[sreeram-11]

Overview

Fixes the CUDA/HIP/MUSA build on master.

#30804 replaced ggml_cuda_graph::instance with instances, but invalidate_cuda_graphs() from #26771 still uses graph->instance, which breaks the CUDA/HIP build.

ggml-cuda.cu(499): error: class "ggml_cuda_graph" has no member "instance"

Fixed with graph->instances.size().

Additional information

Failing CI on 280141e (before):

• CI (CUDA, ubuntu) / cuda: https://github.com/ggml-org/llama.cpp/actions/runs/39986763775/job/120527836370
• CI (CUDA, ubuntu) / hip: https://github.com/ggml-org/llama.cpp/actions/runs/39986763775/job/120527836014
• HIP quality check: https://github.com/ggml-org/llama.cpp/actions/runs/39986763886/job/120527836771

The CI on this PR is the regression test. Expecting it to pass.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES - AI was used to help locate the leftover instance reference and check that no other uses of the removed fields remain. I reviewed and understand the change.

[CISC]

Please fix newlines in common.cuh as well.

[jakelandau]

Same error on Windows (MSVC 19.51, CUDA 13.4, -DCMAKE_CUDA_ARCHITECTURES=120, Ninja) at 06230cb — matches #31056. Applied this PR's change locally: full build passes (all targets incl. llama-server/llama-cli) and the binaries run (llama-cli --version OK).

[ServeurpersoCom]

Quick merge to unbreak the build. The rocm failure is a runner issue, not this change: the HIP build step stops at 47% with no compile error, then test-llama-archs is missing. The hip job, which compiles ggml-cuda.cu, is green.

[ServeurpersoCom]

Built it here on Linux (CUDA) and on Windows (MSVC): both compile and run again with this change, so the build is fixed.