Cannot reduce embeddings with cuML: NotImplementedError: Metric 'cosine' not supported for dense inputs with build_algo='nn_descent'

[6801318d8d]

I tried to use cuML to speed up topic generation as described here.

Thus from my NVIDIA SDK WSL I run:

from cuml.manifold import UMAP

In order to visualize the documents, I would like to reduce the embeddings as described here:

reduced_embeddings = UMAP(n_neighbors=10, n_components=2, min_dist=0.0, metric='cosine').fit_transform(embeddings)

But I obtain the following error:

---------------------------------------------------------------------------
NotImplementedError                       Traceback (most recent call last)
Cell In[164], line 1
----> 1 reduced_embeddings = UMAP(n_neighbors=10, n_components=2, min_dist=0.0, metric='cosine').fit_transform(embeddings)

File ~/.venvs/rapids_python3.12/lib/python3.12/site-packages/cuml/internals/api_decorators.py:213, in _make_decorator_function.<locals>.decorator_function.<locals>.decorator_closure.<locals>.wrapper(*args, **kwargs)
    210     set_api_output_dtype(output_dtype)
    212 if process_return:
--> 213     ret = func(*args, **kwargs)
    214 else:
    215     return func(*args, **kwargs)

File ~/.venvs/rapids_python3.12/lib/python3.12/site-packages/cuml/internals/api_decorators.py:365, in enable_device_interop.<locals>.dispatch(self, *args, **kwargs)
    363 if hasattr(self, "dispatch_func"):
    364     func_name = gpu_func.__name__
--> 365     return self.dispatch_func(func_name, gpu_func, *args, **kwargs)
    366 else:
    367     return gpu_func(self, *args, **kwargs)

File ~/.venvs/rapids_python3.12/lib/python3.12/site-packages/cuml/internals/api_decorators.py:215, in _make_decorator_function.<locals>.decorator_function.<locals>.decorator_closure.<locals>.wrapper(*args, **kwargs)
    213         ret = func(*args, **kwargs)
    214     else:
--> 215         return func(*args, **kwargs)
    217 return cm.process_return(ret)

File base.pyx:739, in cuml.internals.base.UniversalBase.dispatch_func()

File umap.pyx:741, in cuml.manifold.umap.UMAP.fit_transform()

File ~/.venvs/rapids_python3.12/lib/python3.12/site-packages/cuml/internals/api_decorators.py:213, in _make_decorator_function.<locals>.decorator_function.<locals>.decorator_closure.<locals>.wrapper(*args, **kwargs)
    210     set_api_output_dtype(output_dtype)
    212 if process_return:
--> 213     ret = func(*args, **kwargs)
    214 else:
    215     return func(*args, **kwargs)

File ~/.venvs/rapids_python3.12/lib/python3.12/site-packages/cuml/internals/api_decorators.py:365, in enable_device_interop.<locals>.dispatch(self, *args, **kwargs)
    363 if hasattr(self, "dispatch_func"):
    364     func_name = gpu_func.__name__
--> 365     return self.dispatch_func(func_name, gpu_func, *args, **kwargs)
    366 else:
    367     return gpu_func(self, *args, **kwargs)

File ~/.venvs/rapids_python3.12/lib/python3.12/site-packages/cuml/internals/api_decorators.py:215, in _make_decorator_function.<locals>.decorator_function.<locals>.decorator_closure.<locals>.wrapper(*args, **kwargs)
    213         ret = func(*args, **kwargs)
    214     else:
--> 215         return func(*args, **kwargs)
    217 return cm.process_return(ret)

File base.pyx:739, in cuml.internals.base.UniversalBase.dispatch_func()

File umap.pyx:660, in cuml.manifold.umap.UMAP.fit()

File umap.pyx:499, in cuml.manifold.umap.UMAP._build_umap_params()

File umap_utils.pyx:211, in cuml.manifold.umap_utils.coerce_metric()

NotImplementedError: Metric 'cosine' not supported for dense inputs with build_algo='nn_descent'

[MaartenGr]

This relates to cuml specifically, so I would advise opening up an issue there. Small tip, it would help them out if you could also list the version of cuml that you are using (my guess is that you would need the latest version).