Downloads · 30 days
41
17% of all-time downloads
tonythethompson/madlad400-3b-mt-onnx
madlad400-3b-mt-onnx is a translation model from tonythethompson. Use it when you need text moved from one language to another. The card lists the license as apache-2.0.
This repository contains quantized ONNX encoder and decoder files for MADLAD-400-3B-MT, a multilingual machine translation model based on the T5 architecture.
Downloads · 30 days
41
17% of all-time downloads
All-time downloads
237
Public
Repo size
20.8 GB
Likes
0
Public
Click a slice to open those files.
.data23.7 GB · 50%
From the Hugging Face model README
This repository contains quantized ONNX encoder and decoder files for MADLAD-400-3B-MT, a multilingual machine translation model based on the T5 architecture.
| Field | Value |
|---|---|
| Upstream model | google/madlad400-3b-mt |
| ONNX artifact source | ISoloist1/madlad400-3b-mt-onnx |
| Upstream source revision | fa184c675da0b5c9e1c8694fccd4e12e2d422094 |
| ONNX artifact source revision | 0dfc0b48e063b55508704ff75e21416c8679f663 |
| Packaging source revision | 67037ad42f58d6c0fc3dafaa45f3ec97a46e7eb9 |
| Export tool/script | ONNX Runtime quantization pipeline (mirrored from ISoloist1 ONNX bundle) |
| Quantization recipe | ONNX Runtime INT8 encoder/decoder graphs |
Export tooling, precision, and quantization are recorded in the Source table above. This packaging mirror does not publish independent parity benchmarks; validate on your target execution provider before production use.
| File | Description |
|---|---|
encoder_model_quantized.onnx | Quantized ONNX encoder model |
decoder_model_quantized.onnx | Quantized ONNX decoder model |
config.json | T5 model configuration |
tokenizer_config.json | T5 tokenizer configuration |
spiece.model | SentencePiece tokenizer model |
Use these ONNX model files for multilingual machine translation experiments or integration work.
The upstream model card lists allenai/MADLAD-400 and describes training on large-scale public multilingual data, with additional parallel data for the machine translation model.
The upstream model card reports evaluation in the research paper using WMT, NTREX, FLORES-200, and Gatones. This repository does not document ONNX-specific evaluation or parity results.
Apache-2.0, matching the upstream source model.
trt_rtx_fp16/: three ONNX graphs exported from the upstream FP32 model and converted with NVIDIA ModelOpt.trt_rtx_fp16_graphsurgeon/: GraphSurgeon cleanup of the FP16 bundle.All three graphs use separate external-weight sidecars. Download each .onnx file together with its corresponding .onnx_data or .onnx.data sidecar. The existing INT8 files remain unchanged. See verification_report.json for checks and limits. ONNX Runtime CUDAExecutionProvider checks exercised the encoder, decoder, and decoder-with-past paths; the TensorRT-RTX EP itself was not available, so TensorRT-RTX compatibility is not claimed.
The earlier trt_rtx_fp16/ and trt_rtx_fp16_graphsurgeon/ variants failed translation correctness on DirectML and TensorRT-RTX: their encoder output was all zero. They remain as historical artifacts; their prior synthetic CUDA checks did not establish correct destination-provider translations.
trt_rtx_mixed_fp16_fp32/: ModelOpt-derived weights with FP32 protection for RMSNorm, residual additions, and output projections.trt_rtx_mixed_fp16_fp32_graphsurgeon/: the corresponding repaired GraphSurgeon derivative.Both corrected candidates passed five token-for-token translation regression prompts on CPU, DirectML, and TensorRT-RTX on an RTX 5070, with profiled GPU execution. Six weight sidecars are byte-identical to their source variants. This is a limited smoke test, not general translation-quality or Trackdub application acceptance; existing application TensorRT guards are unchanged.
See verification evidence, repair report and conversion script. Original INT8 and historical candidate model files are unchanged.