remove uses of 'rapidsai' conda channel

[jameslamb]

Recently, as this project crossed the boundary between is 25.01.x and 25.03.x release series, we observed some conda install legate-boost in CI falling back to published packages instead of using those built in CI.

Here I tried a couple approaches to prevent that:

• explicitly passing ${CPP_CHANNEL}::legate-boost in CI scripts
  • (where CPP_CHANNEL is an absolute filepath to downloaded conda packages)
  • this failed because conda install does not appear to support that
• using strict channel priority
  • this failed in hard-to-understand ways, I think because candidate legate / cupynumeric packages are spread over 3 different channels (labels are equivalent to channels in some ways)

In the time I was fiddling with this, these builds started working again... I'm not sure why.

Anyway, the changes here should at least partially help. Removing 2 unnecessary channels, which should reduce the risk of the types of conflicts that led to falling back to the wrong package in CI.

• legate/label/branch-25.01 (no longer necessary as of Update legate to 25.03 #199)
• rapidsai (was never necessary here, just copied in from other RAPIDS repos)

[jameslamb]

It looks like this pattern (passing an absolute filepath in the {channel}::package syntax) is not supported.

Multi-download failed. Reason: Transfer finalized, status: 404 [https://conda.anaconda.org/local-conda-packages/noarch/repodata.json] 3085 bytes

(build link)

Tried a couple versions of that locally too, including appending file://.

docker run \
    --rm \
    --env RAPIDS_BUILD_TYPE=nightly \
    --env RAPIDS_NIGHTLY_DATE=2025-02-09 \
    --env RAPIDS_REF_NAME=branch-25.04 \
    --env RAPIDS_REPOSITORY=rapidsai/cudf \
    --env RAPIDS_SHA=428dc188cab5a51c1e15fb90c93a231ad95b7be2 \
    -it rapidsai/ci-conda \
    bash

CPP_CHANNEL=$(rapids-download-conda-from-s3 cpp)

conda install \
    --channel ${CPP_CHANNEL} \
    "file://${CPP_CHANNEL}::libcudf"

Similar result.

UnavailableInvalidChannel: HTTP 404 NOT FOUND for channel cpp_channel <https://conda.anaconda.org/cpp_channel>

The channel is not accessible or is invalid.

I have some other ideas we can try here.

[jameslamb]

I tried using strict channel priority and it does look like the builds are using the locally-download packages... but conda chose the CPU-only build on a test machine with a GPU 😭

...
  + legate-boost    25.03.00dev6  cuda12_py310_0_cpu                /tmp/local-conda-packages      345kB

(build link)

Going to try something else here.