vLLM moves the `get_cuda_view_from_cpu_tensor` and `silu_and_mul_per_block_quant` CUDA kernels from its legacy `_C` extension to the `_C_stable_libtorch` extension, aiming to improve compatibility with future PyTorch versions.
Migrates two key CUDA kernels (`cuda_view` and fused SiLU+Mul with per-block quantization) to the libtorch stable ABI extension, reducing the build's dependency on unstable PyTorch internals.
Operators running vLLM on newer PyTorch releases are less likely to hit build or runtime crashes caused by incompatible internal APIs. This change makes the inference server more resilient to upstream PyTorch changes, simplifying future upgrades. The migration also cleans up the build system by removing duplicated source files. Watch for any performance regressions in the migrated quantization kernels and verify that the torch 2.10 fallback path for `cuda_view` functions correctly in production.