Remove dask, pytorch, and nvidia channels

[vyasr]

Contributes to rapidsai/build-planning#166

[jameslamb]

Support this change generally, but I think that sed change needs to look a little different.

[jameslamb]

The arm64, CUDA 12.0.01 build-rapids jobs are all failing, like this:

#12 12.76 error    libmamba Could not solve for environment specs
#12 12.76     The following packages are incompatible
#12 12.76     ├─ cuda-version =12.0 * is requested and can be installed;
#12 12.76     └─ rapids =25.8 * is not installable because it requires
#12 12.76        └─ libcugraph_etl =25.8 *, which requires
#12 12.76           └─ libcudf =25.8 * but there are no viable options
#12 12.76              ├─ libcudf 25.8.0a6 would require
#12 12.76              │  └─ cuda-version >=11,<12.0a0 *, which conflicts with any installable versions previously reported;
#12 12.76              └─ libcudf [25.08.00a330|25.08.00a331|...|25.8.0a6] would require
#12 12.76                 └─ libkvikio =25.8 * but there are no viable options
#12 12.76                    ├─ libkvikio 25.8.0a0 would require
#12 12.76                    │  └─ cuda-version >=11,<12.0a0 *, which conflicts with any installable versions previously reported;
#12 12.76                    └─ libkvikio [25.08.00a32|25.08.00a34|...|25.8.0a0] would require
#12 12.76                       └─ libcufile-dev =* * but there are no viable options
#12 12.76                          ├─ libcufile-dev [1.10.0.4|1.10.1.7] would require
#12 12.76                          │  └─ cuda-version >=12.5,<12.6.0a0 *, which conflicts with any installable versions previously reported;
#12 12.76                          ├─ libcufile-dev [1.11.0.15|1.11.1.6] would require
#12 12.76                          │  └─ cuda-version >=12.6,<12.7.0a0 *, which conflicts with any installable versions previously reported;
#12 12.76                          ├─ libcufile-dev [1.13.0.11|1.13.1.3] would require
#12 12.76                          │  └─ cuda-version >=12.8,<12.9.0a0 *, which conflicts with any installable versions previously reported;
#12 12.76                          ├─ libcufile-dev [1.14.0.30|1.14.1.1] would require
#12 12.76                          │  └─ cuda-version >=12.9,<12.10.0a0 *, which conflicts with any installable versions previously reported;
#12 12.76                          ├─ libcufile-dev 1.7.2.10 would require
#12 12.76                          │  └─ cuda-version >=12.2,<12.3.0a0 *, which conflicts with any installable versions previously reported;
#12 12.76                          ├─ libcufile-dev 1.8.1.2 would require
#12 12.76                          │  └─ cuda-version >=12.3,<12.4.0a0 *, which conflicts with any installable versions previously reported;
#12 12.76                          └─ libcufile-dev [1.9.0.20|1.9.1.3] would require
#12 12.76                             └─ cuda-version >=12.4,<12.5.0a0 *, which conflicts with any installable versions previously reported.

(build link)

My read of that is "there are not libcufile-dev arm64, CUDA 12.0 packages available from the channels we're looking for... maybe we were previously getting those from the nvidia channel".

[vyasr]

Well that's problematic. It means we've been pulling invalid environments so far. cufile packages didn't exist on arm until 12.2, so if we were getting them from nvidia (likely the correct diagnosis) we were getting packages that weren't actually constrained to the right cuda-version. I wonder why we have these though; our cufile dependency should be conditional on the cuda version.

[vyasr]

Ah the problem is kvikio's recipe, not cudf's. I have to look a bit closer to figure out the correct fix here though; kvikio has a tighter connection to cufile than cudf, so I don't know if simply removing the dependency is the right answer here.

[jameslamb]

Yeah was just going to say... I just looked and see libkvikio has an unconditional runtime dependency on libcufile-dev:

https://github.com/rapidsai/kvikio/blob/6d767f75dc7553065746855e22e3d207a6527a5d/conda/recipes/libkvikio/recipe.yaml#L88

[vyasr]

OK I've put in a temporary patch in rapidsai/kvikio#753 to get around this issue. We'll need a follow up that sets up the correct constraints, probably with a variant build.

[jameslamb]

🎉

Addressed

[vyasr]

/merge