init

convert/reader.go

+package convert
+
+import (
+	"errors"
+	"io"
+	"io/fs"
+	"strings"
+)
+
+type Tensor interface {
+	Name() string
+	Shape() []uint64
+	Kind() uint32
+	SetRepacker(repacker)
+	WriteTo(io.Writer) (int64, error)
+}
+
+type tensorBase struct {
+	name  string
+	shape []uint64
+	repacker
+}
+
+func (t tensorBase) Name() string {
+	return t.name
+}
+
+func (t tensorBase) Shape() []uint64 {
+	return t.shape
+}
+
+const (
+	tensorKindF32 uint32 = iota
+	tensorKindF16
+)
+
+func (t tensorBase) Kind() uint32 {
+	if strings.HasSuffix(t.name, ".block_sparse_moe.gate.weight") {
+		return 0
+	}
+
+	switch len(t.shape) {
+	case 0:
+		panic("invalid tensor shape")
+	case 1:
+		return tensorKindF32
+	default:
+		return tensorKindF16
+	}
+}
+
+func (t *tensorBase) SetRepacker(fn repacker) {
+	t.repacker = fn
+}
+
+type repacker func(string, []float32, []uint64) ([]float32, error)
+
+func parseTensors(fsys fs.FS) ([]Tensor, error) {
+	patterns := []struct {
+		Pattern string
+		Func    func(fs.FS, ...string) ([]Tensor, error)
+	}{
+		{"model-*-of-*.safetensors", parseSafetensors},
+		{"model.safetensors", parseSafetensors},
+		{"pytorch_model-*-of-*.bin", parseTorch},
+		{"pytorch_model.bin", parseTorch},
+		{"consolidated.*.pth", parseTorch},
+	}
+
+	for _, pattern := range patterns {
+		matches, err := fs.Glob(fsys, pattern.Pattern)
+		if err != nil {
+			return nil, err
+		}
+
+		if len(matches) > 0 {
+			return pattern.Func(fsys, matches...)
+		}
+	}
+
+	return nil, errors.New("unknown tensor format")
+}

convert/reader_safetensors.go

+package convert
+
+import (
+	"bytes"
+	"encoding/binary"
+	"encoding/json"
+	"fmt"
+	"io"
+	"io/fs"
+	"slices"
+
+	"github.com/d4l3k/go-bfloat16"
+	"github.com/x448/float16"
+	"golang.org/x/exp/maps"
+)
+
+type safetensorMetadata struct {
+	Type    string   `json:"dtype"`
+	Shape   []uint64 `json:"shape"`
+	Offsets []int64  `json:"data_offsets"`
+}
+
+func parseSafetensors(fsys fs.FS, ps ...string) ([]Tensor, error) {
+	var ts []Tensor
+	for _, p := range ps {
+		f, err := fsys.Open(p)
+		if err != nil {
+			return nil, err
+		}
+		defer f.Close()
+
+		var n int64
+		if err := binary.Read(f, binary.LittleEndian, &n); err != nil {
+			return nil, err
+		}
+
+		b := bytes.NewBuffer(make([]byte, 0, n))
+		if _, err = io.CopyN(b, f, n); err != nil {
+			return nil, err
+		}
+
+		var headers map[string]safetensorMetadata
+		if err := json.NewDecoder(b).Decode(&headers); err != nil {
+			return nil, err
+		}
+
+		keys := maps.Keys(headers)
+		slices.Sort(keys)
+
+		for _, key := range keys {
+			if value := headers[key]; value.Type != "" {
+				ts = append(ts, safetensor{
+					fs:     fsys,
+					path:   p,
+					dtype:  value.Type,
+					offset: safetensorsPad(n, value.Offsets[0]),
+					size:   safetensorsPad(n, value.Offsets[1]) - safetensorsPad(n, value.Offsets[0]),
+					tensorBase: &tensorBase{
+						name:  key,
+						shape: value.Shape,
+					},
+				})
+			}
+		}
+	}
+
+	return ts, nil
+}
+
+// safetensorsPad returns the padded size of the safetensors file given a length n and offset s
+func safetensorsPad(n, offset int64) int64 {
+	return 8 + n + offset
+}
+
+type safetensor struct {
+	fs     fs.FS
+	path   string
+	dtype  string
+	offset int64
+	size   int64
+	*tensorBase
+}
+
+func (st safetensor) WriteTo(w io.Writer) (int64, error) {
+	f, err := st.fs.Open(st.path)
+	if err != nil {
+		return 0, err
+	}
+	defer f.Close()
+
+	if seeker, ok := f.(io.Seeker); ok {
+		if _, err := seeker.Seek(st.offset, io.SeekStart); err != nil {
+			return 0, err
+		}
+	} else {
+		if _, err := io.CopyN(io.Discard, f, st.offset); err != nil {
+			return 0, err
+		}
+	}
+
+	var f32s []float32
+	switch st.dtype {
+	case "F32":
+		f32s = make([]float32, st.size/4)
+		if err = binary.Read(f, binary.LittleEndian, f32s); err != nil {
+			return 0, err
+		}
+	case "F16":
+		u16s := make([]uint16, st.size/2)
+		if err = binary.Read(f, binary.LittleEndian, u16s); err != nil {
+			return 0, err
+		}
+
+		f32s = make([]float32, len(u16s))
+		for i := range u16s {
+			f32s[i] = float16.Frombits(u16s[i]).Float32()
+		}
+
+	case "BF16":
+		u8s := make([]uint8, st.size)
+		if err = binary.Read(f, binary.LittleEndian, u8s); err != nil {
+			return 0, err
+		}
+
+		f32s = bfloat16.DecodeFloat32(u8s)
+	default:
+		return 0, fmt.Errorf("unknown data type: %s", st.dtype)
+	}
+
+	if st.repacker != nil {
+		f32s, err = st.repacker(st.Name(), f32s, st.Shape())
+		if err != nil {
+			return 0, err
+		}
+	}
+
+	switch st.Kind() {
+	case tensorKindF32:
+		return 0, binary.Write(w, binary.LittleEndian, f32s)
+	case tensorKindF16:
+		f16s := make([]uint16, len(f32s))
+		for i := range f32s {
+			f16s[i] = float16.Fromfloat32(f32s[i]).Bits()
+		}
+
+		return 0, binary.Write(w, binary.LittleEndian, f16s)
+	default:
+		return 0, fmt.Errorf("unknown storage type: %d", st.Kind())
+	}
+}

convert/reader_torch.go

+package convert
+
+import (
+	"io"
+	"io/fs"
+
+	"github.com/nlpodyssey/gopickle/pytorch"
+	"github.com/nlpodyssey/gopickle/types"
+)
+
+func parseTorch(fsys fs.FS, ps ...string) ([]Tensor, error) {
+	var ts []Tensor
+	for _, p := range ps {
+		pt, err := pytorch.Load(p)
+		if err != nil {
+			return nil, err
+		}
+
+		for _, k := range pt.(*types.Dict).Keys() {
+			t := pt.(*types.Dict).MustGet(k)
+
+			var shape []uint64
+			for dim := range t.(*pytorch.Tensor).Size {
+				shape = append(shape, uint64(dim))
+			}
+
+			ts = append(ts, torch{
+				storage: t.(*pytorch.Tensor).Source,
+				tensorBase: &tensorBase{
+					name:  k.(string),
+					shape: shape,
+				},
+			})
+		}
+	}
+
+	return ts, nil
+}
+
+type torch struct {
+	storage pytorch.StorageInterface
+	*tensorBase
+}
+
+func (pt torch) WriteTo(w io.Writer) (int64, error) {
+	return 0, nil
+}

