Clement Chan

I have the same question for this.

This may not affect the quality much since the length of padding is very short (about 23ms). I haven't compared the results.

> Hi, > I struggled to find any literature on this myself, you should find some in the Tacotron paper. > All the rest, like scheduling, are rules of thumb...

> You can observe this in tensorboard in the last layer: if you heads (or at least one ehad) in the last layer do not look roughly diagonal, then the...

@bkumardevan07 > Also, have one question why query concatenation is necessary? because We are anyway using skip connections which will add query information. In transformer architecture, multihead attention is surrounded...

所用的 paddle 版本？ 加载的时候传的参数是？ 这里存在一些问题。是因为保存用的是 fluid.save 而加载的时候提示的是 Load_program_state 只能读取那三个 api 保存的结果。

其实需要保存的东西一般来说会有 pdopt 优化器 pdmode 模型结构定义 pdparams 参数文件 load_program_state 是一个较老的接口了，它适用于加载分散的参数文件。每个参数一个文件存在文件夹中。 我们现在推荐的用法是用 paddle.save paddle.load 之类的接口。用于保存 paddle.nn.Layer 的的 state_dict 和优化器的 state_dict. 会得到 pdparams 和 pdopt 还有一套用于静态图的 paddle.static.save paddle.static.load 接口。用于保存 program(paddle 的计算图)和参数等。会得到 pdparams, pdopt...

Is there some necessary context initialization step before tensors on "cuda:1" to be passed to triton kernels?

We found that the error raises from `getPointer` (a function defined in string template) in `python/triton/compiler/make_launcher.py`. ```c int status = cuPointerGetAttribute(&dev_ptr, CU_POINTER_ATTRIBUTE_DEVICE_POINTER, ptr_info.dev_ptr); if (status == CUDA_ERROR_INVALID_VALUE) {{ PyErr_Format(PyExc_ValueError, "Pointer...

> +1 this is effecting me as well. Seems to normally work on cuda:0, also is stochastic on cuda != 0. ~Also your fix of set_device with `torch.cuda` does not...