embedder", |
| "original_bytes": 262144, |
| "parameter": "patch_embedder.weight", |
| "re_quantize_for_matmul": false, |
| "stored_bytes": 73728, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 5242880, |
| "logical_shape": [ |
| 2048, |
| 2560 |
| ], |
| "module": "text_embedder.linear_1", |
| "original_bytes": 10485760, |
| "parameter": "text_embedder.linear_1.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2949120, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "text_embedder.linear_2", |
| "original_bytes": 8388608, |
| "parameter": "text_embedder.linear_2.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.0.attn.to_q", |
| "original_bytes": 8388608, |
| "parameter": "blocks.0.attn.to_q.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.0.attn.to_k", |
| "original_bytes": 8388608, |
| "parameter": "blocks.0.attn.to_k.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.0.attn.to_v", |
| "original_bytes": 8388608, |
| "parameter": "blocks.0.attn.to_v.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.0.attn.to_out", |
| "original_bytes": 8388608, |
| "parameter": "blocks.0.attn.to_out.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 6144, |
| 2048 |
| ], |
| "module": "blocks.0.ffn.gate_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.0.ffn.gate_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 6144, |
| 2048 |
| ], |
| "module": "blocks.0.ffn.up_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.0.ffn.up_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 2048, |
| 6144 |
| ], |
| "module": "blocks.0.ffn.down_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.0.ffn.down_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.1.attn.to_q", |
| "original_bytes": 8388608, |
| "parameter": "blocks.1.attn.to_q.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.1.attn.to_k", |
| "original_bytes": 8388608, |
| "parameter": "blocks.1.attn.to_k.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.1.attn.to_v", |
| "original_bytes": 8388608, |
| "parameter": "blocks.1.attn.to_v.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.1.attn.to_out", |
| "original_bytes": 8388608, |
| "parameter": "blocks.1.attn.to_out.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 6144, |
| 2048 |
| ], |
| "module": "blocks.1.ffn.gate_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.1.ffn.gate_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 6144, |
| 2048 |
| ], |
| "module": "blocks.1.ffn.up_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.1.ffn.up_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 2048, |
| 6144 |
| ], |
| "module": "blocks.1.ffn.down_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.1.ffn.down_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.2.attn.to_q", |
| "original_bytes": 8388608, |
| "parameter": "blocks.2.attn.to_q.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.2.attn.to_k", |
| "original_bytes": 8388608, |
| "parameter": "blocks.2.attn.to_k.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.2.attn.to_v", |
| "original_bytes": 8388608, |
| "parameter": "blocks.2.attn.to_v.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.2.attn.to_out", |
| "original_bytes": 8388608, |
| "parameter": "blocks.2.attn.to_out.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 6144, |
| 2048 |
| ], |
| "module": "blocks.2.ffn.gate_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.2.ffn.gate_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 6144, |
| 2048 |
| ], |
| "module": "blocks.2.ffn.up_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.2.ffn.up_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 2048, |
| 6144 |
| ], |
| "module": "blocks.2.ffn.down_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.2.ffn.down_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.3.attn.to_q", |
| "original_bytes": 8388608, |
| "parameter": "blocks.3.attn.to_q.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.3.attn.to_k", |
| "original_bytes": 8388608, |
| "parameter": "blocks.3.attn.to_k.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.3.attn.to_v", |
| "original_bytes": 8388608, |
| "parameter": "blocks.3.attn.to_v.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.3.attn.to_out", |
| "original_bytes": 8388608, |
| "parameter": "blocks.3.attn.to_out.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 6144, |
| 2048 |
| ], |
| "module": "blocks.3.ffn.gate_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.3.ffn.gate_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 6144, |
| 2048 |
| ], |
| "module": "blocks.3.ffn.up_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.3.ffn.up_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 2048, |
| 6144 |
| ], |
| "module": "blocks.3.ffn.down_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.3.ffn.down_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.4.attn.to_q", |
| "original_bytes": 8388608, |
