Laith Sakka

Thanks benchmarks looks good to me!

are numbers in the summary up to date? @pramodsatya

> Updated numbers in the summary, please let me know if any further changes are needed, thanks! Thanks, the numbers look amazing!

This one should have been solved by @pedroerp update.

I found the issue in this, in the code bellow: we mutable_data_ptr_impl does not seem to requires that storage is initialized because mutable_data will initiailize the sotrage as mentioned int...

I updated the code but it does not seem like mutable_data() actually allocate nothing lol and returns nullptr!

if i remove _validate_sparse_coo_tensor_args things pass, ``` Tensor _sparse_coo_tensor_unsafe(const Tensor& indices, const Tensor& values_, at::IntArrayRef size, c10::optional dtype, c10::optional layout, c10::optional device, c10::optional pin_memory, c10::optional is_coalesced) { if (at::globalContext().checkSparseTensorInvariants()) {...

what happen in the check is that here in the code bellow cpu_min_indices and cpu_max_indices_accessor do not have underlying data allocated! ``` Tensor min_indices = std::get(indices.min(/* dim */ 1, /*...

calling Tensor min_indices = std::get(indices.min(/* dim */ 1, /* keepdim */ false)); on indices return a tensor with storage data being nullptr

also indices passed to Tensor sparse_coo_tensor(const Tensor& indices, const Tensor& values_, c10::optional dtype, …) also is a tensor with out storage.