Last active
July 18, 2024 15:22
-
-
Save gregtatum/573021bf5e7cc8aa53dea832eec2d1df to your computer and use it in GitHub Desktop.
Analyze Marian Models - https://docs.google.com/document/d/18PAGLEiG2Emplbkm-XLxu5iPqO17g2tTlRPo9M6VAT4/edit
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| import argparse | |
| import gzip | |
| import numpy as np | |
| import yaml | |
| import struct | |
| """ | |
| This script takes in a Marian model, either of the npz format or the bin format, and | |
| outputs information about the weights encoded in the file. | |
| Usage: | |
| python analyze-model.py en-ca/model.enca.intgemm.alphas.bin | |
| python analyze-model.py en-it/teacher-1/model.npz | |
| """ | |
| def analyze_npz_file(model_path: str): | |
| if model_path.endswith('.gz'): | |
| with gzip.open(model_path, 'rb') as f: | |
| model = np.load(f) | |
| else: | |
| model = np.load(model_path) | |
| print("Contents of the NPZ file:") | |
| for array_name in model.files: | |
| print(array_name) | |
| for array_name in model.files: | |
| array = model[array_name] | |
| print(f"\nArray name: {array_name}") | |
| print(f"Shape: {array.shape}") | |
| print(f"Data type: {array.dtype}") | |
| # print(f"Contents: \n{array}") | |
| # The yaml is encoded | |
| yaml_bytes = model.get('special:model.yml') | |
| if yaml_bytes: | |
| print_yaml(yaml_bytes) | |
| else: | |
| print("No model.yml was found.") | |
| class Header: | |
| def __init__(self, name_length, data_type, shape_length, data_length): | |
| self.name_length: int = name_length | |
| self.data_type: int = data_type | |
| self.shape_length: int = shape_length | |
| self.data_length: int = data_length | |
| class TypeClass: | |
| signed_type = 0x0100 | |
| unsigned_type = 0x0200 | |
| float_type = 0x0400 | |
| packed_type = 0x0800 | |
| avx2_type = 0x1000 | |
| avx512_type = 0x2000 | |
| intgemm_type = 0x4000 | |
| ssse3_type = 0x0010 # Not in browsermt/marian-dev | |
| sse2_type = 0x0020 # Not in browsermt/marian-dev | |
| # https://github.com/marian-nmt/marian-dev/blob/2d067afb9ce5e3a0b6c32585706affc6e7295920/src/common/types.h#L298 | |
| type_mapping = { | |
| TypeClass.signed_type + 1: "int8", | |
| TypeClass.signed_type + 2: "int16", | |
| TypeClass.signed_type + 4: "int32", | |
| TypeClass.signed_type + 8: "int64", | |
| TypeClass.unsigned_type + 1: "uint8", | |
| TypeClass.unsigned_type + 2: "uint16", | |
| TypeClass.unsigned_type + 4: "uint32", | |
| TypeClass.unsigned_type + 8: "uint64", | |
| TypeClass.float_type + 2: "float16", | |
| TypeClass.float_type + 4: "float32", | |
| TypeClass.float_type + 8: "float64", | |
| TypeClass.packed_type + 2: "packed16", | |
| TypeClass.packed_type + 1 + TypeClass.avx2_type: "packed8avx2", | |
| TypeClass.packed_type + 1 + TypeClass.avx512_type: "packed8avx512", | |
| TypeClass.intgemm_type + 1: "intgemm8", | |
| TypeClass.intgemm_type + 2: "intgemm16", | |
| # The following is not in browsermt/marian-dev | |
| TypeClass.intgemm_type + 1 + TypeClass.ssse3_type: "intgemm8ssse3", | |
| TypeClass.intgemm_type + 1 + TypeClass.avx2_type: "intgemm8avx2", | |
| TypeClass.intgemm_type + 1 + TypeClass.avx512_type: "intgemm8avx512", | |
| TypeClass.intgemm_type + 1 + TypeClass.avx512_type + 4096: "intgemm8avx512vnni", | |
| TypeClass.intgemm_type + 2 + TypeClass.sse2_type: "intgemm16sse2", | |
| TypeClass.intgemm_type + 2 + TypeClass.avx2_type: "intgemm16avx2", | |
| TypeClass.intgemm_type + 2 + TypeClass.avx512_type: "intgemm16avx512" | |
| } | |
| def analyze_bin_file(model_path: str): | |
| if model_path.endswith('.gz'): | |
| with gzip.open(model_path, 'rb') as file: | |
| print("gzip open") | |
| analyze_bin_file_impl(file) | |
| else: | |
