ONNX exports of ai-sage GigaAM checkpoints for RecapIt
These exports were made for RecapIt with ai-sage's reference code and checkpoints. The int8 files were quantized by RecapIt. This repository is not affiliated with ai-sage or Sber.
Source checkpoints and reference code:
- ai-sage/GigaAM-Multilingual, branch large_ctc, revision 3905cd51c3ed4e88c8edf33f3302969ba480a327; pytorch_model.bin SHA-256 c3fabefb50b41f08f4d7ad44e02c26c37d242882704cdcca2ebd98e45eff73d1.
- ai-sage/GigaAM-v3, branch e2e_ctc, revision cec030b4c4f35d928e4a9044a3bdb29ebd499fac; pytorch_model.bin SHA-256 9801f83ef3979779982c4f83ec1c234d7571e8c1a5aa2bfc0b09ceb97a16cd75.
Multilingual: reference to_onnx(dtype=torch.float32), opset 17, onnx_export_mode; external tensors repacked byte-identically into one .data file, Constant tensor retained separately. E2E: torch.onnx.export of forward_for_export, opset 17, dynamic time axis, dynamo=False. CPU fp32; torch 2.15.0, torchaudio 2.11.1, onnx 1.20.1, onnxruntime 1.31.1.
Int8: onnxruntime 1.31.1 quantize_dynamic(input, output, weight_type=QuantType.QInt8), all remaining arguments at that version's defaults; no calibration data and no separate preprocessing pass. Resolved arguments and per-export output identities are in the offline quantize-record.json hosting evidence. Vocabulary: reference decoder token/id order, one UTF-8 LF line per token, , space represented by ▁, exactly once at the reference blank id. Vocabulary decoding matches the reference decoder on 20 test clips per variant using the same greedy token ids.
MIT notices reproduce salute-developers/GigaAM LICENSE at 7447938d791c4f3e643386ee22c33777004293a5 byte for byte. The source cards at the revisions above declare license: mit. MIT License — Copyright (c) 2024 GigaChat Team. See LICENSE for the full notice. The modeling_gigaam.py source is not redistributed here.
Hosted file SHA-256s (excluding this self-referential card; the offline hosting-record.json records this card's size and SHA-256 too):
LICENSE 1070 f00de6715714c7a63d08639cdbfaa40224eefc407302614bd19f1a8b98c875aa
gigaam-multilingual-large-ctc-fp32/Constant_1354_attr__value 2560000 22a9d295f7bfbbef1ca5f4bd02f1f5774d05db838ee7b60db358d287d65cf334
gigaam-multilingual-large-ctc-fp32/LICENSE 1070 f00de6715714c7a63d08639cdbfaa40224eefc407302614bd19f1a8b98c875aa
gigaam-multilingual-large-ctc-fp32/multilingual_large_ctc.onnx 3469754 98580de2d39f495d14f4e185730b3541e09eacb0738b59e44e2adbf1ae941d09
gigaam-multilingual-large-ctc-fp32/multilingual_large_ctc.onnx.data 2341277696 c02c9ca63700a069c702417a0030095f3e33f8ace52ef61ea2b7880e2eae23a9
gigaam-multilingual-large-ctc-fp32/multilingual_vocab.txt 393 4d130287892e1099fedfb3f93c4b4cf8a263151158801680b28977d1be4133f4
gigaam-multilingual-large-ctc-int8/LICENSE 1070 f00de6715714c7a63d08639cdbfaa40224eefc407302614bd19f1a8b98c875aa
gigaam-multilingual-large-ctc-int8/multilingual_large_ctc.int8.onnx 591644659 66f3878487748d89069164039cdaf9b50efa53e83778fc4c5cbd35aa92e06c71
gigaam-multilingual-large-ctc-int8/multilingual_vocab.txt 393 4d130287892e1099fedfb3f93c4b4cf8a263151158801680b28977d1be4133f4
gigaam-v3-e2e-ctc-fp32/LICENSE 1070 f00de6715714c7a63d08639cdbfaa40224eefc407302614bd19f1a8b98c875aa
gigaam-v3-e2e-ctc-fp32/v3_e2e_ctc.onnx 885950080 9d01993e010bd6eb45615b75d8a8cd42002ca6191e21b6561b1ad15c39a77e03
gigaam-v3-e2e-ctc-fp32/v3_e2e_ctc_vocab.txt 2007 142de7570b3de5b3035ce111a89c228e80e6085273731d944093ddf24fa539cd
gigaam-v3-e2e-ctc-int8/LICENSE 1070 f00de6715714c7a63d08639cdbfaa40224eefc407302614bd19f1a8b98c875aa
gigaam-v3-e2e-ctc-int8/v3_e2e_ctc.int8.onnx 224893256 bbcd16c97eae09c77fd4131ae0178f200f090ccfb6d400c9b52c909094342b2a
gigaam-v3-e2e-ctc-int8/v3_e2e_ctc_vocab.txt 2007 142de7570b3de5b3035ce111a89c228e80e6085273731d944093ddf24fa539cd