Merge branch 'main' into drifkin/array-head-count-simple

fs/ggml/gguf_test.go

+package ggml
+
+import (
+	"bytes"
+	"os"
+	"slices"
+	"testing"
+
+	"github.com/google/go-cmp/cmp"
+)
+
+func TestWriteGGUF(t *testing.T) {
+	w, err := os.CreateTemp(t.TempDir(), "*.bin")
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer w.Close()
+
+	if err := WriteGGUF(w, KV{
+		"general.alignment": uint32(16),
+	}, []*Tensor{
+		{Name: "test.0", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(slices.Repeat([]byte{0}, 2*3*4))},
+		{Name: "test.1", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(slices.Repeat([]byte{0}, 2*3*4))},
+		{Name: "test.2", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(slices.Repeat([]byte{0}, 2*3*4))},
+		{Name: "test.3", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(slices.Repeat([]byte{0}, 2*3*4))},
+		{Name: "test.4", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(slices.Repeat([]byte{0}, 2*3*4))},
+		{Name: "test.5", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(slices.Repeat([]byte{0}, 2*3*4))},
+	}); err != nil {
+		t.Fatal(err)
+	}
+
+	r, err := os.Open(w.Name())
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer r.Close()
+
+	ff, _, err := Decode(r, 0)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	if diff := cmp.Diff(ff.KV(), KV{
+		"general.alignment":       uint32(16),
+		"general.parameter_count": uint64(36),
+	}); diff != "" {
+		t.Errorf("Mismatch (-want +got):\n%s", diff)
+	}
+
+	if diff := cmp.Diff(ff.Tensors(), Tensors{
+		Offset: 336,
+		items: []*Tensor{
+			{Name: "test.0", Offset: 0, Shape: []uint64{2, 3}},
+			{Name: "test.1", Offset: 32, Shape: []uint64{2, 3}},
+			{Name: "test.2", Offset: 64, Shape: []uint64{2, 3}},
+			{Name: "test.3", Offset: 96, Shape: []uint64{2, 3}},
+			{Name: "test.4", Offset: 128, Shape: []uint64{2, 3}},
+			{Name: "test.5", Offset: 160, Shape: []uint64{2, 3}},
+		},
+	}, cmp.AllowUnexported(Tensors{})); diff != "" {
+		t.Errorf("Mismatch (-want +got):\n%s", diff)
+	}
+}

fs/ggml/type.go

 package ggml
 
-import "fmt"
+import (
+	"fmt"
+	"log/slog"
+	"strings"
+)
 
-type fileType uint32
+// FileType is the Go equivalent to llama_ftype used for gguf file typing
+type FileType uint32
 
 const (
-	fileTypeF32 fileType = iota
-	fileTypeF16
-	fileTypeQ4_0
-	fileTypeQ4_1
-	fileTypeQ4_1_F16
-	fileTypeQ4_2 // unused
-	fileTypeQ4_3 // unused
-	fileTypeQ8_0
-	fileTypeQ5_0
-	fileTypeQ5_1
-	fileTypeQ2_K
-	fileTypeQ3_K_S
-	fileTypeQ3_K_M
-	fileTypeQ3_K_L
-	fileTypeQ4_K_S
-	fileTypeQ4_K_M
-	fileTypeQ5_K_S
-	fileTypeQ5_K_M
-	fileTypeQ6_K
-	fileTypeIQ2_XXS
-	fileTypeIQ2_XS
-	fileTypeQ2_K_S
-	fileTypeIQ3_XS
-	fileTypeIQ3_XXS
-	fileTypeIQ1_S
-	fileTypeIQ4_NL
-	fileTypeIQ3_S
-	fileTypeIQ3_M
-	fileTypeIQ2_S
-	fileTypeIQ2_M
-	fileTypeIQ4_XS
-	fileTypeIQ1_M
-	fileTypeBF16
+	FileTypeF32 FileType = iota
+	FileTypeF16
+	FileTypeQ4_0
+	FileTypeQ4_1
+	fileTypeQ4_1_F16 // unused by GGML
+	fileTypeQ4_2     // unused by GGML
+	fileTypeQ4_3     // unused by GGML
+	FileTypeQ8_0
+	FileTypeQ5_0
+	FileTypeQ5_1
+	FileTypeQ2_K
+	FileTypeQ3_K_S
+	FileTypeQ3_K_M
+	FileTypeQ3_K_L
+	FileTypeQ4_K_S
+	FileTypeQ4_K_M
+	FileTypeQ5_K_S
+	FileTypeQ5_K_M
+	FileTypeQ6_K
+	fileTypeIQ2_XXS // not supported by ollama
+	fileTypeIQ2_XS  // not supported by ollama
+	FileTypeQ2_K_S
+	fileTypeIQ3_XS  // not supported by ollama
+	fileTypeIQ3_XXS // not supported by ollama
+	fileTypeIQ1_S   // not supported by ollama
+	fileTypeIQ4_NL  // not supported by ollama
+	fileTypeIQ3_S   // not supported by ollama
+	fileTypeIQ3_M   // not supported by ollama
+	fileTypeIQ2_S   // not supported by ollama
+	fileTypeIQ2_M   // not supported by ollama
+	fileTypeIQ4_XS  // not supported by ollama
+	fileTypeIQ1_M   // not supported by ollama
+	FileTypeBF16
+	fileTypeQ4_0_4_4 // unused by GGML
+	fileTypeQ4_0_4_8 // unused by GGML
+	fileTypeQ4_0_8_8 // unused by GGML
+	fileTypeTQ1_0    // not supported by ollama
+	fileTypeTQ2_0    // not supported by ollama
 
-	fileTypeUnknown
+	FileTypeUnknown = 1024
 )
 
-func ParseFileType(s string) (fileType, error) {
+// ParseFileType parses the provided GGUF file type
+// Only Ollama supported types are considered valid
+func ParseFileType(s string) (FileType, error) {
 	switch s {
 	case "F32":
-		return fileTypeF32, nil
+		return FileTypeF32, nil
 	case "F16":
-		return fileTypeF16, nil
+		return FileTypeF16, nil
 	case "Q4_0":
-		return fileTypeQ4_0, nil
+		return FileTypeQ4_0, nil
 	case "Q4_1":
-		return fileTypeQ4_1, nil
-	case "Q4_1_F16":
-		return fileTypeQ4_1_F16, nil
+		return FileTypeQ4_1, nil
 	case "Q8_0":
-		return fileTypeQ8_0, nil
+		return FileTypeQ8_0, nil
 	case "Q5_0":
-		return fileTypeQ5_0, nil
+		return FileTypeQ5_0, nil
 	case "Q5_1":
-		return fileTypeQ5_1, nil
+		return FileTypeQ5_1, nil
 	case "Q2_K":
-		return fileTypeQ2_K, nil
+		return FileTypeQ2_K, nil
 	case "Q3_K_S":
-		return fileTypeQ3_K_S, nil
+		return FileTypeQ3_K_S, nil
 	case "Q3_K_M":
-		return fileTypeQ3_K_M, nil
+		return FileTypeQ3_K_M, nil
 	case "Q3_K_L":
-		return fileTypeQ3_K_L, nil
+		return FileTypeQ3_K_L, nil
 	case "Q4_K_S":
-		return fileTypeQ4_K_S, nil
-	case "Q4_K_M":
-		return fileTypeQ4_K_M, nil
+		return FileTypeQ4_K_S, nil
+	case "Q4_K_M", "Q4_K":
+		return FileTypeQ4_K_M, nil
 	case "Q5_K_S":
-		return fileTypeQ5_K_S, nil
-	case "Q5_K_M":
-		return fileTypeQ5_K_M, nil
+		return FileTypeQ5_K_S, nil
+	case "Q5_K_M", "Q5_K":
+		return FileTypeQ5_K_M, nil
 	case "Q6_K":
-		return fileTypeQ6_K, nil
-	case "IQ2_XXS":
-		return fileTypeIQ2_XXS, nil
-	case "IQ2_XS":
-		return fileTypeIQ2_XS, nil
+		return FileTypeQ6_K, nil
 	case "Q2_K_S":
-		return fileTypeQ2_K_S, nil
-	case "IQ3_XS":
-		return fileTypeIQ3_XS, nil
-	case "IQ3_XXS":
-		return fileTypeIQ3_XXS, nil
-	case "IQ1_S":
-		return fileTypeIQ1_S, nil
-	case "IQ4_NL":
-		return fileTypeIQ4_NL, nil
-	case "IQ3_S":
-		return fileTypeIQ3_S, nil
-	case "IQ3_M":
-		return fileTypeIQ3_M, nil
-	case "IQ2_S":
-		return fileTypeIQ2_S, nil
-	case "IQ2_M":
-		return fileTypeIQ2_M, nil
-	case "IQ4_XS":
-		return fileTypeIQ4_XS, nil
-	case "IQ1_M":
-		return fileTypeIQ1_M, nil
+		return FileTypeQ2_K_S, nil
 	case "BF16":
-		return fileTypeBF16, nil
+		return FileTypeBF16, nil
 	default:
-		return fileTypeUnknown, fmt.Errorf("unknown fileType: %s", s)
+		supportedFileTypes := []FileType{
+			FileTypeF32,
+			FileTypeF16,
+			FileTypeQ4_K_S,
+			FileTypeQ4_K_M,
+			FileTypeQ8_0,
+			// fsggml.FileTypeBF16, // TODO
+		}
+		strs := make([]string, len(supportedFileTypes))
+		for i := range supportedFileTypes {
+			strs[i] = supportedFileTypes[i].String()
+		}
+
+		return FileTypeUnknown, fmt.Errorf("unsupported quantization type %s - supported types are %s", s, strings.Join(strs, ", "))
 	}
 }
 
