llm: speed up gguf decoding by a lot (#5246)

Previously, some costly things were causing the loading of GGUF files
and their metadata and tensor information to be VERY slow:

  * Too many allocations when decoding strings
  * Hitting disk for each read of each key and value, resulting in a
    not-okay amount of syscalls/disk I/O.

The show API is now down to 33ms from 800ms+ for llama3 on a macbook pro
m3.

This commit also prevents collecting large arrays of values when
decoding GGUFs (if desired). When such keys are encountered, their
values are null, and are encoded as such in JSON.

Also, this fixes a broken test that was not encoding valid GGUF.

README.md

@@ -182,6 +182,12 @@ $ ollama run llama3 "Summarize this file: $(cat README.md)"
  Ollama is a lightweight, extensible framework for building and running language models on the local machine. It provides a simple API for creating, running, and managing models, as well as a library of pre-built models that can be easily used in a variety of applications.
 ```
 
+### Show model information
+
+```
+ollama show llama3
+```
+
 ### List models on your computer
 
 ```

api/types.go

@@ -608,6 +608,19 @@ func FormatParams(params map[string][]string) (map[string]interface{}, error) {
 		} else {
 			field := valueOpts.FieldByName(opt.Name)
 			if field.IsValid() && field.CanSet() {
+				if reflect.PointerTo(field.Type()) == reflect.TypeOf((*TriState)(nil)) {
+					boolVal, err := strconv.ParseBool(vals[0])
+					if err != nil {
+						return nil, fmt.Errorf("invalid bool value %s", vals)
+					}
+					if boolVal {
+						out[key] = TriStateTrue
+					} else {
+						out[key] = TriStateFalse
+					}
+					continue
+				}
+
 				switch field.Kind() {
 				case reflect.Float32:
 					floatVal, err := strconv.ParseFloat(vals[0], 32)

api/types_test.go

@@ -2,6 +2,7 @@ package api
 
 import (
 	"encoding/json"
+	"fmt"
 	"math"
 	"testing"
 	"time"
@@ -141,3 +142,65 @@ func TestUseMmapParsingFromJSON(t *testing.T) {
 		})
 	}
 }
+
+func TestUseMmapFormatParams(t *testing.T) {
+	tests := []struct {
+		name string
+		req  map[string][]string
+		exp  TriState
+		err  error
+	}{
+		{
+			name: "True",
+			req: map[string][]string{
+				"use_mmap": []string{"true"},
+			},
+			exp: TriStateTrue,
+			err: nil,
+		},
+		{
+			name: "False",
+			req: map[string][]string{
+				"use_mmap": []string{"false"},
+			},
+			exp: TriStateFalse,
+			err: nil,
+		},
+		{
+			name: "Numeric True",
+			req: map[string][]string{
+				"use_mmap": []string{"1"},
+			},
+			exp: TriStateTrue,
+			err: nil,
+		},
+		{
+			name: "Numeric False",
+			req: map[string][]string{
+				"use_mmap": []string{"0"},
+			},
+			exp: TriStateFalse,
+			err: nil,
+		},
+		{
+			name: "invalid string",
+			req: map[string][]string{
+				"use_mmap": []string{"foo"},
+			},
+			exp: TriStateUndefined,
+			err: fmt.Errorf("invalid bool value [foo]"),
+		},
+	}
+
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			resp, err := FormatParams(test.req)
+			require.Equal(t, err, test.err)
+			respVal, ok := resp["use_mmap"]
+			if test.exp != TriStateUndefined {
+				assert.True(t, ok, "resp: %v", resp)
+				assert.Equal(t, test.exp, respVal)
+			}
+		})
+	}
+}

cmd/cmd.go

@@ -287,38 +287,12 @@ func createBlob(cmd *cobra.Command, client *api.Client, path string) (string, er
 }
 
 func RunHandler(cmd *cobra.Command, args []string) error {
-	client, err := api.ClientFromEnvironment()
-	if err != nil {
-		return err
-	}
-
-	name := args[0]
-
-	// check if the model exists on the server
-	show, err := client.Show(cmd.Context(), &api.ShowRequest{Name: name})
-	var statusError api.StatusError
-	switch {
-	case errors.As(err, &statusError) && statusError.StatusCode == http.StatusNotFound:
-		if err := PullHandler(cmd, []string{name}); err != nil {
-			return err
-		}
-
-		show, err = client.Show(cmd.Context(), &api.ShowRequest{Name: name})
-		if err != nil {
-			return err
-		}
-	case err != nil:
-		return err
-	}
-
 	interactive := true
 
 	opts := runOptions{
-		Model:       args[0],
-		WordWrap:    os.Getenv("TERM") == "xterm-256color",
-		Options:     map[string]interface{}{},
-		MultiModal:  slices.Contains(show.Details.Families, "clip"),
-		ParentModel: show.Details.ParentModel,
+		Model:    args[0],
+		WordWrap: os.Getenv("TERM") == "xterm-256color",
+		Options:  map[string]interface{}{},
 	}
 
 	format, err := cmd.Flags().GetString("format")
@@ -362,11 +336,38 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 	}
 	opts.WordWrap = !nowrap
 
-	if !interactive {
-		return generate(cmd, opts)
+	// Fill out the rest of the options based on information about the
+	// model.
+	client, err := api.ClientFromEnvironment()
+	if err != nil {
+		return err
 	}
 
