add model_format config var

Go is required and not installed by default.

.dockerignore

@@ -1,8 +1,4 @@
-build
-llama/build
-.venv
 .vscode
 ollama
 app
-web
-.env
+llm/llama.cpp/ggml

Dockerfile

@@ -1,15 +1,21 @@
-FROM golang:1.20
+FROM golang:alpine
+
 WORKDIR /go/src/github.com/jmorganca/ollama
+RUN apk add --no-cache git build-base cmake
+
 COPY . .
-RUN CGO_ENABLED=1 go build -ldflags '-linkmode external -extldflags "-static"' .
+RUN go generate ./... && go build -ldflags '-linkmode external -extldflags "-static"' .
 
 FROM alpine
-COPY --from=0 /go/src/github.com/jmorganca/ollama/ollama /bin/ollama
-EXPOSE 11434
+ENV OLLAMA_HOST 0.0.0.0
+RUN apk add --no-cache libstdc++
+
 ARG USER=ollama
 ARG GROUP=ollama
-RUN addgroup -g 1000 $GROUP && adduser -u 1000 -DG $GROUP $USER
+RUN addgroup $GROUP && adduser -D -G $GROUP $USER
+
+COPY --from=0 /go/src/github.com/jmorganca/ollama/ollama /bin/ollama
+
 USER $USER:$GROUP
 ENTRYPOINT ["/bin/ollama"]
-ENV OLLAMA_HOST 0.0.0.0
 CMD ["serve"]

README.md

@@ -165,10 +165,11 @@ Ollama bundles model weights, configurations, and data into a single package, de
 
 ## Building
 
-Install `cmake`:
+Install `cmake` and `go`:
 
 ```
 brew install cmake
+brew install go
 ```
 
 Then generate dependencies and build:

format/openssh.go

@@ -10,15 +10,11 @@ package format
 
 import (
 	"crypto"
-	"crypto/ecdsa"
 	"crypto/ed25519"
-	"crypto/elliptic"
 	"crypto/rand"
-	"crypto/rsa"
 	"encoding/binary"
 	"encoding/pem"
 	"fmt"
-	"math/big"
 
 	"golang.org/x/crypto/ssh"
 )
@@ -41,25 +37,6 @@ type openSSHPrivateKey struct {
 	Rest    []byte `ssh:"rest"`
 }
 
-type openSSHRSAPrivateKey struct {
-	N       *big.Int
-	E       *big.Int
-	D       *big.Int
-	Iqmp    *big.Int
-	P       *big.Int
-	Q       *big.Int
-	Comment string
-	Pad     []byte `ssh:"rest"`
-}
-
-type openSSHECDSAPrivateKey struct {
-	Curve   string
-	Pub     []byte
-	D       *big.Int
-	Comment string
-	Pad     []byte `ssh:"rest"`
-}
-
 type openSSHEd25519PrivateKey struct {
 	Pub     []byte
 	Priv    []byte
@@ -85,64 +62,6 @@ func OpenSSHPrivateKey(key crypto.PrivateKey, comment string) (*pem.Block, error
 	}
 
 	switch k := key.(type) {
-	case *rsa.PrivateKey:
-		e := new(big.Int).SetInt64(int64(k.E))
-
-		key := openSSHRSAPrivateKey{
-			N:       k.N,
-			E:       e,
-			D:       k.D,
-			Iqmp:    k.Precomputed.Qinv,
-			P:       k.Primes[0],
-			Q:       k.Primes[1],
-			Comment: comment,
-		}
-
-		pk1.Keytype = ssh.KeyAlgoRSA
-		pk1.Rest = ssh.Marshal(key)
-
-		w.PubKey = ssh.Marshal(struct {
-			KeyType string
-			E       *big.Int
-			N       *big.Int
-		}{
-			ssh.KeyAlgoRSA, e, k.N,
-		})
-	case *ecdsa.PrivateKey:
-		var curve, keytype string
-		switch name := k.Curve.Params().Name; name {
-		case "P-256":
-			curve = "nistp256"
-			keytype = ssh.KeyAlgoECDSA256
-		case "P-384":
-			curve = "nistp384"
-			keytype = ssh.KeyAlgoECDSA384
-		case "P-521":
-			curve = "nistp521"
-			keytype = ssh.KeyAlgoECDSA521
-		default:
-			return nil, fmt.Errorf("ssh: unknown curve %q", name)
-		}
-
-		pub := elliptic.Marshal(k.Curve, k.X, k.Y)
-
-		key := openSSHECDSAPrivateKey{
-			Curve:   curve,
-			Pub:     pub,
-			D:       k.D,
-			Comment: comment,
-		}
-
-		pk1.Keytype = keytype
-		pk1.Rest = ssh.Marshal(key)
-
-		w.PubKey = ssh.Marshal(struct {
-			KeyType string
-			Curve   string
-			Pub     []byte
-		}{
-			keytype, curve, pub,
-		})
 	case ed25519.PrivateKey:
 		pub, priv := k[32:], k
 		key := openSSHEd25519PrivateKey{

llm/llama.cpp/generate_darwin_amd64.go

@@ -6,5 +6,5 @@ package llm
 //go:generate git -C ggml apply ../ggml_patch/0002-34B-model-support.patch
 //go:generate git -C ggml apply ../ggml_patch/0003-metal-fix-synchronization-in-new-matrix-multiplicati.patch
 //go:generate git -C ggml apply ../ggml_patch/0004-metal-add-missing-barriers-for-mul-mat-2699.patch
-//go:generate cmake --fresh -S ggml -B ggml/build/cpu -DLLAMA_ACCELERATE=on -DLLAMA_K_QUANTS=on -DCMAKE_SYSTEM_PROCESSOR=x86_64 -DCMAKE_OSX_ARCHITECTURES=x86_64 -DCMAKE_OSX_DEPLOYMENT_TARGET=10.11
+//go:generate cmake --fresh -S ggml -B ggml/build/cpu -DLLAMA_ACCELERATE=on -DLLAMA_K_QUANTS=on -DCMAKE_SYSTEM_PROCESSOR=x86_64 -DCMAKE_OSX_ARCHITECTURES=x86_64 -DCMAKE_OSX_DEPLOYMENT_TARGET=11.0
 //go:generate cmake --build ggml/build/cpu --target server --config Release

server/images.go

@@ -116,6 +116,7 @@ type LayerReader struct {
 type ConfigV2 struct {
 	ModelFamily llm.ModelFamily `json:"model_family"`
 	ModelType   string          `json:"model_type"`
+	ModelFormat string          `json:"model_format"`
 	FileType    string          `json:"file_type"`
 	RootFS      RootFS          `json:"rootfs"`
 
@@ -335,6 +336,7 @@ func CreateModel(ctx context.Context, name string, path string, fn func(resp api
 
 					config.ModelFamily = ggml.ModelFamily()
 					config.ModelType = ggml.ModelType().String()
+					config.ModelFormat = ggml.Name()
 					config.FileType = ggml.FileType().String()
 
 					// reset the file
@@ -366,9 +368,10 @@ func CreateModel(ctx context.Context, name string, path string, fn func(resp api
 					return err
 				}
 
-				// copie the model metadata
+				// copy the model metadata
 				config.ModelFamily = source.ModelFamily
 				config.ModelType = source.ModelType
+				config.ModelFormat = source.ModelFormat
 				config.FileType = source.FileType
 
 				for _, l := range mf.Layers {