-func (t fileType) String() string {
+func (t FileType) String() string {
 	switch t {
-	case fileTypeF32:
+	case FileTypeF32:
 		return "F32"
-	case fileTypeF16:
+	case FileTypeF16:
 		return "F16"
-	case fileTypeQ4_0:
+	case FileTypeQ4_0:
 		return "Q4_0"
-	case fileTypeQ4_1:
+	case FileTypeQ4_1:
 		return "Q4_1"
-	case fileTypeQ4_1_F16:
-		return "Q4_1_F16"
-	case fileTypeQ8_0:
+	case FileTypeQ8_0:
 		return "Q8_0"
-	case fileTypeQ5_0:
+	case FileTypeQ5_0:
 		return "Q5_0"
-	case fileTypeQ5_1:
+	case FileTypeQ5_1:
 		return "Q5_1"
-	case fileTypeQ2_K:
+	case FileTypeQ2_K:
 		return "Q2_K"
-	case fileTypeQ3_K_S:
+	case FileTypeQ3_K_S:
 		return "Q3_K_S"
-	case fileTypeQ3_K_M:
+	case FileTypeQ3_K_M:
 		return "Q3_K_M"
-	case fileTypeQ3_K_L:
+	case FileTypeQ3_K_L:
 		return "Q3_K_L"
-	case fileTypeQ4_K_S:
+	case FileTypeQ4_K_S:
 		return "Q4_K_S"
-	case fileTypeQ4_K_M:
+	case FileTypeQ4_K_M:
 		return "Q4_K_M"
-	case fileTypeQ5_K_S:
+	case FileTypeQ5_K_S:
 		return "Q5_K_S"
-	case fileTypeQ5_K_M:
+	case FileTypeQ5_K_M:
 		return "Q5_K_M"
-	case fileTypeQ6_K:
+	case FileTypeQ6_K:
 		return "Q6_K"
-	case fileTypeIQ2_XXS:
-		return "IQ2_XXS"
-	case fileTypeIQ2_XS:
-		return "IQ2_XS"
-	case fileTypeQ2_K_S:
+	case FileTypeQ2_K_S:
 		return "Q2_K_S"
-	case fileTypeIQ3_XS:
-		return "IQ3_XS"
-	case fileTypeIQ3_XXS:
-		return "IQ3_XXS"
-	case fileTypeIQ1_S:
-		return "IQ1_S"
-	case fileTypeIQ4_NL:
-		return "IQ4_NL"
-	case fileTypeIQ3_S:
-		return "IQ3_S"
-	case fileTypeIQ3_M:
-		return "IQ3_M"
-	case fileTypeIQ2_S:
-		return "IQ2_S"
-	case fileTypeIQ4_XS:
-		return "IQ4_XS"
-	case fileTypeIQ2_M:
-		return "IQ2_M"
-	case fileTypeIQ1_M:
-		return "IQ1_M"
-	case fileTypeBF16:
+	case FileTypeBF16:
 		return "BF16"
 	default:
 		return "unknown"
 	}
 }
 
-func (t fileType) Value() uint32 {
+func (t FileType) Value() uint32 {
 	return uint32(t)
 }
+
+func (ftype FileType) ToTensorType() TensorType {
+	switch ftype {
+	case FileTypeF32:
+		return TensorTypeF32
+	case FileTypeF16:
+		return TensorTypeF16
+	case FileTypeQ4_0:
+		return TensorTypeQ4_0
+	case FileTypeQ4_1:
+		return TensorTypeQ4_1
+	case FileTypeQ8_0:
+		return TensorTypeQ8_0
+	case FileTypeQ5_0:
+		return TensorTypeQ5_0
+	case FileTypeQ5_1:
+		return TensorTypeQ5_1
+	case FileTypeQ2_K:
+		return TensorTypeQ2_K
+	case FileTypeQ3_K_S:
+		return TensorTypeQ3_K
+	case FileTypeQ3_K_M:
+		return TensorTypeQ3_K
+	case FileTypeQ3_K_L:
+		return TensorTypeQ3_K
+	case FileTypeQ4_K_S:
+		return TensorTypeQ4_K
+	case FileTypeQ4_K_M:
+		return TensorTypeQ4_K
+	case FileTypeQ5_K_S:
+		return TensorTypeQ5_K
+	case FileTypeQ5_K_M:
+		return TensorTypeQ5_K
+	case FileTypeQ6_K:
+		return TensorTypeQ6_K
+	case FileTypeQ2_K_S:
+		return TensorTypeQ2_K
+	case FileTypeBF16:
+		return TensorTypeBF16
+	default:
+		slog.Warn("unsupported file type", "type", ftype)
+		return 0 // F32
+	}
+}
+
+// TensorType is equivalent to ggml_type for individual tensor types
+// Note: these are not the same as FileType
+type TensorType uint32
+
+const (
+	TensorTypeF32 TensorType = iota
+	TensorTypeF16
+	TensorTypeQ4_0
+	TensorTypeQ4_1
+	tensorTypeQ4_2 // unused by GGML
+	tensorTypeQ4_3 // unused by GGML
+	TensorTypeQ5_0
+	TensorTypeQ5_1
+	TensorTypeQ8_0
+	TensorTypeQ8_1
+	TensorTypeQ2_K
+	TensorTypeQ3_K
+	TensorTypeQ4_K
+	TensorTypeQ5_K
+	TensorTypeQ6_K
+	TensorTypeQ8_K
+	tensorTypeIQ2_XXS // not supported by ollama
+	tensorTypeIQ2_XS  // not supported by ollama
+	tensorTypeIQ3_XXS // not supported by ollama
+	tensorTypeIQ1_S   // not supported by ollama
+	tensorTypeIQ4_NL  // not supported by ollama
+	tensorTypeIQ3_S   // not supported by ollama
+	tensorTypeIQ2_S   // not supported by ollama
+	tensorTypeIQ4_XS  // not supported by ollama
+	TensorTypeI8
+	TensorTypeI16
+	TensorTypeI32
+	TensorTypeI64
+	TensorTypeF64
+	tensorTypeIQ1_M // not supported by ollama
+	TensorTypeBF16
+	tensorTypeQ4_0_4_4   // unused by GGML
+	tensorTypeQ4_0_4_8   // unused by GGML
+	tensorTypeQ4_0_8_8   // unused by GGML
+	tensorTypeTQ1_0      // not supported by ollama
+	tensorTypeTQ2_0      // not supported by ollama
+	tensorTypeIQ4_NL_4_4 // unused by GGML
+	tensorTypeIQ4_NL_4_8 // unused by GGML
+	tensorTypeIQ4_NL_8_8 // unused by GGML
+)
+
+// ParseFileType parses the provided GGUF file type
+// Only Ollama supported types are considered valid
+func ParseTensorType(s string) (TensorType, error) {
+	switch s {
+	case "F32":
+		return TensorTypeF32, nil
+	case "F16":
+		return TensorTypeF16, nil
+	case "Q4_0":
+		return TensorTypeQ4_0, nil
+	case "Q4_1":
+		return TensorTypeQ4_1, nil
+	case "Q5_0":
+		return TensorTypeQ5_0, nil
+	case "Q5_1":
+		return TensorTypeQ5_1, nil
+	case "Q8_0":
+		return TensorTypeQ8_0, nil
+	case "Q8_1":
+		return TensorTypeQ8_1, nil
+	case "Q2_K":
+		return TensorTypeQ2_K, nil
+	case "Q3_K":
+		return TensorTypeQ3_K, nil
+	case "Q4_K":
+		return TensorTypeQ4_K, nil
+	case "Q5_K":
+		return TensorTypeQ5_K, nil
+	case "Q6_K":
+		return TensorTypeQ6_K, nil
+	case "Q8_K":
+		return TensorTypeQ8_K, nil
+	case "F64":
+		return TensorTypeF64, nil
+	case "BF16":
+		return TensorTypeBF16, nil
+	default:
+		return 0, fmt.Errorf("unsupported quantization type %s", s)
+	}
+}
+
+func (t TensorType) IsQuantized() bool {
+	switch t {
+	case TensorTypeF32, TensorTypeF16, TensorTypeBF16:
+		return false
+	default:
+		return true
+	}
+}
+
+func (t TensorType) RowSize(ne uint64) uint64 {
+	return t.TypeSize() * ne / t.BlockSize()
+}
+
+func (t TensorType) String() string {
+	switch t {
+	case TensorTypeF32:
+		return "F32"
+	case TensorTypeF16:
+		return "F16"
+	case TensorTypeQ4_0:
+		return "Q4_0"
+	case TensorTypeQ4_1:
+		return "Q4_1"
+	case TensorTypeQ5_0:
+		return "Q5_0"
+	case TensorTypeQ5_1:
+		return "Q5_1"
+	case TensorTypeQ8_0:
+		return "Q8_0"
+	case TensorTypeQ8_1:
+		return "Q8_1"
+	case TensorTypeQ2_K:
+		return "Q2_K"
+	case TensorTypeQ3_K:
+		return "Q3_K"
+	case TensorTypeQ4_K:
+		return "Q4_K"
+	case TensorTypeQ5_K:
+		return "Q5_K"
+	case TensorTypeQ6_K:
+		return "Q6_K"
+	case TensorTypeQ8_K:
+		return "Q8_K"
+	case TensorTypeF64:
+		return "F64"
+	case TensorTypeBF16:
+		return "BF16"
+	default:
+		return "unknown"
+	}
+}