-	return generateInteractive(cmd, opts)
+	name := args[0]
+	info, err := func() (*api.ShowResponse, error) {
+		showReq := &api.ShowRequest{Name: name}
+		info, err := client.Show(cmd.Context(), showReq)
+		var se api.StatusError
+		if errors.As(err, &se) && se.StatusCode == http.StatusNotFound {
+			if err := PullHandler(cmd, []string{name}); err != nil {
+				return nil, err
+			}
+			return client.Show(cmd.Context(), &api.ShowRequest{Name: name})
+		}
+		return info, err
+	}()
+	if err != nil {
+		return err
+	}
+
+	opts.MultiModal = slices.Contains(info.Details.Families, "clip")
+	opts.ParentModel = info.Details.ParentModel
+	opts.Messages = append(opts.Messages, info.Messages...)
+
+	if interactive {
+		return generateInteractive(cmd, opts)
+	}
+	return generate(cmd, opts)
 }
 
 func errFromUnknownKey(unknownKeyErr error) error {

cmd/interactive.go

@@ -31,65 +31,40 @@ const (
 )
 
 func loadModel(cmd *cobra.Command, opts *runOptions) error {
-	client, err := api.ClientFromEnvironment()
-	if err != nil {
-		return err
-	}
-
 	p := progress.NewProgress(os.Stderr)
 	defer p.StopAndClear()
 
 	spinner := progress.NewSpinner("")
 	p.Add("", spinner)
 
-	showReq := api.ShowRequest{Name: opts.Model}
-	showResp, err := client.Show(cmd.Context(), &showReq)
+	client, err := api.ClientFromEnvironment()
 	if err != nil {
 		return err
 	}
-	opts.MultiModal = slices.Contains(showResp.Details.Families, "clip")
-	opts.ParentModel = showResp.Details.ParentModel
-
-	if len(showResp.Messages) > 0 {
-		opts.Messages = append(opts.Messages, showResp.Messages...)
-	}
 
 	chatReq := &api.ChatRequest{
-		Model:    opts.Model,
-		Messages: []api.Message{},
+		Model:     opts.Model,
+		KeepAlive: opts.KeepAlive,
 	}
 
-	if opts.KeepAlive != nil {
-		chatReq.KeepAlive = opts.KeepAlive
-	}
-
-	err = client.Chat(cmd.Context(), chatReq, func(resp api.ChatResponse) error {
+	return client.Chat(cmd.Context(), chatReq, func(resp api.ChatResponse) error {
 		p.StopAndClear()
-		if len(opts.Messages) > 0 {
-			for _, msg := range opts.Messages {
-				switch msg.Role {
-				case "user":
-					fmt.Printf(">>> %s\n", msg.Content)
-				case "assistant":
-					state := &displayResponseState{}
-					displayResponse(msg.Content, opts.WordWrap, state)
-					fmt.Println()
-					fmt.Println()
-				}
+		for _, msg := range opts.Messages {
+			switch msg.Role {
+			case "user":
+				fmt.Printf(">>> %s\n", msg.Content)
+			case "assistant":
+				state := &displayResponseState{}
+				displayResponse(msg.Content, opts.WordWrap, state)
+				fmt.Println()
+				fmt.Println()
 			}
 		}
 		return nil
 	})
-	if err != nil {
-		return err
-	}
-
-	return nil
 }
 
 func generateInteractive(cmd *cobra.Command, opts runOptions) error {
-	opts.Messages = make([]api.Message, 0)
-
 	err := loadModel(cmd, &opts)
 	if err != nil {
 		return err

gpu/assets.go

@@ -77,20 +77,27 @@ func cleanupTmpDirs() {
 			continue
 		}
 		raw, err := os.ReadFile(filepath.Join(d, "ollama.pid"))
-		if err == nil {
-			pid, err := strconv.Atoi(string(raw))
-			if err == nil {
-				if proc, err := os.FindProcess(pid); err == nil && !errors.Is(proc.Signal(syscall.Signal(0)), os.ErrProcessDone) {
-					// Another running ollama, ignore this tmpdir
-					continue
-				}
-			}
-		} else {
-			slog.Debug("failed to open ollama.pid", "path", d, "error", err)
-		}
-		err = os.RemoveAll(d)
 		if err != nil {
-			slog.Debug("unable to cleanup stale tmpdir", "path", d, "error", err)
+			slog.Warn("failed to read ollama.pid", "path", d, "error", err)
+			// No pid, ignore this tmpdir
+			continue
+		}
+
+		pid, err := strconv.Atoi(string(raw))
+		if err != nil {
+			slog.Warn("failed to parse pid", "path", d, "error", err)
+			continue
+		}
+
+		proc, err := os.FindProcess(pid)
+		if err == nil && !errors.Is(proc.Signal(syscall.Signal(0)), os.ErrProcessDone) {
+			slog.Warn("found running ollama", "pid", pid, "path", d)
+			// Another running ollama, ignore this tmpdir
+			continue
+		}
+
+		if err := os.Remove(d); err != nil {
+			slog.Warn("unable to cleanup stale tmpdir", "path", d, "error", err)
 		}
 	}
 }

llm/ggla.go

@@ -53,7 +53,7 @@ func (llm *ggla) Tensors() Tensors {
 	return llm.tensors
 }
 
-func (llm *ggla) decode(rs io.ReadSeeker) error {
+func (llm *ggla) decode(rs io.ReadSeeker) (retErr error) {
 	var r uint32
 	if err := binary.Read(rs, binary.LittleEndian, &r); err != nil {
 		return err
@@ -69,9 +69,18 @@ func (llm *ggla) decode(rs io.ReadSeeker) error {
 	for {
 		var dims uint32
 		if err := binary.Read(rs, binary.LittleEndian, &dims); err != nil {
+			if errors.Is(err, io.EOF) {
+				return nil
+			}
 			return err
 		}
 