| "parameter": "blocks.4.attn.to_q.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.4.attn.to_k", |
| "original_bytes": 8388608, |
| "parameter": "blocks.4.attn.to_k.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.4.attn.to_v", |
| "original_bytes": 8388608, |
| "parameter": "blocks.4.attn.to_v.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.4.attn.to_out", |
| "original_bytes": 8388608, |
| "parameter": "blocks.4.attn.to_out.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 6144, |
| 2048 |
| ], |
| "module": "blocks.4.ffn.gate_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.4.ffn.gate_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 6144, |
| 2048 |
| ], |
| "module": "blocks.4.ffn.up_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.4.ffn.up_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 2048, |
| 6144 |
| ], |
| "module": "blocks.4.ffn.down_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.4.ffn.down_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.5.attn.to_q", |
| "original_bytes": 8388608, |
| "parameter": "blocks.5.attn.to_q.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.5.attn.to_k", |
| "original_bytes": 8388608, |
| "parameter": "blocks.5.attn.to_k.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.5.attn.to_v", |
| "original_bytes": 8388608, |
| "parameter": "blocks.5.attn.to_v.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.5.attn.to_out", |
| "original_bytes": 8388608, |
| "parameter": "blocks.5.attn.to_out.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 6144, |
| 2048 |
| ], |
| "module": "blocks.5.ffn.gate_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.5.ffn.gate_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 6144, |
| 2048 |
| ], |
| "module": "blocks.5.ffn.up_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.5.ffn.up_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 2048, |
| 6144 |
| ], |
| "module": "blocks.5.ffn.down_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.5.ffn.down_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.6.attn.to_q", |
| "original_bytes": 8388608, |
| "parameter": "blocks.6.attn.to_q.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.6.attn.to_k", |
| "original_bytes": 8388608, |
| "parameter": "blocks.6.attn.to_k.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.6.attn.to_v", |
| "original_bytes": 8388608, |
| "parameter": "blocks.6.attn.to_v.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.6.attn.to_out", |
| "original_bytes": 8388608, |
| "parameter": "blocks.6.attn.to_out.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 6144, |
| 2048 |
| ], |
| "module": "blocks.6.ffn.gate_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.6.ffn.gate_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 6144, |
| 2048 |
| ], |
| "module": "blocks.6.ffn.up_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.6.ffn.up_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 2048, |
| 6144 |
| ], |
| "module": "blocks.6.ffn.down_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.6.ffn.down_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.7.attn.to_q", |
| "original_bytes": 8388608, |
| "parameter": "blocks.7.attn.to_q.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.7.attn.to_k", |
| "original_bytes": 8388608, |
| "parameter": "blocks.7.attn.to_k.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.7.attn.to_v", |
| "original_bytes": 8388608, |
| "parameter": "blocks.7.attn.to_v.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.7.attn.to_out", |
| "original_bytes": 8388608, |
| "parameter": "blocks.7.attn.to_out.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 6144, |
| 2048 |
| ], |
| "module": "blocks.7.ffn.gate_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.7.ffn.gate_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 6144, |
| 2048 |
| ], |
| "module": "blocks.7.ffn.up_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.7.ffn.up_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 2048, |
| 6144 |
| ], |
| "module": "blocks.7.ffn.down_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.7.ffn.down_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.8.attn.to_q", |
| "original_bytes": 8388608, |
| "parameter": "blocks.8.attn.to_q.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.8.attn.to_k", |
| "original_bytes": 8388608, |
| "parameter": "blocks.8.attn.to_k.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.8.attn.to_v", |
| "original_bytes": 8388608, |
| "parameter": "blocks.8.attn.to_v.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.8.attn.to_out", |
| "original_bytes": 8388608, |
| "parameter": "blocks.8.attn.to_out.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 6144, |
| 2048 |
| ], |
| "module": "blocks.8.ffn.gate_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.8.ffn.gate_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 6144, |
| 2048 |
| ], |
| "module": "blocks.8.ffn.up_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.8.ffn.up_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 2048, |
