Add pynvml as a dependency for dask-cudf

[pentschev]

Description

#17250 started using pynvml but did not add the proper dependency, this change fixes the missing dependency.

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[jakirkham]

Thanks Peter! 🙏

Do we want to drop this try...except... then?

cudf/python/dask_cudf/dask_cudf/io/parquet.py

Lines 43 to 60 in 332cc06

	def _get_device_size():
	try:
	# Use PyNVML to find the worker device size.
	import pynvml
	pynvml.nvmlInit()
	index = os.environ.get("CUDA_VISIBLE_DEVICES", "0").split(",")[0]
	if index and not index.isnumeric():
	# This means index is UUID. This works for both MIG and non-MIG device UUIDs.
	handle = pynvml.nvmlDeviceGetHandleByUUID(str.encode(index))
	else:
	# This is a device index
	handle = pynvml.nvmlDeviceGetHandleByIndex(int(index))
	return pynvml.nvmlDeviceGetMemoryInfo(handle).total
	except (ImportError, ValueError):
	# Fall back to a conservative 8GiB default
	return 8 * 1024**3

[rjzamora]

Do we want to drop this try...except... then?

The ImportError shouldn't happen anymore, but the ValueError could happen if CUDA_VISIBLE_DEVICES isn't set correctly (e.g. if Dask-CUDA isn't in use).

[pentschev]

Removed the ImportError catch in da92408

[jakirkham]

Pushed this small change to include pynvml in the dask-cudf Conda package. Hope that is ok 🙂

[jakirkham]

Thanks Peter and Rick! 🙏

[jakirkham]

/merge