[torch.compile] Enable attention and allreduce fusion without custom ops enabled (#24604)

Signed-off-by: Luka Govedič <lgovedic@redhat.com>
Signed-off-by: Luka Govedič <ProExpertProg@users.noreply.github.com>
This commit is contained in:
Luka Govedič
2025-10-17 08:10:23 -06:00
committed by GitHub
parent be429d0cfd
commit bd7157a071
28 changed files with 1519 additions and 721 deletions
@@ -145,7 +145,11 @@ void rms_norm_dynamic_per_token_quant(
if (scale_ub.has_value()) {
TORCH_CHECK(out.dtype() == kFp8Type);
}
TORCH_CHECK(weight.dtype() == input.dtype());
TORCH_CHECK(scales.dtype() == torch::kFloat32);
if (residual) {
TORCH_CHECK(residual->scalar_type() == input.scalar_type());
}
VLLM_DISPATCH_FLOATING_TYPES(
input.scalar_type(), "rms_norm_dynamic_per_token_quant_dispatch", [&] {