| 6144 |
| ], |
| "module": "blocks.8.ffn.down_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.8.ffn.down_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.9.attn.to_q", |
| "original_bytes": 8388608, |
| "parameter": "blocks.9.attn.to_q.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.9.attn.to_k", |
| "original_bytes": 8388608, |
| "parameter": "blocks.9.attn.to_k.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.9.attn.to_v", |
| "original_bytes": 8388608, |
| "parameter": "blocks.9.attn.to_v.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.9.attn.to_out", |
| "original_bytes": 8388608, |
| "parameter": "blocks.9.attn.to_out.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 6144, |
| 2048 |
| ], |
| "module": "blocks.9.ffn.gate_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.9.ffn.gate_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 6144, |
| 2048 |
| ], |
| "module": "blocks.9.ffn.up_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.9.ffn.up_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 2048, |
| 6144 |
| ], |
| "module": "blocks.9.ffn.down_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.9.ffn.down_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.10.attn.to_q", |
| "original_bytes": 8388608, |
| "parameter": "blocks.10.attn.to_q.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.10.attn.to_k", |
| "original_bytes": 8388608, |
| "parameter": "blocks.10.attn.to_k.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.10.attn.to_v", |
| "original_bytes": 8388608, |
| "parameter": "blocks.10.attn.to_v.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.10.attn.to_out", |
| "original_bytes": 8388608, |
| "parameter": "blocks.10.attn.to_out.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 6144, |
| 2048 |
| ], |
| "module": "blocks.10.ffn.gate_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.10.ffn.gate_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 6144, |
| 2048 |
| ], |
| "module": "blocks.10.ffn.up_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.10.ffn.up_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 2048, |
| 6144 |
| ], |
| "module": "blocks.10.ffn.down_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.10.ffn.down_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.11.attn.to_q", |
| "original_bytes": 8388608, |
| "parameter": "blocks.11.attn.to_q.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.11.attn.to_k", |
| "original_bytes": 8388608, |
| "parameter": "blocks.11.attn.to_k.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.11.attn.to_v", |
| "original_bytes": 8388608, |
| "parameter": "blocks.11.attn.to_v.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.11.attn.to_out", |
| "original_bytes": 8388608, |
| "parameter": "blocks.11.attn.to_out.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 6144, |
| 2048 |
| ], |
| "module": "blocks.11.ffn.gate_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.11.ffn.gate_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 6144, |
| 2048 |
| ], |
| "module": "blocks.11.ffn.up_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.11.ffn.up_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 2048, |
| 6144 |
| ], |
| "module": "blocks.11.ffn.down_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.11.ffn.down_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.12.attn.to_q", |
| "original_bytes": 8388608, |
| "parameter": "blocks.12.attn.to_q.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.12.attn.to_k", |
| "original_bytes": 8388608, |
| "parameter": "blocks.12.attn.to_k.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.12.attn.to_v", |
| "original_bytes": 8388608, |
| "parameter": "blocks.12.attn.to_v.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.12.attn.to_out", |
| "original_bytes": 8388608, |
| "parameter": "blocks.12.attn.to_out.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 6144, |
| 2048 |
| ], |
| "module": "blocks.12.ffn.gate_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.12.ffn.gate_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 6144, |
| 2048 |
| ], |
| "module": "blocks.12.ffn.up_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.12.ffn.up_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 2048, |
| 6144 |
| ], |
| "module": "blocks.12.ffn.down_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.12.ffn.down_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.13.attn.to_q", |
| "original_bytes": 8388608, |
| "parameter": "blocks.13.attn.to_q.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.13.attn.to_k", |
| "original_bytes": 8388608, |
| "parameter": "blocks.13.attn.to_k.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.13.attn.to_v", |
| "original_bytes": 8388608, |
| "parameter": "blocks.13.attn.to_v.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.13.attn.to_out", |