go.mod

@@ -11,7 +11,7 @@ require (
 	github.com/spf13/cobra v1.7.0
 	github.com/stretchr/testify v1.9.0
 	github.com/x448/float16 v0.8.4
-	golang.org/x/sync v0.11.0
+	golang.org/x/sync v0.12.0
 )
 
 require (
@@ -70,12 +70,12 @@ require (
 	github.com/twitchyliquid64/golang-asm v0.15.1 // indirect
 	github.com/ugorji/go/codec v1.2.12 // indirect
 	golang.org/x/arch v0.8.0 // indirect
-	golang.org/x/crypto v0.33.0
+	golang.org/x/crypto v0.36.0
 	golang.org/x/exp v0.0.0-20250218142911-aa4b98e5adaa
-	golang.org/x/net v0.35.0 // indirect
-	golang.org/x/sys v0.30.0
-	golang.org/x/term v0.29.0
-	golang.org/x/text v0.22.0
+	golang.org/x/net v0.38.0 // indirect
+	golang.org/x/sys v0.31.0
+	golang.org/x/term v0.30.0
+	golang.org/x/text v0.23.0
 	google.golang.org/protobuf v1.34.1
 	gopkg.in/yaml.v3 v3.0.1 // indirect
 )

go.sum

@@ -214,8 +214,8 @@ golang.org/x/crypto v0.0.0-20190308221718-c2843e01d9a2/go.mod h1:djNgcEr1/C05ACk
 golang.org/x/crypto v0.0.0-20190510104115-cbcb75029529/go.mod h1:yigFU9vqHzYiE8UmvKecakEJjdnWj3jj499lnFckfCI=
 golang.org/x/crypto v0.0.0-20191011191535-87dc89f01550/go.mod h1:yigFU9vqHzYiE8UmvKecakEJjdnWj3jj499lnFckfCI=
 golang.org/x/crypto v0.0.0-20200622213623-75b288015ac9/go.mod h1:LzIPMQfyMNhhGPhUkYOs5KpL4U8rLKemX1yGLhDgUto=
-golang.org/x/crypto v0.33.0 h1:IOBPskki6Lysi0lo9qQvbxiQ+FvsCC/YWOecCHAixus=
-golang.org/x/crypto v0.33.0/go.mod h1:bVdXmD7IV/4GdElGPozy6U7lWdRXA4qyRVGJV57uQ5M=
+golang.org/x/crypto v0.36.0 h1:AnAEvhDddvBdpY+uR+MyHmuZzzNqXSe/GvuDeob5L34=
+golang.org/x/crypto v0.36.0/go.mod h1:Y4J0ReaxCR1IMaabaSMugxJES1EpwhBHhv2bDHklZvc=
 golang.org/x/exp v0.0.0-20180321215751-8460e604b9de/go.mod h1:CJ0aWSM057203Lf6IL+f9T1iT9GByDxfZKAQTCR3kQA=
 golang.org/x/exp v0.0.0-20180807140117-3d87b88a115f/go.mod h1:CJ0aWSM057203Lf6IL+f9T1iT9GByDxfZKAQTCR3kQA=
 golang.org/x/exp v0.0.0-20190121172915-509febef88a4/go.mod h1:CJ0aWSM057203Lf6IL+f9T1iT9GByDxfZKAQTCR3kQA=
@@ -257,8 +257,8 @@ golang.org/x/net v0.0.0-20200822124328-c89045814202/go.mod h1:/O7V0waA8r7cgGh81R
 golang.org/x/net v0.0.0-20201021035429-f5854403a974/go.mod h1:sp8m0HH+o8qH0wwXwYZr8TS3Oi6o0r6Gce1SSxlDquU=
 golang.org/x/net v0.0.0-20210405180319-a5a99cb37ef4/go.mod h1:p54w0d4576C0XHj96bSt6lcn1PtDYWL6XObtHCRCNQM=
 golang.org/x/net v0.0.0-20210614182718-04defd469f4e/go.mod h1:9nx3DQGgdP8bBQD5qxJ1jj9UTztislL4KSBs9R2vV5Y=
-golang.org/x/net v0.35.0 h1:T5GQRQb2y08kTAByq9L4/bz8cipCdA8FbRTXewonqY8=
-golang.org/x/net v0.35.0/go.mod h1:EglIi67kWsHKlRzzVMUD93VMSWGFOMSZgxFjparz1Qk=
+golang.org/x/net v0.38.0 h1:vRMAPTMaeGqVhG5QyLJHqNDwecKTomGeqbnfZyKlBI8=
+golang.org/x/net v0.38.0/go.mod h1:ivrbrMbzFq5J41QOQh0siUuly180yBYtLp+CKbEaFx8=
 golang.org/x/oauth2 v0.0.0-20180821212333-d2e6202438be/go.mod h1:N/0e6XlmueqKjAGxoOufVs8QHGRruUQn6yWY3a++T0U=
 golang.org/x/oauth2 v0.0.0-20200107190931-bf48bf16ab8d/go.mod h1:gOpvHmFTYa4IltrdGE7lF6nIHvwfUNPOp7c8zoXwtLw=
 golang.org/x/sync v0.0.0-20180314180146-1d60e4601c6f/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
@@ -268,8 +268,8 @@ golang.org/x/sync v0.0.0-20190423024810-112230192c58/go.mod h1:RxMgew5VJxzue5/jJ
 golang.org/x/sync v0.0.0-20190911185100-cd5d95a43a6e/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
 golang.org/x/sync v0.0.0-20201020160332-67f06af15bc9/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
 golang.org/x/sync v0.0.0-20210220032951-036812b2e83c/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
-golang.org/x/sync v0.11.0 h1:GGz8+XQP4FvTTrjZPzNKTMFtSXH80RAzG+5ghFPgK9w=
-golang.org/x/sync v0.11.0/go.mod h1:Czt+wKu1gCyEFDUtn0jG5QVvpJ6rzVqr5aXyt9drQfk=
+golang.org/x/sync v0.12.0 h1:MHc5BpPuC30uJk597Ri8TV3CNZcTLu6B6z4lJy+g6Jw=
+golang.org/x/sync v0.12.0/go.mod h1:1dzgHSNfp02xaA81J2MS99Qcpr2w7fw1gpm99rleRqA=
 golang.org/x/sys v0.0.0-20180830151530-49385e6e1522/go.mod h1:STP8DvDyc/dI5b8T5hshtkjS+E42TnysNCUPdjciGhY=
 golang.org/x/sys v0.0.0-20190215142949-d0b11bdaac8a/go.mod h1:STP8DvDyc/dI5b8T5hshtkjS+E42TnysNCUPdjciGhY=
 golang.org/x/sys v0.0.0-20190312061237-fead79001313/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs=
@@ -285,17 +285,17 @@ golang.org/x/sys v0.0.0-20210510120138-977fb7262007/go.mod h1:oPkhp1MJrh7nUepCBc
 golang.org/x/sys v0.0.0-20210630005230-0f9fa26af87c/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
 golang.org/x/sys v0.5.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
 golang.org/x/sys v0.6.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
-golang.org/x/sys v0.30.0 h1:QjkSwP/36a20jFYWkSue1YwXzLmsV5Gfq7Eiy72C1uc=
-golang.org/x/sys v0.30.0/go.mod h1:/VUhepiaJMQUp4+oa/7Zr1D23ma6VTLIYjOOTFZPUcA=
+golang.org/x/sys v0.31.0 h1:ioabZlmFYtWhL+TRYpcnNlLwhyxaM9kWTDEmfnprqik=
+golang.org/x/sys v0.31.0/go.mod h1:BJP2sWEmIv4KK5OTEluFJCKSidICx8ciO85XgH3Ak8k=
 golang.org/x/term v0.0.0-20201126162022-7de9c90e9dd1/go.mod h1:bj7SfCRtBDWHUb9snDiAeCFNEtKQo2Wmx5Cou7ajbmo=
-golang.org/x/term v0.29.0 h1:L6pJp37ocefwRRtYPKSWOWzOtWSxVajvz2ldH/xi3iU=
-golang.org/x/term v0.29.0/go.mod h1:6bl4lRlvVuDgSf3179VpIxBF0o10JUpXWOnI7nErv7s=
+golang.org/x/term v0.30.0 h1:PQ39fJZ+mfadBm0y5WlL4vlM7Sx1Hgf13sMIY2+QS9Y=
+golang.org/x/term v0.30.0/go.mod h1:NYYFdzHoI5wRh/h5tDMdMqCqPJZEuNqVR5xJLd/n67g=
 golang.org/x/text v0.3.0/go.mod h1:NqM8EUOU14njkJ3fqMW+pc6Ldnwhi/IjpwHt7yyuwOQ=
 golang.org/x/text v0.3.3/go.mod h1:5Zoc/QRtKVWzQhOtBMvqHzDpF6irO9z98xDceosuGiQ=
 golang.org/x/text v0.3.5/go.mod h1:5Zoc/QRtKVWzQhOtBMvqHzDpF6irO9z98xDceosuGiQ=
 golang.org/x/text v0.3.6/go.mod h1:5Zoc/QRtKVWzQhOtBMvqHzDpF6irO9z98xDceosuGiQ=
-golang.org/x/text v0.22.0 h1:bofq7m3/HAFvbF51jz3Q9wLg3jkvSPuiZu/pD1XwgtM=
-golang.org/x/text v0.22.0/go.mod h1:YRoo4H8PVmsu+E3Ou7cqLVH8oXWIHVoX0jqUWALQhfY=
+golang.org/x/text v0.23.0 h1:D71I7dUrlY+VX0gQShAThNGHFxZ13dGLBHQLVl1mJlY=
+golang.org/x/text v0.23.0/go.mod h1:/BLNzu4aZCJ1+kcD0DNRotWKage4q2rGVAg4o22unh4=
 golang.org/x/tools v0.0.0-20180525024113-a5b4c53f6e8b/go.mod h1:n7NCudcB/nEzxVGmLbDWY5pfWTLqBcC2KZ6jyYvM4mQ=
 golang.org/x/tools v0.0.0-20180917221912-90fa682c2a6e/go.mod h1:n7NCudcB/nEzxVGmLbDWY5pfWTLqBcC2KZ6jyYvM4mQ=
 golang.org/x/tools v0.0.0-20190114222345-bf090417da8b/go.mod h1:n7NCudcB/nEzxVGmLbDWY5pfWTLqBcC2KZ6jyYvM4mQ=