convert/testdata/gemma-2b-it.json

+{
+  "general.architecture": "gemma",
+  "general.file_type": "1",
+  "general.quantization_version": "2",
+  "gemma.block_count": "18",
+  "gemma.context_length": "8192",
+  "gemma.embedding_length": "2048",
+  "gemma.feed_forward_length": "16384",
+  "gemma.attention.head_count": "8",
+  "gemma.attention.head_count_kv": "1",
+  "gemma.attention.key_length": "256",
+  "gemma.attention.value_length": "256",
+  "gemma.attention.layer_norm_rms_epsilon": "1e-06",
+  "tokenizer.ggml.model": "llama",
+  "tokenizer.ggml.add_bos_token": "true",
+  "tokenizer.ggml.add_eos_token": "false",
+  "tokenizer.ggml.bos_token_id": "2",
+  "tokenizer.ggml.eos_token_id": "1",
+  "tokenizer.ggml.padding_token_id": "0",
+  "tokenizer.ggml.unknown_token_id": "3",
+  "tokenizer.ggml.scores": "0872465d173867d755d3ee728f882b9dc2057a0bfd596fe1e3d131522f1250d8",
+  "tokenizer.ggml.token_type": "485e40bf3d715a4764818fc097d6a2a41db872d82ee714bc500872a3437ff48d",
+  "tokenizer.ggml.tokens": "c6e66de1841f04de8b8d236d461ab720a4c9b9b5414dc293a09c6e10eab45fda",
+  "token_embd.weight": "17b87ab2c01c80657855a5413d0457b4a041afaeda0cc785080e44e2f04acf07",
+  "blk.0.attn_k.weight": "28ac0da05754ad2714ae95da28a5ad191192140b30b8fd22d108d4700c9d989f",
+  "blk.0.attn_norm.weight": "3f9d5675d1ab0eb8a816719dac9fab81f2e95c52be02c34263339acbc087febb",
+  "blk.0.attn_output.weight": "703295c2c63990ff896778685c678f145298886f680f3ed5dc2a7ad54c293265",
+  "blk.0.attn_q.weight": "69c2d0e4870e9d722a190d356203c9605575a16863466c3d1747966ef1cf5791",
+  "blk.0.attn_v.weight": "95219c9c07b5ffe9a9a01e456d845eef2b11f4fc12c93dbbba479db395444c13",
+  "blk.0.ffn_down.weight": "a2feb5eb3d572c57c5bafbf0ab506862df1160fe40965dcfe4b9fd855c08bed7",
+  "blk.0.ffn_gate.weight": "fcca072c445c31f4dc4d5dfaa785b1bdf7271342442099b74fd17268b5829fbf",
+  "blk.0.ffn_norm.weight": "7621f95dbd245cade6fffd6b08797d69d8e3954e960f0b5551b90d967ab95448",
+  "blk.0.ffn_up.weight": "14a9bcdd451403c67136391e1b6e53b3b1830f00199bd911dbcc56d8749c14f4",
+  "blk.1.attn_k.weight": "c70f73c5df20579cb44d971164b48b5f0d8d5abdb38b381e7a8b880ba12aa406",
+  "blk.1.attn_norm.weight": "88b6b91f93a1ef83425a7c7dc2a2fbd3b22704a04c64a80061df376ac8c33626",
+  "blk.1.attn_output.weight": "f031a537490c452be3b3bb51e6b7949a636405756e160976a1c070a792ea00ee",
+  "blk.1.attn_q.weight": "bdb23214b1cf9cfd30f863a0a5868e52c6809d93b7e8f44df096a94204d9896a",
+  "blk.1.attn_v.weight": "e9bbc0b05f2c872fb1403f8f938cd1612b502229ee401f12593b1164c61acc00",
+  "blk.1.ffn_down.weight": "5ff53811038b661a7b8f2bfdf213bebfb185ec1a6060b662f063714f33584d79",
+  "blk.1.ffn_gate.weight": "205085c8c951a5c7543b1495183cd96028fb49f67464b3e9862a2693a6077a33",
+  "blk.1.ffn_norm.weight": "798f354fc85afce9625f5d10093a585a966831698a0560e6c9b97ce659eb4b22",
+  "blk.1.ffn_up.weight": "db92dc5684cb6e90940e13f4d1da555ed20ba4f8cab1e990ddfd7553e2e91315",
+  "blk.2.attn_k.weight": "ef5ce360c4eed6d00d03ca4761e0f8e4b0af4509978468314be14f3d46621044",
+  "blk.2.attn_norm.weight": "6dadbc05dbd0d3fabb4216affa60a3de1378a82d2859dc90b338cbe70f50d455",
+  "blk.2.attn_output.weight": "6bbf87a966f691bbfd7c8d25629aa4e6710107bd431a667434861febb391edc5",
+  "blk.2.attn_q.weight": "4e575c09ae2de417ce9057ce8b073680e860a24aae13a472b68f101b760752e5",
+  "blk.2.attn_v.weight": "cd33f7f01141e9439afdaf2ea1aaced9feaa335e32a58daa136ebd555d4d96f4",
+  "blk.2.ffn_down.weight": "b970ff1b0b6494165defe2fbfa1d31425766ed71e64de9ec4e66ac3955c8bc5f",
+  "blk.2.ffn_gate.weight": "dbb3e1360402e0e369b101995bb686b73f95d4a7673f061be85d64d15dfb0061",
+  "blk.2.ffn_norm.weight": "bfb7980105d8ac9647710454f57a5cdac50598a0f6f4884e16f1d94b00844687",
+  "blk.2.ffn_up.weight": "50ef89339b275a438b664686f6227dd9b6e43853ed6856ec9e33ef4bbd90bda1",
+  "blk.3.attn_k.weight": "be942ea98151434eebcd2c1da4b00e0146152fe524a530689b1fd491cb833d21",
+  "blk.3.attn_norm.weight": "0df2f218daf609c289fb7c60c5f375fa99c0d4e04381ad5a494a19144edd8e20",
+  "blk.3.attn_output.weight": "c2184aaf86aa2cb8f47be49f60b165834e97205f39c6ee1dfd19fd4411a156ce",
+  "blk.3.attn_q.weight": "4f86e2a0a4221c1c84ff9c409ac89893cb95d7208cf65bf1e98e24e01125f991",
+  "blk.3.attn_v.weight": "abfdb8a60c349dadde641d1afc9542025e24fbf41a3238bfa9675e0b1f1e4b68",
+  "blk.3.ffn_down.weight": "58821a8d87008d47d122427911c6fad5272aca70c448bbae223256a74bacd07e",
+  "blk.3.ffn_gate.weight": "776e051f1a0ddd5c4934e69186683a75ca9a3c8c0f61911bba321fed1dd287d2",
+  "blk.3.ffn_norm.weight": "7f380f29335e28be90bfcfae6f6d69fdf5751211b36d2dd62aa5541ed113e4f2",
+  "blk.3.ffn_up.weight": "fc5ae8d488894cbd4951059675468d227da27871d26e925c9941863841c097ee",
+  "blk.4.attn_k.weight": "14833b078cc4c5137bdd5fdc0538047974ca147a99b0282e1b144440c78bc1db",
+  "blk.4.attn_norm.weight": "0a69957d4a15599fb80ad4753558020804925221457d9a5052926754d3768065",
+  "blk.4.attn_output.weight": "887a49b6130fb6297cf10767207c3dd97191b2cf63723449af9c27bca8dbeda0",
+  "blk.4.attn_q.weight": "51fd577b76764824dd6f0d4891c137ebe4736f591b5ca2793c5fff2be49abbde",
+  "blk.4.attn_v.weight": "1a623c43cf9c509d1b7ea0d1a5c04d0af4809665f9f9e93b7d6dba8c5df178fa",
+  "blk.4.ffn_down.weight": "5d61e8856d8941d2b1fd138116d015f63840d0fa1e31e20e20a5ceca1536ceec",
+  "blk.4.ffn_gate.weight": "06640f7273764f8ca5df7e386547417916b6cd7d565a8343153113239a94b0a1",
+  "blk.4.ffn_norm.weight": "91a6c6c41b894228e361435ecbc5058dca34d4911a23da5b56de219299c964d3",
+  "blk.4.ffn_up.weight": "d016dac1055e36d6a10b6317e57f98a904709ea892ef3194342f4d2f6326561e",
+  "blk.5.attn_k.weight": "987146afe124131500808cc0da33c06d207433656d41df6e6d8c99118a83bac5",