| "original_bytes": 8388608, |
| "parameter": "blocks.13.attn.to_out.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 6144, |
| 2048 |
| ], |
| "module": "blocks.13.ffn.gate_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.13.ffn.gate_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 6144, |
| 2048 |
| ], |
| "module": "blocks.13.ffn.up_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.13.ffn.up_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 2048, |
| 6144 |
| ], |
| "module": "blocks.13.ffn.down_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.13.ffn.down_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.14.attn.to_q", |
| "original_bytes": 8388608, |
| "parameter": "blocks.14.attn.to_q.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.14.attn.to_k", |
| "original_bytes": 8388608, |
| "parameter": "blocks.14.attn.to_k.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.14.attn.to_v", |
| "original_bytes": 8388608, |
| "parameter": "blocks.14.attn.to_v.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.14.attn.to_out", |
| "original_bytes": 8388608, |
| "parameter": "blocks.14.attn.to_out.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 6144, |
| 2048 |
| ], |
| "module": "blocks.14.ffn.gate_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.14.ffn.gate_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 6144, |
| 2048 |
| ], |
| "module": "blocks.14.ffn.up_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.14.ffn.up_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 2048, |
| 6144 |
| ], |
| "module": "blocks.14.ffn.down_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.14.ffn.down_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.15.attn.to_q", |
| "original_bytes": 8388608, |
| "parameter": "blocks.15.attn.to_q.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.15.attn.to_k", |
| "original_bytes": 8388608, |
| "parameter": "blocks.15.attn.to_k.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.15.attn.to_v", |
| "original_bytes": 8388608, |
| "parameter": "blocks.15.attn.to_v.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.15.attn.to_out", |
| "original_bytes": 8388608, |
| "parameter": "blocks.15.attn.to_out.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 6144, |
| 2048 |
| ], |
| "module": "blocks.15.ffn.gate_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.15.ffn.gate_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 6144, |
| 2048 |
| ], |
| "module": "blocks.15.ffn.up_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.15.ffn.up_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 2048, |
| 6144 |
| ], |
| "module": "blocks.15.ffn.down_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.15.ffn.down_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.16.attn.to_q", |
| "original_bytes": 8388608, |
| "parameter": "blocks.16.attn.to_q.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.16.attn.to_k", |
| "original_bytes": 8388608, |
| "parameter": "blocks.16.attn.to_k.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.16.attn.to_v", |
| "original_bytes": 8388608, |
| "parameter": "blocks.16.attn.to_v.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.16.attn.to_out", |
| "original_bytes": 8388608, |
| "parameter": "blocks.16.attn.to_out.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 6144, |
| 2048 |
| ], |
| "module": "blocks.16.ffn.gate_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.16.ffn.gate_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 6144, |
| 2048 |
| ], |
| "module": "blocks.16.ffn.up_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.16.ffn.up_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 2048, |
| 6144 |
| ], |
| "module": "blocks.16.ffn.down_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.16.ffn.down_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.17.attn.to_q", |
| "original_bytes": 8388608, |
| "parameter": "blocks.17.attn.to_q.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.17.attn.to_k", |
| "original_bytes": 8388608, |
| "parameter": "blocks.17.attn.to_k.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.17.attn.to_v", |
| "original_bytes": 8388608, |
| "parameter": "blocks.17.attn.to_v.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.17.attn.to_out", |
| "original_bytes": 8388608, |
| "parameter": "blocks.17.attn.to_out.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 6144, |
| 2048 |
| ], |
| "module": "blocks.17.ffn.gate_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.17.ffn.gate_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 6144, |
| 2048 |
| ], |
| "module": "blocks.17.ffn.up_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.17.ffn.up_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 2048, |
| 6144 |
| ], |
| "module": "blocks.17.ffn.down_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.17.ffn.down_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.18.attn.to_q", |
| "original_bytes": 8388608, |
