[breaking][CI] Use CTK 12.4

[hcho3]

Closes #10696
Closes #10370

Marking as breaking, as we are dropping support for CUDA 11.8.

• Apply abseil/abseil-cpp@cfde5f7 to patch Abseil C++ to be compatible with latest CUDA compiler.
• Work around false positive unix.BlockInCriticalSection from clang-tidy. See [clang][analyzer] Possible false positive from unix.BlockInCriticalSection llvm/llvm-project#104241

[trivialfis]

Thank you for working on this! Using 12.4 is going to help with the external memory work significantly. Out of curiosity, what's the kernel driver version of the AMI? Is it using kernel-open? Ubuntu package:

$ apt search nvidia-driver
...
nvidia-headless-555-open/jammy 555.58.02-0ubuntu0~gpu22.04.1 amd64
  NVIDIA headless metapackage (open kernel module)
...

It's required to use HMM addressing:

$ nvidia-smi -q | grep Addressing

    Addressing Mode                       : HMM
    Addressing Mode                       : HMM

Currently, this test is skipped on the CI:

xgboost/tests/cpp/data/test_ellpack_page_raw_format.cu

Line 71 in 773ded6

if (common::SupportsPageableMem()) {

It's OK if we don't want to update the AMI for now, we are still experimenting with the feature and might not support it in the end. Only bringing this up to share the need for newer CUDA versions.

As for the GRPC, I use:

>>> import grpc
>>> grpc.__version__
'1.62.2'

from conda-forge, seems to work well with CUDA 12.4/5, if it helps.

[hcho3]

@trivialfis Here is the kernel driver version:

$ nvidia-smi
Tue Aug 13 19:57:56 2024
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 550.90.07              Driver Version: 550.90.07      CUDA Version: 12.4     |
|-----------------------------------------+------------------------+----------------------+
| GPU  Name                 Persistence-M | Bus-Id          Disp.A | Volatile Uncorr. ECC |
| Fan  Temp   Perf          Pwr:Usage/Cap |           Memory-Usage | GPU-Util  Compute M. |
|                                         |                        |               MIG M. |
|=========================================+========================+======================|
|   0  Tesla T4                       On  |   00000000:00:1E.0 Off |                    0 |
| N/A   32C    P8             14W /   70W |       1MiB /  15360MiB |      0%      Default |
|                                         |                        |                  N/A |
+-----------------------------------------+------------------------+----------------------+

+-----------------------------------------------------------------------------------------+
| Processes:                                                                              |
|  GPU   GI   CI        PID   Type   Process name                              GPU Memory |
|        ID   ID                                                               Usage      |
|=========================================================================================|
|  No running processes found                                                             |
+-----------------------------------------------------------------------------------------+

The AMI is based on Amazon Linux 2, so it probably doesn't use kernel-open.

[terrytangyuan]

What release is this targeting?

[hcho3]

The next release, which is going to be 3.0.0 (due to the rewrite for the JVM packages).

[hcho3]

@terrytangyuan BTW, we would love to get your input on #10639

[terrytangyuan]

Sounds great!

[trivialfis]

Hmm, RMM is bringing in a different CCCL. Might have to patch it by back porting the one liner fix.