integration/embed_test.go

@@ -34,13 +34,15 @@ func cosineSimilarity[V float32 | float64](v1, v2 []V) V {
 func TestAllMiniLMEmbeddings(t *testing.T) {
 	ctx, cancel := context.WithTimeout(context.Background(), 2*time.Minute)
 	defer cancel()
+	client, _, cleanup := InitServerConnection(ctx, t)
+	defer cleanup()
 
 	req := api.EmbeddingRequest{
 		Model:  "all-minilm",
 		Prompt: "why is the sky blue?",
 	}
 
-	res, err := embeddingTestHelper(ctx, t, req)
+	res, err := embeddingTestHelper(ctx, client, t, req)
 
 	if err != nil {
 		t.Fatalf("error: %v", err)
@@ -62,13 +64,15 @@ func TestAllMiniLMEmbeddings(t *testing.T) {
 func TestAllMiniLMEmbed(t *testing.T) {
 	ctx, cancel := context.WithTimeout(context.Background(), 2*time.Minute)
 	defer cancel()
+	client, _, cleanup := InitServerConnection(ctx, t)
+	defer cleanup()
 
 	req := api.EmbedRequest{
 		Model: "all-minilm",
 		Input: "why is the sky blue?",
 	}
 
-	res, err := embedTestHelper(ctx, t, req)
+	res, err := embedTestHelper(ctx, client, t, req)
 
 	if err != nil {
 		t.Fatalf("error: %v", err)
@@ -98,13 +102,15 @@ func TestAllMiniLMEmbed(t *testing.T) {
 func TestAllMiniLMBatchEmbed(t *testing.T) {
 	ctx, cancel := context.WithTimeout(context.Background(), 2*time.Minute)
 	defer cancel()
+	client, _, cleanup := InitServerConnection(ctx, t)
+	defer cleanup()
 
 	req := api.EmbedRequest{
 		Model: "all-minilm",
 		Input: []string{"why is the sky blue?", "why is the grass green?"},
 	}
 
-	res, err := embedTestHelper(ctx, t, req)
+	res, err := embedTestHelper(ctx, client, t, req)
 
 	if err != nil {
 		t.Fatalf("error: %v", err)
@@ -144,6 +150,8 @@ func TestAllMiniLMBatchEmbed(t *testing.T) {
 func TestAllMiniLMEmbedTruncate(t *testing.T) {
 	ctx, cancel := context.WithTimeout(context.Background(), 2*time.Minute)
 	defer cancel()
+	client, _, cleanup := InitServerConnection(ctx, t)
+	defer cleanup()
 
 	truncTrue, truncFalse := true, false
 
@@ -182,7 +190,7 @@ func TestAllMiniLMEmbedTruncate(t *testing.T) {
 	res := make(map[string]*api.EmbedResponse)
 
 	for _, req := range reqs {
-		response, err := embedTestHelper(ctx, t, req.Request)
+		response, err := embedTestHelper(ctx, client, t, req.Request)
 		if err != nil {
 			t.Fatalf("error: %v", err)
 		}
@@ -198,7 +206,7 @@ func TestAllMiniLMEmbedTruncate(t *testing.T) {
 	}
 
 	// check that truncate set to false returns an error if context length is exceeded
-	_, err := embedTestHelper(ctx, t, api.EmbedRequest{
+	_, err := embedTestHelper(ctx, client, t, api.EmbedRequest{
 		Model:    "all-minilm",
 		Input:    "why is the sky blue?",
 		Truncate: &truncFalse,
@@ -210,9 +218,7 @@ func TestAllMiniLMEmbedTruncate(t *testing.T) {
 	}
 }
 
-func embeddingTestHelper(ctx context.Context, t *testing.T, req api.EmbeddingRequest) (*api.EmbeddingResponse, error) {
-	client, _, cleanup := InitServerConnection(ctx, t)
-	defer cleanup()
+func embeddingTestHelper(ctx context.Context, client *api.Client, t *testing.T, req api.EmbeddingRequest) (*api.EmbeddingResponse, error) {
 	if err := PullIfMissing(ctx, client, req.Model); err != nil {
 		t.Fatalf("failed to pull model %s: %v", req.Model, err)
 	}
@@ -226,9 +232,7 @@ func embeddingTestHelper(ctx context.Context, t *testing.T, req api.EmbeddingReq
 	return response, nil
 }
 
-func embedTestHelper(ctx context.Context, t *testing.T, req api.EmbedRequest) (*api.EmbedResponse, error) {
-	client, _, cleanup := InitServerConnection(ctx, t)
-	defer cleanup()
+func embedTestHelper(ctx context.Context, client *api.Client, t *testing.T, req api.EmbedRequest) (*api.EmbedResponse, error) {
 	if err := PullIfMissing(ctx, client, req.Model); err != nil {
 		t.Fatalf("failed to pull model %s: %v", req.Model, err)
 	}

integration/model_arch_test.go

@@ -48,17 +48,6 @@ var (
 	}
 )
 