+  "blk.5.attn_norm.weight": "6b354938966f2608a2fb8d0f5b363ed0d8b0967c2ec8d0abd5c625b413042ded",
+  "blk.5.attn_output.weight": "cdcbfe02c6ff79d5326882b017a02099f5af71beedf6b1b3eb4de01e3a844536",
+  "blk.5.attn_q.weight": "b910d0cff781d3efb42eab0a302f46f286b2de717079175680d5b42bf8c309c8",
+  "blk.5.attn_v.weight": "66d3a279f747412f9f4b0e8abad44540c122ab2e811a7ee74c1f33bc36caade9",
+  "blk.5.ffn_down.weight": "c9b0efd2212981f16d956d8571f054b68780ad01f4917033647e359b557a4653",
+  "blk.5.ffn_gate.weight": "fe96b94109ca141c01f6a04788e20783019ca6ec334aa1f3134810bdb499e557",
+  "blk.5.ffn_norm.weight": "aa7b016e832e7055a36c6e20de58ea1936f995f390401fff1c5fc65906064e49",
+  "blk.5.ffn_up.weight": "555ce27c4873d3375394f38ad3b45e3d8848f9d5642dc1602383d0f0a33c2a14",
+  "blk.6.attn_k.weight": "88280d461db324c4f36475ce396793063e61a27283ec64511b0480890fb5b3b4",
+  "blk.6.attn_norm.weight": "af8f460c411f660d33196286d208f1845fd5a2b45f7b56549a4df31e7515447a",
+  "blk.6.attn_output.weight": "dd9996fb0a256e8375ad3917705258a33fce006bcea0f536caae420a77974d8b",
+  "blk.6.attn_q.weight": "7a4841541191e037cfb9b07930c4d8cab451809658b182f0ada6ccde9615c003",
+  "blk.6.attn_v.weight": "ae81e6a592b64d701a9d40233e986039a56cba8d8d24f61aea93c6393cf3078a",
+  "blk.6.ffn_down.weight": "622dd1ce1706355cbc659a8ab2c4509678ffe0f3ad34258e5e25ed2a5d951bcd",
+  "blk.6.ffn_gate.weight": "8389a735c0bd5591010f8ced9805a2a12c749f6df0d3c18ad4d05c2a302e7168",
+  "blk.6.ffn_norm.weight": "621f5346400382474d61358397bd58fb1459b07c53e376e4bca15e08b3f9b3fb",
+  "blk.6.ffn_up.weight": "8d834e4c42f13c251dfee36cf89e12f1bd400680d00d5c2e6cac0459e9ce2f7f",
+  "blk.7.attn_k.weight": "8bd0412de65a3e64901ef8fe6a28c95e116bf39dc9aa22f0126b9d36688e5ea7",
+  "blk.7.attn_norm.weight": "056d8e56be4e87d6dc6f900762f0dc6fde07bfdc50dd85bfc510415e2bba3f3d",
+  "blk.7.attn_output.weight": "27972eda51da53d416ff95aed78149a2c5a287b47d2cd46f2f544ca692ecb3bb",
+  "blk.7.attn_q.weight": "41eca977b9371f7932800c11a9c45b931310196919e2a0651b847703b180fc7f",
+  "blk.7.attn_v.weight": "13c74fd7e07f08883a09fb070a1fe5bbdd2341b4cb8d1cac07c4b637049b5774",
+  "blk.7.ffn_down.weight": "9e75db42468800849a9a7da603d0072c5e86c8ed2b4d8b20a312a51fb86a7a10",
+  "blk.7.ffn_gate.weight": "db6bdc3117f910088aaf7db51f2da63ea5bd933de36af5599c215bfb26f7db2b",
+  "blk.7.ffn_norm.weight": "48bb82b49bfc8679a1e77f282ee182d952db7a3c11be7ef9a102ee2ddd8011e2",
+  "blk.7.ffn_up.weight": "feebea87175817a0f3585ec0af09dc873d94c203581ae97a712eb356d3b49efe",
+  "blk.8.attn_k.weight": "d5640ad71b6af68d88e17bf8e7fc26c907d2262605457a84247dd9afc2884d69",
+  "blk.8.attn_norm.weight": "75b850c481a69083ae09d0207ba7317b37c735a39fcf5fef5400e6c84fb1257f",
+  "blk.8.attn_output.weight": "cbd669dbdea2bdd90f9f0cc97566b3dffff3c56cecb4f47290ceef30da83b2d6",
+  "blk.8.attn_q.weight": "9edcb63087a431bac361822497e6ecdaa06d9ea4a1a754e36da7ba9f8db81c7c",
+  "blk.8.attn_v.weight": "3fb72c2c4f95a83626aa3e30062f9450b09ab37c7871e229f18bbc5cf744633c",
+  "blk.8.ffn_down.weight": "bd69d2c9172974fff154441b237b4787fb53b2d185325442d5048130ef5bc4ef",
+  "blk.8.ffn_gate.weight": "d04689c80553edd011d1cbaa5d570fffa7fa91e88b66cf1352d89ab60b72f908",
+  "blk.8.ffn_norm.weight": "e49984183b735b7f2c4e4730c289eed9394056d2e283a00fd83ea0915df31a73",
+  "blk.8.ffn_up.weight": "8fe62a1ce8e847e567add6c6f6bf2922bc467495b5eb4c116b3cb85b85b3b211",
+  "blk.9.attn_k.weight": "d90904959e5004cf0d6e729c6bff18cc33c094798b802473c1ec55ab8d276183",
+  "blk.9.attn_norm.weight": "79277f290cc07411115d8fa138045edf4a17b3416ab2145409cbe8ab829fd4ee",
+  "blk.9.attn_output.weight": "5a21bf2e1f09a81405025f96d4153ffb630158e17269cff8ffff935c38ceb1a7",
+  "blk.9.attn_q.weight": "51b1d0febc3b350945be4504f55afa4347517bde0f710e1a4b88e6b17e71e7c7",
+  "blk.9.attn_v.weight": "aab7e1db0a8b50a03036356791ffce736ab010d15674c96eaef8049d80076054",
+  "blk.9.ffn_down.weight": "cbf43ec84becb40c9359a181ab0e641fd7faae7d34b549501f7cfb7afdc3d764",
+  "blk.9.ffn_gate.weight": "dce0e8661c778327bed7f03b6790d26710764188aed9dc746e6e05863891fa57",
+  "blk.9.ffn_norm.weight": "6d41642104f995c77bf31122b13237caebda3e7fcccb1367ce91db36b015e923",
+  "blk.9.ffn_up.weight": "82fe4c67bf24e7b2d6f6e05f7b1234c2bf90c3932951091a9066211b8e15ecbb",
+  "blk.10.attn_k.weight": "f6a9ed8fd8d3229b5d03175c413ffc56a07f2ce7236271986361dd3d8993f9aa",
+  "blk.10.attn_norm.weight": "cebbef89f0326ca8e02df3867a571e4d61c20c2a12f295f98ae590d62bc86010",
+  "blk.10.attn_output.weight": "34f5efb86accb4f06347d83a32558ea8eab3039d128969161a741ebacbb656ff",
+  "blk.10.attn_q.weight": "1e0efe27df2d5d50f7157253ba2cfd436d6781c3dc78ca176d0c16a210b5b763",
+  "blk.10.attn_v.weight": "8f085bf50a2b0f83cd6cdda3c8ef5a9e204a36348ed95871aac725d1f68640cf",
+  "blk.10.ffn_down.weight": "bf3b3cb4cace435809ac7b4cc933f20853af12f1f272d3dcefe7f19c0f203b8b",
+  "blk.10.ffn_gate.weight": "d3df7a1413b1c5adf1a1dcda9e5225a15c89874bae53bb6137ad1ea42fca2d34",
+  "blk.10.ffn_norm.weight": "a1da603b0480471b5ed8e862148cecd5fed918f8304d6933ab0bdb25b8d2fb8f",
+  "blk.10.ffn_up.weight": "bffbba605922e972dc47dda88a0b4659aa52236c76e5fe861a949e6d9a367492",
+  "blk.11.attn_k.weight": "9f31c63d66cd32c29b1eb8bb829d0c8525ce2ae936e0eefdaab6335a2d12a3df",
+  "blk.11.attn_norm.weight": "0bde1a266d8b2e8f202bb7e2e88b19147ca83021901f6d3cae77a4df5548c754",
+  "blk.11.attn_output.weight": "e10725c7cf746ed4a7e472cf7aea6cb564e5db6a1d5197adc980d650a387ccea",
+  "blk.11.attn_q.weight": "05ee758a7d065802630f8c65dca424364c1c8825e389aa33f9405c45e8a50cce",
+  "blk.11.attn_v.weight": "0c3ae7090f11775d24c51120db6e305db6aff706493e7ee123dcab74485ba789",
+  "blk.11.ffn_down.weight": "7ba40b8e12c09c5fb2006b77a771cb01ce894e88a3b3e1877f927a5b89c91709",
