Tianxing Wang

How to make clangd provide intelligence for .cu files

3

I have a very strange problem for my clangd. It can provide intellgence for my cpp files correctly, but when it comes to .cu files, it don't work. Here's my...

nvcc fatal : Unknown option '-mavx'

5

I am configuring the project on wsl2. when I load the CMakeList.txt in the fbgemm_gpu folder, It gives the following error. ``` Cannot get compiler information: Compiler exited with error...

question about merge-path SpMM implementation

2

Hello, I am reading your paper and code and I have some problems about the merge-path SpMM. (1) In my understanding, `limit[i]` meaning paritition i should start from row limit[i]...

what does nPtr[i] mean in your Algorithm 1?

1

hello, I' reading your paper and I wonder what is stored in nPtr[i]? Can you explain to me what target is in alg1 line13 `target = min(nPtr[i] +ePerGPU,nPtr[numNodes])` ? 

Hi, thanks for your awesome demo of speculative sample. Some of your code maybe outdated in new version of transformer. In the `KVCacheModel` class, Bloom model' k cache shape is...

[BUG] ValueError: Tensors must be contiguous when using deepspeed.initialize

2

**Describe the bug** I'm running the BingBertSquad example in `DeepSpeedExamples/training/BingBertSquad`, I use the `google-bert/bert-large-uncased` model from hugging face. I use `bash run_squad_deepspeed.sh 4 ckpt/bert_large_uncased/pytorch_model.bin /dataset /output` to finetune model. In...