add new gemma model (#11204)

* update patches

* cherry pick metal mean kernel

* cherry pick cuda mean kernel

* gemma3n

convert/convert.go

@@ -190,6 +190,8 @@ func ConvertModel(fsys fs.FS, f *os.File) error {
 		conv = &gemma2Model{}
 	case "Gemma3ForCausalLM", "Gemma3ForConditionalGeneration":
 		conv = &gemma3Model{Architecture: p.Architectures[0]}
+	case "Gemma3nForConditionalGeneration":
+		conv = &gemma3nModel{}
 	case "Phi3ForCausalLM":
 		conv = &phi3Model{}
 	case "Qwen2ForCausalLM":

convert/convert_gemma3n.go

+package convert
+
+import (
+	"slices"
+	"strings"
+
+	"github.com/ollama/ollama/fs/ggml"
+	"github.com/pdevine/tensor"
+	"github.com/pdevine/tensor/native"
+	"gonum.org/v1/gonum/stat/distuv"
+)
+
+type gemma3nModel struct {
+	ModelParameters
+
+	TextModel struct {
+		ActivationSparsityPattern []float32 `json:"activation_sparsity_pattern"`
+		AltupActiveIdx            uint32    `json:"altup_active_idx"`
+		AltupCoefClip             float32   `json:"altup_coef_clip"`
+		AltupCorrectScale         bool      `json:"altup_correct_scale"`
+		AltupLRMultiplier         float32   `json:"altup_lr_multiplier"`
+		AltupNumInputs            uint32    `json:"altup_num_inputs"`
+		HeadDim                   uint32    `json:"head_dim"`
+		HiddenSize                uint32    `json:"hidden_size"`
+		HiddenSizePerLayerInput   uint32    `json:"hidden_size_per_layer_input"`
+		IntermediateSize          uint32    `json:"intermediate_size"`
+		LaurelRank                uint32    `json:"laurel_rank"`
+		MaxPositionEmbeddings     uint32    `json:"max_position_embeddings"`
+		NumAttentionHeads         uint32    `json:"num_attention_heads"`
+		NumHiddenLayers           uint32    `json:"num_hidden_layers"`
+		NumKeyValueHeads          uint32    `json:"num_key_value_heads"`
+		NumKVSharedLayers         uint32    `json:"num_kv_shared_layers"`
+		RMSNormEPS                float32   `json:"rms_norm_eps"`
+		RopeLocalBaseFreq         float32   `json:"rope_local_base_freq"`
+		RopeTheta                 float32   `json:"rope_theta"`
+		SlidingWindow             uint32    `json:"sliding_window"`
+		LayerTypes                []string  `json:"layer_types"`
+	} `json:"text_config"`
+	VisionModel struct{} `json:"vision_config"`
+}
+
+func (m *gemma3nModel) KV(t *Tokenizer) ggml.KV {
+	kv := m.ModelParameters.KV(t)
+	kv["general.architecture"] = "gemma3n"
+	kv["gemma3n.activation_sparsity_scale"] = slices.Collect(func(yield func(float32) bool) {
+		norm := distuv.Normal{Mu: 0, Sigma: 1}
+		for _, v := range m.TextModel.ActivationSparsityPattern {
+			if !yield(float32(norm.Quantile(float64(v)))) {
+				break
+			}
+		}
+	})
+	kv["gemma3n.altup.active_idx"] = m.TextModel.AltupActiveIdx
+	kv["gemma3n.altup.correct_scale"] = m.TextModel.AltupCorrectScale
+	kv["gemma3n.altup.lr_multiplier"] = m.TextModel.AltupLRMultiplier
+	kv["gemma3n.altup.num_inputs"] = m.TextModel.AltupNumInputs
+	kv["gemma3n.attention.head_count_kv"] = m.TextModel.NumKeyValueHeads
+	kv["gemma3n.attention.head_count"] = m.TextModel.NumAttentionHeads
+	kv["gemma3n.attention.layer_norm_rms_epsilon"] = m.TextModel.RMSNormEPS
+	kv["gemma3n.attention.sliding_window"] = m.TextModel.SlidingWindow
+	kv["gemma3n.attention.sliding_window_pattern"] = slices.Collect(func(yield func(bool) bool) {
+		for _, t := range m.TextModel.LayerTypes {
+			if !yield(t == "sliding_attention") {
+				break
+			}
+		}
+	})
+	kv["gemma3n.attention.shared_kv_layers"] = m.TextModel.NumKVSharedLayers
+	kv["gemma3n.block_count"] = m.TextModel.NumHiddenLayers
+	kv["gemma3n.context_length"] = m.TextModel.MaxPositionEmbeddings
+	kv["gemma3n.embedding_length_per_layer_input"] = m.TextModel.HiddenSizePerLayerInput
