Commit a83d68b56 for llama.cpp

commit a83d68b562b19fba1f8836edc93eb76d9ca1ce25
Author: Tarek Dakhran <tarek@liquid.ai>
Date:   Fri Nov 20 19:55:44 2026 +0100

    metal : allow the norm chain in either operand of the fused MUL/ADD (#32300)

    Assisted-by: Claude Opus 5.5

diff --git a/ggml/src/ggml-metal/ggml-metal-fusion.cpp b/ggml/src/ggml-metal/ggml-metal-fusion.cpp
index 6766b4bd9..7b55a301e 100644
--- a/ggml/src/ggml-metal/ggml-metal-fusion.cpp
+++ b/ggml/src/ggml-metal/ggml-metal-fusion.cpp
@@ -135,17 +135,15 @@ static bool ggml_metal_fusion_check_norm(
     }

     for (int j = 1; j < (int) fusion->ops.size(); j++) {
-        // the fused MUL/ADD must read the previous node as src0
-        if (nodes[j]->src[0] != nodes[j - 1]) {
-            return false;
-        }
+        // MUL/ADD are commutative, the previous node can be either operand
+        const ggml_tensor * src = nodes[j]->src[0] == nodes[j - 1] ? nodes[j]->src[1] : nodes[j]->src[0];

         // the weight/bias must have the same row width as the norm input
-        if (nodes[j]->src[1]->ne[0] != nodes[0]->ne[0]) {
+        if (src->ne[0] != nodes[0]->ne[0]) {
             return false;
         }

-        if (!ggml_is_contiguous_rows(nodes[j]->src[1])) {
+        if (!ggml_is_contiguous_rows(src)) {
             return false;
         }

diff --git a/ggml/src/ggml-metal/ggml-metal-ops.cpp b/ggml/src/ggml-metal/ggml-metal-ops.cpp
index 1ae7f3d5d..a7a1ef61e 100644
--- a/ggml/src/ggml-metal/ggml-metal-ops.cpp
+++ b/ggml/src/ggml-metal/ggml-metal-ops.cpp
@@ -4698,16 +4698,17 @@ int ggml_metal_op_norm(ggml_metal_op_t ctx, int idx) {

             for (int i = 1; i < n_fuse; i++) {
                 const ggml_tensor * fn = ctx->node(idx + i);
+                const ggml_tensor * fs = fn->src[0] == ctx->node(idx + i - 1) ? fn->src[1] : fn->src[0];

-                bid_fuse[i - 1] = ggml_metal_get_buffer_id(fn->src[1]);
+                bid_fuse[i - 1] = ggml_metal_get_buffer_id(fs);

-                args.nef1[i] = fn->src[1]->ne[1];
-                args.nef2[i] = fn->src[1]->ne[2];
-                args.nef3[i] = fn->src[1]->ne[3];
+                args.nef1[i] = fs->ne[1];
+                args.nef2[i] = fs->ne[2];
+                args.nef3[i] = fs->ne[3];

-                args.nbf1[i] = fn->src[1]->nb[1];
-                args.nbf2[i] = fn->src[1]->nb[2];
-                args.nbf3[i] = fn->src[1]->nb[3];
+                args.nbf1[i] = fs->nb[1];
+                args.nbf2[i] = fs->nb[2];
+                args.nbf3[i] = fs->nb[3];
             }

             if (debug_fusion > 1) {
diff --git a/tests/fusion/MTL.csv b/tests/fusion/MTL.csv
index 295740c7c..b939d6683 100644
--- a/tests/fusion/MTL.csv
+++ b/tests/fusion/MTL.csv
@@ -63,7 +63,7 @@ deepseek4           ,0   ,any     ,MUL+ADD                     ,      8
 deepseek4           ,0   ,any     ,RMS_NORM+MUL                ,     20
 deepseek4           ,0   ,any     ,UNARY+SQRT+ADD+ARGSORT+GET_ROWS+SUM_ROWS+CLAMP+DIV,      4
 deepseek41          ,0   ,any     ,MUL+ADD                     ,     12
-deepseek41          ,0   ,any     ,RMS_NORM+MUL                ,     29
+deepseek41          ,0   ,any     ,RMS_NORM+MUL                ,     31
 deepseek41          ,0   ,any     ,UNARY+SQRT+ADD+ARGSORT+GET_ROWS+SUM_ROWS+SCALE+DIV,      6
 domynedge           ,0   ,any     ,RMS_NORM+MUL                ,      8
 domynedge           ,0   ,any     ,RMS_NORM+MUL+ADD            ,      3
@@ -101,8 +101,8 @@ gemma2              ,0   ,any     ,RMS_NORM+MUL                ,      5
 gemma2              ,0   ,any     ,RMS_NORM+MUL+ADD            ,      4
 gemma3              ,0   ,any     ,RMS_NORM+MUL                ,      9
 gemma3              ,0   ,any     ,RMS_NORM+MUL+ADD            ,      4
-gemma4              ,0   ,any     ,RMS_NORM+MUL                ,     47
-gemma4              ,0   ,any     ,RMS_NORM+MUL+ADD            ,     20
+gemma4              ,0   ,any     ,RMS_NORM+MUL                ,     37
+gemma4              ,0   ,any     ,RMS_NORM+MUL+ADD            ,     30
 glm-dsa             ,0   ,any     ,ADD+ADD                     ,      1
 glm-dsa             ,0   ,any     ,MUL+ADD                     ,      2
 glm-dsa             ,0   ,any     ,NORM+MUL+ADD                ,      2
@@ -258,7 +258,8 @@ muse-glimmer        ,0   ,any     ,RMS_NORM+MUL                ,     10
 muse-glimmer        ,0   ,any     ,RMS_NORM+MUL+ADD            ,      3
 muse-spark          ,0   ,prefill ,MUL+ADD                     ,      2
 muse-spark          ,0   ,decode  ,MUL+ADD                     ,      4
-muse-spark          ,0   ,any     ,RMS_NORM+MUL                ,     14
+muse-spark          ,0   ,any     ,RMS_NORM+MUL                ,     11
+muse-spark          ,0   ,any     ,RMS_NORM+MUL+ADD            ,      3
 muse-spark          ,0   ,any     ,UNARY+ADD+ARGSORT+GET_ROWS+SUM_ROWS+CLAMP+DIV,      2
 nanbeige            ,0   ,any     ,RMS_NORM+MUL                ,      5
 nemotron            ,0   ,any     ,NORM+MUL+ADD                ,      5
diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp
index 42e356ae0..11bfa42a9 100644
--- a/tests/test-backend-ops.cpp
+++ b/tests/test-backend-ops.cpp
@@ -4135,6 +4135,7 @@ struct test_rms_norm_mul_add : public test_case {
     const bool weight_broadcast;
     const ggml_type type_w;   // type of b and d, GGML_TYPE_COUNT -> same as a
     const ggml_type type_add; // type of c, GGML_TYPE_COUNT -> same as b
+    const bool swap;          // the norm chain is the second operand of MUL and ADD

