Row: ENG-DOCS-ABI-ROCM-REFRESH
Refresh the public ABI and ROCm instructions against current source. README and the C API reference still name ABI 23 although include/vllm.h defines 26. README invokes examples/server although CMake names the output vllm-server. docs/BUILD.md still describes ROCm as one registered kernel and an unverified allocator fix, despite native kernels and recorded board measurements.
Scope: README.md, docs/BUILD.md, docs/reference/c-api.md, and a scoped documentation spec. Add a September news item for the shipped ABI additions and measured ROCm native EXL3 path, with current limitations. Simplify the affected prose. Preserve benchmark values and lifecycle states. Exclude the GLM and FP8 placement corrections already covered by PRs #2374 and #2236.
Verification is CPU-only: source assertions, documentation structure and reference checks, applicable record gates, and independent review. No model execution, new benchmark, GPU, or runtime change applies.
Row:
ENG-DOCS-ABI-ROCM-REFRESHRefresh the public ABI and ROCm instructions against current source. README and the C API reference still name ABI 23 although include/vllm.h defines 26. README invokes examples/server although CMake names the output vllm-server. docs/BUILD.md still describes ROCm as one registered kernel and an unverified allocator fix, despite native kernels and recorded board measurements.
Scope: README.md, docs/BUILD.md, docs/reference/c-api.md, and a scoped documentation spec. Add a September news item for the shipped ABI additions and measured ROCm native EXL3 path, with current limitations. Simplify the affected prose. Preserve benchmark values and lifecycle states. Exclude the GLM and FP8 placement corrections already covered by PRs #2374 and #2236.
Verification is CPU-only: source assertions, documentation structure and reference checks, applicable record gates, and independent review. No model execution, new benchmark, GPU, or runtime change applies.