+		defer func() {
+			if errors.Is(retErr, io.EOF) {
+				retErr = io.ErrUnexpectedEOF
+			}
+		}()
+
 		var namesize uint32
 		if err := binary.Read(rs, binary.LittleEndian, &namesize); err != nil {
 			return err
@@ -108,7 +117,7 @@ func (llm *ggla) decode(rs io.ReadSeeker) error {
 			return err
 		}
 
-		if _, err := rs.Seek((offset+31)&-32, io.SeekStart); err != nil {
+		if _, err := rs.Seek((offset+31)&-32-offset, io.SeekCurrent); err != nil {
 			return err
 		}
 

llm/ggml.go

@@ -6,6 +6,8 @@ import (
 	"fmt"
 	"io"
 	"strings"
+
+	"github.com/ollama/ollama/util/bufioutil"
 )
 
 type GGML struct {
@@ -69,6 +71,30 @@ func (kv KV) HeadCountKV() uint64 {
 	return 1
 }
 
+func (kv KV) EmbeddingHeadCount() uint64 {
+	if heads := kv.HeadCount(); heads > 0 {
+		return kv.EmbeddingLength() / kv.HeadCount()
+	}
+
+	return 0
+}
+
+func (kv KV) EmbeddingHeadCountK() uint64 {
+	if k := kv.u64(fmt.Sprintf("%s.attention.key_length", kv.Architecture())); k > 0 {
+		return k
+	}
+
+	return kv.EmbeddingHeadCount()
+}
+
+func (kv KV) EmbeddingHeadCountV() uint64 {
+	if v := kv.u64(fmt.Sprintf("%s.attention.value_length", kv.Architecture())); v > 0 {
+		return v
+	}
+
+	return kv.EmbeddingHeadCount()
+}
+
 func (kv KV) GQA() uint64 {
 	return kv.HeadCount() / kv.HeadCountKV()
 }
@@ -254,7 +280,18 @@ func DetectGGMLType(b []byte) string {
 	}
 }
 
-func DecodeGGML(rs io.ReadSeeker) (*GGML, int64, error) {
+// DecodeGGML decodes a GGML model from the given reader.
+//
+// It collects array values for arrays with a size less than or equal to
+// maxArraySize. If maxArraySize is 0, the default value of 1024 is used. If
+// the maxArraySize is negative, all arrays are collected.
+func DecodeGGML(rs io.ReadSeeker, maxArraySize int) (*GGML, int64, error) {
+	if maxArraySize == 0 {
+		maxArraySize = 1024
+	}
+
+	rs = bufioutil.NewBufferedSeeker(rs, 32<<10)
+
 	var magic uint32
 	if err := binary.Read(rs, binary.LittleEndian, &magic); err != nil {
 		return nil, 0, err
@@ -267,17 +304,15 @@ func DecodeGGML(rs io.ReadSeeker) (*GGML, int64, error) {
 	case FILE_MAGIC_GGLA:
 		c = &containerGGLA{}
 	case FILE_MAGIC_GGUF_LE:
-		c = &containerGGUF{ByteOrder: binary.LittleEndian}
+		c = &containerGGUF{ByteOrder: binary.LittleEndian, maxArraySize: maxArraySize}
 	case FILE_MAGIC_GGUF_BE:
-		c = &containerGGUF{ByteOrder: binary.BigEndian}
+		c = &containerGGUF{ByteOrder: binary.BigEndian, maxArraySize: maxArraySize}
 	default:
 		return nil, 0, errors.New("invalid file magic")
 	}
 
 	model, err := c.Decode(rs)
-	if errors.Is(err, io.EOF) {
-		// noop
-	} else if err != nil {
+	if err != nil {
 		return nil, 0, err
 	}
 
@@ -297,7 +332,10 @@ func (llm GGML) GraphSize(context, batch uint64) (partialOffload, fullOffload ui
 	embedding := llm.KV().EmbeddingLength()
 	heads := llm.KV().HeadCount()
 	headsKV := llm.KV().HeadCountKV()
-	vocab := uint64(len(llm.KV()["tokenizer.ggml.tokens"].([]any)))
+	vocab := uint64(llm.KV()["tokenizer.ggml.tokens"].(*array).size)
+
+	embeddingHeads := llm.KV().EmbeddingHeadCount()
+	embeddingHeadsK := llm.KV().EmbeddingHeadCountK()
 
 	layers := llm.Tensors().Layers()
 
@@ -308,7 +346,7 @@ func (llm GGML) GraphSize(context, batch uint64) (partialOffload, fullOffload ui
 		partialOffload = 4 * batch * embedding
 		partialOffload += max(
 			// 4*batch*(4+6*embedding+context*(2*heads)+llm.KV().GQA()),
-			4*batch*(1+embedding+max(context, embedding))+embedding*embedding*9/16+4*context*(batch*heads+embedding/heads*headsKV),
+			4*batch*(1+embedding+max(context, embedding))+embedding*embedding*9/16+4*context*(batch*heads+embeddingHeads*headsKV),
 			4*batch*(embedding+vocab)+embedding*vocab*105/128,
 		)
 