| with open(model_path, 'rb') as file: | |
| analyze_bin_file_impl(file) | |
| def print_yaml(data): | |
| bytes = bytearray(data) | |
| yaml_str = bytes.replace(b'\x00', b'').decode('utf-8') | |
| yaml_content = yaml.safe_load(yaml_str) | |
| # Display the YAML content | |
| print("\nYAML:\n") | |
| print(yaml.dump(yaml_content, default_flow_style=False)) | |
| def analyze_bin_file_impl(file): | |
| # Read the binary file version | |
| # The format character for Q is: "unsigned long long" | |
| # https://docs.python.org/3/library/struct.html#format-characters | |
| binary_file_version = struct.unpack('Q', file.read(struct.calcsize('Q')))[0] | |
| print(f'Binary File Version: {binary_file_version}') | |
| # Read the number of headers | |
| header_size = struct.unpack('Q', file.read(struct.calcsize('Q')))[0] | |
| headers: list[Header] = [] | |
| for _ in range(header_size): | |
| header_data = file.read(struct.calcsize('QQQQ')) | |
| header = Header(*struct.unpack('QQQQ', header_data)) | |
| headers.append(header) | |
| # Read all names | |
| names: list[str] = [] | |
| for header in headers: | |
| name = file.read(header.name_length).decode('utf-8').strip('\x00') | |
| names.append(name) | |
| # Read all shapes | |
| shapes = [] | |
| for header in headers: | |
| shape = struct.unpack(f'{header.shape_length}i', file.read(header.shape_length * struct.calcsize('i'))) | |
| shapes.append(shape) | |
| # TODO - This could be incorrect, as it was ChatGPT generated. The next value | |
| # could be read to determine the offset amount. The data is padded to a 256 offset. | |
| current_pos = file.tell() | |
| next_pos = ((current_pos + struct.calcsize('Q')) // 256 + 1) * 256 | |
| offset = next_pos - current_pos - struct.calcsize('Q') | |
| file.seek(offset + struct.calcsize('Q'), 1) | |
| data = [] | |
| for header in headers: | |
| data.append(file.read(header.data_length)) | |
| print("Data:\n") | |
| for header, name, shape, vector in zip(headers, names, shapes, data): | |
| data_sample = " ".join([f"{n:02x}" for n in vector[:20]]) | |
| type_name = type_mapping.get(header.data_type, "unknown") | |
| print( | |
| name.ljust(35), type_name.ljust(20), f"{shape}".ljust(15), data_sample,) | |
| if name.startswith("special:") and name.endswith(".yml"): | |
| print_yaml(vector) | |
| def main() -> None: | |
| parser = argparse.ArgumentParser( | |
| description=__doc__, | |
| # Preserves whitespace in the help text. | |
| formatter_class=argparse.RawTextHelpFormatter, | |
| ) | |
| parser.add_argument("model") | |
| args = parser.parse_args() | |
| model_path: str = args.model | |
| if model_path.endswith(".bin") or model_path.endswith(".bin.gz"): | |
| print("Analyzing the file as a Marian bin file") | |
| analyze_bin_file(model_path) | |
| else: | |
| print("Analyzing the file as a numpy npz file") | |
| analyze_npz_file(model_path) | |
| if __name__ == "__main__": | |
| main() |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| ➤ python analyze-model.py en-it/bicleaner-fix/exported/model.enit.intgemm.alphas.bin.gz | |
| Analyzing the file as a Marian bin file | |
| gzip open | |
| Binary File Version: 1 | |
| Data: | |
| Wemb intgemm8 (32000, 256) 07 03 fb fe 05 07 06 f9 04 04 04 fb 03 fd 01 0d 06 f2 ff fe | |
| Wemb_QuantMultA intgemm8 (1, 1) 7f 15 dd 78 41 07 06 f9 04 04 04 fb 03 fd 01 0d 06 f2 ff fe | |
| decoder_ff_logit_out_b float32 (1, 32000) 5b 55 b2 c0 57 d9 a6 c1 ce 24 c4 40 8d 46 d1 c0 17 bc 07 40 | |
| decoder_l1_context_Wk intgemm8 (256, 256) d0 fb 1f d2 06 19 0b ee f7 ef 0e d8 d1 fa 00 ce fa 26 f2 e7 | |
| decoder_l1_context_Wk_QuantMultA float32 (1, 1) af 1a 46 41 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 | |