-func getTimeouts(t *testing.T) (soft time.Duration, hard time.Duration) {
-	deadline, hasDeadline := t.Deadline()
-	if !hasDeadline {
-		return 8 * time.Minute, 10 * time.Minute
-	} else if deadline.Compare(time.Now().Add(2*time.Minute)) <= 0 {
-		t.Skip("too little time")
-		return time.Duration(0), time.Duration(0)
-	}
-	return -time.Since(deadline.Add(-2 * time.Minute)), -time.Since(deadline.Add(-20 * time.Second))
-}
-
 func TestModelsGenerate(t *testing.T) {
 	softTimeout, hardTimeout := getTimeouts(t)
 	slog.Info("Setting timeouts", "soft", softTimeout, "hard", hardTimeout)

integration/quantization_test.go

+//go:build integration && models
+
+package integration
+
+import (
+	"bytes"
+	"context"
+	"fmt"
+	"log/slog"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/ollama/ollama/api"
+)
+
+func TestQuantization(t *testing.T) {
+	sourceModels := []string{
+		"qwen2.5:0.5b-instruct-fp16",
+	}
+	quantizations := []string{
+		"Q8_0",
+		"Q4_K_S",
+		"Q4_K_M",
+		"Q4_K",
+	}
+	softTimeout, hardTimeout := getTimeouts(t)
+	started := time.Now()
+	slog.Info("Setting timeouts", "soft", softTimeout, "hard", hardTimeout)
+	ctx, cancel := context.WithTimeout(context.Background(), hardTimeout)
+	defer cancel()
+	client, _, cleanup := InitServerConnection(ctx, t)
+	defer cleanup()
+
+	for _, base := range sourceModels {
+		if err := PullIfMissing(ctx, client, base); err != nil {
+			t.Fatalf("pull failed %s", err)
+		}
+		for _, quant := range quantizations {
+			newName := fmt.Sprintf("%s__%s", base, quant)
+			t.Run(newName, func(t *testing.T) {
+				if time.Now().Sub(started) > softTimeout {
+					t.Skip("skipping remaining tests to avoid excessive runtime")
+				}
+				req := &api.CreateRequest{
+					Model:        newName,
+					Quantization: quant,
+					From:         base,
+				}
+				fn := func(resp api.ProgressResponse) error {
+					// fmt.Print(".")
+					return nil
+				}
+				t.Logf("quantizing: %s -> %s", base, quant)
+				if err := client.Create(ctx, req, fn); err != nil {
+					t.Fatalf("create failed %s", err)
+				}
+				defer func() {
+					req := &api.DeleteRequest{
+						Model: newName,
+					}
+					t.Logf("deleting: %s -> %s", base, quant)
+					if err := client.Delete(ctx, req); err != nil {
+						t.Logf("failed to clean up %s: %s", req.Model, err)
+					}
+				}()
+				// Check metadata on the model
+				resp, err := client.Show(ctx, &api.ShowRequest{Name: newName})
+				if err != nil {
+					t.Fatalf("unable to show model: %s", err)
+				}
+				if !strings.Contains(resp.Details.QuantizationLevel, quant) {
+					t.Fatalf("unexpected quantization for %s:\ngot: %s", newName, resp.Details.QuantizationLevel)
+				}
+
+				stream := true
+				genReq := api.GenerateRequest{
+					Model:     newName,
+					Prompt:    "why is the sky blue?",
+					KeepAlive: &api.Duration{Duration: 3 * time.Second},
+					Options: map[string]any{
+						"seed":        42,
+						"temperature": 0.0,
+					},
+					Stream: &stream,
+				}
+				t.Logf("verifying: %s -> %s", base, quant)
+
+				// Some smaller quantizations can cause models to have poor quality
+				// or get stuck in repetition loops, so we stop as soon as we have any matches
+				anyResp := []string{"rayleigh", "scattering", "day", "sun", "moon", "color", "nitrogen", "oxygen"}
+				reqCtx, reqCancel := context.WithCancel(ctx)
+				atLeastOne := false
+				var buf bytes.Buffer
+				genfn := func(response api.GenerateResponse) error {
+					buf.Write([]byte(response.Response))
+					fullResp := strings.ToLower(buf.String())
+					for _, resp := range anyResp {
+						if strings.Contains(fullResp, resp) {
+							atLeastOne = true
+							t.Log(fullResp)
+							reqCancel()
+							break
+						}
+					}
+					return nil
+				}
+
+				done := make(chan int)
+				var genErr error
+				go func() {
+					genErr = client.Generate(reqCtx, &genReq, genfn)
+					done <- 0
+				}()
+
+				select {
+				case <-done:
+					if genErr != nil && !atLeastOne {
+						t.Fatalf("failed with %s request prompt %s ", genReq.Model, genReq.Prompt)
+					}
+				case <-ctx.Done():
+					t.Error("outer test context done while waiting for generate")
+				}
+
+				t.Logf("passed")
+
+			})
+		}
+	}
+}

integration/utils_test.go

@@ -217,6 +217,7 @@ func InitServerConnection(ctx context.Context, t *testing.T) (*api.Client, strin
 					slog.Error("failed to open server log", "logfile", lifecycle.ServerLogFile, "error", err)
 					return
 				}
+				defer fp.Close()
 				data, err := io.ReadAll(fp)
 				if err != nil {
 					slog.Error("failed to read server log", "logfile", lifecycle.ServerLogFile, "error", err)
@@ -358,3 +359,14 @@ func skipUnderMinVRAM(t *testing.T, gb uint64) {
 		}
 	}
 }
+
+func getTimeouts(t *testing.T) (soft time.Duration, hard time.Duration) {
+	deadline, hasDeadline := t.Deadline()
+	if !hasDeadline {
+		return 8 * time.Minute, 10 * time.Minute
+	} else if deadline.Compare(time.Now().Add(2*time.Minute)) <= 0 {
+		t.Skip("too little time")
+		return time.Duration(0), time.Duration(0)
+	}
+	return -time.Since(deadline.Add(-2 * time.Minute)), -time.Since(deadline.Add(-20 * time.Second))
+}

kvcache/causal.go

@@ -239,7 +239,7 @@ func (c *Causal) findStartLoc() (int, error) {
 		}
 	}
 
-	return 0, fmt.Errorf("%w (length: %v)", ErrKvCacheFull, len(c.cells))
+	return 0, fmt.Errorf("%w (cache: %v batch: %v)", ErrKvCacheFull, len(c.cells), c.curBatchSize)
 }
 
 func (c *Causal) updateSlidingWindow() {

llama/build-info.cpp

 int LLAMA_BUILD_NUMBER = 0;
-char const *LLAMA_COMMIT = "2016f07bd106c73699ecbaace80f55db5ed95dac";
+char const *LLAMA_COMMIT = "e1e8e0991ffd9e99a445c6812bb519d5bac9f4b5";
 char const *LLAMA_COMPILER = "";
 char const *LLAMA_BUILD_TARGET = "";

llama/llama.cpp/common/common.h

@@ -342,6 +342,8 @@ struct common_params {
 
     // multimodal models (see examples/llava)
     struct common_params_model mmproj;
+    bool mmproj_use_gpu = true;     // use GPU for multimodal model
+    bool no_mmproj = false;         // explicitly disable multimodal model
     std::vector<std::string> image; // path to image file(s)
 
     // embedding

llama/llama.cpp/common/json-schema-to-grammar.cpp

@@ -16,6 +16,9 @@ using json = nlohmann::ordered_json;
 static std::string build_repetition(const std::string & item_rule, int min_items, int max_items, const std::string & separator_rule = "") {
     auto has_max = max_items != std::numeric_limits<int>::max();
 
+    if (max_items == 0) {
+        return "";
+    }
     if (min_items == 0 && max_items == 1) {
         return item_rule + "?";
     }

llama/llama.cpp/examples/llava/clip-impl.h

@@ -2,8 +2,6 @@
 #include "gguf.h"
 #include "clip.h"
 
