metal : use function constant for RMS_NORM + SCALE

Replaces the runtime use_scale karg with a Metal function constant. The
norm+mul kernel is compiled with FC_norm_use_scale=false for MUL fusion and
FC_norm_use_scale=true for SCALE fusion, so the fused kernel has no runtime
branch.

Assisted-by: pi:llama.cpp/DeepSeek-V4-Flash-Vision-Exp

ggml/src/ggml-metal/ggml-metal-device.cpp

@@ -1983,7 +1983,48 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_norm(ggml_metal_
 
     ggml_metal_pipeline_with_params res = ggml_metal_library_get_pipeline(lib, name);
     if (!res.pipeline) {
-        res = ggml_metal_library_compile_pipeline(lib, base, name, nullptr);
+        ggml_metal_cv_t cv = ggml_metal_cv_init();
+        ggml_metal_cv_set_bool(cv, false, FC_NORM + 0);
+
+        res = ggml_metal_library_compile_pipeline(lib, base, name, cv);
+
+        ggml_metal_cv_free(cv);
+    }
+
+    res.smem = 32*sizeof(float);
+
+    return res;
+}
+
+ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_norm_scale(ggml_metal_library_t lib, const ggml_tensor * op) {
+    assert(op->op == GGML_OP_NORM || op->op == GGML_OP_RMS_NORM);
+
+    GGML_ASSERT(ggml_is_contiguous_rows(op->src[0]));
+
+    char base[256];
+    char name[256];
+
+    const char * suffix = "";
+    if (op->ne[0] % 4 == 0) {
+        suffix = "_4";
+    }
+
+    switch (op->op) {
+        case GGML_OP_NORM:     snprintf(base, 256, "kernel_norm_mul_f32%s", suffix);     break;
+        case GGML_OP_RMS_NORM: snprintf(base, 256, "kernel_rms_norm_mul_f32%s", suffix); break;
+        default: GGML_ABORT("fatal error");
+    }
+
+    snprintf(name, 256, "%s_use_scale", base);
+
+    ggml_metal_pipeline_with_params res = ggml_metal_library_get_pipeline(lib, name);
+    if (!res.pipeline) {
+        ggml_metal_cv_t cv = ggml_metal_cv_init();
+        ggml_metal_cv_set_bool(cv, true, FC_NORM + 0);
+
+        res = ggml_metal_library_compile_pipeline(lib, base, name, cv);
+
+        ggml_metal_cv_free(cv);
     }
 
     res.smem = 32*sizeof(float);

ggml/src/ggml-metal/ggml-metal-device.h

@@ -154,6 +154,7 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_bin_one
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_l2_norm           (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_group_norm        (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_norm              (ggml_metal_library_t lib, const struct ggml_tensor * op, int32_t n_fuse);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_norm_scale        (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_rope              (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_im2col            (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_conv_transpose_1d (ggml_metal_library_t lib, const struct ggml_tensor * op);

ggml/src/ggml-metal/ggml-metal-impl.h

@@ -114,6 +114,7 @@
 #define FC_SUM_ROWS                    1400
 #define FC_UPSCALE                     1500
 #define FC_GATED_DELTA_NET             1600
+#define FC_NORM                        1700
 
 // op-specific constants
 #define OP_FLASH_ATTN_EXT_NQPSG 8
@@ -613,7 +614,6 @@ typedef struct {
     uint64_t nbf2[3];
     uint64_t nbf3[3];
     float    scale_val;
-    int32_t  use_scale;
 } ggml_metal_kargs_norm;
 
 typedef struct {

ggml/src/ggml-metal/ggml-metal-ops.cpp

@@ -4086,10 +4086,10 @@ int ggml_metal_op_norm(ggml_metal_op_t ctx, int idx) {
         /*.nbf2   =*/ { nb02 },
         /*.nbf3   =*/ { nb03 },
         /*.scale_val =*/ 1.0f,
-        /*.use_scale =*/ 0,
     };
 
     int n_fuse = 1;
+    bool fused_norm_scale = false;
 
     ggml_metal_buffer_id bid_fuse[2] = { bid_src0, bid_src0 };
 
@@ -4131,12 +4131,12 @@ int ggml_metal_op_norm(ggml_metal_op_t ctx, int idx) {
 
         if (fusion && fusion->id == GGML_METAL_FUSION_NORM_SCALE) {
             n_fuse = n;
+            fused_norm_scale = true;
 
             ctx->count_fusions(fusion);
 
             const ggml_tensor * scale_node = ctx->node(idx + 1);
             args.scale_val = ggml_get_op_params_f32(scale_node, 0);
-            args.use_scale = 1;
 
             if (debug_fusion > 1) {
                 GGML_LOG_DEBUG("%s: fuse: %s + SCALE\n", __func__, ggml_op_name(op->op));
@@ -4156,7 +4156,9 @@ int ggml_metal_op_norm(ggml_metal_op_t ctx, int idx) {
         }
     }
 
-    auto pipeline = ggml_metal_library_get_pipeline_norm(lib, op, n_fuse);
+    auto pipeline = fused_norm_scale ?
+        ggml_metal_library_get_pipeline_norm_scale(lib, op) :
+        ggml_metal_library_get_pipeline_norm(lib, op, n_fuse);
 
     int nth = 32; // SIMD width
 

ggml/src/ggml-metal/kernels/norm.metal

@@ -1,5 +1,7 @@
 #include "common.h"
 
+constant bool FC_norm_use_scale [[function_constant(FC_NORM + 0)]];
+
 // F == 1 : norm (no fuse)
 // F == 2 : norm + mul
 // F == 3 : norm + mul + add
@@ -80,7 +82,7 @@ kernel void kernel_norm_fuse_impl(
             y[i00] = (y[i00]*scale);
         }
         if (F == 2) {
-            if (args.use_scale) {
+            if (FC_norm_use_scale) {
                 y[i00] = (y[i00]*scale) * args.scale_val;
             } else {
                 y[i00] = (y[i00]*scale)*f0[i00];
@@ -159,7 +161,7 @@ kernel void kernel_rms_norm_fuse_impl(
             y[i00] = (x[i00]*scale);
         }
         if (F == 2) {
-            if (args.use_scale) {
+            if (FC_norm_use_scale) {
                 y[i00] = (x[i00]*scale) * args.scale_val;
             } else {
                 y[i00] = (x[i00]*scale)*f0[i00];