+  "blk.11.ffn_gate.weight": "db76388a023b98097972d354ba1c6a5e26efdeb1c596b9c28bf2cd8f6596975e",
+  "blk.11.ffn_norm.weight": "a38c3ae1b89a68ddc7b72c99c5b28be7fe3787c4fad9904d0c43d64eaf00c474",
+  "blk.11.ffn_up.weight": "13c8142f9cf1eddc658babf978daf3515c4ccc45f849f3e7e3930aa18a8480a0",
+  "blk.12.attn_k.weight": "f03241c36ac87cb57429a2ef22186b8d7d0b590a8b173beb01fa13d93772f3b1",
+  "blk.12.attn_norm.weight": "4568f654e6d65104d586e7c16ba960c83428698ce103022b7e0be15e2884e13b",
+  "blk.12.attn_output.weight": "04867603f82f91e41306e09b33ecda0104b3ee4834061f2c0bbdc8da33c72509",
+  "blk.12.attn_q.weight": "70fe04b9a8e08b6100cc8d6b58bf4cbbad15ca1de82d63baca5d352ba6c4cbae",
+  "blk.12.attn_v.weight": "15cb28db61a86c98687991d7e611bc92a1fcc6007f3432149cfb5fe518a4f65e",
+  "blk.12.ffn_down.weight": "6d10c790a4e3dc44c2dc36d96251ae97cdf30a4fa04d4c43e31bfbd038e6a7b7",
+  "blk.12.ffn_gate.weight": "3462a2d8f6b4743b25e24da51b90018ac2858d05ac7e582bcb69063cfdac1104",
+  "blk.12.ffn_norm.weight": "1f96392c1faa34e34ae5dea55a6a86c5aa4c79758952075d53d28de89dd88456",
+  "blk.12.ffn_up.weight": "d22eacc612a7411953d948483c5fb201e11722955ee0754da866e7bec578ac6d",
+  "blk.13.attn_k.weight": "5864977e6b733ea942647d6feed5c76156c48c200649c22e4e11b9e5860e57f3",
+  "blk.13.attn_norm.weight": "87e053535144723db4145aa5402acc54331b7696752d852bb9fc542ff33f0fb5",
+  "blk.13.attn_output.weight": "078145f5ad83f8b14f97a869346f7fd1583b24d1e3edadaa95d3da4242973f8f",
+  "blk.13.attn_q.weight": "3b8caf35504cbc4d1a7dd6e011a95760703b7f71e2218b030b1254f811362dd7",
+  "blk.13.attn_v.weight": "4fdf8365a603e043e5b40c4a21c84ac167f9be62794178f9d8a608dfe5653bf9",
+  "blk.13.ffn_down.weight": "a07d3abbfcacf48ba028df2cab895be32cc15022d23389a745286e79c1b1d1fd",
+  "blk.13.ffn_gate.weight": "1d2ab39666aa2909acc96787432a3ed13b19d25170f74665fadff9b17bbaffb1",
+  "blk.13.ffn_norm.weight": "4f2e809fda5f3eadf52578ee50e0ba36e53be91e55dce418c12dfe595f5f18e7",
+  "blk.13.ffn_up.weight": "8783d2720c2c37ca176a5801e0b3ef1f9cc9cf3ef1cd37af423aaf6b2a27e2bd",
+  "blk.14.attn_k.weight": "ce9428e2b55d43ae0c6690dbd56182f99adc427694ba8236b405cc8ea5035e86",
+  "blk.14.attn_norm.weight": "6abb35f9db8251d6ae954bda147c6ada2371b0574d11702e828f3c6ac99b7cc0",
+  "blk.14.attn_output.weight": "fe3880916d0ceb5bff672c88bbefb7060a545be609bf049beb2024b38221836d",
+  "blk.14.attn_q.weight": "7c8ad81be6f4a350931fd108b5f7c9e366e8c26ef62d1d85ffef5dca8fd893f8",
+  "blk.14.attn_v.weight": "e4bdedffacbebe38567a0734dfd67db90e911d9a9669fcde9a7c4ad8a0066c52",
+  "blk.14.ffn_down.weight": "ef6694dff1e05820aac0cd2b22f39ac7788b4967afc9250775575554c66aab2c",
+  "blk.14.ffn_gate.weight": "db63c4179e2db704bc505e2b4696e055b593e295a1b7c4c586fc793bdd5aab19",
+  "blk.14.ffn_norm.weight": "2796a62d832a9710148f95d533320492a33e712b2e5218659c548705bd11684d",
+  "blk.14.ffn_up.weight": "3f78c78d8c2d54df45f799d4ff902316628af296834afe4ceed63d4a324ff03e",
+  "blk.15.attn_k.weight": "6e810ee3859e07695645ee0c9a5efc7962668984a5f0a9325f47e462743b447c",
+  "blk.15.attn_norm.weight": "0956b576ae96db0b28cb09f761f801cfd9281432284664f0fe181c8d9c55d1ec",
+  "blk.15.attn_output.weight": "03a17f7e94208177aace5cc41b7f54670ba57873b7274ff6e23caf58cce110ca",
+  "blk.15.attn_q.weight": "b8edafe7d2216a6f8b4ae4905a906475490e6ea418f6e1d3cec563dbdc6fab91",
+  "blk.15.attn_v.weight": "f8ae8cae0f4cfa34a459824eba57350c3c248104ba5607e7d9dc7d7c39aaf4a6",
+  "blk.15.ffn_down.weight": "8d02eb439da852246d2ca67e9b7b6de0b090b80744355e64728a23e41926505b",
+  "blk.15.ffn_gate.weight": "ed5bf361c67db8731f186b775826f21c33bdb521111fd2d922539719a770239f",
+  "blk.15.ffn_norm.weight": "5942ca3c73209ac9a0c8bfd9b4aab7f7be7aee9aa12d9c35833493b44af76767",
+  "blk.15.ffn_up.weight": "f4bebf4ad99ec5f911327dec347be6c595814885309c7bc5647ce28c7f4d1cf5",
+  "blk.16.attn_k.weight": "756a534c19364448e0958b8948fe33891c6ccda0fbb4dfa2024e1f532a87804b",
+  "blk.16.attn_norm.weight": "386b7b9e4e6509f6af9c022d942b6c6c6cc136aeed8751ecb037c74d7c4bfb93",
+  "blk.16.attn_output.weight": "3ba1a766a25830b84d7c22178203635f9c5624caad290bc5e5d73da5d5e7a2ec",
+  "blk.16.attn_q.weight": "d39b0c91e1fda7685d50a0f7cc8d18c44b5bdc90a142c7fda0bc329cca1afa74",
+  "blk.16.attn_v.weight": "98b33fcb0ee3483cff1b06ecb44d7b7ffb4d34c268248e4d73dfdf82b2065b2f",
+  "blk.16.ffn_down.weight": "14006f5e4acb2f9416271ae562e299359cd2585739c7fc77ccbca54495563948",
+  "blk.16.ffn_gate.weight": "12f8abae2d301d8f88bedb6af98b1daecc7b0b8d05148594f931f30958d77aca",
+  "blk.16.ffn_norm.weight": "129a15a046ee96d06de288bd43c80f77a6b0fb3a159c7367154c6e4aaf362672",
+  "blk.16.ffn_up.weight": "b4a5911a45f3871ef1d4efb7dc7108645a564b70f818eccf45beebef2e844ee9",
+  "blk.17.attn_k.weight": "5e1bfcff0146ebdde3817b656952892eb671e14e75afc92fa53f84f8eecbec4c",
+  "blk.17.attn_norm.weight": "60bc988fab7c4b29ee9de599df41a8de00caa94fcd74677da011fac82f60f465",
+  "blk.17.attn_output.weight": "ba49b40d6a0b5685f749c24b0edbed3adc44dbe13b5d5e5fa1e56169fc746555",
+  "blk.17.attn_q.weight": "82bb415d24efcd14d03ace03f907bb70db6a204c76a0bdd1892e0fba165db87d",
+  "blk.17.attn_v.weight": "73dbe54beb91a899884e275ea81ffc5187a20cb7d5b68d5c299b783096999d94",
+  "blk.17.ffn_down.weight": "7c086166241e0664f8963fd1ca4ed74c737abfb2525ec20f8435821ff50158f3",
+  "blk.17.ffn_gate.weight": "51a32f78244d42a539f619c5ce661db9e6cf41636280a826d439b5444edcd28c",
+  "blk.17.ffn_norm.weight": "c4bb247fccd1ecc84875028af63dd20aaf5cbd17eb94a9bc36679c09285dccab",
+  "blk.17.ffn_up.weight": "b5886182790bc6fbadd63de9bc4ffee416f3b69a66280d197ab8c18edf769abf",
+  "output_norm.weight": "481f3097d0a20412e35b3a739b1b958487bcd41ff67744baa3c9acbddd2ee4d4"
+}