@@ -316,15 +354,15 @@ func (llm GGML) GraphSize(context, batch uint64) (partialOffload, fullOffload ui
 			// mixtral 8x22b
 			ff := uint64(llm.KV()["llama.feed_forward_length"].(uint32))
 			partialOffload = max(
-				3*ffnGateExpsWeight.Size()+4*batch*(2*ff+headsKV+embedding+context+embedding/heads*headsKV),
-				4*(context*batch*heads+context*embedding/heads*headsKV+batch*1024+embedding/heads*headsKV*batch),
+				3*ffnGateExpsWeight.Size()+4*batch*(2*ff+headsKV+embedding+context+embeddingHeads*headsKV),
+				4*(context*batch*heads+context*embeddingHeads*headsKV+batch*1024+embeddingHeads*headsKV*batch),
 			)
 		} else if ffnGateWeight, ok := layers["blk.0"]["ffn_gate.0.weight"]; ok {
 			// mixtral 8x7b
 			ffnGateWeight1 := ffnGateWeight.Shape[1]
 			fullOffload = 4 * batch * (2 + 3*embedding + context*(1+heads) + 2*headsKV + ffnGateWeight1)
 			partialOffload = max(
-				4*batch*(3+embedding/heads*headsKV+embedding+context*(1+heads)+ffnGateWeight1)+(embedding*embedding+3*embedding*headsKV*ffnGateWeight1)*9/16,
+				4*batch*(3+embeddingHeads*headsKV+embedding+context*(1+heads)+ffnGateWeight1)+(embedding*embedding+3*embedding*headsKV*ffnGateWeight1)*9/16,
 				4*batch*(1+2*embedding+context*(1+heads))+embedding*(6*context*headsKV/heads+embedding*9/16),
 			)
 		}
@@ -368,15 +406,14 @@ func (llm GGML) GraphSize(context, batch uint64) (partialOffload, fullOffload ui
 			fullOffload,
 		)
 	case "deepseek2":
-		keys := uint64(llm.KV()["deepseek2.attention.key_length"].(uint32))
 		fullOffload = max(
 			4*batch*(3*embedding+vocab),
-			4*batch*(3*embedding+2+context*(1+headsKV)+2*keys*headsKV),
+			4*batch*(3*embedding+2+context*(1+headsKV)+2*embeddingHeadsK*headsKV),
 		)
 
 		partialOffload = max(
 			4*batch*(3*embedding+vocab)+embedding*vocab*105/128,
-			4*batch*(2*embedding+1+2*keys*headsKV+context+context*headsKV)+4*keys*context*headsKV+embedding*keys*headsKV*9/16,
+			4*batch*(2*embedding+1+2*embeddingHeadsK*headsKV+context+context*headsKV)+4*embeddingHeadsK*context*headsKV+embedding*embeddingHeadsK*headsKV*9/16,
 		)
 	}
 

llm/ggml_test.go

@@ -0,0 +1 @@
+package llm

llm/gguf.go

@@ -3,11 +3,10 @@ package llm
 import (
 	"bytes"
 	"encoding/binary"
+	"encoding/json"
 	"fmt"
 	"io"
 	"strings"
-
-	"log/slog"
 )
 
 type containerGGUF struct {
@@ -29,6 +28,12 @@ type containerGGUF struct {
 		NumTensor uint64
 		NumKV     uint64
 	}
+
+	maxArraySize int
+}
+
+func (c *containerGGUF) canCollectArray(size int) bool {
+	return c.maxArraySize < 0 || size <= c.maxArraySize
 }
 
 func (c *containerGGUF) Name() string {
@@ -54,7 +59,6 @@ func (c *containerGGUF) Decode(rs io.ReadSeeker) (model, error) {
 	}
 
 	model := newGGUF(c)
-	slog.Debug(fmt.Sprintf("model = %#v", model))
 	if err := model.Decode(rs); err != nil {
 		return nil, err
 	}
@@ -85,6 +89,8 @@ type gguf struct {
 	tensors []*Tensor
 
 	parameters uint64
+
+	scratch [16 << 10]byte
 }
 
 func newGGUF(container *containerGGUF) *gguf {
@@ -181,34 +187,34 @@ func (llm *gguf) Decode(rs io.ReadSeeker) error {
 	}
 
 	// decode tensors
-	for i := 0; uint64(i) < llm.numTensor(); i++ {
+	for range llm.numTensor() {
 		name, err := readGGUFString(llm, rs)
 		if err != nil {
-			return err
+			return fmt.Errorf("failed to read tensor name: %w", err)
 		}
 
 		// dims is the number of dimensions in the tensor
 		dims, err := readGGUF[uint32](llm, rs)
 		if err != nil {
-			return err
+			return fmt.Errorf("failed to read tensor dimensions: %w", err)
 		}
 
 		shape := [4]uint64{1, 1, 1, 1}
 		for i := 0; uint32(i) < dims; i++ {
 			shape[i], err = readGGUF[uint64](llm, rs)
 			if err != nil {
-				return err
+				return fmt.Errorf("failed to read tensor shape: %w", err)
 			}
 		}
 
 		kind, err := readGGUF[uint32](llm, rs)
 		if err != nil {
-			return err
+			return fmt.Errorf("failed to read tensor kind: %w", err)
 		}
 
 		offset, err := readGGUF[uint64](llm, rs)
 		if err != nil {
-			return err
+			return fmt.Errorf("failed to read tensor offset: %w", err)
 		}
 
 		tensor := Tensor{
@@ -230,24 +236,19 @@ func (llm *gguf) Decode(rs io.ReadSeeker) error {
 		alignment = 32
 	}
 
-	offset, err := rs.Seek(0, io.SeekCurrent)
-	if err != nil {
-		return err
-	}
-
-	padding := llm.padding(offset, int64(alignment))
-	if _, err := rs.Seek(padding, io.SeekCurrent); err != nil {
-		return err
-	}
-
 	for _, tensor := range llm.tensors {
-		if _, err := rs.Seek(int64(tensor.Size()), io.SeekCurrent); err != nil {
-			return err
+		offset, err := rs.Seek(0, io.SeekCurrent)
+		if err != nil {
+			return fmt.Errorf("failed to get current offset: %w", err)
 		}
 