     std::string op_desc(ggml_tensor * t) override {
         GGML_UNUSED(t);
@@ -4151,6 +4152,9 @@ struct test_rms_norm_mul_add : public test_case {
         if (type_add != GGML_TYPE_COUNT) {
             s += "," + VAR_TO_STR(type_add);
         }
+        if (swap) {
+            s += "," + VAR_TO_STR(swap);
+        }
         return s;
     }

@@ -4158,9 +4162,10 @@ struct test_rms_norm_mul_add : public test_case {
             std::array<int64_t, 4> ne = {64, 5, 4, 3},
             float eps = 1e-6f, bool broadcast = false, bool multi_add = false, bool post_mul = false,
             bool alias_rms_input = false, bool weight_broadcast = false, ggml_type type_w = GGML_TYPE_COUNT,
-            ggml_type type_add = GGML_TYPE_COUNT)
+            ggml_type type_add = GGML_TYPE_COUNT, bool swap = false)
         : type(type), ne(ne), eps(eps), broadcast(broadcast), multi_add(multi_add), post_mul(post_mul),
-          alias_rms_input(alias_rms_input), weight_broadcast(weight_broadcast), type_w(type_w), type_add(type_add) {}
+          alias_rms_input(alias_rms_input), weight_broadcast(weight_broadcast), type_w(type_w), type_add(type_add),
+          swap(swap) {}

     ggml_tensor * build_graph(ggml_context * ctx) override {
         std::array<int64_t, 4> broadcast_dims = {ne[0]*2, ne[1]*3, ne[2]*3, ne[3]*4};
@@ -4191,8 +4196,9 @@ struct test_rms_norm_mul_add : public test_case {
         if (multi_add) {
             a = ggml_add(ctx, ggml_add(ctx, a, b), c);
         }
-        ggml_tensor * mul = ggml_mul(ctx, ggml_rms_norm(ctx, a, eps), b);
-        ggml_tensor * out = alias_rms_input ? ggml_add_inplace(ctx, a, mul) : ggml_add(ctx, mul, c);
+        ggml_tensor * norm = ggml_rms_norm(ctx, a, eps);
+        ggml_tensor * mul  = swap ? ggml_mul(ctx, b, norm) : ggml_mul(ctx, norm, b);
+        ggml_tensor * out  = alias_rms_input ? ggml_add_inplace(ctx, a, mul) : swap ? ggml_add(ctx, c, mul) : ggml_add(ctx, mul, c);
         if (post_mul) {
             out = ggml_mul(ctx, out, d);
         }
@@ -11771,6 +11777,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
         for (uint32_t n : { 64, 1025 }) {
             test_cases.emplace_back(new test_rms_norm_mul_add(GGML_TYPE_F32, { n, 5, 4, 3 }, eps, false));
             test_cases.emplace_back(new test_rms_norm_mul_add(GGML_TYPE_F32, { n, 5, 4, 3 }, eps, true));
+            test_cases.emplace_back(new test_rms_norm_mul_add(GGML_TYPE_F32, { n, 5, 4, 3 }, eps, false, false, false, false, false, GGML_TYPE_COUNT, GGML_TYPE_COUNT, true));
             test_cases.emplace_back(new test_rms_norm_mul_add(GGML_TYPE_BF16, { n, 5, 4, 3 }, eps, false, false, false, false, false, GGML_TYPE_F32));
             test_cases.emplace_back(new test_rms_norm_mul_add(GGML_TYPE_BF16, { n, 5, 4, 3 }, eps, true, false, false, false, false, GGML_TYPE_F32));
             test_cases.emplace_back(new test_rms_norm_mul_add(GGML_TYPE_BF16, { n, 5, 4, 3 }, eps, true, false, false, false, false, GGML_TYPE_F32, GGML_TYPE_BF16));