-#include "clip.h"
-
 #include <climits>
 #include <cstdarg>
 #include <string>
@@ -17,33 +15,31 @@
 #define KEY_FTYPE               "general.file_type"
 #define KEY_NAME                "general.name"
 #define KEY_DESCRIPTION         "general.description"
-#define KEY_HAS_TEXT_ENC        "clip.has_text_encoder"
-#define KEY_HAS_VIS_ENC         "clip.has_vision_encoder"
-#define KEY_HAS_LLAVA_PROJ      "clip.has_llava_projector"
-#define KEY_HAS_MINICPMV_PROJ   "clip.has_minicpmv_projector"
-#define KEY_HAS_GLM_PROJ        "clip.has_glm_projector"
 #define KEY_MINICPMV_VERSION    "clip.minicpmv_version"
-#define KEY_HAS_QWEN2VL_MERGER  "clip.has_qwen2vl_merger"
 #define KEY_USE_GELU            "clip.use_gelu"
 #define KEY_USE_SILU            "clip.use_silu"
-#define KEY_N_EMBD              "clip.%s.embedding_length"
-#define KEY_N_FF                "clip.%s.feed_forward_length"
-#define KEY_N_BLOCK             "clip.%s.block_count"
-#define KEY_N_HEAD              "clip.%s.attention.head_count"
-#define KEY_LAYER_NORM_EPS      "clip.%s.attention.layer_norm_epsilon"
-#define KEY_PROJ_DIM            "clip.%s.projection_dim"
-#define KEY_TOKENS              "tokenizer.ggml.tokens"
-#define KEY_N_POSITIONS         "clip.text.context_length"
+#define KEY_N_EMBD              "clip.vision.embedding_length"
+#define KEY_N_FF                "clip.vision.feed_forward_length"
+#define KEY_N_BLOCK             "clip.vision.block_count"
+#define KEY_N_HEAD              "clip.vision.attention.head_count"
+#define KEY_LAYER_NORM_EPS      "clip.vision.attention.layer_norm_epsilon"
+#define KEY_PROJ_DIM            "clip.vision.projection_dim"
 #define KEY_IMAGE_SIZE          "clip.vision.image_size"
 #define KEY_PATCH_SIZE          "clip.vision.patch_size"
 #define KEY_IMAGE_MEAN          "clip.vision.image_mean"
 #define KEY_IMAGE_STD           "clip.vision.image_std"
-#define KEY_PROJ_TYPE           "clip.projector_type"
 #define KEY_FEATURE_LAYER       "clip.vision.feature_layer"
+#define KEY_PROJ_SCALE_FACTOR   "clip.vision.projector.scale_factor"
+#define KEY_PROJ_TYPE           "clip.projector_type"
+
+#define KEY_USE_GLU_MLP         "clip.use_glu_mlp"  // for qwen2.5vl
+#define KEY_USE_RMS_NORM        "clip.use_rms_norm" // for qwen2.5vl
 
 #define KEY_MM_PATCH_MERGE_TYPE   "clip.vision.mm_patch_merge_type"
 #define KEY_IMAGE_GRID_PINPOINTS  "clip.vision.image_grid_pinpoints"
 #define KEY_IMAGE_CROP_RESOLUTION "clip.vision.image_crop_resolution"
+#define KEY_WIN_ATTN_PATTERN      "clip.vision.n_wa_pattern"
+#define KEY_ATTN_WINDOW_SIZE      "clip.vision.window_size"
 
 
 //
@@ -60,7 +56,9 @@
 #define TN_ATTN_V          "%s.blk.%d.attn_v.%s"
 #define TN_ATTN_OUTPUT     "%s.blk.%d.attn_out.%s"
 #define TN_FFN_DOWN        "%s.blk.%d.ffn_down.%s"
+#define TN_FFN_GATE        "%s.blk.%d.ffn_gate.%s"
 #define TN_FFN_UP          "%s.blk.%d.ffn_up.%s"
+#define TN_FFN_GATE        "%s.blk.%d.ffn_gate.%s"
 #define TN_LN_1            "%s.blk.%d.ln1.%s"
 #define TN_LN_2            "%s.blk.%d.ln2.%s"
 #define TN_LN_PRE          "%s.pre_ln.%s"
@@ -72,6 +70,8 @@
 #define TN_IMAGE_NEWLINE   "model.image_newline"
 #define TN_MM_INP_PROJ     "mm.input_projection.weight" // gemma3
 #define TN_MM_SOFT_EMB_N   "mm.soft_emb_norm.weight"    // gemma3
+#define TN_MM_PROJECTOR    "mm.model.fc.weight"         // idefics3
+#define TN_TOK_IMG_BREAK   "v.token_embd.img_break"     // pixtral
 
 // mimicpmv
 #define TN_MINICPMV_POS_EMBD_K "resampler.pos_embed_k"
@@ -87,18 +87,19 @@
 #define TN_GLM_ADAPTER_D_H_2_4H "adapter.linear.dense_h_to_4h.%s"
 #define TN_GLM_ADAPTER_GATE     "adapter.linear.gate.%s"
 #define TN_GLM_ADAPTER_D_4H_2_H "adapter.linear.dense_4h_to_h.%s"
-#define TN_GLM_BOI_W            "adapter.boi"
-#define TN_GLM_EOI_W            "adapter.eoi"
 
 enum projector_type {
     PROJECTOR_TYPE_MLP,
     PROJECTOR_TYPE_MLP_NORM,
     PROJECTOR_TYPE_LDP,
     PROJECTOR_TYPE_LDPV2,
-    PROJECTOR_TYPE_RESAMPLER,
+    PROJECTOR_TYPE_MINICPMV,
     PROJECTOR_TYPE_GLM_EDGE,
-    PROJECTOR_TYPE_MERGER,
+    PROJECTOR_TYPE_QWEN2VL,
     PROJECTOR_TYPE_GEMMA3,
+    PROJECTOR_TYPE_IDEFICS3,
+    PROJECTOR_TYPE_PIXTRAL,
+    PROJECTOR_TYPE_QWEN25VL,
     PROJECTOR_TYPE_UNKNOWN,
 };
 
@@ -106,10 +107,13 @@ static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
     { PROJECTOR_TYPE_MLP,       "mlp" },
     { PROJECTOR_TYPE_LDP,       "ldp" },
     { PROJECTOR_TYPE_LDPV2,     "ldpv2"},
-    { PROJECTOR_TYPE_RESAMPLER, "resampler"},
+    { PROJECTOR_TYPE_MINICPMV,  "resampler"},
     { PROJECTOR_TYPE_GLM_EDGE,  "adapter"},
-    { PROJECTOR_TYPE_MERGER,    "qwen2vl_merger"},
+    { PROJECTOR_TYPE_QWEN2VL,   "qwen2vl_merger"},
+    { PROJECTOR_TYPE_QWEN25VL,  "qwen2.5vl_merger"},
     { PROJECTOR_TYPE_GEMMA3,    "gemma3"},
+    { PROJECTOR_TYPE_IDEFICS3,  "idefics3"},
+    { PROJECTOR_TYPE_PIXTRAL,   "pixtral"},
 };
 
 static projector_type clip_projector_type_from_string(const std::string & str) {

llama/llama.cpp/examples/llava/clip.h

@@ -47,7 +47,7 @@ CLIP_API struct clip_ctx * clip_init(const char * fname, struct clip_context_par
 CLIP_API void clip_free(struct clip_ctx * ctx);
 
 CLIP_API size_t clip_embd_nbytes(const struct clip_ctx * ctx);
-CLIP_API size_t clip_embd_nbytes_by_img(const struct clip_ctx * ctx, int img_h, int img_w);
+CLIP_API size_t clip_embd_nbytes_by_img(const struct clip_ctx * ctx, int img_w, int img_h);
 
 CLIP_API int32_t clip_get_image_size (const struct clip_ctx * ctx);
 CLIP_API int32_t clip_get_patch_size (const struct clip_ctx * ctx);
@@ -59,9 +59,20 @@ CLIP_API const char * clip_patch_merge_type(const struct clip_ctx * ctx);
 CLIP_API const int32_t * clip_image_grid(const struct clip_ctx * ctx);
 CLIP_API size_t get_clip_image_grid_size(const struct clip_ctx * ctx);
 
-CLIP_API int clip_n_patches        (const struct clip_ctx * ctx);
-CLIP_API int clip_n_patches_by_img (const struct clip_ctx * ctx, struct clip_image_f32 * img);
-CLIP_API int clip_n_mmproj_embd    (const struct clip_ctx * ctx);
+GGML_DEPRECATED(CLIP_API int clip_n_patches(const struct clip_ctx * ctx),
+    "use clip_n_output_tokens instead");
+GGML_DEPRECATED(CLIP_API int clip_n_patches_by_img(const struct clip_ctx * ctx, struct clip_image_f32 * img),
+    "use clip_n_output_tokens instead");
+
+CLIP_API int clip_n_output_tokens(const struct clip_ctx * ctx, struct clip_image_f32 * img);
+
+// for M-RoPE, this will be the number of token positions in X and Y directions
+// for other models, X will be the total number of tokens and Y will be 1
+CLIP_API int clip_n_output_tokens_x(const struct clip_ctx * ctx, struct clip_image_f32 * img);
+CLIP_API int clip_n_output_tokens_y(const struct clip_ctx * ctx, struct clip_image_f32 * img);
+
+// this should be equal to the embedding dimension of the text model
+CLIP_API int clip_n_mmproj_embd(const struct clip_ctx * ctx);
 
 CLIP_API int clip_uhd_num_image_embeds_col(struct clip_ctx * ctx_clip);
 CLIP_API void clip_add_load_image_size(struct clip_ctx * ctx_clip, struct clip_image_size * load_image_size);
@@ -114,8 +125,6 @@ CLIP_API bool clip_is_qwen2vl(const struct clip_ctx * ctx);
 CLIP_API bool clip_is_llava(const struct clip_ctx * ctx);
 CLIP_API bool clip_is_gemma3(const struct clip_ctx * ctx);
 
