Row: BACKEND-ROCM
Reconstruct the contributor's opt-in fused norm-quant epilogue on current
main. Exact VT_NORM_QUANT_FUSED=1 lets the ROCm RMSNorm producer emit Q8_K
activation blocks and lets a matching dense K-quant consumer reuse them.
The default remains off. A producer token must bind the activation pointer,
shape, stride, dtype, stream, and scratch lifetime. Nonmatching consumers must
invalidate or refuse stale scratch rather than reuse it.
The intended correctness gate compares fused scratch with the standalone Q8_K
quantizer and the CPU oracle, covers tied maxima and zero rows, verifies the
default-off and stale-token routes, and proves production reachability by
mutation.
This issue excludes inherited #2782/#2790 commits, separate skinny-GEMM work,
and historical performance claims that are not reproduced on the current
tree.
Depends on landed #2782 and #2790 infrastructure.
Row:
BACKEND-ROCMReconstruct the contributor's opt-in fused norm-quant epilogue on current
main. ExactVT_NORM_QUANT_FUSED=1lets the ROCm RMSNorm producer emit Q8_Kactivation blocks and lets a matching dense K-quant consumer reuse them.
The default remains off. A producer token must bind the activation pointer,
shape, stride, dtype, stream, and scratch lifetime. Nonmatching consumers must
invalidate or refuse stale scratch rather than reuse it.
The intended correctness gate compares fused scratch with the standalone Q8_K
quantizer and the CPU oracle, covers tied maxima and zero rows, verifies the
default-off and stale-token routes, and proves production reachability by
mutation.
This issue excludes inherited #2782/#2790 commits, separate skinny-GEMM work,
and historical performance claims that are not reproduced on the current
tree.
Depends on landed #2782 and #2790 infrastructure.