+	kv["gemma3n.embedding_length"] = m.TextModel.HiddenSize
+	kv["gemma3n.feed_forward_length"] = m.TextModel.IntermediateSize
+	kv["gemma3n.head_dim"] = m.TextModel.HeadDim
+	kv["gemma3n.laurel_rank"] = m.TextModel.LaurelRank
+	kv["gemma3n.num_kv_shared_layers"] = m.TextModel.NumKVSharedLayers
+	kv["gemma3n.rope.freq_base_local"] = m.TextModel.RopeLocalBaseFreq
+	kv["gemma3n.rope.freq_base"] = m.TextModel.RopeTheta
+	return kv
+}
+
+func (m *gemma3nModel) Tensors(ts []Tensor) []*ggml.Tensor {
+	out, ts := mergeTensors(ts,
+		merge{"altup_proj.*.weight", "altup_proj.weight"},
+		merge{"altup_unembd_proj.*.weight", "altup_unembd_proj.weight"},
+	)
+
+	for _, t := range ts {
+		switch {
+		case strings.Contains(t.Name(), "audio_tower"),
+			strings.Contains(t.Name(), "embed_audio"),
+			strings.Contains(t.Name(), "vision_tower"),
+			strings.Contains(t.Name(), "embed_vision"):
+			// TODO: handle audio and vision towers
+			continue
+		case strings.Contains(t.Name(), "altup_predict_coef"),
+			strings.Contains(t.Name(), "altup_correct_coef"):
+			if m.TextModel.AltupCoefClip > 0 {
+				t.SetRepacker(func(name string, data []float32, shape []uint64) (_ []float32, err error) {
+					dims := make([]int, len(shape))
+					for i := range shape {
+						dims[i] = int(shape[i])
+					}
+
+					var t tensor.Tensor = tensor.New(tensor.WithShape(dims...), tensor.WithBacking(data))
+
+					t, err = tensor.Clamp(t, -m.TextModel.AltupCoefClip, m.TextModel.AltupCoefClip)
+					if err != nil {
+						return nil, err
+					}
+
+					if err := t.Reshape(t.Shape().TotalSize()); err != nil {
+						return nil, err
+					}
+
+					return native.VectorF32(t.(*tensor.Dense))
+				})
+			}
+		}
+
+		out = append(out, &ggml.Tensor{
+			Name:     t.Name(),
+			Kind:     t.Kind(),
+			Shape:    t.Shape(),
+			WriterTo: t,
+		})
+	}
+
+	return out
+}
+
+func (m *gemma3nModel) Replacements() []string {
+	return []string{
+		"model.language_model.embed_tokens_per_layer", "per_layer_token_embd",
+		"model.language_model.embed_tokens", "token_embd",
+		"model.language_model.per_layer_model_projection", "per_layer_model_proj",
+		"model.language_model.per_layer_projection_norm", "per_layer_proj_norm", "model.language_model.altup_projections", "altup_proj",
+		"model.language_model.altup_unembed_projections", "altup_unembd_proj",
+		"model.language_model.norm", "output_norm",
+		"model.language_model.layers", "blk",
+
+		"input_layernorm", "attn_norm",
+		"self_attn.q_proj", "attn_q",
+		"self_attn.q_norm", "attn_q_norm",
+		"self_attn.k_proj", "attn_k",
+		"self_attn.k_norm", "attn_k_norm",
+		"self_attn.v_proj", "attn_v",
+		"self_attn.o_proj", "attn_output",
+		"post_attention_layernorm", "post_attention_norm",
+		"pre_feedforward_layernorm", "ffn_norm",
+		"mlp.gate_proj", "ffn_gate",
+		"mlp.up_proj", "ffn_up",
+		"mlp.down_proj", "ffn_down",
+		"post_feedforward_layernorm", "post_ffw_norm",
+		"per_layer_input_gate", "inp_gate",
+		"per_layer_projection", "proj",
+		"post_per_layer_input_norm", "post_norm",
+		"altup.", "altup_",
+		"modality_router", "router",
+		"prediction_coefs", "predict_coef",
+		"correction_coefs", "correct_coef",
+		"correct_output_scale", "correct_scale.weight",
+		"laurel.", "laurel_",
+		"linear_left", "l",
+		"linear_right", "r",
+		"post_laurel_norm", "post_norm",
+	}
+}

fs/config.go

@@ -10,4 +10,5 @@ type Config interface {
 	Strings(string, ...[]string) []string
 	Ints(string, ...[]int32) []int32
 	Floats(string, ...[]float32) []float32
+	Bools(string, ...[]bool) []bool
 }

fs/ggml/ggml.go

@@ -166,6 +166,11 @@ func (kv KV) Floats(key string, defaultValue ...[]float32) []float32 {
 	return val.values
 }
 
