Skip to content

LLM inference in C/C++

C++ ggml

ggml-cpu : fix rms_norm_back wrong output under in-place aliasing (#24305)

Yash Raj Pandeyfd3271e0b4d8
diff --git a/ggml/src/ggml-cpu/ops.cpp b/ggml/src/ggml-cpu/ops.cpp
index 3a1912ae91..becac9d6ef 100644
--- a/ggml/src/ggml-cpu/ops.cpp
+++ b/ggml/src/ggml-cpu/ops.cpp
@@ -4008,12 +4008,12 @@ static void ggml_compute_forward_rms_norm_back_f32(
// dx := scale(dx, rrms)
float * dx = (float *) ((char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3);
- // dx[i00] = (x*(-sum_xdz/sum_eps) + dz) / sqrtf(mean_eps)
- ggml_vec_cpy_f32 (ne00, dx, x);
- // ggml_vec_scale_f32(ne00, dx, -mean_xdz/mean_eps);
- ggml_vec_scale_f32(ne00, dx, (float)(-sum_xdz)/sum_eps);
- ggml_vec_acc_f32 (ne00, dx, dz);
- ggml_vec_scale_f32(ne00, dx, rrms);
+ // dx[i00] = (dz + x*(-sum_xdz/sum_eps)) * rrms
+ // note: https://github.com/ggml-org/ggml/issues/1491
+ const float scale_x = (float) (-sum_xdz) / sum_eps;
+ for (int64_t i00 = 0; i00 < ne00; i00++) {
+ dx[i00] = (dz[i00] + x[i00] * scale_x) * rrms;
+ }
}
}
}