Merge upstream ollama/ollama main branch while preserving CUDA 3.7 support

- Added support for new gpt-oss model from upstream
- Preserved CUDA Compute Capability 3.7 (Tesla K80) support
- Kept CUDA 11 configuration alongside CUDA 12
- Maintained all documentation specific to ollama37 fork
- Integrated new tool parsing improvements
- Added new backend methods and patches from upstream

convert/convert.go

@@ -202,6 +202,8 @@ func ConvertModel(fsys fs.FS, f *os.File) error {
 		conv = &bertModel{}
 	case "CohereForCausalLM":
 		conv = &commandrModel{}
+	case "GptOssForCausalLM":
+		conv = &gptossModel{}
 	default:
 		return fmt.Errorf("unsupported architecture %q", p.Architectures[0])
 	}

convert/convert_gptoss.go

@@ -0,0 +1,178 @@
+package convert
+
+import (
+	"bytes"
+	"cmp"
+	"encoding/binary"
+	"io"
+	"slices"
+	"strings"
+
+	"github.com/ollama/ollama/fs/ggml"
+	"github.com/pdevine/tensor"
+	"github.com/pdevine/tensor/native"
+)
+
+type gptossModel struct {
+	ModelParameters
+	HiddenLayers         uint32  `json:"num_hidden_layers"`
+	HiddenSize           uint32  `json:"hidden_size"`
+	IntermediateSize     uint32  `json:"intermediate_size"`
+	AttentionHeads       uint32  `json:"num_attention_heads"`
+	KeyValueHeads        uint32  `json:"num_key_value_heads"`
+	HeadDim              uint32  `json:"head_dim"`
+	Experts              uint32  `json:"num_experts"`
+	ExpertsPerToken      uint32  `json:"experts_per_token"`
+	RMSNormEpsilon       float32 `json:"rms_norm_eps"`
+	InitialContextLength uint32  `json:"initial_context_length"`
+	RopeTheta            float32 `json:"rope_theta"`
+	RopeScalingFactor    float32 `json:"rope_scaling_factor"`
+	SlidingWindow        uint32  `json:"sliding_window"`
+}
+
+var _ ModelConverter = (*gptossModel)(nil)
+
+func (m *gptossModel) KV(t *Tokenizer) ggml.KV {
+	kv := m.ModelParameters.KV(t)
+	kv["general.architecture"] = "gptoss"
+	kv["general.file_type"] = uint32(4)
+	kv["gptoss.context_length"] = uint32(m.RopeScalingFactor * float32(m.InitialContextLength))
+	kv["gptoss.block_count"] = m.HiddenLayers
+	kv["gptoss.embedding_length"] = m.HiddenSize
+	kv["gptoss.feed_forward_length"] = m.IntermediateSize
+	kv["gptoss.expert_count"] = m.Experts
+	kv["gptoss.expert_used_count"] = m.ExpertsPerToken
+	kv["gptoss.attention.head_count"] = m.AttentionHeads
+	kv["gptoss.attention.head_count_kv"] = m.KeyValueHeads
+	kv["gptoss.attention.key_length"] = m.HeadDim
+	kv["gptoss.attention.value_length"] = m.HeadDim
+	kv["gptoss.attention.layer_norm_rms_epsilon"] = cmp.Or(m.RMSNormEpsilon, 1e-5)
+	kv["gptoss.attention.sliding_window"] = m.SlidingWindow
+	kv["gptoss.rope.freq_base"] = m.RopeTheta
+	kv["gptoss.rope.scaling.factor"] = m.RopeScalingFactor
+	kv["gptoss.rope.scaling.original_context_length"] = m.InitialContextLength
+	kv["tokenizer.ggml.bos_token_id"] = uint32(199998) // <|startoftext|>
+	kv["tokenizer.ggml.add_bos_token"] = false
+	kv["tokenizer.ggml.eos_token_id"] = uint32(199999) // <|endoftext|>