+func (kv KV) Bools(key string, defaultValue ...[]bool) []bool {
+	val, _ := keyValue(kv, key, &array[bool]{values: append(defaultValue, []bool(nil))[0]})
+	return val.values
+}
+
 func (kv KV) OllamaEngineRequired() bool {
 	return slices.Contains([]string{
 		"gemma3",

fs/ggml/gguf.go

@@ -609,6 +609,10 @@ func ggufWriteKV(ws io.WriteSeeker, k string, v any) error {
 		err = writeGGUFArray(ws, ggufTypeString, v)
 	case *array[string]:
 		err = writeGGUFArray(ws, ggufTypeString, v.values)
+	case []bool:
+		err = writeGGUFArray(ws, ggufTypeBool, v)
+	case *array[bool]:
+		err = writeGGUFArray(ws, ggufTypeBool, v.values)
 	default:
 		return fmt.Errorf("improper type for '%s'", k)
 	}

go.mod

@@ -25,6 +25,7 @@ require (
 	github.com/pdevine/tensor v0.0.0-20240510204454-f88f4562727c
 	golang.org/x/image v0.22.0
 	golang.org/x/tools v0.30.0
+	gonum.org/v1/gonum v0.15.0
 )
 
 require (
@@ -44,7 +45,6 @@ require (
 	github.com/xtgo/set v1.0.0 // indirect
 	go4.org/unsafe/assume-no-moving-gc v0.0.0-20231121144256-b99613f794b6 // indirect
 	golang.org/x/xerrors v0.0.0-20200804184101-5ec99f83aff1 // indirect
-	gonum.org/v1/gonum v0.15.0 // indirect
 	gorgonia.org/vecf32 v0.9.0 // indirect
 	gorgonia.org/vecf64 v0.9.0 // indirect
 )

llama/patches/0005-solar-pro.patch

@@ -150,7 +150,7 @@ index 4cce5166..7f6617fa 100644
  llama_model_loader::llama_model_loader(
          const std::string & fname,
 diff --git a/src/llama-model.cpp b/src/llama-model.cpp
-index 3a4e72a3..831b68c0 100644
+index 3a4e72a3..db62973f 100644
 --- a/src/llama-model.cpp
 +++ b/src/llama-model.cpp
 @@ -1402,6 +1402,21 @@ void llama_model::load_hparams(llama_model_loader & ml) {

llama/patches/0008-ensure-KV-cache-is-fully-defragmented.patch

@@ -22,10 +22,10 @@ multiple batches of processing until everything is complete.
  4 files changed, 59 insertions(+), 79 deletions(-)
 
 diff --git a/src/llama-context.cpp b/src/llama-context.cpp
-index c22687e4..c5948e8f 100644
+index dca22d8b..1f3a3956 100644
 --- a/src/llama-context.cpp
 +++ b/src/llama-context.cpp
-@@ -950,9 +950,12 @@ int llama_context::decode(llama_batch & inp_batch) {
+@@ -947,9 +947,12 @@ int llama_context::decode(llama_batch & inp_batch) {
  
          // find KV slot
          if (!kv_self->find_slot(ubatch)) {
@@ -41,7 +41,7 @@ index c22687e4..c5948e8f 100644
          }
  
          ggml_backend_sched_reset(sched.get());
-@@ -1967,9 +1970,12 @@ void llama_context::opt_epoch_iter(
+@@ -1965,9 +1968,12 @@ void llama_context::opt_epoch_iter(
  
              // TODO: not sure if this is needed
              if (!kv_self->find_slot(ubatch)) {

llama/patches/0015-add-argsort-and-cuda-copy-for-i32.patch

@@ -10,10 +10,10 @@ Subject: [PATCH] add argsort and cuda copy for i32
  3 files changed, 192 insertions(+), 2 deletions(-)
 
 diff --git a/ggml/src/ggml-cpu/ops.cpp b/ggml/src/ggml-cpu/ops.cpp
-index becdae07..7a44b6cf 100644
+index 955fec59..654e2f28 100644
 --- a/ggml/src/ggml-cpu/ops.cpp
 +++ b/ggml/src/ggml-cpu/ops.cpp
-@@ -6890,6 +6890,45 @@ static void ggml_compute_forward_argsort_f32(
+@@ -6822,6 +6822,45 @@ static void ggml_compute_forward_argsort_f32(
      }
  }
  