convert/tokenizer.go

@@ -3,19 +3,150 @@ package convert
 import (
 	"cmp"
 	"crypto/sha256"
+	"encoding/hex"
 	"encoding/json"
+	"errors"
 	"fmt"
+	"io/fs"
 	"log/slog"
 	"os"
 	"slices"
+)
 
-	"golang.org/x/exp/maps"
+const (
+	_ int32 = iota
+	tokenTypeNormal
+	tokenTypeUnknown
+	tokenTypeControl
+	tokenTypeUserDefined
+	tokenTypeUnused
+	tokenTypeByte
 )
 
 type Tokenizer struct {
+	*Vocabulary
+	SpecialVocabulary []*SpecialVocabulary
+	Merges            []string
+
+	Pre      string
+	Template string
+}
+
+func parseTokenizer(fsys fs.FS, specialTokenTypes []string) (*Tokenizer, error) {
+	v, err := parseVocabulary(fsys)
+	if err != nil {
+		return nil, err
+	}
+
+	t := &Tokenizer{
+		Vocabulary: v,
+		Pre:        "default",
+	}
+
+	addedTokens := make(map[string]token)
+	if f, err := fsys.Open("tokenizer.json"); errors.Is(err, os.ErrNotExist) {
+	} else if err != nil {
+		return nil, err
+	} else {
+		defer f.Close()
+
+		var tt tokenizer
+		if err := json.NewDecoder(f).Decode(&tt); err != nil {
+			return nil, err
+		}
+
+		for _, t := range tt.AddedTokens {
+			addedTokens[t.Content] = t
+		}
+
+		t.Merges = tt.Model.Merges
+
+		sha256sum := sha256.New()
+		for _, pt := range tt.PreTokenizer.PreTokenizers {
+			switch pt.Type {
+			case "Split":
+				if pt.Pattern.Regex != "" {
+					// create a checksum of all Split pretokenizers which should be sufficient
+					// to identify the pretokenizer
+					sha256sum.Write([]byte(pt.Pattern.Regex))
+				}
+			}
+		}
+
+		switch digest := hex.EncodeToString(sha256sum.Sum(nil)); digest {
+		case "d98f9631be1e9607a9848c26c1f9eac1aa9fc21ac6ba82a2fc0741af9780a48f":
+			t.Pre = "llama-bpe"
+		case "03df5c5863ad70781dcfdef491ead25140f895fe8010964be0daefe27be32b02":
+			t.Pre = "deepseek-llm"
+		case "21cde974d587f0d54dc8d56b183cc1e6239600172035c68fbd6d4b9f8da0576e":
+			t.Pre = "deepseek-coder"
+		case "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855":
+			// noop, empty pretokenizer
+		default:
+			slog.Warn("unknown pretokenizer, using default", "digest", digest)
+		}
+	}
+
+	if f, err := fsys.Open("tokenizer_config.json"); errors.Is(err, os.ErrNotExist) {
+	} else if err != nil {
+		return nil, err
+	} else {
+		defer f.Close()
+
+		var p map[string]json.RawMessage
+		if err := json.NewDecoder(f).Decode(&p); err != nil {
+			return nil, err
+		}
+
+		if template, ok := p["chat_template"]; ok {
+			if err := json.Unmarshal(template, &t.Template); err != nil {
+				return nil, err
+			}
+		}
+
+		for _, st := range specialTokenTypes {
+			sv := SpecialVocabulary{Type: st}
+			if bts, ok := p[fmt.Sprintf("add_%s_token", st)]; ok {
+				if err := json.Unmarshal(bts, &sv.AddToken); err != nil {
+					return nil, err
+				}
+			}
+
+			if bts, ok := p[fmt.Sprintf("%s_token", st)]; ok {
+				var content string
+				if err := json.Unmarshal(bts, &content); err != nil {
+					var mm map[string]any
+					if err := json.Unmarshal(bts, &mm); err != nil {
+						continue
+					}
+
+					content, ok = mm["content"].(string)
+					if !ok {
+						continue
+					}
+				}
+
+				sv.Content = content
+			}
+
+			if id, ok := addedTokens[sv.Content]; ok {
+				sv.ID = id.ID
+				t.SpecialVocabulary = append(t.SpecialVocabulary, &sv)
+			}
+		}
+	}
+
+	return t, nil
+}
+
+type tokenizer struct {
 	Version     string  `json:"version"`
-	AddedTokens []Token        `json:"added_tokens"`
-	Model       TokenizerModel `json:"model"`
+	AddedTokens []token `json:"added_tokens"`
+	Model       struct {
+		Type   string         `json:"type"`
+		Vocab  map[string]int `json:"vocab"`
+		Merges []string       `json:"merges"`
+	} `json:"model"`
 
 	PreTokenizer struct {
 		PreTokenizers []struct {
@@ -27,83 +158,108 @@ type Tokenizer struct {
 	} `json:"pre_tokenizer"`
 }
 
-type TokenizerModel struct {
-	Type   string         `json:"type"`
-	Vocab  map[string]int `json:"vocab"`
-	Merges []string       `json:"merges"`
-	Tokens []Token
-}
-
-type Token struct {
+type token struct {
 	ID          int    `json:"id"`
 	Content     string `json:"content"`
 	Special     bool   `json:"special"`
 	UserDefined bool
 }
 