-		padding := llm.padding(int64(tensor.Size()), int64(alignment))
+		padding := llm.padding(offset, int64(alignment))
 		if _, err := rs.Seek(padding, io.SeekCurrent); err != nil {
-			return err
+			return fmt.Errorf("failed to seek to init padding: %w", err)
+		}
+
+		if _, err := rs.Seek(int64(tensor.Size()), io.SeekCurrent); err != nil {
+			return fmt.Errorf("failed to seek to tensor: %w", err)
 		}
 	}
 
@@ -285,22 +286,48 @@ func readGGUFV1String(llm *gguf, r io.Reader) (string, error) {
 	return b.String(), nil
 }
 
+func discardGGUFString(llm *gguf, r io.Reader) error {
+	buf := llm.scratch[:8]
+	_, err := io.ReadFull(r, buf)
+	if err != nil {
+		return err
+	}
+
+	size := int(llm.ByteOrder.Uint64(buf))
+	for size > 0 {
+		n, err := r.Read(llm.scratch[:min(size, cap(llm.scratch))])
+		if err != nil {
+			return err
+		}
+		size -= n
+	}
+	return nil
+}
+
 func readGGUFString(llm *gguf, r io.Reader) (string, error) {
 	if llm.Version == 1 {
 		return readGGUFV1String(llm, r)
 	}
 
-	var length uint64
-	if err := binary.Read(r, llm.ByteOrder, &length); err != nil {
+	buf := llm.scratch[:8]
+	_, err := io.ReadFull(r, buf)
+	if err != nil {
 		return "", err
 	}
 
-	var b bytes.Buffer
-	if _, err := io.CopyN(&b, r, int64(length)); err != nil {
+	length := int(llm.ByteOrder.Uint64(buf))
+	if length > len(llm.scratch) {
+		buf = make([]byte, length)
+	} else {
+		buf = llm.scratch[:length]
+	}
+	clear(buf)
+
+	_, err = io.ReadFull(r, buf)
+	if err != nil {
 		return "", err
 	}
-
-	return b.String(), nil
+	return string(buf), nil
 }
 
 func writeGGUFString(llm *gguf, w io.Writer, s string) error {
@@ -316,7 +343,16 @@ func writeGGUFString(llm *gguf, w io.Writer, s string) error {
 	return err
 }
 
-func readGGUFV1Array(llm *gguf, r io.Reader) (a []any, err error) {
+type array struct {
+	size   int
+	values []any
+}
+
+func (a *array) MarshalJSON() ([]byte, error) {
+	return json.Marshal(a.values)
+}
+
+func readGGUFV1Array(llm *gguf, r io.Reader) (*array, error) {
 	t, err := readGGUF[uint32](llm, r)
 	if err != nil {
 		return nil, err
@@ -327,7 +363,12 @@ func readGGUFV1Array(llm *gguf, r io.Reader) (a []any, err error) {
 		return nil, err
 	}
 
-	for i := 0; uint32(i) < n; i++ {
+	a := &array{size: int(n)}
+	if llm.canCollectArray(int(n)) {
+		a.values = make([]any, 0, int(n))
+	}
+
+	for i := range n {
 		var e any
 		switch t {
 		case ggufTypeUint8:
@@ -361,13 +402,15 @@ func readGGUFV1Array(llm *gguf, r io.Reader) (a []any, err error) {
 			return nil, err
 		}
 
-		a = append(a, e)
+		if a.values != nil {
+			a.values[i] = e
+		}
 	}
 
-	return
+	return a, nil
 }
 
-func readGGUFArray(llm *gguf, r io.Reader) (a []any, err error) {
+func readGGUFArray(llm *gguf, r io.Reader) (*array, error) {
 	if llm.Version == 1 {
 		return readGGUFV1Array(llm, r)
 	}
@@ -382,7 +425,12 @@ func readGGUFArray(llm *gguf, r io.Reader) (a []any, err error) {
 		return nil, err
 	}
 
-	for i := 0; uint64(i) < n; i++ {
+	a := &array{size: int(n)}
+	if llm.canCollectArray(int(n)) {
+		a.values = make([]any, int(n))
+	}
+
+	for i := range n {
 		var e any
 		switch t {
 		case ggufTypeUint8:
@@ -408,7 +456,11 @@ func readGGUFArray(llm *gguf, r io.Reader) (a []any, err error) {
 		case ggufTypeBool:
 			e, err = readGGUF[bool](llm, r)
 		case ggufTypeString:
-			e, err = readGGUFString(llm, r)
+			if a.values != nil {
+				e, err = readGGUFString(llm, r)
+			} else {
+				err = discardGGUFString(llm, r)
+			}
 		default:
 			return nil, fmt.Errorf("invalid array type: %d", t)
 		}
@@ -416,10 +468,12 @@ func readGGUFArray(llm *gguf, r io.Reader) (a []any, err error) {
 			return nil, err
 		}
 
-		a = append(a, e)
+		if a.values != nil {
+			a.values[i] = e
+		}
 	}
 
-	return
+	return a, nil
 }
 
 func writeGGUFArray[S ~[]E, E any](llm *gguf, w io.Writer, t uint32, s S) error {

llm/memory.go

@@ -115,8 +115,8 @@ func EstimateGPULayers(gpus []gpu.GpuInfo, ggml *GGML, projectors []string, opts
 		slog.Warn("model missing blk.0 layer size")
 	}
 