@@ -59,7 +59,7 @@ index becdae07..7a44b6cf 100644
  void ggml_compute_forward_argsort(
      const ggml_compute_params * params,
      ggml_tensor * dst) {
-@@ -6901,6 +6940,10 @@ void ggml_compute_forward_argsort(
+@@ -6833,6 +6872,10 @@ void ggml_compute_forward_argsort(
              {
                  ggml_compute_forward_argsort_f32(params, dst);
              } break;
@@ -195,7 +195,7 @@ index 607ded85..53b02634 100644
 +    }
  }
 diff --git a/ggml/src/ggml-cuda/cpy.cu b/ggml/src/ggml-cuda/cpy.cu
-index 2d46176e..47383486 100644
+index d027271f..4abd01d7 100644
 --- a/ggml/src/ggml-cuda/cpy.cu
 +++ b/ggml/src/ggml-cuda/cpy.cu
 @@ -38,6 +38,13 @@ static __device__ void cpy_1_f16_f32(const char * cxi, char * cdsti) {
@@ -257,7 +257,7 @@ index 2d46176e..47383486 100644
  static __device__ void cpy_blck_f32_q8_0(const char * cxi, char * cdsti) {
      const float * xi = (const float *) cxi;
      block_q8_0 * dsti = (block_q8_0 *) cdsti;
-@@ -631,6 +676,8 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
+@@ -633,6 +678,8 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
          ggml_cpy_f16_f16_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
      } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32) {
          ggml_cpy_f16_f32_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
@@ -266,7 +266,7 @@ index 2d46176e..47383486 100644
      } else {
          GGML_ABORT("%s: unsupported type combination (%s to %s)\n", __func__,
                  ggml_type_name(src0->type), ggml_type_name(src1->type));
-@@ -686,6 +733,8 @@ void* ggml_cuda_cpy_fn(const ggml_tensor * src0, ggml_tensor * src1) {
+@@ -688,6 +735,8 @@ void* ggml_cuda_cpy_fn(const ggml_tensor * src0, ggml_tensor * src1) {
          return (void*) cpy_f32_f16<cpy_1_f32_f16>;
      } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32) {
          return (void*) cpy_f32_f16<cpy_1_f16_f32>;

