Add support for new models and fix GitHub issues

- Add Gemma3n model support with text generation capabilities
- Add new CUDA mean operations for improved performance
- Add macOS documentation and performance tests
- Update LLAMA patches for ROCm/CUDA compatibility
- Fix various model conversion and processing issues
- Update CI workflows and build configurations
- Add library model tests and Shakespeare test data

🤖 Generated with [Claude Code](https://claude.ai/code)

Co-Authored-By: Claude <noreply@anthropic.com>

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

@@ -0,0 +1,165 @@
+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"`
+		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.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",
+	}
+}

convert/convert_mixtral.go

@@ -2,9 +2,6 @@ package convert
 
 import (
 	"fmt"
-	"io"
-	"slices"
-	"strings"
 
 	"github.com/ollama/ollama/fs/ggml"
 )
@@ -30,65 +27,38 @@ func (p *mixtralModel) KV(t *Tokenizer) ggml.KV {
 }
 
 func (p *mixtralModel) Tensors(ts []Tensor) []*ggml.Tensor {
-	oldnew := []string{
-		"model.layers", "blk",
-		"w1", "ffn_gate_exps",
-		"w2", "ffn_down_exps",
-		"w3", "ffn_up_exps",
-	}
-
-	for i := range p.NumLocalExperts {
-		oldnew = append(oldnew, fmt.Sprintf(".block_sparse_moe.experts.%d.", i), ".")
-	}
-
-	// group experts of the same layer (model.layers.%d) and type (w[123]) into a single tensor
-	namer := strings.NewReplacer(oldnew...)
-	experts := make(map[string]experts)
-
-	// merge experts into a single tensor while removing them from ts
-	ts = slices.DeleteFunc(ts, func(t Tensor) bool {
-		if !strings.Contains(t.Name(), ".block_sparse_moe.experts.") {
-			return false
-		}
-
-		name := namer.Replace(t.Name())
-		experts[name] = append(experts[name], t)
-		return true
-	})
-
-	var out []*ggml.Tensor
-	for n, e := range experts {
-		// TODO(mxyng): sanity check experts
-		out = append(out, &ggml.Tensor{
-			Name:     n,
-			Kind:     e[0].Kind(),
-			Shape:    append([]uint64{uint64(len(e))}, e[0].Shape()...),
-			WriterTo: e,
+	merges := make([]merge, 0, p.NumHiddenLayers*6)
+	for i := range p.NumHiddenLayers {
+		merges = append(merges, merge{
+			fmt.Sprintf("blk.%d.*.w1.weight", i),
+			fmt.Sprintf("blk.%d.ffn_gate_exps.weight", i),
+		}, merge{
+			fmt.Sprintf("blk.%d.*.w1.bias", i),
+			fmt.Sprintf("blk.%d.ffn_gate_exps.bias", i),
+		}, merge{
+			fmt.Sprintf("blk.%d.*.w2.weight", i),
+			fmt.Sprintf("blk.%d.ffn_up_exps.weight", i),
+		}, merge{
+			fmt.Sprintf("blk.%d.*.w2.bias", i),
+			fmt.Sprintf("blk.%d.ffn_up_exps.bias", i),
+		}, merge{
+			fmt.Sprintf("blk.%d.*.w3.weight", i),
+			fmt.Sprintf("blk.%d.ffn_down_exps.weight", i),
+		}, merge{
+			fmt.Sprintf("blk.%d.*.w3.bias", i),
+			fmt.Sprintf("blk.%d.ffn_down_exps.bias", i),
 		})
 	}
 
+	out, ts := mergeTensors(ts, merges...)
 	return append(out, p.llamaModel.Tensors(ts)...)
 }
 
 func (p *mixtralModel) Replacements() []string {
 	return append(
 		p.llamaModel.Replacements(),
+		"model.layers", "blk",
 		"block_sparse_moe.gate", "ffn_gate_inp",
+		"block_sparse_moe.experts.", ".",
 	)
 }
-
-type experts []Tensor
-
-func (e experts) WriteTo(w io.Writer) (int64, error) {
-	// TODO(mxyng): experts _should_ be numerically sorted by expert but this should check
-	for _, t := range e {
-		// the canonical merged experts tensor stacks all experts along a new, 0 axis,
-		// e.g. `tensor.Stack(0, e[0], e[1:]...)`, which requires allocating temporary buffers
-		// this accomplishes the same thing by writing each expert tensor in sequence
-		if _, err := t.WriteTo(w); err != nil {
-			return 0, err
-		}
-	}
-
-	return 0, nil
-}

convert/tensor.go

@@ -2,7 +2,9 @@ package convert
 
 import (
 	"cmp"
+	"io"
 	"iter"
+	"path"
 	"slices"
 	"strings"
 
