Valery Chernov
Valery Chernov
The main idea is to construct some handshake mechanism which can help to check that TVM executor obtained from pre-compiled libraries and other files is corresponded to given model. There...
**Description**: Update python tests for TVM EP for checking different scenarios of handshake mechanism. **Motivation and Context** Recently new feature was added to TVM EP (handshake mechanism). There is test...
**Description**: Replace hand-made conversion from wstring to string by embedded ONNX Runtime method. **Motivation and Context** I developed build on Windows for TVM EP and faced with problem of wstring-string...
This is a draft of implementation of 'set_output_zero_copy' method on VirtualMachine side. Brief description of approach. 1. There is python API function 'set_output' which save external outputs in VM outputs_...
**Need OCTOAI_API_KEY to use it** Add octoai (https://octoai.cloud/) as back-end for models. Greedy_until path is enabled (while likelihood is WIP)
**DRAFT** It is studying way to implement it as new accuracy evaluation approach Pattern for matching with correct answer for gsm8k task was updated. It allows better catching correct answers...
**DRAFT** It is studying way to implement it as new accuracy evaluation approach Pattern for matching with correct answer for TriviaQA task was updated. It allows better catching correct answers...
For models like llama2 the special confirmation and access are needed. Such models are still can be tested in lm-evaluation-harness by put token through model_args. But "use_auth_token" and "token" are...
Implementation of Col2Im operation from ONNX opset on TVM side