CANN: add operator fusion support for ADD + RMS_NORM (#17512)

This commit implements operator fusion for ADD + RMS_NORM operations
in the CANN backend to reduce memory access overhead and improve
performance. The fusion is controlled by the GGML_CANN_OPERATOR_FUSION
environment variable (default: false).

Changes:
- Implement ggml_cann_op_add_rms_norm_fused() using ACLNN AddRmsNorm
- Add ggml_cann_can_fuse() to check fusion eligibility
- Integrate fusion logic into computation graph evaluation
- Add test cases for ADD + RMS_NORM fusion
- Update documentation with new environment variable

The fusion combines ADD and RMS_NORM into a single kernel call,
which is more efficient than executing them separately.

ggml/src/ggml-cann/aclnn_ops.cpp

@@ -26,6 +26,7 @@
 #include "ggml.h"
 
 #include <aclnnop/aclnn_add.h>
+#include <aclnnop/aclnn_add_rms_norm.h>
 #include <aclnnop/aclnn_addcdiv.h>
 #include <aclnnop/aclnn_argmax.h>
 #include <aclnnop/aclnn_avgpool2d.h>
@@ -3805,3 +3806,57 @@ void ggml_cann_ssm_conv(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
                             cubeMathType);
 }
 
+
+void ggml_cann_op_add_rms_norm_fused(ggml_backend_cann_context & ctx,
+                                     ggml_tensor *               add_node,
+                                     ggml_tensor *               rms_norm_node) {
+    // Get the two input tensors for ADD operation
+    ggml_tensor * x1 = add_node->src[0];
+    ggml_tensor * x2 = add_node->src[1];
+
+    // Create ACL tensors for the two ADD inputs
+    acl_tensor_ptr acl_x1 = ggml_cann_create_tensor(x1);
+    acl_tensor_ptr acl_x2 = ggml_cann_create_tensor(x2);
+
+    // Get epsilon parameter from rms_norm_tensor
+    float eps;
+    memcpy(&eps, rms_norm_node->op_params, sizeof(float));
+
+    // Build gamma tensor (RMS normalization scaling factor)
+    // Gamma should match the normalized dimensions (last dimension of x1)
+    size_t acl_gamma_nb[GGML_MAX_DIMS];
+    acl_gamma_nb[0] = ggml_type_size(rms_norm_node->type);
+    for (int i = 1; i < GGML_MAX_DIMS; i++) {
+        acl_gamma_nb[i] = acl_gamma_nb[i - 1] * x1->ne[i - 1];
+    }
+    acl_tensor_ptr acl_gamma =
+        get_cache_acl_tensor(ctx, &ctx.rms_norm_one_tensor_cache.cache, ctx.rms_norm_one_tensor_cache.size, x1->ne,
+                             acl_gamma_nb, rms_norm_node->type,
+                             1,    // dims - only the last dimension
+                             1.0f  // value
+        );
+
+    // Build rstdOut tensor (output for normalized standard deviation)
+    // Shape should be the dimensions that are NOT normalized
+    int64_t acl_rstd_ne[] = { 1, x1->ne[1], x1->ne[2], x1->ne[3] };
+    size_t  acl_rstd_nb[GGML_MAX_DIMS - 1];
+    acl_rstd_nb[0] = sizeof(float);
+    for (int i = 1; i < GGML_MAX_DIMS - 1; i++) {
+        acl_rstd_nb[i] = acl_rstd_nb[i - 1] * acl_rstd_ne[i - 1];
+    }
+    acl_tensor_ptr acl_rstd =
+        get_cache_acl_tensor(ctx, &ctx.rms_norm_zero_tensor_cache.cache, ctx.rms_norm_zero_tensor_cache.size,
+                             acl_rstd_ne, acl_rstd_nb, GGML_TYPE_F32, GGML_MAX_DIMS,
+                             0.0f  // value
+        );
+
+    acl_tensor_ptr acl_xout = ggml_cann_create_tensor(add_node);
+
+    // Create yOut tensor (final output after RMS normalization)
+    acl_tensor_ptr acl_yout = ggml_cann_create_tensor(rms_norm_node);
+
+    // Call fused ADD + RMS_NORM operator
+    GGML_CANN_CALL_ACLNN_OP(ctx, AddRmsNorm, acl_x1.get(), acl_x2.get(), acl_gamma.get(),
+                            eps,  // double type
+                            acl_yout.get(), acl_rstd.get(), acl_xout.get());
+}