Multi-GPU support with dask #179
Conversation
for more information, see https://pre-commit.ci
for more information, see https://pre-commit.ci
|
We should look into the cost of allocating ahead of time for all operations that are currently in-place |
|
Median out of core is a bad choice. Uses way more memory and is slower. Loose Loose |
There was a problem hiding this comment.
I don’t think I can really review this.
I checked out parts like the PCA code, but generally the whole codebase lacks some abstractions that would get rid of visual noise while exposing the intent, especially for patterns like invoking a kernel with block sizes specified (mentioned here: https://github.com/scverse/rapids_singlecell/pull/179/files#r1838497326)
I also still want to see docstrings for the kernels! (ideally coupled with an abstraction for the kernel pattern)
Let’s tackle both of these abstractions before the next feature PR. I’m happy to meet with you to design them.
If the first abstraction (calling kernels) and the big code moves (e.g. PCA) were done before this PR, I could actually see what changes happened and what’s new and therefore review this, but like this it’s just too much to wrap my head around.
PS: there are a bunch of unaddressed comments still above
| @@ -260,6 +266,21 @@ def in_bounds( | |||
| ) | |||
|
|
|||
|
|
|||
| def _hvg_expm1(X): | |||
There was a problem hiding this comment.
what does that have to do with HVG? isn’t it just expm1?
also seems like it could be cleaner with singledispatch
There was a problem hiding this comment.
singledispatch doesnt work with cupy
This comment was marked as duplicate.
This comment was marked as duplicate.
Sorry, something went wrong.
| int n_cells) { | ||
| int cell = blockDim.x * blockIdx.x + threadIdx.x; | ||
| if(cell >= n_cells){ | ||
| return; |
There was a problem hiding this comment.
same as in the other PR: no error handling? is this expected to be called with invalid inputs?
| if isinstance(X, cp.ndarray): | ||
| X = cp.log1p(X) | ||
| else: | ||
| elif sparse.issparse(X): | ||
| X = X.log1p() | ||
|
|
||
| elif isinstance(X, DaskArray): | ||
| if isinstance(X._meta, cp.ndarray): | ||
| X = X.map_blocks(lambda x: cp.log1p(x), meta=_meta_dense(X.dtype)) | ||
| elif isinstance(X._meta, sparse.csr_matrix): | ||
| X = X.map_blocks(lambda x: x.log1p(), meta=_meta_sparse(X.dtype)) |
There was a problem hiding this comment.
this should become a helper like the expm1 above
| import dask | ||
| import dask.array as da | ||
|
|
||
| if isinstance(X._meta, sparse.csr_matrix): |
There was a problem hiding this comment.
please figure out a way to reuse code here.
Co-authored-by: Philipp A. <[email protected]>
for more information, see https://pre-commit.ci
This adds dask support
Functions to add:
calculate_qc_metricsnormalize_totallog1phighly_variable_geneswithseuratandcell_rangerscalePCAneighbors