| "parameter": "blocks.18.attn.to_q.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.18.attn.to_k", |
| "original_bytes": 8388608, |
| "parameter": "blocks.18.attn.to_k.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.18.attn.to_v", |
| "original_bytes": 8388608, |
| "parameter": "blocks.18.attn.to_v.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.18.attn.to_out", |
| "original_bytes": 8388608, |
| "parameter": "blocks.18.attn.to_out.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 6144, |
| 2048 |
| ], |
| "module": "blocks.18.ffn.gate_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.18.ffn.gate_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 6144, |
| 2048 |
| ], |
| "module": "blocks.18.ffn.up_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.18.ffn.up_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 2048, |
| 6144 |
| ], |
| "module": "blocks.18.ffn.down_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.18.ffn.down_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.19.attn.to_q", |
| "original_bytes": 8388608, |
| "parameter": "blocks.19.attn.to_q.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.19.attn.to_k", |
| "original_bytes": 8388608, |
| "parameter": "blocks.19.attn.to_k.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.19.attn.to_v", |
| "original_bytes": 8388608, |
| "parameter": "blocks.19.attn.to_v.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.19.attn.to_out", |
| "original_bytes": 8388608, |
| "parameter": "blocks.19.attn.to_out.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 6144, |
| 2048 |
| ], |
| "module": "blocks.19.ffn.gate_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.19.ffn.gate_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 6144, |
| 2048 |
| ], |
| "module": "blocks.19.ffn.up_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.19.ffn.up_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 2048, |
| 6144 |
| ], |
| "module": "blocks.19.ffn.down_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.19.ffn.down_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.20.attn.to_q", |
| "original_bytes": 8388608, |
| "parameter": "blocks.20.attn.to_q.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.20.attn.to_k", |
| "original_bytes": 8388608, |
| "parameter": "blocks.20.attn.to_k.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.20.attn.to_v", |
| "original_bytes": 8388608, |
| "parameter": "blocks.20.attn.to_v.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.20.attn.to_out", |
| "original_bytes": 8388608, |
| "parameter": "blocks.20.attn.to_out.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 6144, |
| 2048 |
| ], |
| "module": "blocks.20.ffn.gate_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.20.ffn.gate_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 6144, |
| 2048 |
| ], |
| "module": "blocks.20.ffn.up_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.20.ffn.up_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 2048, |
| 6144 |
| ], |
| "module": "blocks.20.ffn.down_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.20.ffn.down_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.21.attn.to_q", |
| "original_bytes": 8388608, |
| "parameter": "blocks.21.attn.to_q.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.21.attn.to_k", |
| "original_bytes": 8388608, |
| "parameter": "blocks.21.attn.to_k.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.21.attn.to_v", |
| "original_bytes": 8388608, |
| "parameter": "blocks.21.attn.to_v.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.21.attn.to_out", |
| "original_bytes": 8388608, |
| "parameter": "blocks.21.attn.to_out.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 6144, |
| 2048 |
| ], |
| "module": "blocks.21.ffn.gate_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.21.ffn.gate_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 6144, |
| 2048 |
| ], |
| "module": "blocks.21.ffn.up_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.21.ffn.up_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 2048, |
| 6144 |
| ], |
| "module": "blocks.21.ffn.down_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.21.ffn.down_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.22.attn.to_q", |
| "original_bytes": 8388608, |
| "parameter": "blocks.22.attn.to_q.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.22.attn.to_k", |
| "original_bytes": 8388608, |
| "parameter": "blocks.22.attn.to_k.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.22.attn.to_v", |
| "original_bytes": 8388608, |
| "parameter": "blocks.22.attn.to_v.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.22.attn.to_out", |
| "original_bytes": 8388608, |
| "parameter": "blocks.22.attn.to_out.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 6144, |
| 2048 |
| ], |
| "module": "blocks.22.ffn.gate_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.22.ffn.gate_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 6144, |