llama/patches/0019-metal-add-mean-kernel-14267.patch

+From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
+From: Georgi Gerganov <ggerganov@gmail.com>
+Date: Thu, 19 Jun 2025 08:05:21 +0300
+Subject: [PATCH] metal : add mean kernel (#14267)
+
+* metal : add mean kernel
+
+ggml-ci
+
+* cont : dedup implementation
+
+ggml-ci
+---
+ ggml/src/ggml-metal/ggml-metal.m     | 33 ++++++++++++++++---
+ ggml/src/ggml-metal/ggml-metal.metal | 48 ++++++++++++++++++++++------
+ 2 files changed, 67 insertions(+), 14 deletions(-)
+
+diff --git a/ggml/src/ggml-metal/ggml-metal.m b/ggml/src/ggml-metal/ggml-metal.m
+index ee4f2dcb..f20f5615 100644
+--- a/ggml/src/ggml-metal/ggml-metal.m
++++ b/ggml/src/ggml-metal/ggml-metal.m
+@@ -489,6 +489,7 @@ enum ggml_metal_kernel_type {
+     GGML_METAL_KERNEL_TYPE_COS,
+     GGML_METAL_KERNEL_TYPE_NEG,
+     GGML_METAL_KERNEL_TYPE_SUM_ROWS,
++    GGML_METAL_KERNEL_TYPE_MEAN,
+     GGML_METAL_KERNEL_TYPE_POOL_2D_AVG_F32,
+     GGML_METAL_KERNEL_TYPE_POOL_2D_MAX_F32,
+     GGML_METAL_KERNEL_TYPE_ARGMAX,
+@@ -1436,6 +1437,7 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
+         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_COS,                             cos,                             true);
+         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_NEG,                             neg,                             true);
+         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SUM_ROWS,                        sum_rows,                        true);
++        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MEAN,                            mean,                            true);
+         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGMAX,                          argmax,                          true);
+         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_POOL_2D_AVG_F32,                 pool_2d_avg_f32,                 true);
+         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_POOL_2D_MAX_F32,                 pool_2d_max_f32,                 true);
+@@ -1634,6 +1636,7 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
+         case GGML_OP_LOG:
+             return false; // TODO: implement
+         case GGML_OP_SUM_ROWS:
++        case GGML_OP_MEAN:
+         case GGML_OP_SOFT_MAX:
+         case GGML_OP_GROUP_NORM:
+             return has_simdgroup_reduction && ggml_is_contiguous(op->src[0]);
+@@ -2362,11 +2365,30 @@ static bool ggml_metal_encode_node(
+                 [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+             } break;
+         case GGML_OP_SUM_ROWS:
++        case GGML_OP_MEAN:
+             {
+                 GGML_ASSERT(src0->nb[0] == ggml_type_size(src0->type));
+ 
+-                id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SUM_ROWS].pipeline;
++                id<MTLComputePipelineState> pipeline = nil;
++
++                switch (dst->op) {
++                    case GGML_OP_SUM_ROWS:
++                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SUM_ROWS].pipeline;
++                        break;
++                    case GGML_OP_MEAN:
++                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MEAN].pipeline;
++                        break;
++                    default:
++                        GGML_ABORT("fatal error");
++                }
++
++                int nth = 32; // SIMD width
++
++                while (nth < ne00 && nth < (int) pipeline.maxTotalThreadsPerThreadgroup) {
++                    nth *= 2;
++                }
+ 
++                nth = MIN(nth, ne00);
+ 
+                 ggml_metal_kargs_sum_rows args = {
+                    /*.ne00 =*/ ne00,
+@@ -2396,11 +2418,12 @@ static bool ggml_metal_encode_node(
+                 };
+ 
+                 [encoder setComputePipelineState:pipeline];
+-                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+-                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+-                [encoder setBytes:&args length:sizeof(args) atIndex:2];
++                [encoder setBytes:&args length:sizeof(args) atIndex:0];
++                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1];
++                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
++                [encoder setThreadgroupMemoryLength:32*sizeof(float) atIndex:0];
+ 
+-                [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
++                [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+             } break;
+         case GGML_OP_SOFT_MAX:
+             {
+diff --git a/ggml/src/ggml-metal/ggml-metal.metal b/ggml/src/ggml-metal/ggml-metal.metal
+index 9cfddf45..08e8d807 100644
+--- a/ggml/src/ggml-metal/ggml-metal.metal
++++ b/ggml/src/ggml-metal/ggml-metal.metal
+@@ -956,31 +956,61 @@ kernel void kernel_neg(
+     dst[tpig] = -src0[tpig];
+ }
+ 
++template <bool norm>
+ kernel void kernel_sum_rows(
++        constant ggml_metal_kargs_sum_rows & args,
+         device const float * src0,
+         device       float * dst,
+-        constant ggml_metal_kargs_sum_rows & args,
+-        uint3 tpig[[thread_position_in_grid]]) {
+-    int64_t i3 = tpig.z;
+-    int64_t i2 = tpig.y;
+-    int64_t i1 = tpig.x;
++        threadgroup  float * shmem_f32 [[threadgroup(0)]],
++        uint3   tgpig[[threadgroup_position_in_grid]],
++        ushort3 tpitg[[thread_position_in_threadgroup]],
++        ushort  sgitg[[simdgroup_index_in_threadgroup]],
++        ushort  tiisg[[thread_index_in_simdgroup]],
++        ushort3   ntg[[threads_per_threadgroup]]) {
++    int64_t i3 = tgpig.z;
++    int64_t i2 = tgpig.y;
++    int64_t i1 = tgpig.x;
+ 
+     if (i3 >= args.ne03 || i2 >= args.ne02 || i1 >= args.ne01) {
+         return;
+     }
+ 
++    if (sgitg == 0) {
++        shmem_f32[tiisg] = 0.0f;
++    }
++
+     device const float * src_row = (device const float *) ((device const char *) src0 + i1*args.nb01 + i2*args.nb02 + i3*args.nb03);
+     device       float * dst_row = (device       float *) ((device       char *) dst  + i1*args.nb1  + i2*args.nb2  + i3*args.nb3);
+ 
+-    float row_sum = 0;
++    float sumf = 0;
+ 
+-    for (int64_t i0 = 0; i0 < args.ne00; i0++) {
+-        row_sum += src_row[i0];
++    for (int64_t i0 = tpitg.x; i0 < args.ne00; i0 += ntg.x) {
++        sumf += src_row[i0];
+     }
+ 
+-    dst_row[0] = row_sum;
++    sumf = simd_sum(sumf);
++
++    threadgroup_barrier(mem_flags::mem_threadgroup);
++
++    if (tiisg == 0) {
++        shmem_f32[sgitg] = sumf;
++    }
++
++    threadgroup_barrier(mem_flags::mem_threadgroup);
++
++    sumf = shmem_f32[tiisg];
++    sumf = simd_sum(sumf);
++
++    if (tpitg.x == 0) {
++        dst_row[0] = norm ? sumf / args.ne00 : sumf;
++    }
+ }
+ 
++typedef decltype(kernel_sum_rows<false>) kernel_sum_rows_t;
++
++template [[host_name("kernel_sum_rows")]] kernel kernel_sum_rows_t kernel_sum_rows<false>;
++template [[host_name("kernel_mean")]]     kernel kernel_sum_rows_t kernel_sum_rows<true>;
++
+ template<typename T>
+ kernel void kernel_soft_max(
+         device const  char * src0,

