BladeDISC

BladeDISC copied to clipboard

tensorflow version is 1.15.5， cuda version cuda-10.1 when set TF_ENABLE_TAO true, coredump as following： 2023-07-17 21:34:47.554155: I external/org_tensorflow/tensorflow/cc/saved_model/loader.cc:288] Restoring SavedModel bundle. *** Aborted at 1689600887 (unix time) try "date -d...

houjincheng1992

just like `CUDA_LAUNCH_BLOCKING` in PyTorch, BladeDISC needs a flag to sync on stream by force.

Yancey1989

TorchBench CI has detected a performance signal. Affected Tests: - eval-cuda-fp32: - functorch_dp_cifar10[disc (latency)] 1.779 -> 1.679, +5.6211% - functorch_maml_omniglot[blade (latency)] 0.71 -> 0.595, +16.1972% - hf_Bert_mini[blade (latency)] 0.546 ->...

zzpmiracle

Guo-Peilin

qiuxiafei

When I used bladedisc, I noticed that bladedisc would launch different gemm kernel for different input shapes. May I ask what rules bladedisc uses to select gemm kernel？ Thanks

LiZerun

Diffusion model benchmark

6

We have support diffusers in https://github.com/alibaba/BladeDISC/issues/867 . This issue tracks performance of all the diffuser pipelines. For the concern of performance, we use BlaDNN to tuning models during runtime. The...

zzpmiracle

The link of "our website" is invalid, please have a check. **Describe the bug** A clear and concise description of what the bug is. **To Reproduce** Steps to reproduce the...

ywangwxd

[WIP] Initial impl of HF Diffusers adapter

1

deeply

Guo-Peilin