| 2048 |
| ], |
| "module": "blocks.22.ffn.up_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.22.ffn.up_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 2048, |
| 6144 |
| ], |
| "module": "blocks.22.ffn.down_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.22.ffn.down_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.23.attn.to_q", |
| "original_bytes": 8388608, |
| "parameter": "blocks.23.attn.to_q.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.23.attn.to_k", |
| "original_bytes": 8388608, |
| "parameter": "blocks.23.attn.to_k.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.23.attn.to_v", |
| "original_bytes": 8388608, |
| "parameter": "blocks.23.attn.to_v.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 4194304, |
| "logical_shape": [ |
| 2048, |
| 2048 |
| ], |
| "module": "blocks.23.attn.to_out", |
| "original_bytes": 8388608, |
| "parameter": "blocks.23.attn.to_out.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 2359296, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 6144, |
| 2048 |
| ], |
| "module": "blocks.23.ffn.gate_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.23.ffn.gate_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 6144, |
| 2048 |
| ], |
| "module": "blocks.23.ffn.up_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.23.ffn.up_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| }, |
| { |
| "group_size": 64, |
| "logical_numel": 12582912, |
| "logical_shape": [ |
| 2048, |
| 6144 |
| ], |
| "module": "blocks.23.ffn.down_proj", |
| "original_bytes": 25165824, |
| "parameter": "blocks.23.ffn.down_proj.weight", |
| "re_quantize_for_matmul": true, |
| "stored_bytes": 7077888, |
| "use_quantized_matmul": false, |
| "weights_dtype": "uint4", |
| "wrapper_class": "SDNQLinear" |
| } |
| ], |
| "logical_params": 1318191104, |
| "original_parameter_bytes": 2636382208, |
| "stored_bytes": 741482496 |
| }, |
| "lingbot_video_code_revision": "a2bb04b78edd848500dc27a26e035a95442ae186", |
| "moe_experts": { |
| "compression_ratio": 0.0, |
| "entries": [], |
| "logical_numel": 0, |
| "original_bytes": 0, |
| "stored_bytes": 0, |
| "tensor_count": 0 |
| }, |
| "quantization_seconds": 1.5678673509974033, |
| "quantized": { |
| "logical_params": 1318191104, |
| "original_byte_coverage_percent": 94.42866487012878, |
| "original_parameter_bytes_covered": 2636382208, |
| "parameter_coverage_percent": 97.12783780249565, |
| "stored_bytes": 741482496 |
| }, |
| "recipe": { |
| "expert_runtime": "SDNQ dequantize to BF16 plus SGLang Triton fused MoE", |
| "generic_use_quantized_matmul": false, |
| "group_size": 0, |
| "name": "uint4-static-transformer-only-with-3d-expert-adapter", |
| "quant_conv": false, |
| "quant_embedding": false, |
| "use_dynamic_quantization": false, |
| "use_hadamard": false, |
| "use_svd": false, |
| "weights_dtype": "uint4" |
| }, |
| "saved_component_bytes": 897107220, |
| "sdnq_describe": "v0.2.1-17-gd841c38", |
| "sdnq_revision": "d841c383ff7be38728d4df829e17af4f15d4fd66", |
| "source_repo_id": "robbyant/lingbot-video-dense-1.3b", |
| "source_revision": "f9789a7d9b4772a47aba62d4eb5282ddefd1da21", |
| "torch": "2.8.0+cu128", |
| "total": { |
| "logical_params": 1357171264, |
| "original_parameter_bytes": 2791929984 |
| }, |
| "unquantized": { |
| "largest_parameters": [ |
| { |
| "bytes": 100663296, |
| "dtype": "float32", |
| "numel": 25165824, |
| "parameter": "time_modulation.1.weight", |
| "shape": [ |
| 12288, |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 33554432, |
| "dtype": "float32", |
| "numel": 8388608, |
| "parameter": "norm_out_modulation.1.weight", |
| "shape": [ |
| 4096, |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 16777216, |
| "dtype": "float32", |
| "numel": 4194304, |
| "parameter": "time_embedder.linear_2.weight", |
| "shape": [ |
| 2048, |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 2097152, |
| "dtype": "float32", |
| "numel": 524288, |
| "parameter": "time_embedder.linear_1.weight", |
| "shape": [ |
| 2048, |
| 256 |
| ] |
| }, |
| { |
| "bytes": 262144, |
| "dtype": "bfloat16", |
| "numel": 131072, |
| "parameter": "proj_out.weight", |
| "shape": [ |
| 64, |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 49152, |
| "dtype": "float32", |
| "numel": 12288, |
| "parameter": "time_modulation.1.bias", |
| "shape": [ |
| 12288 |
| ] |
| }, |
| { |
| "bytes": 49152, |
| "dtype": "float32", |
| "numel": 12288, |
| "parameter": "blocks.0.scale_shift_table", |
| "shape": [ |
| 1, |
| 12288 |
| ] |
| }, |
| { |
| "bytes": 49152, |
| "dtype": "float32", |
| "numel": 12288, |
| "parameter": "blocks.1.scale_shift_table", |
| "shape": [ |
| 1, |
| 12288 |
| ] |
| }, |
| { |
| "bytes": 49152, |
| "dtype": "float32", |
| "numel": 12288, |
| "parameter": "blocks.2.scale_shift_table", |
| "shape": [ |
| 1, |
| 12288 |
| ] |
| }, |
| { |
| "bytes": 49152, |
| "dtype": "float32", |
| "numel": 12288, |
| "parameter": "blocks.3.scale_shift_table", |