| decoder_l1_context_Wo intgemm8 (256, 256) fb b4 ee df 1a 0e 1c 03 09 1b 12 13 26 eb 1c 11 10 e4 e0 dd | |
| decoder_l1_context_Wo_QuantMultA float32 (1, 1) df 96 a1 41 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 | |
| decoder_l1_context_Wo_ln_bias float32 (1, 256) eb fc 14 3e 7b 7f d4 3e a0 4f e5 3d 4a 83 8a 3e 90 a2 e1 3d | |
| decoder_l1_context_Wo_ln_scale float32 (1, 256) 4b 1f 96 3f 99 30 54 3f bf df 91 3f 01 8c 98 3f 74 55 93 3f | |
| decoder_l1_context_Wq intgemm8 (256, 256) ff 13 04 0b fd fc 1d 05 f6 fa 18 1b 0e 01 18 3a fd 13 01 f5 | |
| decoder_l1_context_Wq_QuantMultA float32 (1, 1) a4 69 76 41 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 | |
| decoder_l1_context_Wv intgemm8 (256, 256) 18 d7 18 dd 00 f3 0d f8 03 16 01 fd fd fd de f9 01 e7 14 df | |
| decoder_l1_context_Wv_QuantMultA float32 (1, 1) af 1a 46 41 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 | |
| decoder_l1_context_bk float32 (1, 256) 14 d7 4e 40 a0 92 74 40 d5 c0 6c c0 a9 8f 4d 40 73 4c 80 40 | |
| decoder_l1_context_bo float32 (1, 256) 4a c1 05 bf 89 f1 5c c0 1a 3f 91 be 5b b9 87 c0 23 06 1b 3e | |
| decoder_l1_context_bq float32 (1, 256) a1 80 86 be 0d e2 42 c0 9b 01 28 40 78 08 eb 3e fe cd 9e bf | |
| decoder_l1_context_bv float32 (1, 256) 1f ab 4b bf ed ac 35 3f 34 a5 5e 3f 88 69 a0 bf 75 35 7b bf | |
| decoder_l1_ffn_W1 intgemm8 (256, 1536) 07 05 fb 02 06 fd 04 03 0a 03 01 07 f6 03 f8 02 02 f9 01 04 | |
| decoder_l1_ffn_W1_QuantMultA float32 (1, 1) b3 04 5b 41 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 | |
| decoder_l1_ffn_W2 intgemm8 (1536, 256) 02 00 07 f1 01 06 02 f8 f7 03 fc 06 05 fd 09 07 fc 07 fe 0c | |
| decoder_l1_ffn_W2_QuantMultA float32 (1, 1) 62 d8 26 41 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 | |
| decoder_l1_ffn_b1 float32 (1, 1536) 16 bc 09 c0 bf e0 87 c0 67 9a 00 c0 df a0 8e c0 e4 5f 53 be | |
| decoder_l1_ffn_b2 float32 (1, 256) b4 21 cd bf 8b 9e 94 3f 0a b1 d4 3f bc 2f a2 c0 88 5a 0e c0 | |
| decoder_l1_ffn_ffn_ln_bias float32 (1, 256) 1c 51 2f 3e 4d be b2 bd 19 03 a5 bd 24 bd e3 3d 8c ef 03 3e | |
| decoder_l1_ffn_ffn_ln_scale float32 (1, 256) 7b 8d 08 40 b8 21 a8 3f ab f3 be 3f 7f ca b3 3f c9 11 d6 3f | |
| decoder_l1_rnn_W intgemm8 (256, 256) db 0b 02 f7 01 16 09 08 f6 03 f1 0c 00 f5 ee fb f2 13 f4 e7 | |
| decoder_l1_rnn_W_QuantMultA float32 (1, 1) df 1f 4b 41 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 | |
| decoder_l1_rnn_Wf intgemm8 (256, 256) f6 f2 06 01 0b 0a 06 ff 15 0d 1f 14 27 e7 f1 0d 07 f9 06 f2 | |
| decoder_l1_rnn_Wf_QuantMultA float32 (1, 1) df 1f 4b 41 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 | |
| decoder_l1_rnn_bf float32 (1, 256) c5 d0 31 c0 24 19 64 40 04 af 0d 40 15 92 56 40 9e 74 ad bf | |
| decoder_l1_rnn_ffn_ln_bias float32 (1, 256) 9c e3 25 3e 7c f0 e2 3d a4 26 34 3d b6 97 1d 3e f0 10 30 bb | |
| decoder_l1_rnn_ffn_ln_scale float32 (1, 256) 85 e6 3e 3f ed 0c 1d 3f d0 85 47 3f 91 da 89 3f 16 c3 4f 3f | |
| decoder_l2_context_Wk intgemm8 (256, 256) 0a 17 07 03 f2 f8 1b f3 ff 10 f7 06 08 ea db ff f4 f4 07 27 | |
| decoder_l2_context_Wk_QuantMultA float32 (1, 1) af 1a 46 41 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 | |
| decoder_l2_context_Wo intgemm8 (256, 256) f2 04 ce 06 f8 f1 0c 17 06 ff e6 0a 15 12 0c 1e 35 1f f5 1b | |
| decoder_l2_context_Wo_QuantMultA float32 (1, 1) 1f f2 c2 41 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 | |
| decoder_l2_context_Wo_ln_bias float32 (1, 256) 84 8d b8 bd a5 14 82 be 2f 25 9d be 45 24 e6 3c 30 d9 b9 3d | |
| decoder_l2_context_Wo_ln_scale float32 (1, 256) a7 ca b6 3f c2 13 c2 3f 56 cc 8e 3f 28 8b cc 3f dd b9 b5 3f | |
| decoder_l2_context_Wq intgemm8 (256, 256) 02 03 ed ff 09 fb fe e5 13 0e 10 04 f0 fe ff 1a 05 1b 06 e4 | |