-func (t *Token) Type() int32 {
-	switch {
-	case t.Special:
-		return tokenTypeControl
-	case t.UserDefined:
-		return tokenTypeUserDefined
-	default:
-		return tokenTypeNormal
-	}
-}
-
-func (t *Tokenizer) maxID() int {
-	return max(
-		slices.Max(maps.Values(t.Model.Vocab)),
-		slices.MaxFunc(t.AddedTokens, func(a, b Token) int {
-			return cmp.Compare(a.ID, b.ID)
-		}).ID,
-	)
+type Vocabulary struct {
+	Model  string
+	Tokens []string
+	Scores []float32
+	Types  []int32
 }
 
-func parseTokens(dirpath string) (pre string, tokens []Token, merges []string, err error) {
-	f, err := os.Open(dirpath)
+func parseVocabularyFromTokenizer(fsys fs.FS) (*Vocabulary, error) {
+	f, err := fsys.Open("tokenizer.json")
 	if err != nil {
-		panic(err)
+		return nil, err
 	}
 	defer f.Close()
 
-	var t Tokenizer
+	var t tokenizer
 	if err := json.NewDecoder(f).Decode(&t); err != nil {
-		return "", nil, nil, err
+		return nil, err
 	}
 
-	tokens = make([]Token, t.maxID()+1)
+	var tokens []token
 	for k, v := range t.Model.Vocab {
-		tokens[v] = Token{ID: v, Content: k, Special: false, UserDefined: false}
+		tokens = append(tokens, token{
+			ID:      v,
+			Content: k,
+		})
 	}
 
-	for _, v := range t.AddedTokens {
-		v.UserDefined = true
-		tokens[v.ID] = v
+	for _, t := range t.AddedTokens {
+		t.UserDefined = true
+		tokens = append(tokens, t)
 	}
 
-	sha256sum := sha256.New()
-	for _, pt := range t.PreTokenizer.PreTokenizers {
-		switch pt.Type {
-		case "Split":
-			if pt.Pattern.Regex != "" {
-				sha256sum.Write([]byte(pt.Pattern.Regex))
+	slices.SortFunc(tokens, func(i, j token) int {
+		return cmp.Compare(i.ID, j.ID)
+	})
+
+	v := Vocabulary{Model: "gpt2"}
+	for _, t := range tokens {
+		v.Tokens = append(v.Tokens, t.Content)
+		v.Scores = append(v.Scores, float32(t.ID))
+
+		switch {
+		case t.Special:
+			v.Types = append(v.Types, tokenTypeControl)
+		case t.UserDefined:
+			v.Types = append(v.Types, tokenTypeUserDefined)
+		default:
+			v.Types = append(v.Types, tokenTypeNormal)
 		}
 	}
+
+	return &v, nil
+}
+
+func parseVocabulary(fsys fs.FS) (*Vocabulary, error) {
+	patterns := []struct {
+		Pattern string
+		Func    func(fs.FS) (*Vocabulary, error)
+	}{
+		{"tokenizer.model", parseSentencePiece},
+		{"tokenizer.json", parseVocabularyFromTokenizer},
 	}
 
-	switch digest := fmt.Sprintf("%x", sha256sum.Sum(nil)); digest {
-	case "d98f9631be1e9607a9848c26c1f9eac1aa9fc21ac6ba82a2fc0741af9780a48f":
-		pre = "llama-bpe"
-	case "03df5c5863ad70781dcfdef491ead25140f895fe8010964be0daefe27be32b02":
-		pre = "deepseek-llm"
-	case "21cde974d587f0d54dc8d56b183cc1e6239600172035c68fbd6d4b9f8da0576e":
-		pre = "deepseek-coder"
-	default:
-		slog.Warn("unknown pretokenizer, using default", "digest", digest)
-		pre = "default"
+	for _, pattern := range patterns {
+		if _, err := fs.Stat(fsys, pattern.Pattern); errors.Is(err, os.ErrNotExist) {
+			continue
+		} else if err != nil {
+			return nil, err
+		}
+
+		return pattern.Func(fsys)
+	}
+
+	return nil, errors.New("unknown tensor format")
+}
+
+type SpecialVocabulary struct {
+	Type     string
+	ID       int
+	Content  string
+	AddToken bool
+}
+
+func (sv SpecialVocabulary) Key() string {
+	switch t := sv.Type; t {
+	case "bos", "eos", "cls", "mask":
+		return t
+	case "unk":
+		return "unknown"
+	case "sep":
+		//nolint:misspell // this is an upstream typo
+		return "seperator"
+	case "pad":
+		return "padding"
 	}
 
-	return pre, tokens, t.Model.Merges, nil
+	panic("unknown special vocabulary type")
 }

convert/tokenizer_spm.go

+package convert
+
+import (
+	"cmp"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"io/fs"
+	"os"
+	"slices"
+
+	"google.golang.org/protobuf/proto"
+
+	"github.com/ollama/ollama/convert/sentencepiece"
+)
+
+func parseSentencePiece(fsys fs.FS) (*Vocabulary, error) {
+	bts, err := fs.ReadFile(fsys, "tokenizer.model")
+	if err != nil {
+		return nil, err
+	}
+
+	var spm sentencepiece.ModelProto
+	if err := proto.Unmarshal(bts, &spm); err != nil {
+		return nil, err
+	}
+
+	v := Vocabulary{Model: "llama"}
+	for _, piece := range spm.GetPieces() {
+		v.Tokens = append(v.Tokens, piece.GetPiece())
+		v.Scores = append(v.Scores, piece.GetScore())
+
+		switch t := piece.GetType(); t {
+		case sentencepiece.ModelProto_SentencePiece_UNKNOWN,
+			sentencepiece.ModelProto_SentencePiece_CONTROL,
+			sentencepiece.ModelProto_SentencePiece_UNUSED,
+			sentencepiece.ModelProto_SentencePiece_BYTE:
+			v.Types = append(v.Types, int32(t))
+		default:
+			v.Types = append(v.Types, int32(sentencepiece.ModelProto_SentencePiece_NORMAL))
+		}
+	}
+
+	f, err := fsys.Open("added_tokens.json")
+	if errors.Is(err, os.ErrNotExist) {
+		return &v, nil
+	} else if err != nil {
+		return nil, err
+	}
+	defer f.Close()
+
+	var atm map[string]int
+	if err := json.NewDecoder(f).Decode(&atm); err != nil {
+		return nil, err
+	}
+
+	type t struct {
+		id      int
+		content string
+	}
+
+	var ts []t
+	for content, id := range atm {
+		ts = append(ts, t{id, content})
+	}
+
+	slices.SortFunc(ts, func(i, j t) int {
+		return cmp.Compare(i.id, j.id)
+	})
+
+	n := len(v.Tokens)
+	for i, t := range ts {
+		if t.id != i+n {
+			return nil, fmt.Errorf("invalid token id: %d", t.id)
+		}
+
+		v.Tokens = append(v.Tokens, t.content)
+		v.Scores = append(v.Scores, -1000.0)
+		v.Types = append(v.Types, tokenTypeUserDefined)
+	}
+
+	return &v, nil
+}

docs/development.md

@@ -104,7 +104,7 @@ like to use. For example, to compile an optimized binary for an Intel i9-9880H,
 you might use:
 
 ```
-OLLAMA_CUSTOM_CPU_DEFS="-DLLAMA_AVX=on -DLLAMA_AVX2=on -DLLAMA_F16C=on -DLLAMA_FMA=on" go generate ./...
+OLLAMA_CUSTOM_CPU_DEFS="-DGGML_AVX=on -DGGML_AVX2=on -DGGML_F16C=on -DGGML_FMA=on" go generate ./...
 go build .
 ```
 
@@ -114,15 +114,18 @@ If you have Docker available, you can build linux binaries with `./scripts/build
 
 ### Windows
 
-Note: The windows build for Ollama is still under development.
+Note: The Windows build for Ollama is still under development.
 