ml/backend.go

@@ -253,6 +253,7 @@ type Tensor interface {
 
 	Neg(ctx Context) Tensor
 	Add(ctx Context, t2 Tensor) Tensor
+	Sub(ctx Context, t2 Tensor) Tensor
 	Mul(ctx Context, t2 Tensor) Tensor
 	Div(ctx Context, t2 Tensor) Tensor
 
@@ -276,6 +277,7 @@ type Tensor interface {
 	Tanh(ctx Context) Tensor
 	GELU(ctx Context) Tensor
 	SILU(ctx Context) Tensor
+	RELU(ctx Context) Tensor
 	Sigmoid(ctx Context) Tensor
 
 	Reshape(ctx Context, shape ...int) Tensor
@@ -297,6 +299,12 @@ type Tensor interface {
 
 	TopK(ctx Context, k int) Tensor
 	Argsort(ctx Context) Tensor
+	Mean(ctx Context) Tensor
+	Variance(ctx Context) Tensor
+	Stddev(ctx Context) Tensor
+	Sqr(ctx Context) Tensor
+	Sqrt(ctx Context) Tensor
+	Clamp(ctx Context, min, max float32) Tensor
 }
 
 // ScaledDotProductAttention implements a fused attention

ml/backend/ggml/ggml.go

@@ -297,7 +297,9 @@ func New(modelPath string, params ml.BackendParams) (ml.Backend, error) {
 			if _, ok := meta.Tensors().GroupLayers()["output"]; !ok && t.Name == "token_embd.weight" {
 				createTensor(tensor{source: t, target: "output.weight"}, output.bts, blocks)
 			}
-		case contains(t.Name, "cls", "output", "output_norm"):
+		case contains(t.Name, "cls", "output", "output_norm",
+			"altup_proj", "altup_unembd_proj",
+			"per_layer_token_embd", "per_layer_model_proj", "per_layer_proj_norm"):
 			createTensor(tensor{source: t}, output.bts, blocks)
 		case strings.HasPrefix(t.Name, "v.") || strings.HasPrefix(t.Name, "mm."):
 			// TODO: assign vision tensors to the gpu if possible
@@ -893,6 +895,13 @@ func (t *Tensor) Add(ctx ml.Context, t2 ml.Tensor) ml.Tensor {
 	}
 }
 
+func (t *Tensor) Sub(ctx ml.Context, t2 ml.Tensor) ml.Tensor {
+	return &Tensor{
+		b: t.b,
+		t: C.ggml_sub(ctx.(*Context).ctx, t.t, t2.(*Tensor).t),
+	}
+}
+
 func (t *Tensor) Repeat(ctx ml.Context, dim, n int) ml.Tensor {
 	if dim < 0 || dim >= C.GGML_MAX_DIMS {
 		panic("invalid dimension")
@@ -1200,6 +1209,13 @@ func (t *Tensor) SILU(ctx ml.Context) ml.Tensor {
 	}
 }
 