-	// fp16 k,v = (1 (k) + 1 (v)) * sizeof(float16) * n_ctx * n_layer * n_embd / n_head * n_head_kv
-	var kv uint64 = 2 * 2 * uint64(opts.NumCtx) * ggml.KV().BlockCount() * ggml.KV().EmbeddingLength() / ggml.KV().HeadCount() * ggml.KV().HeadCountKV()
+	// fp16 k,v = sizeof(float16) * n_ctx * n_layer * (n_embd_head_k + n_embd_head_v) * n_head_kv
+	var kv uint64 = 2 * uint64(opts.NumCtx) * ggml.KV().BlockCount() * (ggml.KV().EmbeddingHeadCountK() + ggml.KV().EmbeddingHeadCountV()) * ggml.KV().HeadCountKV()
 
 	// KV is proportional to the number of layers
 	layerSize += kv / ggml.KV().BlockCount()

llm/memory_test.go

@@ -22,13 +22,14 @@ func TestEstimateGPULayers(t *testing.T) {
 	defer f.Close()
 	gguf := NewGGUFV3(binary.LittleEndian)
 	inputLayerCount := 5
+
 	tensors := []Tensor{
-		{Name: "blk.0.attn.weight", Kind: uint32(0), Offset: uint64(0), Shape: []uint64{1, 1, 1, 1}, WriterTo: &bytes.Reader{}},
-		{Name: "blk.1.attn.weight", Kind: uint32(0), Offset: uint64(0), Shape: []uint64{1, 1, 1, 1}, WriterTo: &bytes.Reader{}},
-		{Name: "blk.2.attn.weight", Kind: uint32(0), Offset: uint64(0), Shape: []uint64{1, 1, 1, 1}, WriterTo: &bytes.Reader{}},
-		{Name: "blk.3.attn.weight", Kind: uint32(0), Offset: uint64(0), Shape: []uint64{1, 1, 1, 1}, WriterTo: &bytes.Reader{}},
-		{Name: "blk.4.attn.weight", Kind: uint32(0), Offset: uint64(0), Shape: []uint64{1, 1, 1, 1}, WriterTo: &bytes.Reader{}},
-		{Name: "output.weight", Kind: uint32(0), Offset: uint64(0), Shape: []uint64{1, 1, 1, 1}, WriterTo: &bytes.Reader{}},
+		{Name: "blk.0.attn.weight", Kind: uint32(0), Offset: uint64(0), Shape: []uint64{1, 1, 1, 1}, WriterTo: bytes.NewReader(make([]byte, 32))},
+		{Name: "blk.1.attn.weight", Kind: uint32(0), Offset: uint64(0), Shape: []uint64{1, 1, 1, 1}, WriterTo: bytes.NewReader(make([]byte, 32))},
+		{Name: "blk.2.attn.weight", Kind: uint32(0), Offset: uint64(0), Shape: []uint64{1, 1, 1, 1}, WriterTo: bytes.NewReader(make([]byte, 32))},
+		{Name: "blk.3.attn.weight", Kind: uint32(0), Offset: uint64(0), Shape: []uint64{1, 1, 1, 1}, WriterTo: bytes.NewReader(make([]byte, 32))},
+		{Name: "blk.4.attn.weight", Kind: uint32(0), Offset: uint64(0), Shape: []uint64{1, 1, 1, 1}, WriterTo: bytes.NewReader(make([]byte, 32))},
+		{Name: "output.weight", Kind: uint32(0), Offset: uint64(0), Shape: []uint64{1, 1, 1, 1}, WriterTo: bytes.NewReader(make([]byte, 32))},
 	}
 	assert.Len(t, tensors, inputLayerCount+1)
 	err = gguf.Encode(f, KV{
@@ -45,8 +46,10 @@ func TestEstimateGPULayers(t *testing.T) {
 	}, tensors)
 	require.NoError(t, err)
 
-	ggml, err := LoadModel(f.Name())
-	require.NoError(t, err)
+	ggml, err := LoadModel(f.Name(), 0)
+	if err != nil {
+		t.Fatal(err)
+	}
 
 	// Simple CPU scenario
 	gpus := []gpu.GpuInfo{

llm/server.go

@@ -60,7 +60,12 @@ type llmServer struct {
 	sem *semaphore.Weighted
 }
 
-func LoadModel(model string) (*GGML, error) {
+// LoadModel will load a model from disk. The model must be in the GGML format.
+//
+// It collects array values for arrays with a size less than or equal to
+// maxArraySize. If maxArraySize is 0, the default value of 1024 is used. If
+// the maxArraySize is negative, all arrays are collected.
+func LoadModel(model string, maxArraySize int) (*GGML, error) {
 	if _, err := os.Stat(model); err != nil {
 		return nil, err
 	}
@@ -71,7 +76,7 @@ func LoadModel(model string) (*GGML, error) {
 	}
 	defer f.Close()
 
-	ggml, _, err := DecodeGGML(f)
+	ggml, _, err := DecodeGGML(f, maxArraySize)
 	return ggml, err
 }
 
@@ -81,7 +86,17 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 	var err error
 	var cpuRunner string
 	var estimate MemoryEstimate
-	var systemMemory uint64
+	var systemTotalMemory uint64
+	var systemFreeMemory uint64
+
+	systemMemInfo, err := gpu.GetCPUMem()
+	if err != nil {
+		slog.Error("failed to lookup system memory", "error", err)
+	} else {
+		systemTotalMemory = systemMemInfo.TotalMemory
+		systemFreeMemory = systemMemInfo.FreeMemory
+		slog.Debug("system memory", "total", format.HumanBytes2(systemTotalMemory), "free", systemFreeMemory)
+	}
 