-Install required tools:
+First, install required tools:
 
 - MSVC toolchain - C/C++ and cmake as minimal requirements
 - Go version 1.22 or higher
 - MinGW (pick one variant) with GCC.
   - [MinGW-w64](https://www.mingw-w64.org/)
   - [MSYS2](https://www.msys2.org/)
+- The `ThreadJob` Powershell module: `Install-Module -Name ThreadJob -Scope CurrentUser`
+
+Then, build the `ollama` binary:
 
 ```powershell
 $env:CGO_ENABLED="1"

docs/docker.md

@@ -63,7 +63,7 @@ docker run -d --device /dev/kfd --device /dev/dri -v ollama:/root/.ollama -p 114
 Now you can run a model:
 
 ```
-docker exec -it ollama ollama run llama3
+docker exec -it ollama ollama run llama3.1
 ```
 
 ### Try different models

docs/faq.md

@@ -227,7 +227,7 @@ curl http://localhost:11434/api/chat -d '{"model": "mistral"}'
 
 To preload a model using the CLI, use the command:
 ```shell
-ollama run llama3 ""
+ollama run llama3.1 ""
 ```
 
 ## How do I keep a model loaded in memory or make it unload immediately?
@@ -257,3 +257,23 @@ If you wish to override the `OLLAMA_KEEP_ALIVE` setting, use the `keep_alive` AP
 ## How do I manage the maximum number of requests the Ollama server can queue?
 
 If too many requests are sent to the server, it will respond with a 503 error indicating the server is overloaded.  You can adjust how many requests may be queue by setting `OLLAMA_MAX_QUEUE`.
+
+## How does Ollama handle concurrent requests?
+
+Ollama supports two levels of concurrent processing.  If your system has sufficient available memory (system memory when using CPU inference, or VRAM for GPU inference) then multiple models can be loaded at the same time.  For a given model, if there is sufficient available memory when the model is loaded, it is configured to allow parallel request processing.
+
+If there is insufficient available memory to load a new model request while one or more models are already loaded, all new requests will be queued until the new model can be loaded.  As prior models become idle, one or more will be unloaded to make room for the new model.  Queued requests will be processed in order.  When using GPU inference new models must be able to completely fit in VRAM to allow concurrent model loads.
+
+Parallel request processing for a given model results in increasing the context size by the number of parallel requests.  For example, a 2K context with 4 parallel requests will result in an 8K context and additional memory allocation.
+
+The following server settings may be used to adjust how Ollama handles concurrent requests on most platforms:
+
+- `OLLAMA_MAX_LOADED_MODELS` - The maximum number of models that can be loaded concurrently provided they fit in available memory.  The default is 3 * the number of GPUs or 3 for CPU inference.
+- `OLLAMA_NUM_PARALLEL` - The maximum number of parallel requests each model will process at the same time.  The default will auto-select either 4 or 1 based on available memory.
+- `OLLAMA_MAX_QUEUE` - The maximum number of requests Ollama will queue when busy before rejecting additional requests. The default is 512
+
+Note: Windows with Radeon GPUs currently default to 1 model maximum due to limitations in ROCm v5.7 for available VRAM reporting.  Once ROCm v6.2 is available, Windows Radeon will follow the defaults above.  You may enable concurrent model loads on Radeon on Windows, but ensure you don't load more models than will fit into your GPUs VRAM.
+
+## How does Ollama load models on multiple GPUs?
+
+Installing multiple GPUs of the same brand can be a great way to increase your available VRAM to load larger models.  When you load a new model, Ollama evaluates the required VRAM for the model against what is currently available.  If the model will entirely fit on any single GPU, Ollama will load the model on that GPU.  This typically provides the best performance as it reduces the amount of data transfering across the PCI bus during inference.  If the model does not fit entirely on one GPU, then it will be spread across all the available GPUs.
\ No newline at end of file

docs/gpu.md

@@ -8,7 +8,7 @@ Check your compute compatibility to see if your card is supported:
 | Compute Capability | Family              | Cards                                                                                                       |
 | ------------------ | ------------------- | ----------------------------------------------------------------------------------------------------------- |
 | 9.0                | NVIDIA              | `H100`                                                                                                      |
-| 8.9                | GeForce RTX 40xx    | `RTX 4090` `RTX 4080` `RTX 4070 Ti` `RTX 4060 Ti`                                                           |
+| 8.9                | GeForce RTX 40xx    | `RTX 4090` `RTX 4080 SUPER` `RTX 4080` `RTX 4070 Ti SUPER` `RTX 4070 Ti` `RTX 4070 SUPER` `RTX 4070` `RTX 4060 Ti` `RTX 4060`  |
 |                    | NVIDIA Professional | `L4` `L40` `RTX 6000`                                                                                       |
 | 8.6                | GeForce RTX 30xx    | `RTX 3090 Ti` `RTX 3090` `RTX 3080 Ti` `RTX 3080` `RTX 3070 Ti` `RTX 3070` `RTX 3060 Ti` `RTX 3060`         |
 |                    | NVIDIA Professional | `A40` `RTX A6000` `RTX A5000` `RTX A4000` `RTX A3000` `RTX A2000` `A10` `A16` `A2`                          |
@@ -18,7 +18,7 @@ Check your compute compatibility to see if your card is supported:
 |                    | Quadro              | `RTX 8000` `RTX 6000` `RTX 5000` `RTX 4000`                                                                 |
 | 7.0                | NVIDIA              | `TITAN V` `V100` `Quadro GV100`                                                                             |
 | 6.1                | NVIDIA TITAN        | `TITAN Xp` `TITAN X`                                                                                        |
-|                    | GeForce GTX         | `GTX 1080 Ti` `GTX 1080` `GTX 1070 Ti` `GTX 1070` `GTX 1060` `GTX 1050`                                     |
+|                    | GeForce GTX         | `GTX 1080 Ti` `GTX 1080` `GTX 1070 Ti` `GTX 1070` `GTX 1060` `GTX 1050 Ti` `GTX 1050`                       |
 |                    | Quadro              | `P6000` `P5200` `P4200` `P3200` `P5000` `P4000` `P3000` `P2200` `P2000` `P1000` `P620` `P600` `P500` `P520` |
 |                    | Tesla               | `P40` `P4`                                                                                                  |
 | 6.0                | NVIDIA              | `Tesla P100` `Quadro GP100`                                                                                 |
@@ -46,13 +46,24 @@ sudo modprobe nvidia_uvm`
 
 ## AMD Radeon
 Ollama supports the following AMD GPUs:
+
+### Linux Support
 | Family         | Cards and accelerators                                                                                                               |
 | -------------- | ---------------------------------------------------------------------------------------------------------------------------------------------- |
 | AMD Radeon RX  | `7900 XTX` `7900 XT` `7900 GRE` `7800 XT` `7700 XT` `7600 XT` `7600` `6950 XT` `6900 XTX` `6900XT` `6800 XT` `6800` `Vega 64` `Vega 56`    |
 | AMD Radeon PRO | `W7900` `W7800` `W7700` `W7600` `W7500` `W6900X` `W6800X Duo` `W6800X` `W6800` `V620` `V420` `V340` `V320` `Vega II Duo` `Vega II` `VII` `SSG` |
 | AMD Instinct   | `MI300X` `MI300A` `MI300` `MI250X` `MI250` `MI210` `MI200` `MI100` `MI60` `MI50`                                                               |
 
-### Overrides
+### Windows Support
+With ROCm v6.1, the following GPUs are supported on Windows.
+
+| Family         | Cards and accelerators                                                                                                               |
+| -------------- | ---------------------------------------------------------------------------------------------------------------------------------------------- |
+| AMD Radeon RX  | `7900 XTX` `7900 XT` `7900 GRE` `7800 XT` `7700 XT` `7600 XT` `7600` `6950 XT` `6900 XTX` `6900XT` `6800 XT` `6800`    |
+| AMD Radeon PRO | `W7900` `W7800` `W7700` `W7600` `W7500` `W6900X` `W6800X Duo` `W6800X` `W6800` `V620` |
+
+
+### Overrides on Linux
 Ollama leverages the AMD ROCm library, which does not support all AMD GPUs. In
 some cases you can force the system to try to use a similar LLVM target that is
 close.  For example The Radeon RX 5400 is `gfx1034` (also known as 10.3.4)
@@ -63,7 +74,7 @@ would set `HSA_OVERRIDE_GFX_VERSION="10.3.0"` as an environment variable for the
 server.  If you have an unsupported AMD GPU you can experiment using the list of
 supported types below.
 
-At this time, the known supported GPU types are the following LLVM Targets.
+At this time, the known supported GPU types on linux are the following LLVM Targets.
 This table shows some example GPUs that map to these LLVM targets:
 | **LLVM Target** | **An Example GPU** |
 |-----------------|---------------------|

docs/import.md

-# Import a model
+# Import
 
-This guide walks through importing a GGUF, PyTorch or Safetensors model.
+GGUF models and select Safetensors models can be imported directly into Ollama.
 
-## Importing (GGUF)
+## Import GGUF
 
-### Step 1: Write a `Modelfile`
+A binary GGUF file can be imported directly into Ollama through a Modelfile.
 
-Start by creating a `Modelfile`. This file is the blueprint for your model, specifying weights, parameters, prompt templates and more.
-
-```
-FROM ./mistral-7b-v0.1.Q4_0.gguf
+```dockerfile
+FROM /path/to/file.gguf
 ```
 
