Document RAPIDS on Modal notebooks

[ncclementi]

Notice that Modal notebooks defaul has a pre-defined enviornment.

Look into how to get cudf and cuml included in here: https://github.com/modal-labs/modal-client/blob/v1.1.3/modal/experimental/__init__.py#L234-L342

[ncclementi]

I've been trying to get this to work, but I'm running into several issues.

The first approach I triad was to install cudf via uv directly in the notebook as suggested in the docs.

But when I try to use it and run some simple code, it fails, as it's not finding the GPUs, like it's not visible. Even though nvidia-smi shows the GPU

Steps to reproduce:

• Enable a GPU in the notebook
• %uv pip install "cudf-cu13==25.12.*"

import cudf 

s = cudf.Series([1, 2, 3])
s.apply(lambda x: x+1)

---------------------------------------------------------------------------
IndexError                                Traceback (most recent call last)
File /usr/local/lib/python3.12/site-packages/cudf/core/indexed_frame.py:3572, in IndexedFrame._apply(self, func, kernel_class, *args, **kwargs)
   3571     kr = kernel_class(self, func, args)
-> 3572     kernel, retty = kr.get_kernel()
   3573 except Exception as e:

File /usr/local/lib/python3.12/site-packages/cudf/core/udf/udf_kernel_base.py:178, in ApplyKernelBase.get_kernel(self)
    177 def get_kernel(self):
--> 178     return self._compile_or_get_kernel()

File /usr/local/lib/python3.12/site-packages/cudf/core/udf/udf_kernel_base.py:195, in ApplyKernelBase._compile_or_get_kernel(self)
    193     return kernel, masked_or_scalar
--> 195 kernel, scalar_return_type = self.compile_kernel()
    197 np_return_type = (
    198     numpy_support.as_dtype(scalar_return_type)
    199     if scalar_return_type.is_internal
    200     else scalar_return_type.np_dtype
    201 )

File /usr/local/lib/python3.12/site-packages/cudf/core/udf/udf_kernel_base.py:144, in ApplyKernelBase.compile_kernel(self)
    143 with capture_nrt_usage:
--> 144     return_type = self._get_udf_return_type()
    146 self.sig = self._construct_signature(return_type)

File /usr/local/lib/python3.12/site-packages/cudf/core/udf/udf_kernel_base.py:111, in ApplyKernelBase._get_udf_return_type(self)
    110 with _CUDFNumbaConfig():
--> 111     _, output_type = compile_udf(self.func, compile_sig)
    112 if isinstance(output_type, MaskedType):

File /usr/local/lib/python3.12/site-packages/cudf/core/udf/utils.py:207, in compile_udf(udf, type_signature)
    205 # We haven't compiled a function like this before, so need to fall back to
    206 # compilation with Numba
--> 207 ptx_code, return_type = cuda.compile_ptx_for_current_device(
    208     udf, type_signature, device=True
    209 )
    210 if not isinstance(return_type, MaskedType):

File /usr/local/lib/python3.12/site-packages/numba/cuda/compiler.py:390, in compile_ptx_for_current_device(pyfunc, sig, debug, lineinfo, device, fastmath, opt, abi, abi_info)
    388 """Compile a Python function to PTX for a given signature for the current
    389 device's compute capabilility. See :func:`compile_ptx`."""
--> 390 cc = get_current_device().compute_capability
    391 return compile_ptx(pyfunc, sig, debug=debug, lineinfo=lineinfo,
    392                    device=device, fastmath=fastmath, cc=cc, opt=opt,
    393                    abi=abi, abi_info=abi_info)

File /usr/local/lib/python3.12/site-packages/numba/cuda/api.py:443, in get_current_device()
    442 "Get current device associated with the current thread"
--> 443 return current_context().device

File /usr/local/lib/python3.12/site-packages/numba/cuda/cudadrv/devices.py:220, in get_context(devnum)
    217 """Get the current device or use a device by device number, and
    218 return the CUDA context.
    219 """
--> 220 return _runtime.get_or_create_context(devnum)

File /usr/local/lib/python3.12/site-packages/numba/cuda/cudadrv/devices.py:138, in _Runtime.get_or_create_context(self, devnum)
    137 if attached_ctx is None:
--> 138     return self._get_or_create_context_uncached(devnum)
    139 else:

File /usr/local/lib/python3.12/site-packages/numba/cuda/cudadrv/devices.py:158, in _Runtime._get_or_create_context_uncached(self, devnum)
    156 else:
    157     # Get primary context for the active device
--> 158     ctx = self.gpus[ac.devnum].get_primary_context()
    159     # Is active context the primary context?

File /usr/local/lib/python3.12/site-packages/numba/cuda/cudadrv/devices.py:40, in _DeviceList.__getitem__(self, devnum)
     37 '''
     38 Returns the context manager for device *devnum*.
     39 '''
---> 40 return self.lst[devnum]

IndexError: list index out of range

The above exception was the direct cause of the following exception:

ValueError                                Traceback (most recent call last)
Cell In[3], line 2
      1 s = cudf.Series([1, 2, 3])
----> 2 s.apply(lambda x: x+1)

File /usr/local/lib/python3.12/site-packages/cudf/core/series.py:2692, in Series.apply(self, func, convert_dtype, args, by_row, **kwargs)
   2689 elif by_row != "compat":
   2690     raise NotImplementedError("by_row is currently not supported.")
-> 2692 result = self._apply(func, SeriesApplyKernel, *args, **kwargs)
   2693 result.name = self.name
   2694 return result

File /usr/local/lib/python3.12/contextlib.py:81, in ContextDecorator.__call__.<locals>.inner(*args, **kwds)
     78 @wraps(func)
     79 def inner(*args, **kwds):
     80     with self._recreate_cm():
---> 81         return func(*args, **kwds)

File /usr/local/lib/python3.12/site-packages/cudf/core/indexed_frame.py:3574, in IndexedFrame._apply(self, func, kernel_class, *args, **kwargs)
   3572     kernel, retty = kr.get_kernel()
   3573 except Exception as e:
-> 3574     raise ValueError(
   3575         "user defined function compilation failed."
   3576     ) from e
   3578 # Mask and data column preallocated
   3579 ans_col = _return_arr_from_dtype(retty, len(self))

ValueError: user defined function compilation failed.

cc: @charlesfrye