+func (t *Tensor) RELU(ctx ml.Context) ml.Tensor {
+	return &Tensor{
+		b: t.b,
+		t: C.ggml_relu_inplace(ctx.(*Context).ctx, t.t),
+	}
+}
+
 func (t *Tensor) Conv2D(ctx ml.Context, t2 ml.Tensor, s0, s1, p0, p1, d0, d1 int) ml.Tensor {
 	return &Tensor{
 		b: t.b,
@@ -1275,3 +1291,42 @@ func (t *Tensor) Argsort(ctx ml.Context) ml.Tensor {
 		t: C.ggml_argsort(ctx.(*Context).ctx, t.t, C.GGML_SORT_ORDER_ASC),
 	}
 }
+
+func (t *Tensor) Mean(ctx ml.Context) ml.Tensor {
+	return &Tensor{
+		b: t.b,
+		t: C.ggml_mean(ctx.(*Context).ctx, t.t),
+	}
+}
+
+func (t *Tensor) Variance(ctx ml.Context) ml.Tensor {
+	return t.Add(ctx, t.Mean(ctx).Scale(ctx, -1)).
+		Sqr(ctx).
+		SumRows(ctx).
+		Scale(ctx, 1/float64(t.Dim(0)))
+}
+
+func (t *Tensor) Stddev(ctx ml.Context) ml.Tensor {
+	return t.Variance(ctx).Sqrt(ctx)
+}
+
+func (t *Tensor) Sqr(ctx ml.Context) ml.Tensor {
+	return &Tensor{
+		b: t.b,
+		t: C.ggml_sqr(ctx.(*Context).ctx, t.t),
+	}
+}
+
+func (t *Tensor) Sqrt(ctx ml.Context) ml.Tensor {
+	return &Tensor{
+		b: t.b,
+		t: C.ggml_sqrt(ctx.(*Context).ctx, t.t),
+	}
+}
+
+func (t *Tensor) Clamp(ctx ml.Context, min, max float32) ml.Tensor {
+	return &Tensor{
+		b: t.b,
+		t: C.ggml_clamp(ctx.(*Context).ctx, t.t, C.float(min), C.float(max)),
+	}
+}

ml/backend/ggml/ggml/src/ggml-cuda/common.cuh

@@ -362,6 +362,26 @@ static __device__ __forceinline__ half2 warp_reduce_sum(half2 a) {
 #endif // FP16_AVAILABLE
 }
 
+// Row reduction kernel template - compute sum (norm=false) or mean (norm=true)
+template<bool norm>
+static __global__ void reduce_rows_f32(const float * x, float * dst, const int ncols) {
+    const int row = blockIdx.x;
+    const int col = threadIdx.x;
+
+    float sum = 0.0f;
+    for (int i = col; i < ncols; i += blockDim.x) {
+        sum += x[row * ncols + i];
+    }
+
+    sum = warp_reduce_sum(sum);
+
+    if (col != 0) {
+        return;
+    }
+
+    dst[row] = norm ? sum / ncols : sum;
+}
+
 template<int width = WARP_SIZE>
 static __device__ __forceinline__ float warp_reduce_max(float x) {
 #pragma unroll

ml/backend/ggml/ggml/src/ggml-cuda/ggml-cuda.cu

@@ -35,6 +35,7 @@
 #include "ggml-cuda/ssm-scan.cuh"
 #include "ggml-cuda/sum.cuh"
 #include "ggml-cuda/sumrows.cuh"
+#include "ggml-cuda/mean.cuh"
 #include "ggml-cuda/tsembd.cuh"
 #include "ggml-cuda/unary.cuh"
 #include "ggml-cuda/upscale.cuh"
@@ -2322,6 +2323,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_SUM_ROWS:
             ggml_cuda_op_sum_rows(ctx, dst);
             break;
+        case GGML_OP_MEAN:
+            ggml_cuda_op_mean(ctx, dst);
+            break;
         case GGML_OP_SSM_CONV:
             ggml_cuda_op_ssm_conv(ctx, dst);
             break;
@@ -3211,6 +3215,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_POOL_2D:
         case GGML_OP_SUM:
         case GGML_OP_SUM_ROWS:
+        case GGML_OP_MEAN:
         case GGML_OP_ARGSORT:
         case GGML_OP_ACC:
             return true;

ml/backend/ggml/ggml/src/ggml-cuda/mean.cu

+#include "mean.cuh"
+
+void ggml_cuda_op_mean(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0   = dst->src[0];
+    const float *       src0_d = (const float *) src0->data;
+    float *             dst_d  = (float *) dst->data;
+    cudaStream_t        stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    const int64_t ncols = src0->ne[0];
+    const int64_t nrows = ggml_nrows(src0);
+
+    const dim3 block_dims(WARP_SIZE, 1, 1);
+    const dim3 block_nums(nrows, 1, 1);
+    reduce_rows_f32</*norm*/ true><<<block_nums, block_dims, 0, stream>>>(src0_d, dst_d, ncols);
+}

ml/backend/ggml/ggml/src/ggml-cuda/mean.cuh

+#include "common.cuh"
+
+void ggml_cuda_op_mean(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

ml/backend/ggml/ggml/src/ggml-cuda/sumrows.cu

 #include "sumrows.cuh"
 