 	// If the user wants zero GPU layers, reset the gpu list to be CPU/system ram info
 	if opts.NumGPU == 0 {
@@ -91,19 +106,10 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 		cpuRunner = serverForCpu()
 		estimate = EstimateGPULayers(gpus, ggml, projectors, opts)
 	} else {
-		if gpus[0].Library == "metal" {
-			memInfo, err := gpu.GetCPUMem()
-			if err != nil {
-				slog.Error("failed to lookup system memory", "error", err)
-			} else {
-				systemMemory = memInfo.TotalMemory
-				slog.Debug("system memory", "total", format.HumanBytes2(systemMemory))
-			}
-		}
 		estimate = EstimateGPULayers(gpus, ggml, projectors, opts)
 
 		switch {
-		case gpus[0].Library == "metal" && estimate.VRAMSize > systemMemory:
+		case gpus[0].Library == "metal" && estimate.VRAMSize > systemTotalMemory:
 			// disable partial offloading when model is greater than total system memory as this
 			// can lead to locking up the system
 			opts.NumGPU = 0
@@ -211,7 +217,10 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 	}
 
 	// Windows CUDA should not use mmap for best performance
-	if (runtime.GOOS == "windows" && gpus[0].Library == "cuda") || opts.UseMMap == api.TriStateFalse {
+	// Linux  with a model larger than free space, mmap leads to thrashing
+	if (runtime.GOOS == "windows" && gpus[0].Library == "cuda" && opts.UseMMap == api.TriStateUndefined) ||
+		(runtime.GOOS == "linux" && systemFreeMemory < estimate.TotalSize && opts.UseMMap == api.TriStateUndefined) ||
+		opts.UseMMap == api.TriStateFalse {
 		params = append(params, "--no-mmap")
 	}
 
@@ -408,7 +417,7 @@ func projectorMemoryRequirements(filename string) uint64 {
 	}
 	defer file.Close()
 
-	ggml, _, err := DecodeGGML(file)
+	ggml, _, err := DecodeGGML(file, 0)
 	if err != nil {
 		return 0
 	}

scripts/install.sh

@@ -279,7 +279,7 @@ if ! check_gpu nvidia-smi || [ -z "$(nvidia-smi | grep -o "CUDA Version: [0-9]*\
     case $OS_NAME in
         centos|rhel) install_cuda_driver_yum 'rhel' $(echo $OS_VERSION | cut -d '.' -f 1) ;;
         rocky) install_cuda_driver_yum 'rhel' $(echo $OS_VERSION | cut -c1) ;;
-        fedora) [ $OS_VERSION -lt '37' ] && install_cuda_driver_yum $OS_NAME $OS_VERSION || install_cuda_driver_yum $OS_NAME '37';;
+        fedora) [ $OS_VERSION -lt '39' ] && install_cuda_driver_yum $OS_NAME $OS_VERSION || install_cuda_driver_yum $OS_NAME '39';;
         amzn) install_cuda_driver_yum 'fedora' '37' ;;
         debian) install_cuda_driver_apt $OS_NAME $OS_VERSION ;;
         ubuntu) install_cuda_driver_apt $OS_NAME $(echo $OS_VERSION | sed 's/\.//') ;;

server/images.go

@@ -414,17 +414,22 @@ func CreateModel(ctx context.Context, name model.Name, modelFileDir, quantizatio
 							return err
 						}
 
-						layers, err := parseFromFile(ctx, temp, "", fn)
+						layer, err := NewLayer(temp, baseLayer.MediaType)
 						if err != nil {
 							return err
 						}
 
-						if len(layers) != 1 {
-							return errors.New("quantization failed")
+						if _, err := temp.Seek(0, io.SeekStart); err != nil {
+							return err
 						}
 
-						baseLayer.Layer = layers[0].Layer
-						baseLayer.GGML = layers[0].GGML
+						ggml, _, err := llm.DecodeGGML(temp, 0)
+						if err != nil {
+							return err
+						}
+
+						baseLayer.Layer = layer
+						baseLayer.GGML = ggml
 					}
 				}
 

server/model.go

@@ -63,7 +63,7 @@ func parseFromModel(ctx context.Context, name model.Name, fn func(api.ProgressRe
 			}
 			defer blob.Close()
 
-			ggml, _, err := llm.DecodeGGML(blob)
+			ggml, _, err := llm.DecodeGGML(blob, 0)
 			if err != nil {
 				return nil, err
 			}
@@ -176,7 +176,7 @@ func parseFromZipFile(_ context.Context, file *os.File, digest string, fn func(a
 	}
 	defer bin.Close()
 
-	ggml, _, err := llm.DecodeGGML(bin)
+	ggml, _, err := llm.DecodeGGML(bin, 0)
 	if err != nil {
 		return nil, err
 	}
@@ -210,7 +210,7 @@ func parseFromFile(ctx context.Context, file *os.File, digest string, fn func(ap
 
 	var offset int64
 	for offset < stat.Size() {
-		ggml, n, err := llm.DecodeGGML(file)
+		ggml, n, err := llm.DecodeGGML(file, 0)
 		if errors.Is(err, io.EOF) {
 			break
 		} else if err != nil {

server/routes.go

@@ -754,7 +754,11 @@ func GetModelInfo(req api.ShowRequest) (*api.ShowResponse, error) {
 }
 