| decoder_l2_context_Wq_QuantMultA float32 (1, 1) 25 13 1e 41 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 | |
| decoder_l2_context_Wv intgemm8 (256, 256) 18 ef 0e ee f8 f8 07 ff 24 e5 d5 04 08 fa 06 f6 fc 0d f3 e8 | |
| decoder_l2_context_Wv_QuantMultA float32 (1, 1) af 1a 46 41 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 | |
| decoder_l2_context_bk float32 (1, 256) f3 85 46 41 df 4b 38 40 4c 0a 51 c1 db e1 47 be 26 20 66 41 | |
| decoder_l2_context_bo float32 (1, 256) 7d a0 6d bf f9 cb ec 3f 52 9d 5f 3f 1a b8 02 c0 d2 29 b7 bf | |
| decoder_l2_context_bq float32 (1, 256) f5 1b 95 40 4f 55 49 3d 37 49 1b c0 db b2 c3 3f c2 80 59 c0 | |
| decoder_l2_context_bv float32 (1, 256) 53 ed c8 bd b3 af 89 be 58 82 83 be d4 16 6d 3d 7a 77 1d 3e | |
| decoder_l2_ffn_W1 intgemm8 (256, 1536) f8 f8 02 f7 f2 fb 07 fb 01 07 03 01 f4 fe f8 fb fb f2 09 fe | |
| decoder_l2_ffn_W1_QuantMultA float32 (1, 1) 81 1f 69 41 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 | |
| decoder_l2_ffn_W2 intgemm8 (1536, 256) fc fd 01 06 fd 00 f7 fe fc f7 fb ff 01 fe f6 04 07 03 02 ff | |
| decoder_l2_ffn_W2_QuantMultA float32 (1, 1) 13 8b 45 41 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 | |
| decoder_l2_ffn_b1 float32 (1, 1536) ae 9c 88 c0 14 d2 b8 bf b7 b5 da bf 4c 52 18 c0 e7 70 b5 bf | |
| decoder_l2_ffn_b2 float32 (1, 256) 17 4d 79 40 ec c3 2a bf 9d 7a 9c be 0e c7 f2 bf 52 2b da bf | |
| decoder_l2_ffn_ffn_ln_bias float32 (1, 256) e6 0d e3 3d ed be f3 be c9 32 83 be 69 43 b1 be 74 4d b6 3e | |
| decoder_l2_ffn_ffn_ln_scale float32 (1, 256) ff 9c 3c 40 37 90 38 40 e5 e6 3b 40 c7 35 5e 40 47 ae 53 40 | |
| decoder_l2_rnn_W intgemm8 (256, 256) ef ff 05 f3 fb 0d fa 05 00 f9 f9 10 01 0c 00 fe 08 ff fd fe | |
| decoder_l2_rnn_W_QuantMultA float32 (1, 1) a3 a3 26 41 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 | |
| decoder_l2_rnn_Wf intgemm8 (256, 256) 1b fe 03 ee 26 10 08 f9 03 fa 0a 17 f9 05 ff 04 0c 19 08 28 | |
| decoder_l2_rnn_Wf_QuantMultA float32 (1, 1) a3 a3 26 41 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 | |
| decoder_l2_rnn_bf float32 (1, 256) 11 79 ba 3c a7 be 93 3d 21 a3 b9 40 32 25 da bf 50 43 0c c0 | |
| decoder_l2_rnn_ffn_ln_bias float32 (1, 256) 68 7a 75 ba e9 ef 39 be 59 87 3f 3d a6 26 47 3e 6b 2e ee bd | |
| decoder_l2_rnn_ffn_ln_scale float32 (1, 256) 42 36 78 3f 9b 34 d8 3f 0f 58 59 3f 01 6f b0 3f 37 27 ac 3f | |
| encoder_l1_ffn_W1 intgemm8 (256, 1536) 05 e6 1b f7 fb f2 0a ef 19 fd ff 0c f9 25 16 f9 f1 0a 0c 02 | |
| encoder_l1_ffn_W1_QuantMultA float32 (1, 1) ea 5a ba 41 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 | |
| encoder_l1_ffn_W2 intgemm8 (1536, 256) f5 16 2b ff fc ff eb 04 0d 0c e5 fb 0e fd 05 0e f2 f4 10 ff | |
| encoder_l1_ffn_W2_QuantMultA float32 (1, 1) 37 77 21 41 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 | |
| encoder_l1_ffn_b1 float32 (1, 1536) 07 f3 ac bf f2 d9 22 c0 34 ee ac bf d9 38 e1 3e 56 af 57 bf | |
| encoder_l1_ffn_b2 float32 (1, 256) e4 31 c1 be fa 5b a2 bf 03 5d 4a bf 41 40 31 3f 86 20 17 3e | |
| encoder_l1_ffn_ffn_ln_bias float32 (1, 256) 4a ea e8 3c f8 5f 8e 3d a6 97 b7 3d 12 7a 22 bd 43 9d a6 bb | |
| encoder_l1_ffn_ffn_ln_scale float32 (1, 256) 69 b5 72 3f bd c0 7a 3f 5a d7 79 3f 33 cb 82 3f d7 96 7b 3f | |
| encoder_l1_self_Wk intgemm8 (256, 256) 07 0f f5 03 06 01 18 04 0e 21 0c ff f6 ef ff fc f7 05 fe 11 | |
| encoder_l1_self_Wk_QuantMultA float32 (1, 1) 48 7c 23 41 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 | |
| encoder_l1_self_Wo intgemm8 (256, 256) 25 ed ec ca ce 4a ae 0a 1f 1f b8 e7 00 e2 d4 27 56 1f 21 44 | |
| encoder_l1_self_Wo_QuantMultA float32 (1, 1) 7c 0a 74 41 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 | |