+	kv["tokenizer.ggml.eos_token_ids"] = []int32{
+		199999, /* <|endoftext|> */
+		200002, /* <|return|> */
+		200012, /* <|call|> */
+	}
+	kv["tokenizer.ggml.add_eos_token"] = false
+	return kv
+}
+
+func (m *gptossModel) Tensors(ts []Tensor) []*ggml.Tensor {
+	var out []*ggml.Tensor
+	mxfp4s := make(map[string]*mxfp4)
+	for _, t := range ts {
+		if strings.HasSuffix(t.Name(), ".blocks") || strings.HasSuffix(t.Name(), ".scales") {
+			dot := strings.LastIndex(t.Name(), ".")
+			name, suffix := t.Name()[:dot], t.Name()[dot+1:]
+			if _, ok := mxfp4s[name]; !ok {
+				mxfp4s[name] = &mxfp4{}
+			}
+
+			switch suffix {
+			case "blocks":
+				mxfp4s[name].blocks = t
+			case "scales":
+				mxfp4s[name].scales = t
+			}
+		} else {
+			out = append(out, &ggml.Tensor{
+				Name:     t.Name(),
+				Kind:     t.Kind(),
+				Shape:    t.Shape(),
+				WriterTo: t,
+			})
+		}
+	}
+
+	for name, mxfp4 := range mxfp4s {
+		dims := mxfp4.blocks.Shape()
+		out = append(out, &ggml.Tensor{
+			Name:     name,
+			Kind:     uint32(ggml.TensorTypeMXFP4),
+			Shape:    []uint64{dims[0], dims[1], dims[2] * dims[3] * 2},
+			WriterTo: mxfp4,
+		})
+	}
+
+	return out
+}
+
+func (m *gptossModel) Replacements() []string {
+	return []string{
+		// noop replacements so other replacements will not be applied
+		".blocks", ".blocks",
+		".scales", ".scales",
+		// real replacements
+		"block", "blk",
+		"attn.norm", "attn_norm",
+		"attn.qkv", "attn_qkv",
+		"attn.sinks", "attn_sinks",
+		"attn.out", "attn_out",
+		"mlp.norm", "ffn_norm",
+		"mlp.gate", "ffn_gate_inp",
+		"mlp.mlp1_", "ffn_gate_up_exps.",
+		"mlp.mlp2_", "ffn_down_exps.",
+		"embedding", "token_embd",
+		"norm", "output_norm",
+		"unembedding", "output",
+		"scale", "weight",
+	}
+}
+
+type mxfp4 struct {
+	blocks, scales Tensor
+}
+
+func (m *mxfp4) WriteTo(w io.Writer) (int64, error) {
+	var b bytes.Buffer
+	if _, err := m.blocks.WriteTo(&b); err != nil {
+		return 0, err
+	}
+
+	blocksDims := make([]int, len(m.blocks.Shape()))
+	for i, d := range m.blocks.Shape() {
+		blocksDims[i] = int(d)
+	}
+
+	var blocks tensor.Tensor = tensor.New(tensor.WithShape(blocksDims...), tensor.WithBacking(b.Bytes()))
+
+	var s bytes.Buffer
+	if _, err := m.scales.WriteTo(&s); err != nil {
+		return 0, err
+	}
+
+	scalesDims := slices.Repeat([]int{1}, len(m.blocks.Shape()))
+	for i, d := range m.scales.Shape() {
+		scalesDims[i] = int(d)
+	}
+
+	var scales tensor.Tensor = tensor.New(tensor.WithShape(scalesDims...), tensor.WithBacking(s.Bytes()))
+
+	out, err := tensor.Concat(3, scales, blocks)
+	if err != nil {
+		return 0, err
+	}
+
+	out = tensor.Materialize(out)
+
+	if err := out.Reshape(out.Shape().TotalSize()); err != nil {
+		return 0, err
+	}
+
+	u8s, err := native.VectorU8(out.(*tensor.Dense))
+	if err != nil {
+		return 0, err
+	}
+
+	if err := binary.Write(w, binary.LittleEndian, u8s); err != nil {
+		return 0, err
+	}
+
+	return 0, nil
+}