| "shape": [ |
| 1, |
| 12288 |
| ] |
| }, |
| { |
| "bytes": 49152, |
| "dtype": "float32", |
| "numel": 12288, |
| "parameter": "blocks.4.scale_shift_table", |
| "shape": [ |
| 1, |
| 12288 |
| ] |
| }, |
| { |
| "bytes": 49152, |
| "dtype": "float32", |
| "numel": 12288, |
| "parameter": "blocks.5.scale_shift_table", |
| "shape": [ |
| 1, |
| 12288 |
| ] |
| }, |
| { |
| "bytes": 49152, |
| "dtype": "float32", |
| "numel": 12288, |
| "parameter": "blocks.6.scale_shift_table", |
| "shape": [ |
| 1, |
| 12288 |
| ] |
| }, |
| { |
| "bytes": 49152, |
| "dtype": "float32", |
| "numel": 12288, |
| "parameter": "blocks.7.scale_shift_table", |
| "shape": [ |
| 1, |
| 12288 |
| ] |
| }, |
| { |
| "bytes": 49152, |
| "dtype": "float32", |
| "numel": 12288, |
| "parameter": "blocks.8.scale_shift_table", |
| "shape": [ |
| 1, |
| 12288 |
| ] |
| }, |
| { |
| "bytes": 49152, |
| "dtype": "float32", |
| "numel": 12288, |
| "parameter": "blocks.9.scale_shift_table", |
| "shape": [ |
| 1, |
| 12288 |
| ] |
| }, |
| { |
| "bytes": 49152, |
| "dtype": "float32", |
| "numel": 12288, |
| "parameter": "blocks.10.scale_shift_table", |
| "shape": [ |
| 1, |
| 12288 |
| ] |
| }, |
| { |
| "bytes": 49152, |
| "dtype": "float32", |
| "numel": 12288, |
| "parameter": "blocks.11.scale_shift_table", |
| "shape": [ |
| 1, |
| 12288 |
| ] |
| }, |
| { |
| "bytes": 49152, |
| "dtype": "float32", |
| "numel": 12288, |
| "parameter": "blocks.12.scale_shift_table", |
| "shape": [ |
| 1, |
| 12288 |
| ] |
| }, |
| { |
| "bytes": 49152, |
| "dtype": "float32", |
| "numel": 12288, |
| "parameter": "blocks.13.scale_shift_table", |
| "shape": [ |
| 1, |
| 12288 |
| ] |
| }, |
| { |
| "bytes": 49152, |
| "dtype": "float32", |
| "numel": 12288, |
| "parameter": "blocks.14.scale_shift_table", |
| "shape": [ |
| 1, |
| 12288 |
| ] |
| }, |
| { |
| "bytes": 49152, |
| "dtype": "float32", |
| "numel": 12288, |
| "parameter": "blocks.15.scale_shift_table", |
| "shape": [ |
| 1, |
| 12288 |
| ] |
| }, |
| { |
| "bytes": 49152, |
| "dtype": "float32", |
| "numel": 12288, |
| "parameter": "blocks.16.scale_shift_table", |
| "shape": [ |
| 1, |
| 12288 |
| ] |
| }, |
| { |
| "bytes": 49152, |
| "dtype": "float32", |
| "numel": 12288, |
| "parameter": "blocks.17.scale_shift_table", |
| "shape": [ |
| 1, |
| 12288 |
| ] |
| }, |
| { |
| "bytes": 49152, |
| "dtype": "float32", |
| "numel": 12288, |
| "parameter": "blocks.18.scale_shift_table", |
| "shape": [ |
| 1, |
| 12288 |
| ] |
| }, |
| { |
| "bytes": 49152, |
| "dtype": "float32", |
| "numel": 12288, |
| "parameter": "blocks.19.scale_shift_table", |
| "shape": [ |
| 1, |
| 12288 |
| ] |
| }, |
| { |
| "bytes": 49152, |
| "dtype": "float32", |
| "numel": 12288, |
| "parameter": "blocks.20.scale_shift_table", |
| "shape": [ |
| 1, |
| 12288 |
| ] |
| }, |
| { |
| "bytes": 49152, |
| "dtype": "float32", |
| "numel": 12288, |
| "parameter": "blocks.21.scale_shift_table", |
| "shape": [ |
| 1, |
| 12288 |
| ] |
| }, |
| { |
| "bytes": 49152, |
| "dtype": "float32", |
| "numel": 12288, |
| "parameter": "blocks.22.scale_shift_table", |
| "shape": [ |
| 1, |
| 12288 |
| ] |
| }, |
| { |
| "bytes": 49152, |
| "dtype": "float32", |
| "numel": 12288, |
| "parameter": "blocks.23.scale_shift_table", |
| "shape": [ |
| 1, |
| 12288 |
| ] |
| }, |
| { |
| "bytes": 16384, |
| "dtype": "float32", |
| "numel": 4096, |
| "parameter": "norm_out_modulation.1.bias", |
| "shape": [ |
| 4096 |
| ] |
| }, |
| { |
| "bytes": 10240, |
| "dtype": "float32", |
| "numel": 2560, |
| "parameter": "text_embedder.norm.weight", |
| "shape": [ |
| 2560 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "time_embedder.linear_1.bias", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "time_embedder.linear_2.bias", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.0.norm1.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.0.norm_post_attn.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.0.norm2.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.0.norm_post_ffn.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.1.norm1.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.1.norm_post_attn.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.1.norm2.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.1.norm_post_ffn.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.2.norm1.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.2.norm_post_attn.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.2.norm2.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.2.norm_post_ffn.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.3.norm1.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.3.norm_post_attn.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.3.norm2.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.3.norm_post_ffn.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.4.norm1.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.4.norm_post_attn.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.4.norm2.