| encoder_l1_self_Wo_ln_bias float32 (1, 256) 78 ae 13 3b 13 ff ed 3d 04 af a7 bc 30 1a ba bd 87 87 f2 3b | |
| encoder_l1_self_Wo_ln_scale float32 (1, 256) b5 c6 cd 3e 96 3f f9 3e 35 fb 00 3f d2 9d 0c 3f b8 6c 1b 3f | |
| encoder_l1_self_Wq intgemm8 (256, 256) 00 09 00 02 07 13 03 02 08 09 1d 08 06 01 fe ec 05 1c f4 08 | |
| encoder_l1_self_Wq_QuantMultA float32 (1, 1) 48 7c 23 41 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 | |
| encoder_l1_self_Wv intgemm8 (256, 256) 02 fb 00 0d 13 04 0d d2 0d dd 0d e3 20 14 d8 08 d1 26 fc 31 | |
| encoder_l1_self_Wv_QuantMultA float32 (1, 1) 48 7c 23 41 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 | |
| encoder_l1_self_bk float32 (1, 256) 74 9e ae 3e 50 42 b6 bd fd 48 92 3e 54 4c 9d 3d 72 62 82 3c | |
| encoder_l1_self_bo float32 (1, 256) 16 06 41 3f 76 4b 6b bf f4 b6 47 bd 04 40 2d 3f 40 5d de 3e | |
| encoder_l1_self_bq float32 (1, 256) e3 cd d5 3d 5f 76 b6 bd 79 3c ec be 87 f8 0b 3f 6f 5a 38 3f | |
| encoder_l1_self_bv float32 (1, 256) c9 85 2e 3f 44 13 05 c0 be b5 0e 3f 96 9f d8 be 60 41 a8 be | |
| encoder_l2_ffn_W1 intgemm8 (256, 1536) f4 0f fe 0f 33 de 23 01 1b d8 01 07 ee 1b 0f f3 e5 01 13 fa | |
| encoder_l2_ffn_W1_QuantMultA float32 (1, 1) 72 9e 4b 41 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 | |
| encoder_l2_ffn_W2 intgemm8 (1536, 256) f9 f3 fa 12 02 fe ff 17 0d d6 ef d9 16 fa e5 02 0d f5 05 f0 | |
| encoder_l2_ffn_W2_QuantMultA float32 (1, 1) 6f fa 6c 41 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 | |
| encoder_l2_ffn_b1 float32 (1, 1536) dc 9c 43 bf 09 eb 88 bf e0 cf a3 bf 38 62 eb be 0e 22 85 be | |
| encoder_l2_ffn_b2 float32 (1, 256) b6 f5 92 3e d9 50 80 3e 21 da c8 bf 16 54 e0 bf 16 7f 95 bf | |
| encoder_l2_ffn_ffn_ln_bias float32 (1, 256) 42 92 e0 bc cb 26 b3 3c f7 e7 07 3e 76 b5 82 3d 65 8c 6b 3d | |
| encoder_l2_ffn_ffn_ln_scale float32 (1, 256) 5c 6e 64 3f 0f 09 77 3f 83 2d 82 3f 79 57 7b 3f 8e b0 7f 3f | |
| encoder_l2_self_Wk intgemm8 (256, 256) e2 10 f7 fd fc 0a ee fe 0f 0c 0e f0 fb 15 06 01 f8 01 e2 18 | |
| encoder_l2_self_Wk_QuantMultA float32 (1, 1) 4a eb 5e 41 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 | |
| encoder_l2_self_Wo intgemm8 (256, 256) fe f0 04 22 13 cb b0 36 dd f8 04 fa fc 51 2e 0e f1 e7 fa ce | |
| encoder_l2_self_Wo_QuantMultA float32 (1, 1) 40 8f 0d 42 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 | |
| encoder_l2_self_Wo_ln_bias float32 (1, 256) 7b ca a3 bb 8f fd b7 bd ae 3f c3 3c c8 96 9e 3c 1c 6e d3 3d | |
| encoder_l2_self_Wo_ln_scale float32 (1, 256) 57 f5 74 3f 10 5f 80 3f c5 61 83 3f a1 a3 7d 3f 91 ab 83 3f | |
| encoder_l2_self_Wq intgemm8 (256, 256) ce f4 02 f1 22 25 1f 02 f8 1d f7 f4 e2 fa 4a df 02 f1 ef 0c | |
| encoder_l2_self_Wq_QuantMultA float32 (1, 1) 4a eb 5e 41 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 | |
| encoder_l2_self_Wv intgemm8 (256, 256) c3 1d 0f 2f d5 05 07 e7 e6 09 dd 16 29 dc b9 04 38 24 e3 d1 | |
| encoder_l2_self_Wv_QuantMultA float32 (1, 1) 4a eb 5e 41 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 | |
| encoder_l2_self_bk float32 (1, 256) 50 83 35 3e 83 51 83 bd 47 e5 22 be 76 a6 91 be f3 39 cc 3e | |
| encoder_l2_self_bo float32 (1, 256) af 01 85 be 47 50 40 3f 3f 93 62 be 0c ce 51 bf 6b 28 62 be | |
| encoder_l2_self_bq float32 (1, 256) ca 1e ca 3e 93 03 ae be 0d c6 02 3f d1 07 c5 3f 29 3c 9f bf | |
| encoder_l2_self_bv float32 (1, 256) 00 a3 c4 3e 07 ae 3c 3e 99 0f 72 3e 09 8f 6e be 4b ca 6d 3c | |
| encoder_l3_ffn_W1 intgemm8 (256, 1536) 20 19 04 d5 fe ff 1a fc d3 ea ee 0d e0 fb 00 07 13 ff f5 fa | |
| encoder_l3_ffn_W1_QuantMultA float32 (1, 1) cf 95 63 41 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 | |
