openvino
1fb539a2 - [GPU] Add DisableFP16ComSinGenPatternForHiFiGAN and fix md creation of concat onednn (#33580)

Commit
127 days ago
[GPU] Add DisableFP16ComSinGenPatternForHiFiGAN and fix md creation of concat onednn (#33580) ### Description of the issue(symptom, root-cause, how it was resolved) - inf happens when hift model(https://www.modelscope.cn/models/FunAudioLLM/Fun-CosyVoice3-0.5B-2512/summary) with infer_precision=f16. - SineGen of HiFiGAN(https://github.com/FunAudioLLM/CosyVoice/blob/1dcc59676fe3fa863f983ab7820e481560c73be7/cosyvoice/hifigan/generator.py#L157-L189) could make inf in fp16 because of large input value multiplication (e.g. hop_length=480 makes multiply x480) - Added pattern matcher(multiply - interpolate - transpose - sin) to run those node with fp32 - Fixed concat onednn md creation to call proper function after recent code change #### The code and line that caused this issue (if it is not changed directly) - https://github.com/openvinotoolkit/openvino/blob/96863b85dc99233b061959a83ff822bd71862d07/src/plugins/intel_gpu/src/graph/impls/onednn/concatenation_onednn.cpp#L61-L64 #### Reproduction step and snapshot (if applicable. Do not attach for customer model) - $ ./benchmark_app -d GPU.1 -m ~/task/cvs178835/hift_ir/openvino_hift_model.xml -hint none -nstreams 1 -nireq 1 -niter 1 -shape [1,80,500] -i ~/task/cvs178835/hift_x_padding.bin #### Problematic graph - <sine generation pattern cumsum - multiply(x6.28) - transpose - multiply(x480) - interpolate - sin> <img width="1042" height="929" alt="image" src="https://github.com/user-attachments/assets/70a5be4d-9cc5-4f7e-b8ac-99c2b0872ade" /> #### Checklist - [x] Is it a proper fix? (not a workaround) - [x] Did you include test case for this fix, if necessary? - [x] Did you review existing test that can be extended to cover this scenario? Which test did you review? ### Tickets: - 178835
Author
Parents
Loading