@@ -74,3 +76,54 @@ func splitDim(t Tensor, dim int, splits ...split) iter.Seq[*ggml.Tensor] {
 		}
 	}
 }
+
+type merge struct {
+	pattern, name string
+}
+
+// mergeTensors merges tensors that match a given pattern into a single tensor.
+func mergeTensors(unmatched []Tensor, merges ...merge) (out []*ggml.Tensor, _ []Tensor) {
+	var matched []Tensor
+	for i := range merges {
+		matched, unmatched = slicesSplitFunc(unmatched, func(t Tensor) bool {
+			matched, _ := path.Match(merges[i].pattern, t.Name())
+			return matched
+		})
+
+		if len(matched) > 0 {
+			out = append(out, &ggml.Tensor{
+				Name:     merges[i].name,
+				Kind:     matched[0].Kind(),
+				Shape:    append([]uint64{uint64(len(matched))}, matched[0].Shape()...),
+				WriterTo: mergeGroup(matched),
+			})
+		}
+	}
+
+	return out, unmatched
+}
+
+// slicesSplitFunc splits a slice into two slices based on a predicate function.
+func slicesSplitFunc[S ~[]E, E comparable](s S, fn func(e E) bool) (matched, unmatched S) {
+	for _, e := range s {
+		if fn(e) {
+			matched = append(matched, e)
+		} else {
+			unmatched = append(unmatched, e)
+		}
+	}
+
+	return matched, unmatched
+}
+
+type mergeGroup []Tensor
+
+func (g mergeGroup) WriteTo(w io.Writer) (int64, error) {
+	for _, t := range g {
+		if _, err := t.WriteTo(w); err != nil {
+			return 0, err
+		}
+	}
+
+	return 0, nil
+}

convert/tensor_test.go

@@ -9,6 +9,8 @@ import (
 	"strings"
 	"testing"
 
+	"github.com/google/go-cmp/cmp"
+	"github.com/ollama/ollama/fs/ggml"
 	"github.com/pdevine/tensor"
 )
 
@@ -302,3 +304,99 @@ func TestSplitDim(t *testing.T) {
 		}
 	})
 }
+
+func TestMerge(t *testing.T) {
+	unmatched := []Tensor{
+		&fakeTensor{
+			name:  "a.0.b",
+			shape: []uint64{5, 2},
+			data:  []float32{10, 11, 12, 13, 14, 15, 16, 17, 18, 19},
+		},
+		&fakeTensor{
+			name:  "a.1.b",
+			shape: []uint64{5, 2},
+			data:  []float32{20, 21, 22, 23, 24, 25, 26, 27, 28, 29},
+		},
+		&fakeTensor{
+			name:  "c.0.d",
+			shape: []uint64{5, 2},
+			data:  []float32{30, 31, 32, 33, 34, 35, 36, 37, 38, 39},
+		},
+		&fakeTensor{
+			name:  "c.1.d",
+			shape: []uint64{5, 2},
+			data:  []float32{40, 41, 42, 43, 44, 45, 46, 47, 48, 49},
+		},
+		&fakeTensor{
+			name:  "e.0.f",
+			shape: []uint64{5, 2},
+			data:  []float32{50, 51, 52, 53, 54, 55, 56, 57, 58, 59},
+		},
+	}
+
+	checkMatched := func(t *testing.T, n int, matched []*ggml.Tensor) {
+		for i := range n {
+			got := matched[i]
+			if diff := cmp.Diff([]uint64{2, 5, 2}, got.Shape); diff != "" {
+				t.Errorf("unexpected (-want +got):\n%s", diff)
+			}
+
+			var b bytes.Buffer
+			if _, err := got.WriteTo(&b); err != nil {
+				t.Fatal(err)
+			}
+
+			f32s := make([]float32, 20)
+			if err := binary.Read(&b, binary.LittleEndian, &f32s); err != nil {
+				t.Fatal(err)
+			}
+
+			offset := 10 + (i * 20)
+			want := make([]float32, 20)
+			for j := range 20 {
+				want[j] = float32(offset + j)
+			}
+
+			if diff := cmp.Diff(want, f32s); diff != "" {
+				t.Errorf("unexpected data (-want +got):\n%s", diff)
+			}
+		}
+	}
+
+	t.Run("single merge", func(t *testing.T) {
+		matched, unmatched := mergeTensors(unmatched, merge{"a.*.b", "a.b"})
+		if len(unmatched) != 3 {
+			t.Error("expected 3 remaining tensors, got", len(unmatched))
+		}
+
+		if len(matched) != 1 {
+			t.Error("expected 1 merged tensor, got", len(matched))
+		}
+
+		checkMatched(t, 1, matched)
+	})
+
+	t.Run("multiple merges", func(t *testing.T) {
+		matched, unmatched := mergeTensors(unmatched, merge{"a.*.b", "a.b"}, merge{"c.*.d", "c.d"})
+		if len(unmatched) != 1 {
+			t.Error("expected 1 remaining tensors, got", len(unmatched))
+		}
+
+		if len(matched) != 2 {
+			t.Error("expected 2 merged tensor, got", len(matched))
+		}
+
+		checkMatched(t, 2, matched)
+	})
+
+	t.Run("no match", func(t *testing.T) {
+		matched, unmatched := mergeTensors(unmatched, merge{"x.*.y", "x.y"})
+		if len(unmatched) != 5 {
+			t.Error("expected 5 remaining tensors, got", len(unmatched))
+		}
+
+		if len(matched) != 0 {
+			t.Error("expected no merged tensors, got", len(matched))
+		}
+	})
+}