| encoder_l3_ffn_W2 intgemm8 (1536, 256) 06 f5 17 07 03 0a 03 00 13 f9 fd f3 ed ea f2 01 08 01 16 ff | |
| encoder_l3_ffn_W2_QuantMultA float32 (1, 1) 26 2e 6d 41 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 | |
| encoder_l3_ffn_b1 float32 (1, 1536) 22 4e ea bf 5c 70 23 c0 8c 88 fe 3e cb 87 37 c0 0d 27 23 c0 | |
| encoder_l3_ffn_b2 float32 (1, 256) 8d 1c 09 40 91 2f 59 be 68 c1 7f 3f eb ad fb be ba d8 af bf | |
| encoder_l3_ffn_ffn_ln_bias float32 (1, 256) d2 c1 55 bd ee fb 91 3c ef 8d bd 3d 1c 23 ad 3c f0 e4 b6 3d | |
| encoder_l3_ffn_ffn_ln_scale float32 (1, 256) f2 56 9b 3f 9b c7 a0 3f 50 a8 96 3f 25 fc 9c 3f 24 42 a0 3f | |
| encoder_l3_self_Wk intgemm8 (256, 256) f9 eb 05 04 f7 00 e1 ee ea 1d fc 16 22 1a df 34 0d 1c 53 20 | |
| encoder_l3_self_Wk_QuantMultA float32 (1, 1) 51 24 67 41 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 | |
| encoder_l3_self_Wo intgemm8 (256, 256) 28 d9 f9 05 08 e8 24 44 de 3a e1 22 fd 2e e3 fc 0e fe 29 aa | |
| encoder_l3_self_Wo_QuantMultA float32 (1, 1) b3 0c 0b 42 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 | |
| encoder_l3_self_Wo_ln_bias float32 (1, 256) 7b 66 3f bc bd c4 e6 bd dc fa 93 3e 02 2f 2f 3e f2 88 98 3e | |
| encoder_l3_self_Wo_ln_scale float32 (1, 256) 6e 48 83 3f 89 33 84 3f a8 56 83 3f 3e 21 77 3f 6c 2a 86 3f | |
| encoder_l3_self_Wq intgemm8 (256, 256) eb 19 fb 15 de f2 15 0b 15 fd 16 e3 0d 03 0c 2f e3 0e 10 11 | |
| encoder_l3_self_Wq_QuantMultA float32 (1, 1) 51 24 67 41 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 | |
| encoder_l3_self_Wv intgemm8 (256, 256) 16 e9 0c 14 14 1e d3 fc 04 10 f9 ef 14 e3 23 21 2e 0d 0c 09 | |
| encoder_l3_self_Wv_QuantMultA float32 (1, 1) 51 24 67 41 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 | |
| encoder_l3_self_bk float32 (1, 256) 76 cd a9 3d 64 e0 7b be e7 e6 79 3d ab 5e d7 3d 2d f3 5a 3e | |
| encoder_l3_self_bo float32 (1, 256) 6f 4c f0 3d 3b e8 5e 3f 27 79 89 bf f5 65 a4 bf 67 4d 44 bf | |
| encoder_l3_self_bq float32 (1, 256) c5 15 40 3f e8 d8 0a bf 58 85 38 bf d1 70 34 bf 94 d7 5f be | |
| encoder_l3_self_bv float32 (1, 256) 6a 87 3e be ee b9 08 be e7 12 93 be f9 a8 57 bf bd c4 d8 3e | |
| encoder_l4_ffn_W1 intgemm8 (256, 1536) 0e f5 02 fb 02 e4 0f 06 08 13 1d 28 fe 0a fb dc 12 3b 37 eb | |
| encoder_l4_ffn_W1_QuantMultA float32 (1, 1) 5a 21 2f 41 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 | |
| encoder_l4_ffn_W2 intgemm8 (1536, 256) de 09 f2 08 fa 07 06 1d fb 0d 13 0f f5 1f 11 0e dc f6 f1 15 | |
| encoder_l4_ffn_W2_QuantMultA float32 (1, 1) 2b ac 69 41 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 | |
| encoder_l4_ffn_b1 float32 (1, 1536) ce 46 dd bf 92 ef ab 3e 65 80 d5 3f f4 8f c7 bf 80 ed de bf | |
| encoder_l4_ffn_b2 float32 (1, 256) 9a 41 16 40 e8 1b b0 3f 63 8b 22 bf 58 8d 0e bf 1d 63 f7 be | |
| encoder_l4_ffn_ffn_ln_bias float32 (1, 256) 95 bd e9 bd bb f4 07 bd fb a0 d8 3d 90 09 53 3d 0b 82 f9 3b | |
| encoder_l4_ffn_ffn_ln_scale float32 (1, 256) a9 42 8a 3f bb 40 88 3f d8 41 8e 3f 48 16 7f 3f 65 12 91 3f | |
| encoder_l4_self_Wk intgemm8 (256, 256) f1 f8 14 e9 09 f5 e2 f9 04 fe 0b 15 02 fa fc ff f9 fc f6 0e | |
| encoder_l4_self_Wk_QuantMultA float32 (1, 1) 07 23 59 41 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 | |
| encoder_l4_self_Wo intgemm8 (256, 256) 05 01 fd 1b f6 0f 0c 0a f2 f2 fe 04 ff ea 02 12 fc 06 06 ef | |
| encoder_l4_self_Wo_QuantMultA float32 (1, 1) 65 44 ed 41 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 | |
| encoder_l4_self_Wo_ln_bias float32 (1, 256) 15 ff 37 3d 50 ac f4 3c 5d fb 95 3b 15 93 08 3e 56 63 ba 3d | |
| encoder_l4_self_Wo_ln_scale float32 (1, 256) 01 ed 8f 3f f3 52 97 3f a8 17 a2 3f a3 83 92 3f 7b 63 9b 3f | |