convert/convert_test.go

@@ -11,14 +11,13 @@ import (
 	"io"
 	"io/fs"
 	"log/slog"
+	"maps"
 	"os"
 	"path/filepath"
 	"slices"
 	"strings"
 	"testing"
 
-	"golang.org/x/exp/maps"
-
 	"github.com/ollama/ollama/fs/ggml"
 )
 
@@ -137,9 +136,7 @@ func TestConvertModel(t *testing.T) {
 				t.Fatal(err)
 			}
 
-			keys := maps.Keys(expect)
-			slices.Sort(keys)
-			for _, k := range keys {
+			for _, k := range slices.Sorted(maps.Keys(expect)) {
 				if v, ok := actual[k]; !ok {
 					t.Errorf("missing %s", k)
 				} else if v != expect[k] {
@@ -343,9 +340,7 @@ func TestConvertAdapter(t *testing.T) {
 
 			actual := generateResultsJSON(t, r, m.KV(), m.Tensors())
 
-			keys := maps.Keys(c.Expected)
-			slices.Sort(keys)
-			for _, k := range keys {
+			for _, k := range slices.Sorted(maps.Keys(c.Expected)) {
 				if v, ok := actual[k]; !ok {
 					t.Errorf("missing %s", k)
 				} else if v != c.Expected[k] {

convert/reader.go

@@ -31,8 +31,10 @@ func (t tensorBase) Shape() []uint64 {
 }
 
 const (
-	tensorKindF32 uint32 = iota
-	tensorKindF16
+	tensorKindFP32 uint32 = iota
+	tensorKindFP16
+	tensorKindMXFP4 = 4
+	tensorKindBF16  = 30
 )
 
 func (t tensorBase) Kind() uint32 {
@@ -43,16 +45,16 @@ func (t tensorBase) Kind() uint32 {
 		t.name == "v.pre_tile_position_embd.weight" ||
 		t.name == "v.post_tile_position_embd.weight" {
 		// these tensors are always F32
-		return 0
+		return tensorKindFP32
 	}
 
 	switch len(t.shape) {
 	case 0:
 		panic("invalid tensor shape")
 	case 1:
-		return tensorKindF32
+		return tensorKindFP32
 	default:
-		return tensorKindF16
+		return tensorKindFP16
 	}
 }
 

convert/reader_safetensors.go

@@ -8,12 +8,12 @@ import (
 	"fmt"
 	"io"
 	"io/fs"
+	"maps"
 	"slices"
 	"strings"
 
 	"github.com/d4l3k/go-bfloat16"
 	"github.com/x448/float16"
-	"golang.org/x/exp/maps"
 )
 
 type safetensorMetadata struct {
@@ -46,8 +46,7 @@ func parseSafetensors(fsys fs.FS, replacer *strings.Replacer, ps ...string) ([]T
 			return nil, err
 		}
 
-		keys := maps.Keys(headers)
-		slices.Sort(keys)
+		keys := slices.Sorted(maps.Keys(headers))
 
 		names := make(map[string]struct{}, len(keys))
 
@@ -94,6 +93,15 @@ type safetensor struct {
 	*tensorBase
 }
 
+func (st safetensor) Kind() uint32 {
+	kind := st.tensorBase.Kind()
+	if st.dtype == "BF16" && kind != tensorKindFP32 {
+		kind = tensorKindBF16
+	}
+
+	return kind
+}
+
 func (st safetensor) Clone() Tensor {
 	return &safetensor{
 		fs:     st.fs,
@@ -151,6 +159,9 @@ func (st safetensor) WriteTo(w io.Writer) (int64, error) {
 		}
 
 		f32s = bfloat16.DecodeFloat32(u8s)
+	case "U8":
+		// U8 tensors do not support repacking or type conversion.
+		return io.CopyN(w, f, st.size)
 	default:
 		return 0, fmt.Errorf("unknown data type: %s", st.dtype)
 	}
@@ -163,15 +174,18 @@ func (st safetensor) WriteTo(w io.Writer) (int64, error) {
 	}
 
 	switch st.Kind() {
-	case tensorKindF32:
+	case tensorKindFP32:
 		return 0, binary.Write(w, binary.LittleEndian, f32s)
-	case tensorKindF16:
+	case tensorKindFP16:
 		f16s := make([]uint16, len(f32s))
 		for i := range f32s {
 			f16s[i] = float16.Fromfloat32(f32s[i]).Bits()
 		}
 
 		return 0, binary.Write(w, binary.LittleEndian, f16s)
+	case tensorKindBF16:
+		u8s := bfloat16.EncodeFloat32(f32s)
+		return 0, binary.Write(w, binary.LittleEndian, u8s)
 	default:
 		return 0, fmt.Errorf("unknown storage type: %d", st.Kind())
 	}

convert/tokenizer.go

@@ -8,11 +8,10 @@ import (
 	"fmt"
 	"io/fs"
 	"log/slog"
+	"maps"
 	"os"
 	"slices"
 	"strings"
-
-	"golang.org/x/exp/maps"
 )
 
 const (
@@ -260,11 +259,8 @@ func parseVocabularyFromTokenizer(fsys fs.FS) (*Vocabulary, error) {
 		tokens[token.ID] = token
 	}
 
-	keys := maps.Keys(tokens)
-	slices.Sort(keys)
-
 	v := Vocabulary{Model: "gpt2"}
-	for _, k := range keys {
+	for _, k := range slices.Sorted(maps.Keys(tokens)) {
 		token := tokens[k]
 		v.Tokens = append(v.Tokens, token.Content)
 		v.Scores = append(v.Scores, float32(token.ID))