-static __global__ void k_sum_rows_f32(const float * x, float * dst, const int ncols) {
-    const int row = blockIdx.x;
-    const int col = threadIdx.x;
-
-    float sum = 0.0f;
-    for (int i = col; i < ncols; i += blockDim.x) {
-        sum += x[row * ncols + i];
-    }
-
-    sum = warp_reduce_sum(sum);
-
-    if (col == 0) {
-        dst[row] = sum;
-    }
-}
-
 void sum_rows_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
     const dim3 block_dims(WARP_SIZE, 1, 1);
     const dim3 block_nums(nrows, 1, 1);
-    k_sum_rows_f32<<<block_nums, block_dims, 0, stream>>>(x, dst, ncols);
+    reduce_rows_f32</*norm*/false><<<block_nums, block_dims, 0, stream>>>(x, dst, ncols);
 }
 
 void ggml_cuda_op_sum_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
@@ -35,5 +19,8 @@ void ggml_cuda_op_sum_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const int64_t ncols = src0->ne[0];
     const int64_t nrows = ggml_nrows(src0);
 
-    sum_rows_f32_cuda(src0_d, dst_d, ncols, nrows, stream);
+    const dim3 block_dims(WARP_SIZE, 1, 1);
+    const dim3 block_nums(nrows, 1, 1);
+
+    reduce_rows_f32</*norm=*/false><<<block_nums, block_dims, 0, stream>>>(src0_d, dst_d, ncols);
 }

ml/backend/ggml/ggml/src/ggml-cuda/sumrows.cuh

 #include "common.cuh"
 
 void sum_rows_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, cudaStream_t stream);
-
 void ggml_cuda_op_sum_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

ml/backend/ggml/ggml/src/ggml-metal/ggml-metal-embed.metal

@@ -3434,31 +3434,61 @@ kernel void kernel_neg(
     dst[tpig] = -src0[tpig];
 }
 
+template <bool norm>
 kernel void kernel_sum_rows(
+        constant ggml_metal_kargs_sum_rows & args,
         device const float * src0,
         device       float * dst,
-        constant ggml_metal_kargs_sum_rows & args,
-        uint3 tpig[[thread_position_in_grid]]) {
-    int64_t i3 = tpig.z;
-    int64_t i2 = tpig.y;
-    int64_t i1 = tpig.x;
+        threadgroup  float * shmem_f32 [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    int64_t i3 = tgpig.z;
+    int64_t i2 = tgpig.y;
+    int64_t i1 = tgpig.x;
 
     if (i3 >= args.ne03 || i2 >= args.ne02 || i1 >= args.ne01) {
         return;
     }
 
+    if (sgitg == 0) {
+        shmem_f32[tiisg] = 0.0f;
+    }
+
     device const float * src_row = (device const float *) ((device const char *) src0 + i1*args.nb01 + i2*args.nb02 + i3*args.nb03);
     device       float * dst_row = (device       float *) ((device       char *) dst  + i1*args.nb1  + i2*args.nb2  + i3*args.nb3);
 
-    float row_sum = 0;
+    float sumf = 0;
 
-    for (int64_t i0 = 0; i0 < args.ne00; i0++) {
-        row_sum += src_row[i0];
+    for (int64_t i0 = tpitg.x; i0 < args.ne00; i0 += ntg.x) {
+        sumf += src_row[i0];
     }
 
-    dst_row[0] = row_sum;
+    sumf = simd_sum(sumf);
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    if (tiisg == 0) {
+        shmem_f32[sgitg] = sumf;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    sumf = shmem_f32[tiisg];
+    sumf = simd_sum(sumf);
+
+    if (tpitg.x == 0) {
+        dst_row[0] = norm ? sumf / args.ne00 : sumf;
+    }
 }
 
+typedef decltype(kernel_sum_rows<false>) kernel_sum_rows_t;
+
+template [[host_name("kernel_sum_rows")]] kernel kernel_sum_rows_t kernel_sum_rows<false>;
+template [[host_name("kernel_mean")]]     kernel kernel_sum_rows_t kernel_sum_rows<true>;
+
 template<typename T>
 kernel void kernel_soft_max(
         device const  char * src0,