-CLIP_API int get_deepest_feature_layer(const struct clip_ctx * ctx);
-
 CLIP_API bool clip_encode_float_image (struct clip_ctx * ctx, int n_threads, float * img, int h, int w, float * vec);
 
 

llama/llama.cpp/examples/llava/llava.cpp

@@ -112,7 +112,7 @@ static struct clip_image_grid_shape get_anyres_image_grid_shape(const std::pair<
 }
 
 // Take the image segments in a grid configuration and return the embeddings and the number of embeddings into preallocated memory (image_embd_out)
-static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector<float *> & image_embd_v, struct clip_image_grid_shape grid_shape, float * image_embd_out, int * n_img_pos_out) {
+static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector<float *> & image_embd_v, struct clip_image_grid_shape grid_shape, float * image_embd_out, int * n_img_pos_out, clip_image_f32 * img_input) {
     struct {
         struct ggml_context * ctx;
     } model;
@@ -175,7 +175,7 @@ static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector<float *>
 
     model.ctx = ggml_init(params);
 
-    struct ggml_tensor * image_features = ggml_new_tensor_3d(model.ctx, GGML_TYPE_F32, clip_n_mmproj_embd(ctx_clip), clip_n_patches(ctx_clip), num_images - 1); // example: 4096 x 576 x 4
+    struct ggml_tensor * image_features = ggml_new_tensor_3d(model.ctx, GGML_TYPE_F32, clip_n_mmproj_embd(ctx_clip), clip_n_output_tokens(ctx_clip, img_input), num_images - 1); // example: 4096 x 576 x 4
     // ggml_tensor_printf(image_features,"image_features",__LINE__,false,false);
     // fill it with the image embeddings, ignoring the base
     for (size_t i = 1; i < num_images; i++) {
@@ -214,8 +214,8 @@ static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector<float *>
 
     memcpy(image_embd_out, image_embd_v[0], clip_embd_nbytes(ctx_clip)); // main image as global context
     // append without newline tokens (default behavior in llava_arch when not using unpad ):
-    memcpy(image_embd_out + clip_n_patches(ctx_clip) * clip_n_mmproj_embd(ctx_clip), (float*)result->data, clip_embd_nbytes(ctx_clip) * (num_images-1)); // grid patches
-    *n_img_pos_out = static_cast<int>(result->ne[1]+clip_n_patches(ctx_clip));
+    memcpy(image_embd_out + clip_n_output_tokens(ctx_clip, img_input) * clip_n_mmproj_embd(ctx_clip), (float*)result->data, clip_embd_nbytes(ctx_clip) * (num_images-1)); // grid patches
+    *n_img_pos_out = static_cast<int>(result->ne[1]+clip_n_output_tokens(ctx_clip, img_input));
 
     // Debug: Test single segments
     // Current findings: sending base image, sending a segment embedding all works similar to python
@@ -313,7 +313,7 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
                 image_embd + n_img_pos_out * clip_n_mmproj_embd(ctx_clip),
                 image_embd_v[i],
                 clip_embd_nbytes_by_img(ctx_clip, nx, ny));
-            n_img_pos_out += clip_n_patches_by_img(ctx_clip, img_res);
+            n_img_pos_out += clip_n_output_tokens(ctx_clip, img_res);
         }
         *n_img_pos = n_img_pos_out;
         for (size_t i = 0; i < image_embd_v.size(); i++) {
@@ -342,8 +342,8 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
     }
     else if (strcmp(mm_patch_merge_type, "spatial_unpad") != 0) {
         // flat / default llava-1.5 type embedding
-        *n_img_pos = clip_n_patches(ctx_clip);
         clip_image_f32 * img_res = clip_image_f32_get_img(img_res_v.get(), 0);
+        *n_img_pos = clip_n_output_tokens(ctx_clip, img_res);
         bool encoded = clip_image_encode(ctx_clip, n_threads, img_res, image_embd); // image_embd shape is 576 x 4096
         if (!encoded) {
             LOG_ERR("Unable to encode image\n");
@@ -381,7 +381,8 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
         struct clip_image_grid_shape grid_shape = get_anyres_image_grid_shape({img->nx,img->ny}, grid_pinpoints, image_size);
 
         int n_img_pos_out;
-        clip_llava_handle_patches(ctx_clip, image_embd_v, grid_shape, image_embd, &n_img_pos_out);
+        clip_image_f32 * img_input = clip_image_f32_get_img(img_res_v.get(), 0);
+        clip_llava_handle_patches(ctx_clip, image_embd_v, grid_shape, image_embd, &n_img_pos_out, img_input);
         *n_img_pos = n_img_pos_out;
 
         for (size_t i = 0; i < image_embd_v.size(); i++) {

llama/llama.cpp/include/llama.h

@@ -111,6 +111,7 @@ extern "C" {
         LLAMA_VOCAB_PRE_TYPE_TRILLION       = 31,
         LLAMA_VOCAB_PRE_TYPE_BAILINGMOE     = 32,
         LLAMA_VOCAB_PRE_TYPE_LLAMA4         = 33,
+        LLAMA_VOCAB_PRE_TYPE_PIXTRAL        = 34,
     };
 
     enum llama_rope_type {
@@ -1237,6 +1238,7 @@ extern "C" {
         "will be removed in the future (see https://github.com/ggml-org/llama.cpp/pull/9896#discussion_r1800920915)");
 
     /// @details Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
+    /// Setting k <= 0 makes this a noop
     LLAMA_API struct llama_sampler * llama_sampler_init_top_k      (int32_t k);
 
     /// @details Nucleus sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751

llama/llama.cpp/src/llama-arch.cpp

@@ -20,6 +20,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_REFACT,           "refact"           },
     { LLM_ARCH_BERT,             "bert"             },
     { LLM_ARCH_NOMIC_BERT,       "nomic-bert"       },
+    { LLM_ARCH_NOMIC_BERT_MOE,   "nomic-bert-moe"   },
     { LLM_ARCH_JINA_BERT_V2,     "jina-bert-v2"     },
     { LLM_ARCH_BLOOM,            "bloom"            },
     { LLM_ARCH_STABLELM,         "stablelm"         },
@@ -73,7 +74,6 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_WAVTOKENIZER_DEC, "wavtokenizer-dec" },
     { LLM_ARCH_PLM,              "plm"              },
     { LLM_ARCH_BAILINGMOE,       "bailingmoe"       },
-    { LLM_ARCH_MISTRAL3,         "mistral3"         },
     { LLM_ARCH_UNKNOWN,          "(unknown)"        },
 };
 
@@ -109,6 +109,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_EXPERT_WEIGHTS_SCALE,              "%s.expert_weights_scale"              },
     { LLM_KV_EXPERT_WEIGHTS_NORM,               "%s.expert_weights_norm"               },
     { LLM_KV_EXPERT_GATING_FUNC,                "%s.expert_gating_func"                },
+    { LLM_KV_MOE_EVERY_N_LAYERS,                "%s.moe_every_n_layers"                },
     { LLM_KV_POOLING_TYPE,                      "%s.pooling_type"                      },
     { LLM_KV_LOGIT_SCALE,                       "%s.logit_scale"                       },
     { LLM_KV_DECODER_START_TOKEN_ID,            "%s.decoder_start_token_id"            },
@@ -511,6 +512,24 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
         },
     },
+    {
+        LLM_ARCH_NOMIC_BERT_MOE,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_TOKEN_EMBD_NORM, "token_embd_norm" },
+            { LLM_TENSOR_TOKEN_TYPES,     "token_types" },
+            { LLM_TENSOR_ATTN_OUT_NORM,   "blk.%d.attn_output_norm" },
+            { LLM_TENSOR_ATTN_QKV,        "blk.%d.attn_qkv" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_LAYER_OUT_NORM,  "blk.%d.layer_output_norm" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+            { LLM_TENSOR_FFN_GATE_INP,    "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,   "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,     "blk.%d.ffn_up_exps" },
+        },
+    },
     {
         LLM_ARCH_JINA_BERT_V2,
         {
@@ -1587,22 +1606,6 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP_SHEXP,       "blk.%d.ffn_up_shexp" },
         },
     },
-    {
-        LLM_ARCH_MISTRAL3,
-        {
-            { LLM_TENSOR_TOKEN_EMBD,  "token_embd" },
-            { LLM_TENSOR_OUTPUT_NORM, "output_norm" },
-            { LLM_TENSOR_ATTN_NORM,   "blk.%d.attn_norm" },
-            { LLM_TENSOR_ATTN_Q,      "blk.%d.attn_q" },
-            { LLM_TENSOR_ATTN_K,      "blk.%d.attn_k" },
-            { LLM_TENSOR_ATTN_V,      "blk.%d.attn_v" },
-            { LLM_TENSOR_ATTN_OUT,    "blk.%d.attn_output" },
-            { LLM_TENSOR_FFN_NORM,    "blk.%d.ffn_norm" },
-            { LLM_TENSOR_FFN_GATE,    "blk.%d.ffn_gate" },
-            { LLM_TENSOR_FFN_UP,      "blk.%d.ffn_up" },
-            { LLM_TENSOR_FFN_DOWN,    "blk.%d.ffn_down" },
-        }
-    },
     {
         LLM_ARCH_UNKNOWN,
         {