-(Optional) many chat models require a prompt template in order to answer correctly. A default prompt template can be specified with the `TEMPLATE` instruction in the `Modelfile`:
-
-```
-FROM ./mistral-7b-v0.1.Q4_0.gguf
-TEMPLATE "[INST] {{ .Prompt }} [/INST]"
-```
-
-### Step 2: Create the Ollama model
-
-Finally, create a model from your `Modelfile`:
+## Import Safetensors
 
-```
-ollama create example -f Modelfile
-```
+If the model being imported is one of these architectures, it can be imported directly into Ollama through a Modelfile:
 
-### Step 3: Run your model
+ - LlamaForCausalLM
+ - MistralForCausalLM
+ - GemmaForCausalLM
 
-Next, test the model with `ollama run`:
-
-```
-ollama run example "What is your favourite condiment?"
+```dockerfile
+FROM /path/to/safetensors/directory
 ```
 
-## Importing (PyTorch & Safetensors)
+For architectures not directly convertable by Ollama, see llama.cpp's [guide](https://github.com/ggerganov/llama.cpp/blob/master/README.md#prepare-and-quantize) on conversion. After conversion, see [Import GGUF](#import-gguf).
 
-> Importing from PyTorch and Safetensors is a longer process than importing from GGUF. Improvements that make it easier are a work in progress.
+## Automatic Quantization
 
-### Setup
+> [!NOTE]
+> Automatic quantization requires v0.1.35 or higher.
 
-First, clone the `ollama/ollama` repo:
+Ollama is capable of quantizing FP16 or FP32 models to any of the supported quantizations with the `-q/--quantize` flag in `ollama create`.
 
+```dockerfile
+FROM /path/to/my/gemma/f16/model
 ```
-git clone git@github.com:ollama/ollama.git ollama
-cd ollama
-```
-
-and then fetch its `llama.cpp` submodule:
 
 ```shell
-git submodule init
-git submodule update llm/llama.cpp
+$ ollama create -q Q4_K_M mymodel
+transferring model data
+quantizing F16 model to Q4_K_M
+creating new layer sha256:735e246cc1abfd06e9cdcf95504d6789a6cd1ad7577108a70d9902fef503c1bd
+creating new layer sha256:0853f0ad24e5865173bbf9ffcc7b0f5d56b66fd690ab1009867e45e7d2c4db0f
+writing manifest
+success
 ```
 
-Next, install the Python dependencies:
+### Supported Quantizations
 
-```
-python3 -m venv llm/llama.cpp/.venv
-source llm/llama.cpp/.venv/bin/activate
-pip install -r llm/llama.cpp/requirements.txt
-```
-
-Then build the `quantize` tool:
+- `Q4_0`
+- `Q4_1`
+- `Q5_0`
+- `Q5_1`
+- `Q8_0`
 
-```
-make -C llm/llama.cpp quantize
-```
+#### K-means Quantizations
 
-### Clone the HuggingFace repository (optional)
+- `Q3_K_S`
+- `Q3_K_M`
+- `Q3_K_L`
+- `Q4_K_S`
+- `Q4_K_M`
+- `Q5_K_S`
+- `Q5_K_M`
+- `Q6_K`
 
-If the model is currently hosted in a HuggingFace repository, first clone that repository to download the raw model.
+## Template Detection
 
-Install [Git LFS](https://docs.github.com/en/repositories/working-with-files/managing-large-files/installing-git-large-file-storage), verify it's installed, and then clone the model's repository:
-
-```
-git lfs install
-git clone https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.1 model
-```
+> [!NOTE]
+> Template detection requires v0.1.42 or higher.
 
-### Convert the model
+Ollama uses model metadata, specifically `tokenizer.chat_template`, to automatically create a template appropriate for the model you're importing.
 
-> Note: some model architectures require using specific convert scripts. For example, Qwen models require running `convert-hf-to-gguf.py` instead of `convert.py`
-
-```
-python llm/llama.cpp/convert.py ./model --outtype f16 --outfile converted.bin
-```
-
-### Quantize the model
-
-```
-llm/llama.cpp/quantize converted.bin quantized.bin q4_0
+```dockerfile
+FROM /path/to/my/gemma/model
 ```
 
-### Step 3: Write a `Modelfile`
-
-Next, create a `Modelfile` for your model:
-
-```
-FROM quantized.bin
-TEMPLATE "[INST] {{ .Prompt }} [/INST]"
-```
-
-### Step 4: Create the Ollama model
-
-Finally, create a model from your `Modelfile`:
-
-```
-ollama create example -f Modelfile
-```
-
-### Step 5: Run your model
-
-Next, test the model with `ollama run`:
-
-```
-ollama run example "What is your favourite condiment?"
-```
-
-## Publishing your model (optional – early alpha)
-
-Publishing models is in early alpha. If you'd like to publish your model to share with others, follow these steps:
-
-1. Create [an account](https://ollama.com/signup)
-2. Copy your Ollama public key:
-  - macOS: `cat ~/.ollama/id_ed25519.pub | pbcopy`
-  - Windows: `type %USERPROFILE%\.ollama\id_ed25519.pub`
-  - Linux: `cat /usr/share/ollama/.ollama/id_ed25519.pub`
-3. Add your public key to your [Ollama account](https://ollama.com/settings/keys)
-
-Next, copy your model to your username's namespace:
-
-```
-ollama cp example <your username>/example
-```
-
-> Note: model names may only contain lowercase letters, digits, and the characters `.`, `-`, and `_`.
-
-Then push the model:
-
-```
-ollama push <your username>/example
+```shell
+$ ollama create mymodel
+transferring model data
+using autodetected template gemma-instruct
+creating new layer sha256:baa2a0edc27d19cc6b7537578a9a7ba1a4e3214dc185ed5ae43692b319af7b84
+creating new layer sha256:ba66c3309914dbef07e5149a648fd1877f030d337a4f240d444ea335008943cb
+writing manifest
+success
 ```
 
-After publishing, your model will be available at `https://ollama.com/<your username>/example`.
-
-## Quantization reference
-
-The quantization options are as follow (from highest highest to lowest levels of quantization). Note: some architectures such as Falcon do not support K quants.
-
-- `q2_K`
-- `q3_K`
-- `q3_K_S`
-- `q3_K_M`
-- `q3_K_L`
-- `q4_0` (recommended)
-- `q4_1`
-- `q4_K`
-- `q4_K_S`
-- `q4_K_M`
-- `q5_0`
-- `q5_1`
-- `q5_K`
-- `q5_K_S`
-- `q5_K_M`
-- `q6_K`
-- `q8_0`
-- `f16`
+Defining a template in the Modelfile will disable this feature which may be useful if you want to use a different template than the autodetected one.