 func getKVData(digest string, verbose bool) (llm.KV, error) {
-	kvData, err := llm.LoadModel(digest)
+	maxArraySize := 0
+	if verbose {
+		maxArraySize = -1
+	}
+	kvData, err := llm.LoadModel(digest, maxArraySize)
 	if err != nil {
 		return nil, err
 	}
@@ -1101,11 +1105,20 @@ func Serve(ln net.Listener) error {
 	schedCtx, schedDone := context.WithCancel(ctx)
 	sched := InitScheduler(schedCtx)
 	s := &Server{addr: ln.Addr(), sched: sched}
-	r := s.GenerateRoutes()
+
+	http.Handle("/", s.GenerateRoutes())
 
 	slog.Info(fmt.Sprintf("Listening on %s (version %s)", ln.Addr(), version.Version))
 	srvr := &http.Server{
-		Handler: r,
+		// Use http.DefaultServeMux so we get net/http/pprof for
+		// free.
+		//
+		// TODO(bmizerany): Decide if we want to make this
+		// configurable so it is not exposed by default, or allow
+		// users to bind it to a different port. This was a quick
+		// and easy way to get pprof, but it may not be the best
+		// way.
+		Handler: nil,
 	}
 
 	// listen for a ctrl+c and stop any loaded llm

server/sched.go

@@ -144,7 +144,7 @@ func (s *Scheduler) processPending(ctx context.Context) {
 					}
 
 					// Load model for fitting
-					ggml, err := llm.LoadModel(pending.model.ModelPath)
+					ggml, err := llm.LoadModel(pending.model.ModelPath, 0)
 					if err != nil {
 						pending.errCh <- err
 						break

server/sched_test.go

@@ -128,14 +128,14 @@ func newScenario(t *testing.T, ctx context.Context, modelName string, estimatedV
 		"tokenizer.ggml.scores":         []float32{0},
 		"tokenizer.ggml.token_type":     []int32{0},
 	}, []llm.Tensor{
-		{Name: "blk.0.attn.weight", Kind: uint32(0), Offset: uint64(0), Shape: []uint64{1, 1, 1, 1}, WriterTo: &bytes.Reader{}},
-		{Name: "output.weight", Kind: uint32(0), Offset: uint64(0), Shape: []uint64{1, 1, 1, 1}, WriterTo: &bytes.Reader{}},
+		{Name: "blk.0.attn.weight", Kind: uint32(0), Offset: uint64(0), Shape: []uint64{1, 1, 1, 1}, WriterTo: bytes.NewReader(make([]byte, 32))},
+		{Name: "output.weight", Kind: uint32(0), Offset: uint64(0), Shape: []uint64{1, 1, 1, 1}, WriterTo: bytes.NewReader(make([]byte, 32))},
 	})
 	require.NoError(t, err)
 
 	fname := f.Name()
 	model := &Model{Name: modelName, ModelPath: fname}
-	scenario.ggml, err = llm.LoadModel(model.ModelPath)
+	scenario.ggml, err = llm.LoadModel(model.ModelPath, 0)
 	require.NoError(t, err)
 
 	scenario.req = &LlmRequest{

util/bufioutil/buffer_seeker.go

@@ -0,0 +1,34 @@
+package bufioutil
+
+import (
+	"bufio"
+	"io"
+)
+
+type BufferedSeeker struct {
+	rs io.ReadSeeker
+	br *bufio.Reader
+}
+
+func NewBufferedSeeker(rs io.ReadSeeker, size int) *BufferedSeeker {
+	return &BufferedSeeker{
+		rs: rs,
+		br: bufio.NewReaderSize(rs, size),
+	}
+}
+
+func (b *BufferedSeeker) Read(p []byte) (int, error) {
+	return b.br.Read(p)
+}
+
+func (b *BufferedSeeker) Seek(offset int64, whence int) (int64, error) {
+	if whence == io.SeekCurrent {
+		offset -= int64(b.br.Buffered())
+	}
+	n, err := b.rs.Seek(offset, whence)
+	if err != nil {
+		return 0, err
+	}
+	b.br.Reset(b.rs)
+	return n, nil
+}

util/bufioutil/buffer_seeker_test.go

@@ -0,0 +1,64 @@
+package bufioutil
+
+import (
+	"bytes"
+	"io"
+	"strings"
+	"testing"
+)
+
+func TestBufferedSeeker(t *testing.T) {
+	const alphabet = "abcdefghijklmnopqrstuvwxyz"
+
+	bs := NewBufferedSeeker(strings.NewReader(alphabet), 0) // minReadBufferSize = 16
+
+	checkRead := func(buf []byte, expected string) {
+		t.Helper()
+		_, err := bs.Read(buf)
+		if err != nil {
+			t.Fatal(err)
+		}
+		if !bytes.Equal(buf, []byte(expected)) {
+			t.Fatalf("expected %s, got %s", expected, buf)
+		}
+	}
+
+	// Read the first 5 bytes
+	buf := make([]byte, 5)
+
+	checkRead(buf, "abcde")
+
+	// Seek back to the beginning
+	_, err := bs.Seek(0, io.SeekStart)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	// read 'a'
+	checkRead(buf[:1], "a")
+
+	if bs.br.Buffered() == 0 {
+		t.Fatalf("totally unexpected sanity check failed")
+	}
+
+	// Seek past 'b'
+	_, err = bs.Seek(1, io.SeekCurrent)
+	if err != nil {
+		t.Fatal(err)
+	}
+	checkRead(buf, "cdefg")
+
+	// Seek back to the beginning
+	_, err = bs.Seek(0, io.SeekStart)
+	if err != nil {
+		t.Fatal(err)
+	}
+	checkRead(buf, "abcde")
+
+	// Seek to the end
+	_, err = bs.Seek(-5, io.SeekEnd)
+	if err != nil {
+		t.Fatal(err)
+	}
+	checkRead(buf, "vwxyz")
+}