[BE] Drop CUDA < 12.6 support from CI/CD and raise the declared minimum (#192257)

`CUDA_ARCHES = ["12.6", "13.0", "13.2"]`, so nothing below 12.6 is built or
shipped any more. This removes the CI/CD leftovers and makes the minimum
explicit.

## Changes

| file | change |
|---|---|
| `cmake/public/cuda.cmake` | require CUDA **12.6**, was 12.1 |
| `.ci/docker/build.sh` | drop the `pytorch-linux-jammy-cuda12.4-cudnn9-py3-gcc11` image config |
| `.ci/docker/common/install_cuda.sh` | drop `install_124()` and its `12.4` dispatch arm |
| `.ci/docker/manywheel/Dockerfile_2_28` | `ARG BASE_CUDA_VERSION` default `11.8` → `12.6` |
| `docker.Makefile` | `CUDA_VERSION` `12.1.1` → `12.6.3`, `CUDA_VERSION_SHORT` `12.1` → `12.6` |
| `README.md` | Dockerfile prose said CUDA 12.1 |

### Notes on each

- **The 12.4 image config is orphaned.** `grep -rn "cuda12\.4" .github/ .ci/`
  returns only the definition in `build.sh` itself — no workflow and no
  `docker-builds.yml` entry ever requests it, so it has not been built. With it
  gone, `install_124()` has no remaining caller (verified: `grep -rn install_124`
  is empty after the change).
- **`docker.Makefile` was actively broken.** The 12.1 default made
  `CUDA_PATH = cu121`, and the Dockerfile installs torch from
  `download.pytorch.org/whl/$CUDA_PATH` — an index that no longer receives
  builds. 12.6.3 matches the version `install_126()` installs.
- **`ARG BASE_CUDA_VERSION=11.8`** was the only stale default in that file; the
  three other stages (`cuda`, `magma`, `cpu_final`) already default to 12.6.

## Scope

**Raising the cmake minimum narrows the source-build range**: building against
CUDA 12.1–12.5 now fails fast with `PyTorch requires CUDA 12.6 or above` instead
of compiling an untested configuration. That is the intended, user-visible part
of this change — flagging it explicitly for reviewers.

The **C++ `CUDA_VERSION` preprocessor guards are deliberately untouched.** There
are 34 such guard lines under 12060 in `aten/`, `torch/` and `c10/` (26 ×
`12040`, 4 × `12000`, 2 × `12030`, 1 × `12050`, 1 × `11060`). Removing those
branches touches kernel and runtime code and each site needs the correct branch
kept, so it belongs in its own reviewed change.

## Test plan

`bash -n` clean on both edited shell scripts. No dangling references to the
removed image config or install function. Nothing else in the tree selects CUDA
< 12.6.

Pull Request resolved: https://github.com/pytorch/pytorch/pull/192257
Approved by: https://github.com/ngimel, https://github.com/tinglvv

.ci/docker/build.sh

@@ -91,13 +91,6 @@ fi
 # configuration, so we hardcode everything here rather than do it
 # from scratch
 case "$tag" in
-  pytorch-linux-jammy-cuda12.4-cudnn9-py3-gcc11)
-    CUDA_VERSION=12.4
-    ANACONDA_PYTHON_VERSION=3.10
-    GCC_VERSION=11
-    KATEX=yes
-    TRITON=yes
-    ;;
   pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc11)
     CUDA_VERSION=12.8.1
     ANACONDA_PYTHON_VERSION=3.10

.ci/docker/common/install_cuda.sh

@@ -107,21 +107,6 @@ function install_cupti_headers {
   echo "CUPTI ${cupti_version} headers installed to ${target_dir}."
 }
 
-function install_124 {
-  CUDNN_VERSION=9.1.0.70
-  CUSPARSELT_VERSION=0.6.2.3
-  echo "Installing CUDA 12.4.1 and cuDNN ${CUDNN_VERSION} and NCCL and cuSparseLt-${CUSPARSELT_VERSION}"
-  install_cuda 12.4.1 cuda_12.4.1_550.54.15_linux
-
-  install_cudnn 12 $CUDNN_VERSION
-
-  CUDA_VERSION=12.4 bash install_nccl.sh
-
-  CUDA_VERSION=12.4 bash install_cusparselt.sh $CUSPARSELT_VERSION
-
-  ldconfig
-}
-
 function install_126 {
   CUDNN_VERSION=9.10.2.21
   CUSPARSELT_VERSION=0.7.1.0
@@ -257,8 +242,6 @@ function install_134 {
 while test $# -gt 0
 do
     case "$1" in
-    12.4) install_124;
-        ;;
     12.6|12.6.*) install_126;
         ;;
     12.8|12.8.*) install_128;

.ci/docker/manywheel/Dockerfile_2_28

@@ -1,5 +1,5 @@
 # syntax = docker/dockerfile:experimental
-ARG BASE_CUDA_VERSION=11.8
+ARG BASE_CUDA_VERSION=12.6
 ARG GPU_IMAGE=amd64/almalinux:8
 FROM quay.io/pypa/manylinux_2_28_x86_64 as base
 

README.md

@@ -423,7 +423,7 @@ should increase shared memory size either with `--ipc=host` or `--shm-size` comm
 
 **NOTE:** Must be built with a Docker version >= 23.0
 
-The Dockerfile is supplied to build images with CUDA 12.1 support and cuDNN v9.
+The Dockerfile is supplied to build images with CUDA 12.6 support and cuDNN v9.
 You can pass `PYTHON_VERSION=x.y` make variable to specify which Python version is to be used by Miniconda, or leave it
 unset to use the default, as the Dockerfile uses system Python.
 

cmake/public/cuda.cmake

@@ -70,8 +70,8 @@ endif()
 message(STATUS "PyTorch: CUDA detected: " ${CUDA_VERSION})
 message(STATUS "PyTorch: CUDA nvcc is: " ${CUDA_NVCC_EXECUTABLE})
 message(STATUS "PyTorch: CUDA toolkit directory: " ${CUDA_TOOLKIT_ROOT_DIR})
-if(CUDA_VERSION VERSION_LESS 12.1)
-  message(FATAL_ERROR "PyTorch requires CUDA 12.1 or above.")
+if(CUDA_VERSION VERSION_LESS 12.6)
+  message(FATAL_ERROR "PyTorch requires CUDA 12.6 or above.")
 endif()
 
 if(CUDA_FOUND)

docker.Makefile

@@ -8,8 +8,8 @@ $(warning WARNING: No docker user found using results from whoami)
 DOCKER_ORG                = $(shell whoami)
 endif
 
-CUDA_VERSION_SHORT       ?= 12.1
-CUDA_VERSION             ?= 12.1.1
+CUDA_VERSION_SHORT       ?= 12.6
+CUDA_VERSION             ?= 12.6.3
 CUDNN_VERSION            ?= 9
 BASE_RUNTIME              = ubuntu:24.04
 BASE_DEVEL                = ubuntu:24.04