llama/llama.cpp/src/llama-arch.h

@@ -24,6 +24,7 @@ enum llm_arch {
     LLM_ARCH_REFACT,
     LLM_ARCH_BERT,
     LLM_ARCH_NOMIC_BERT,
+    LLM_ARCH_NOMIC_BERT_MOE,
     LLM_ARCH_JINA_BERT_V2,
     LLM_ARCH_BLOOM,
     LLM_ARCH_STABLELM,
@@ -75,7 +76,6 @@ enum llm_arch {
     LLM_ARCH_CHAMELEON,
     LLM_ARCH_SOLAR,
     LLM_ARCH_WAVTOKENIZER_DEC,
-    LLM_ARCH_MISTRAL3,
     LLM_ARCH_PLM,
     LLM_ARCH_BAILINGMOE,
     LLM_ARCH_UNKNOWN,
@@ -113,6 +113,7 @@ enum llm_kv {
     LLM_KV_EXPERT_WEIGHTS_SCALE,
     LLM_KV_EXPERT_WEIGHTS_NORM,
     LLM_KV_EXPERT_GATING_FUNC,
+    LLM_KV_MOE_EVERY_N_LAYERS,
     LLM_KV_POOLING_TYPE,
     LLM_KV_LOGIT_SCALE,
     LLM_KV_DECODER_START_TOKEN_ID,

llama/llama.cpp/src/llama-chat.cpp

@@ -50,8 +50,8 @@ static const std::map<std::string, llm_chat_template> LLM_CHAT_TEMPLATES = {
     { "deepseek3",         LLM_CHAT_TEMPLATE_DEEPSEEK_3        },
     { "command-r",         LLM_CHAT_TEMPLATE_COMMAND_R         },
     { "llama3",            LLM_CHAT_TEMPLATE_LLAMA_3           },
-    { "chatglm3",          LLM_CHAT_TEMPLATE_CHATGML_3         },
-    { "chatglm4",          LLM_CHAT_TEMPLATE_CHATGML_4         },
+    { "chatglm3",          LLM_CHAT_TEMPLATE_CHATGLM_3         },
+    { "chatglm4",          LLM_CHAT_TEMPLATE_CHATGLM_4         },
     { "glmedge",           LLM_CHAT_TEMPLATE_GLMEDGE           },
     { "minicpm",           LLM_CHAT_TEMPLATE_MINICPM           },
     { "exaone3",           LLM_CHAT_TEMPLATE_EXAONE_3          },
@@ -62,6 +62,7 @@ static const std::map<std::string, llm_chat_template> LLM_CHAT_TEMPLATES = {
     { "yandex",            LLM_CHAT_TEMPLATE_YANDEX            },
     { "bailing",           LLM_CHAT_TEMPLATE_BAILING           },
     { "llama4",            LLM_CHAT_TEMPLATE_LLAMA4            },
+    { "smolvlm",           LLM_CHAT_TEMPLATE_SMOLVLM           },
 };
 
 llm_chat_template llm_chat_template_from_str(const std::string & name) {
@@ -81,7 +82,9 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
     if (tmpl_contains("<|im_start|>")) {
         return tmpl_contains("<|im_sep|>")
             ? LLM_CHAT_TEMPLATE_PHI_4
-            : LLM_CHAT_TEMPLATE_CHATML;
+            : tmpl_contains("<end_of_utterance>")
+                ? LLM_CHAT_TEMPLATE_SMOLVLM // SmolVLM uses <|im_start|> as BOS, but it is NOT chatml
+                : LLM_CHAT_TEMPLATE_CHATML;
     } else if (tmpl.find("mistral") == 0 || tmpl_contains("[INST]")) {
         if (tmpl_contains("[SYSTEM_PROMPT]")) {
             return LLM_CHAT_TEMPLATE_MISTRAL_V7;
@@ -119,8 +122,12 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
         }
     } else if (tmpl_contains("<|assistant|>") && tmpl_contains("<|end|>")) {
         return LLM_CHAT_TEMPLATE_PHI_3;
+    } else if (tmpl_contains("[gMASK]<sop>")) {
+        return LLM_CHAT_TEMPLATE_CHATGLM_4;
     } else if (tmpl_contains("<|assistant|>") && tmpl_contains("<|user|>")) {
         return tmpl_contains("</s>") ? LLM_CHAT_TEMPLATE_FALCON_3 : LLM_CHAT_TEMPLATE_GLMEDGE;
+    } else if (tmpl_contains("<|{{ item['role'] }}|>") && tmpl_contains("<|begin_of_image|>")) {
+        return LLM_CHAT_TEMPLATE_GLMEDGE;
     } else if (tmpl_contains("<|user|>") && tmpl_contains("<|endoftext|>")) {
         return LLM_CHAT_TEMPLATE_ZEPHYR;
     } else if (tmpl_contains("bos_token + message['role']")) {
@@ -149,9 +156,7 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
         return LLM_CHAT_TEMPLATE_LLAMA_3;
     } else if (tmpl_contains("[gMASK]sop")) {
         // chatglm3-6b
-        return LLM_CHAT_TEMPLATE_CHATGML_3;
-    } else if (tmpl_contains("[gMASK]<sop>")) {
-        return LLM_CHAT_TEMPLATE_CHATGML_4;
+        return LLM_CHAT_TEMPLATE_CHATGLM_3;
     } else if (tmpl_contains(LU8("<用户>"))) {
         // MiniCPM-3B-OpenHermes-2.5-v2-GGUF
         return LLM_CHAT_TEMPLATE_MINICPM;
@@ -432,7 +437,7 @@ int32_t llm_chat_apply_template(
         if (add_ass) {
             ss << "<|start_header_id|>assistant<|end_header_id|>\n\n";
         }
-    } else if (tmpl == LLM_CHAT_TEMPLATE_CHATGML_3) {
+    } else if (tmpl == LLM_CHAT_TEMPLATE_CHATGLM_3) {
         // chatglm3-6b
         ss << "[gMASK]" << "sop";
         for (auto message : chat) {
@@ -442,7 +447,7 @@ int32_t llm_chat_apply_template(
         if (add_ass) {
             ss << "<|assistant|>";
         }
-    } else if (tmpl == LLM_CHAT_TEMPLATE_CHATGML_4) {
+    } else if (tmpl == LLM_CHAT_TEMPLATE_CHATGLM_4 || tmpl == LLM_CHAT_TEMPLATE_GLMEDGE) {
         ss << "[gMASK]" << "<sop>";
         for (auto message : chat) {
             std::string role(message->role);
@@ -451,14 +456,6 @@ int32_t llm_chat_apply_template(
         if (add_ass) {
             ss << "<|assistant|>";
         }
-    } else if (tmpl == LLM_CHAT_TEMPLATE_GLMEDGE) {
-        for (auto message : chat) {
-            std::string role(message->role);
-            ss << "<|" << role << "|>" << "\n" << message->content;
-        }
-        if (add_ass) {
-            ss << "<|assistant|>";
-        }
     } else if (tmpl == LLM_CHAT_TEMPLATE_MINICPM) {
         // MiniCPM-3B-OpenHermes-2.5-v2-GGUF
         for (auto message : chat) {
@@ -620,7 +617,23 @@ int32_t llm_chat_apply_template(
         if (add_ass) {
             ss << "<|header_start|>assistant<|header_end|>\n\n";
         }
-    }  else {
+    } else if (tmpl == LLM_CHAT_TEMPLATE_SMOLVLM) {
+        // SmolVLM
+        ss << "<|im_start|>"; // uses <|im_start|> as BOS, but the actual content is NOT chatml
+        for (auto message : chat) {
+            std::string role(message->role);
+            if (role == "system") {
+                ss << message->content << "\n\n";
+            } else if (role == "user") {
+                ss << "User: " << message->content << "<end_of_utterance>\n";
+            } else {
+                ss << "Assistant: " << message->content << "<end_of_utterance>\n";
+            }
+        }
+        if (add_ass) {
+            ss << "Assistant:";
+        }
+    } else {
         // template not supported
         return -1;
     }