| encoder_l4_self_Wq intgemm8 (256, 256) d8 e7 1e 06 ff 00 f2 0f fd 07 1a fa 1c fb e9 1f 06 c6 cb ee | |
| encoder_l4_self_Wq_QuantMultA float32 (1, 1) 07 23 59 41 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 | |
| encoder_l4_self_Wv intgemm8 (256, 256) 07 e4 dc f1 e3 e4 29 10 ef 13 10 0e 01 d7 19 df ea f7 ba 26 | |
| encoder_l4_self_Wv_QuantMultA float32 (1, 1) 07 23 59 41 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 | |
| encoder_l4_self_bk float32 (1, 256) cd c4 45 bf e0 96 f1 be c9 06 8f 3e e0 e9 24 bf af 22 42 be | |
| encoder_l4_self_bo float32 (1, 256) 9e 32 b8 bd 9d 33 cd 3e 37 be 8a be 03 c3 fa be 6f 09 6b bd | |
| encoder_l4_self_bq float32 (1, 256) 26 c4 5e be f8 91 5e 40 57 f0 d8 be 15 b6 c2 3f 20 fc e7 be | |
| encoder_l4_self_bv float32 (1, 256) 80 4b 6f 3d 33 8b 2d bc 52 75 d3 3e 1a 21 ba be e8 b8 1a 3f | |
| encoder_l5_ffn_W1 intgemm8 (256, 1536) e7 0f e9 fa 00 20 11 f2 1f 01 10 05 ef 04 09 0d 10 fb fc 03 | |
| encoder_l5_ffn_W1_QuantMultA float32 (1, 1) e4 a9 3d 41 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 | |
| encoder_l5_ffn_W2 intgemm8 (1536, 256) 06 0f f0 28 03 fb f1 ff 14 fc ed 0e 0a f6 f6 0c 03 08 08 f5 | |
| encoder_l5_ffn_W2_QuantMultA float32 (1, 1) c7 0d 20 41 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 | |
| encoder_l5_ffn_b1 float32 (1, 1536) 38 d7 d8 bf 28 e1 5d bf 2b 3a 11 bf 60 21 25 bf 33 27 4c c0 | |
| encoder_l5_ffn_b2 float32 (1, 256) 84 ab e6 be 62 d5 42 3f f6 51 3e bf fd b1 98 3f 1e 14 fa bf | |
| encoder_l5_ffn_ffn_ln_bias float32 (1, 256) 88 7d c5 3a 93 52 17 bb c8 39 a2 3c 75 f4 8f bd a6 52 45 3c | |
| encoder_l5_ffn_ffn_ln_scale float32 (1, 256) 5f 3f 90 3f 64 3e ab 3f 82 c1 b7 3f 7e 11 8d 3f 83 de a6 3f | |
| encoder_l5_self_Wk intgemm8 (256, 256) 03 0f 08 1e f2 f8 0d 0f 00 fe f2 07 06 fe 06 fb f8 07 09 fd | |
| encoder_l5_self_Wk_QuantMultA float32 (1, 1) 02 8b 51 41 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 | |
| encoder_l5_self_Wo intgemm8 (256, 256) 02 f9 ed ef 3d fd f1 1c 03 d5 fe dc e2 df c6 dd 1a cd f2 e4 | |
| encoder_l5_self_Wo_QuantMultA float32 (1, 1) 5d a7 79 41 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 | |
| encoder_l5_self_Wo_ln_bias float32 (1, 256) 27 e7 05 be 04 c1 1c bd 75 ae 8e 3d c2 5e e2 3d 82 c9 6e 3c | |
| encoder_l5_self_Wo_ln_scale float32 (1, 256) 93 67 8d 3f 77 7c 9a 3f c1 97 9e 3f ef a8 9c 3f 36 e2 a3 3f | |
| encoder_l5_self_Wq intgemm8 (256, 256) d1 f9 e9 ef f2 fb f1 ff 05 e3 2a 0e f5 c2 df ea 14 24 fe 2e | |
| encoder_l5_self_Wq_QuantMultA float32 (1, 1) 02 8b 51 41 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 | |
| encoder_l5_self_Wv intgemm8 (256, 256) ed fa 07 db f4 e1 d8 f8 d3 d7 eb 23 eb e0 41 f4 f3 1d d3 fa | |
| encoder_l5_self_Wv_QuantMultA float32 (1, 1) 02 8b 51 41 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 | |
| encoder_l5_self_bk float32 (1, 256) 8d fa 4e 3e 5e a0 72 3e 04 2a 8d bd 90 23 62 3e 70 14 39 3e | |
| encoder_l5_self_bo float32 (1, 256) f5 9a ae 3f 96 9a 6d 3f 85 75 af bd a0 18 bc be f5 c8 4a 3c | |
| encoder_l5_self_bq float32 (1, 256) 0b 35 a9 be 0e 5e 7c bf 71 73 88 3f af 7a 9b bf a5 fe 4d 3e | |
| encoder_l5_self_bv float32 (1, 256) b1 6d 74 be ab 5f b9 bd dc 13 a6 be 0b 1e 61 bf cc bb 94 3f | |
| encoder_l6_ffn_W1 intgemm8 (256, 1536) ef d1 fc fe 01 f3 17 1a 01 09 e6 e7 fe e7 02 de f7 11 21 13 | |
| encoder_l6_ffn_W1_QuantMultA float32 (1, 1) ad 33 44 41 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 | |
| encoder_l6_ffn_W2 intgemm8 (1536, 256) 15 05 e5 16 26 fd 08 fd 03 eb ec ff fa fd 07 14 e5 0b fb 02 | |
| encoder_l6_ffn_W2_QuantMultA float32 (1, 1) 8d cd 86 40 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 | |