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.4.norm_post_ffn.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.5.norm1.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.5.norm_post_attn.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.5.norm2.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.5.norm_post_ffn.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.6.norm1.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.6.norm_post_attn.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.6.norm2.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.6.norm_post_ffn.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.7.norm1.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.7.norm_post_attn.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.7.norm2.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.7.norm_post_ffn.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.8.norm1.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.8.norm_post_attn.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.8.norm2.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.8.norm_post_ffn.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.9.norm1.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.9.norm_post_attn.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.9.norm2.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.9.norm_post_ffn.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.10.norm1.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.10.norm_post_attn.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.10.norm2.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.10.norm_post_ffn.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.11.norm1.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.11.norm_post_attn.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.11.norm2.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.11.norm_post_ffn.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.12.norm1.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.12.norm_post_attn.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.12.norm2.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.12.norm_post_ffn.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.13.norm1.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.13.norm_post_attn.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.13.norm2.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.13.norm_post_ffn.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.14.norm1.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.14.norm_post_attn.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.14.norm2.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.14.norm_post_ffn.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.15.norm1.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.15.norm_post_attn.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.15.norm2.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.15.norm_post_ffn.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.16.norm1.weight", |
| "shape": [ |
| 2048 |
| ] |
| }, |
| { |
| "bytes": 8192, |
| "dtype": "float32", |
| "numel": 2048, |
| "parameter": "blocks.16.norm_post_attn.weight", |
| "shape": [ |
| 2048 |
| ] |
| } |
| ], |
| "logical_params": 38980160, |
| "original_parameter_bytes": 155547776, |
| "parameter_count": 206 |
| } |
| } |
| }, |
| "recipe": { |
| "expert_runtime": "SDNQ dequantize to BF16 plus SGLang Triton fused MoE", |
| "generic_use_quantized_matmul": false, |
| "group_size": 0, |
| "name": "uint4-static-transformer-only-with-3d-expert-adapter", |
| "quant_conv": false, |
| "quant_embedding": false, |
| "use_dynamic_quantization": false, |
| "use_hadamard": false, |
| "use_svd": false, |
| "weights_dtype": "uint4" |
| }, |
| "runtime": { |
| "adapter": "lingbot_sdnq_runtime", |
| "generic_linear": "eager BF16 dequantization plus torch.nn.functional.linear", |
| "moe_experts": "packed SDNQ UINT4, per-call BF16 dequantization, SGLang Triton fused MoE", |
| "persistent_bf16_expert_cache": false |
| }, |
| "schema_version": 1, |
| "source": { |
| "lingbot_video_repo": "https://github.com/Robbyant/lingbot-video", |
| "lingbot_video_revision": "a2bb04b78edd848500dc27a26e035a95442ae186", |
| "repo_id": "robbyant/lingbot-video-dense-1.3b", |
| "revision": "f9789a7d9b4772a47aba62d4eb5282ddefd1da21", |
| "sdnq_describe": "v0.2.1-17-gd841c38", |
| "sdnq_repo": "https://github.com/Disty0/sdnq", |
| "sdnq_revision": "d841c383ff7be38728d4df829e17af4f15d4fd66" |
| }, |
| "target_repo_id": "WaveCut/LingBot-Video-Dense-1.3B-SDNQ-uint4-static", |
| "unquantized_components": { |
| "scheduler": "copied byte-for-byte from the pinned source snapshot", |
| "text_encoder": "copied byte-for-byte from the pinned source snapshot", |
| "tokenizer_or_processor": "copied byte-for-byte from the pinned source snapshot", |
| "vae": "copied byte-for-byte from the pinned source snapshot" |
| } |
| } |
|
|