| encoder_l6_ffn_b1 float32 (1, 1536) b2 36 b1 bf 72 b0 64 c0 55 96 d6 bf 09 2d d1 3f d2 5f de bf | |
| encoder_l6_ffn_b2 float32 (1, 256) 50 3c 1e bf 81 eb a0 bf bc 2f 71 3e 7a 50 8a 40 8c f6 90 3f | |
| encoder_l6_ffn_ffn_ln_bias float32 (1, 256) 72 ad 05 3e 65 99 d0 bc 86 03 d2 bd a6 00 96 be d9 a9 8f bd | |
| encoder_l6_ffn_ffn_ln_scale float32 (1, 256) 4b 45 8c 3f 17 c9 90 3f a0 f6 97 3f 5f 18 7b 3f 7f d6 8e 3f | |
| encoder_l6_self_Wk intgemm8 (256, 256) 15 13 0a 01 05 f8 0c 09 07 f8 f2 07 f9 13 e8 ec 17 f9 ff 09 | |
| encoder_l6_self_Wk_QuantMultA float32 (1, 1) dd cd 3e 41 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 | |
| encoder_l6_self_Wo intgemm8 (256, 256) 2b e8 12 f0 1e f7 e5 00 11 eb fd 0c 08 fa ea f3 05 ea fb 06 | |
| encoder_l6_self_Wo_QuantMultA float32 (1, 1) 32 1a 8f 41 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 | |
| encoder_l6_self_Wo_ln_bias float32 (1, 256) 39 2a ca bd 59 b9 59 bd 42 81 35 be fc 09 a8 bc 81 3f 75 3c | |
| encoder_l6_self_Wo_ln_scale float32 (1, 256) fb 3f a4 3f 45 ae a8 3f cb b5 a6 3f 43 ab a0 3f 9a 06 aa 3f | |
| encoder_l6_self_Wq intgemm8 (256, 256) de f4 17 0c 08 f6 f1 2e 18 f8 13 10 fa 06 ea fd ec 10 d9 01 | |
| encoder_l6_self_Wq_QuantMultA float32 (1, 1) dd cd 3e 41 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 | |
| encoder_l6_self_Wv intgemm8 (256, 256) e2 20 12 d2 04 05 fd fe 16 15 ef 1d 07 02 05 f3 fe 1e 02 f3 | |
| encoder_l6_self_Wv_QuantMultA float32 (1, 1) dd cd 3e 41 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 | |
| encoder_l6_self_bk float32 (1, 256) f7 c2 24 3f cd 1d 43 be de 49 5b 3f ae 2f 9a 3e 0f 55 96 be | |
| encoder_l6_self_bo float32 (1, 256) 5b 3e 48 3e 95 08 a5 bd f0 0b 2e 3f b3 fa 6d bd d6 ab 3a bf | |
| encoder_l6_self_bq float32 (1, 256) 35 e0 c3 3e 45 22 7c bf 0e 6d 81 c0 79 56 61 3f 9c fd 8a bf | |
| encoder_l6_self_bv float32 (1, 256) e5 97 2a bf f2 04 ee 3e 54 6a 8a 3f 00 78 fe be cb 8c 59 3f | |
| none_QuantMultA float32 (1, 1) 50 a4 02 41 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 | |
| special:model.yml int8 (1187,) 62 65 72 74 2d 74 72 61 69 6e 2d 74 79 70 65 2d 65 6d 62 65 | |
| YAML: | |
| bert-train-type-embeddings: true | |
| bert-type-vocab-size: 2 | |
| dec-cell: ssru | |
| dec-cell-base-depth: 2 | |
| dec-cell-high-depth: 1 | |
| dec-depth: 2 | |
| dim-emb: 256 | |
| dim-rnn: 1024 | |
| dim-vocabs: | |
| - 32000 | |
| - 32000 | |
| enc-cell: gru | |
| enc-cell-depth: 1 | |
| enc-depth: 6 | |
| enc-type: bidirectional | |
| factors-combine: sum | |
| factors-dim-emb: 0 | |
| input-types: [] | |
| layer-normalization: false | |
| lemma-dependency: '' | |
| lemma-dim-emb: 0 | |
| output-omit-bias: false | |
| right-left: false | |
| skip: false | |
| tied-embeddings: false | |
| tied-embeddings-all: true | |
| tied-embeddings-src: false | |
| transformer-aan-activation: swish | |
| transformer-aan-depth: 2 | |
| transformer-aan-nogate: false | |
| transformer-decoder-autoreg: rnn | |
| transformer-dim-aan: 2048 | |
| transformer-dim-ffn: 1536 | |
| transformer-ffn-activation: relu | |
| transformer-ffn-depth: 2 | |
| transformer-guided-alignment-layer: last | |
| transformer-heads: 8 | |
| transformer-no-projection: false | |
| transformer-pool: false | |
| transformer-postprocess: dan | |
| transformer-postprocess-emb: d | |
| transformer-postprocess-top: '' | |
| transformer-preprocess: '' | |
| transformer-tied-layers: [] | |
| transformer-train-position-embeddings: false | |
| type: transformer | |
| ulr: false | |
| ulr-dim-emb: 0 | |
| ulr-trainable-transformation: false | |
| version: v1.10.25; e8a1a253 2021-12-07 17:47:33 +0000 |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment