update llama.cpp submodule to c29af7e (#2868)

Add ollama user to video group

Dockerfile

@@ -132,6 +132,7 @@ ENV OLLAMA_HOST 0.0.0.0
 ENV PATH=/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin
 ENV LD_LIBRARY_PATH=/usr/local/nvidia/lib:/usr/local/nvidia/lib64
 ENV NVIDIA_DRIVER_CAPABILITIES=compute,utility
+ENV NVIDIA_VISIBLE_DEVICES=all
 
 ENTRYPOINT ["/bin/ollama"]
 CMD ["serve"]

README.md

@@ -62,6 +62,8 @@ Here are some example models that can be downloaded:
 | Orca Mini          | 3B         | 1.9GB | `ollama run orca-mini`         |
 | Vicuna             | 7B         | 3.8GB | `ollama run vicuna`            |
 | LLaVA              | 7B         | 4.5GB | `ollama run llava`             |
+| Gemma              | 2B         | 1.4GB | `ollama run gemma:2b`          |
+| Gemma              | 7B         | 4.8GB | `ollama run gemma:7b`          |
 
 > Note: You should have at least 8 GB of RAM available to run the 7B models, 16 GB to run the 13B models, and 32 GB to run the 33B models.
 
@@ -258,19 +260,23 @@ See the [API documentation](./docs/api.md) for all endpoints.
 ### Web & Desktop
 
 - [Bionic GPT](https://github.com/bionic-gpt/bionic-gpt)
+- [Enchanted (macOS native)](https://github.com/AugustDev/enchanted)
 - [HTML UI](https://github.com/rtcfirefly/ollama-ui)
 - [Chatbot UI](https://github.com/ivanfioravanti/chatbot-ollama)
 - [Typescript UI](https://github.com/ollama-interface/Ollama-Gui?tab=readme-ov-file)
 - [Minimalistic React UI for Ollama Models](https://github.com/richawo/minimal-llm-ui)
 - [Open WebUI](https://github.com/open-webui/open-webui)
 - [Ollamac](https://github.com/kevinhermawan/Ollamac)
-- [big-AGI](https://github.com/enricoros/big-agi/blob/main/docs/config-ollama.md)
+- [big-AGI](https://github.com/enricoros/big-AGI/blob/main/docs/config-local-ollama.md)
 - [Cheshire Cat assistant framework](https://github.com/cheshire-cat-ai/core)
 - [Amica](https://github.com/semperai/amica)
 - [chatd](https://github.com/BruceMacD/chatd)
 - [Ollama-SwiftUI](https://github.com/kghandour/Ollama-SwiftUI)
 - [MindMac](https://mindmac.app)
 - [NextJS Web Interface for Ollama](https://github.com/jakobhoeg/nextjs-ollama-llm-ui)
+- [Msty](https://msty.app)
+- [Chatbox](https://github.com/Bin-Huang/Chatbox)
+- [NextChat](https://github.com/ChatGPTNextWeb/ChatGPT-Next-Web) with [Get Started Doc](https://docs.nextchat.dev/models/ollama)
 
 ### Terminal
 
@@ -301,6 +307,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 
 - [LangChain](https://python.langchain.com/docs/integrations/llms/ollama) and [LangChain.js](https://js.langchain.com/docs/modules/model_io/models/llms/integrations/ollama) with [example](https://js.langchain.com/docs/use_cases/question_answering/local_retrieval_qa)
 - [LangChainGo](https://github.com/tmc/langchaingo/) with [example](https://github.com/tmc/langchaingo/tree/main/examples/ollama-completion-example)
+- [LangChain4j](https://github.com/langchain4j/langchain4j) with [example](https://github.com/langchain4j/langchain4j-examples/tree/main/ollama-examples/src/main/java)
 - [LlamaIndex](https://gpt-index.readthedocs.io/en/stable/examples/llm/ollama.html)
 - [LangChain4j](https://github.com/langchain4j/langchain4j/tree/main/langchain4j-ollama)
 - [LiteLLM](https://github.com/BerriAI/litellm)
@@ -315,8 +322,10 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [LangChainDart](https://github.com/davidmigloz/langchain_dart)
 - [Semantic Kernel - Python](https://github.com/microsoft/semantic-kernel/tree/main/python/semantic_kernel/connectors/ai/ollama)
 - [Haystack](https://github.com/deepset-ai/haystack-integrations/blob/main/integrations/ollama.md)
+- [Elixir LangChain](https://github.com/brainlid/langchain)
 - [Ollama for R - rollama](https://github.com/JBGruber/rollama)
 - [Ollama-ex for Elixir](https://github.com/lebrunel/ollama-ex)
+- [Ollama Connector for SAP ABAP](https://github.com/b-tocs/abap_btocs_ollama)
 
 ### Mobile
 
@@ -337,6 +346,9 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Rivet plugin](https://github.com/abrenneke/rivet-plugin-ollama)
 - [Llama Coder](https://github.com/ex3ndr/llama-coder) (Copilot alternative using Ollama)
 - [Obsidian BMO Chatbot plugin](https://github.com/longy2k/obsidian-bmo-chatbot)
+- [Copilot for Obsidian plugin](https://github.com/logancyang/obsidian-copilot)
+- [Obsidian Local GPT plugin](https://github.com/pfrankov/obsidian-local-gpt)
 - [Open Interpreter](https://docs.openinterpreter.com/language-model-setup/local-models/ollama)
 - [twinny](https://github.com/rjmacarthy/twinny) (Copilot and Copilot chat alternative using Ollama)
 - [Wingman-AI](https://github.com/RussellCanfield/wingman-ai) (Copilot code and chat alternative using Ollama and HuggingFace)
+- [Page Assist](https://github.com/n4ze3m/page-assist) (Chrome Extension)

api/client.go

@@ -21,7 +21,7 @@ import (
 
 type Client struct {
 	base *url.URL
-	http http.Client
+	http *http.Client
 }
 
 func checkError(resp *http.Response, body []byte) error {
@@ -66,30 +66,13 @@ func ClientFromEnvironment() (*Client, error) {
 		}
 	}
 
-	client := Client{
+	return &Client{
 		base: &url.URL{
 			Scheme: scheme,
 			Host:   net.JoinHostPort(host, port),
 		},
-	}
-
-	mockRequest, err := http.NewRequest(http.MethodHead, client.base.String(), nil)
-	if err != nil {
-		return nil, err
-	}
-
-	proxyURL, err := http.ProxyFromEnvironment(mockRequest)
-	if err != nil {
-		return nil, err
-	}
-
-	client.http = http.Client{
-		Transport: &http.Transport{
-			Proxy: http.ProxyURL(proxyURL),
-		},
-	}
-
-	return &client, nil
+		http: http.DefaultClient,
+	}, nil
 }
 
 func (c *Client) do(ctx context.Context, method, path string, reqData, respData any) error {

api/types.go

@@ -83,7 +83,7 @@ type Metrics struct {
 	EvalDuration       time.Duration `json:"eval_duration,omitempty"`
 }
 
-// Options specfied in GenerateRequest, if you add a new option here add it to the API docs also
+// Options specified in GenerateRequest, if you add a new option here add it to the API docs also
 type Options struct {
 	Runner
 
@@ -121,7 +121,6 @@ type Runner struct {
 	VocabOnly          bool    `json:"vocab_only,omitempty"`
 	UseMMap            bool    `json:"use_mmap,omitempty"`
 	UseMLock           bool    `json:"use_mlock,omitempty"`
-	EmbeddingOnly      bool    `json:"embedding_only,omitempty"`
 	RopeFrequencyBase  float32 `json:"rope_frequency_base,omitempty"`
 	RopeFrequencyScale float32 `json:"rope_frequency_scale,omitempty"`
 	NumThread          int     `json:"num_thread,omitempty"`
@@ -395,7 +394,6 @@ func DefaultOptions() Options {
 			UseMLock:           false,
 			UseMMap:            true,
 			UseNUMA:            false,
-			EmbeddingOnly:      true,
 		},
 	}
 }

app/lifecycle/updater.go

@@ -34,20 +34,6 @@ type UpdateResponse struct {
 	UpdateVersion string `json:"version"`
 }
 
-func getClient(req *http.Request) http.Client {
-	proxyURL, err := http.ProxyFromEnvironment(req)
-	if err != nil {
-		slog.Warn(fmt.Sprintf("failed to handle proxy: %s", err))
-		return http.Client{}
-	}
-
-	return http.Client{
-		Transport: &http.Transport{
-			Proxy: http.ProxyURL(proxyURL),
-		},
-	}
-}
-
 func IsNewReleaseAvailable(ctx context.Context) (bool, UpdateResponse) {
 	var updateResp UpdateResponse
 
@@ -83,10 +69,9 @@ func IsNewReleaseAvailable(ctx context.Context) (bool, UpdateResponse) {
 	}
 	req.Header.Set("Authorization", signature)
 	req.Header.Set("User-Agent", fmt.Sprintf("ollama/%s (%s %s) Go/%s", version.Version, runtime.GOARCH, runtime.GOOS, runtime.Version()))
-	client := getClient(req)
 
 	slog.Debug("checking for available update", "requestURL", requestURL)
-	resp, err := client.Do(req)
+	resp, err := http.DefaultClient.Do(req)
 	if err != nil {
 		slog.Warn(fmt.Sprintf("failed to check for update: %s", err))
 		return false, updateResp
@@ -101,6 +86,11 @@ func IsNewReleaseAvailable(ctx context.Context) (bool, UpdateResponse) {
 	if err != nil {
 		slog.Warn(fmt.Sprintf("failed to read body response: %s", err))
 	}
+
+	if resp.StatusCode != 200 {
+		slog.Info(fmt.Sprintf("check update error %d - %.96s", resp.StatusCode, string(body)))
+		return false, updateResp
+	}
 	err = json.Unmarshal(body, &updateResp)
 	if err != nil {
 		slog.Warn(fmt.Sprintf("malformed response checking for update: %s", err))
@@ -119,8 +109,8 @@ func DownloadNewRelease(ctx context.Context, updateResp UpdateResponse) error {
 	if err != nil {
 		return err
 	}
-	client := getClient(req)
-	resp, err := client.Do(req)
+
+	resp, err := http.DefaultClient.Do(req)
 	if err != nil {
 		return fmt.Errorf("error checking update: %w", err)
 	}
@@ -151,7 +141,7 @@ func DownloadNewRelease(ctx context.Context, updateResp UpdateResponse) error {
 	cleanupOldDownloads()
 
 	req.Method = http.MethodGet
-	resp, err = client.Do(req)
+	resp, err = http.DefaultClient.Do(req)
 	if err != nil {
 		return fmt.Errorf("error checking update: %w", err)
 	}

app/ollama.iss

@@ -49,9 +49,6 @@ SetupLogging=yes
 CloseApplications=yes
 RestartApplications=no
 
-; Make sure they can at least download llama2 as a minimum
-ExtraDiskSpaceRequired=3826806784
-
 ; https://jrsoftware.org/ishelp/index.php?topic=setup_wizardimagefile
 WizardSmallImageFile=.\assets\setup.bmp
 
@@ -116,7 +113,8 @@ Filename: "{cmd}"; Parameters: "/c timeout 5"; Flags: runhidden
 Type: filesandordirs; Name: "{%TEMP}\ollama*"
 Type: filesandordirs; Name: "{%LOCALAPPDATA}\Ollama"
 Type: filesandordirs; Name: "{%LOCALAPPDATA}\Programs\Ollama"
-Type: filesandordirs; Name: "{%USERPROFILE}\.ollama"
+Type: filesandordirs; Name: "{%USERPROFILE}\.ollama\models"
+Type: filesandordirs; Name: "{%USERPROFILE}\.ollama\history"
 ; NOTE: if the user has a custom OLLAMA_MODELS it will be preserved
 
 [Messages]

cmd/cmd.go

@@ -718,39 +718,36 @@ func initializeKeypair() error {
 	_, err = os.Stat(privKeyPath)
 	if os.IsNotExist(err) {
 		fmt.Printf("Couldn't find '%s'. Generating new private key.\n", privKeyPath)
-		_, privKey, err := ed25519.GenerateKey(rand.Reader)
+		cryptoPublicKey, cryptoPrivateKey, err := ed25519.GenerateKey(rand.Reader)
 		if err != nil {
 			return err
 		}
 
-		privKeyBytes, err := format.OpenSSHPrivateKey(privKey, "")
+		privateKeyBytes, err := ssh.MarshalPrivateKey(cryptoPrivateKey, "")
 		if err != nil {
 			return err
 		}
 
-		err = os.MkdirAll(filepath.Dir(privKeyPath), 0o755)
-		if err != nil {
+		if err := os.MkdirAll(filepath.Dir(privKeyPath), 0o755); err != nil {
 			return fmt.Errorf("could not create directory %w", err)
 		}
 
-		err = os.WriteFile(privKeyPath, pem.EncodeToMemory(privKeyBytes), 0o600)
+		if err := os.WriteFile(privKeyPath, pem.EncodeToMemory(privateKeyBytes), 0o600); err != nil {
+			return err
+		}
+
+		sshPublicKey, err := ssh.NewPublicKey(cryptoPublicKey)
 		if err != nil {
 			return err
 		}
 
-		sshPrivateKey, err := ssh.NewSignerFromKey(privKey)
-		if err != nil {
+		publicKeyBytes := ssh.MarshalAuthorizedKey(sshPublicKey)
+
+		if err := os.WriteFile(pubKeyPath, publicKeyBytes, 0o644); err != nil {
 			return err
 		}
 
-		pubKeyData := ssh.MarshalAuthorizedKey(sshPrivateKey.PublicKey())
-
-		err = os.WriteFile(pubKeyPath, pubKeyData, 0o644)
-		if err != nil {
-			return err
-		}
-
-		fmt.Printf("Your new public key is: \n\n%s\n", string(pubKeyData))
+		fmt.Printf("Your new public key is: \n\n%s\n", publicKeyBytes)
 	}
 	return nil
 }

docs/api.md

@@ -321,7 +321,6 @@ curl http://localhost:11434/api/generate -d '{
     "vocab_only": false,
     "use_mmap": true,
     "use_mlock": false,
-    "embedding_only": false,
     "rope_frequency_base": 1.1,
     "rope_frequency_scale": 0.8,
     "num_thread": 8

docs/faq.md

@@ -113,9 +113,9 @@ If a different directory needs to be used, set the environment variable `OLLAMA_
 
 Refer to the section [above](#how-do-i-configure-ollama-server) for how to set environment variables on your platform.
 
-## Does Ollama send my prompts and answers back to Ollama.ai to use in any way?
+## Does Ollama send my prompts and answers back to ollama.com?
 
-No, Ollama runs entirely locally, and conversation data will never leave your machine.
+No. Ollama runs locally, and conversation data does not leave your machine.
 
 ## How can I use Ollama in Visual Studio Code?
 

docs/import.md

@@ -124,7 +124,10 @@ ollama run example "What is your favourite condiment?"
 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. Run `cat ~/.ollama/id_ed25519.pub` (or `type %USERPROFILE%\.ollama\id_ed25519.pub` on Windows) to view your Ollama public key. Copy this to the clipboard.
+2. Copy your Ollama public key:
+  - macOS: `cat ~/.ollama/id_ed25519.pub`
+  - 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:

docs/tutorials/langchainpy.md

@@ -42,12 +42,12 @@ text_splitter=RecursiveCharacterTextSplitter(chunk_size=500, chunk_overlap=0)
 all_splits = text_splitter.split_documents(data)
 ```
 
-It's split up, but we have to find the relevant splits and then submit those to the model. We can do this by creating embeddings and storing them in a vector database. We can use Ollama directly to instantiate an embedding model. We will use ChromaDB in this example for a vector database. `pip install GPT4All chromadb`
+It's split up, but we have to find the relevant splits and then submit those to the model. We can do this by creating embeddings and storing them in a vector database. We can use Ollama directly to instantiate an embedding model. We will use ChromaDB in this example for a vector database. `pip install chromadb`
 
 ```python
 from langchain.embeddings import OllamaEmbeddings
 from langchain.vectorstores import Chroma
-oembed = OllamaEmbeddings(base_url="http://localhost:11434", model="llama2")
+oembed = OllamaEmbeddings(base_url="http://localhost:11434", model="nomic-embed-text")
 vectorstore = Chroma.from_documents(documents=all_splits, embedding=oembed)
 ```
 
@@ -66,7 +66,7 @@ The next thing is to send the question and the relevant parts of the docs to the
 ```python
 from langchain.chains import RetrievalQA
 qachain=RetrievalQA.from_chain_type(ollama, retriever=vectorstore.as_retriever())
-qachain({"query": question})
+qachain.invoke({"query": question})
 ```
 
 The answer received from this chain was:

examples/python-chat-app/README.md

@@ -0,0 +1,21 @@
+# Ollama Chat App
+
+Build a Llama2 chat app using Streamlit and Ollama.
+
+## Running the Example
+
+1. Ensure you have the `llama2` model installed:
+
+   ```bash
+   ollama pull llama2
+   ```
+2. Install the Python Requirements.
+
+   ```bash
+   pip install -r requirements.txt
+   ```
+3. Run the example:
+
+   ```bash
+   python main.py
+   ```

format/openssh.go

@@ -1,102 +0,0 @@
-// Copyright 2012 The Go Authors. All rights reserved.
-// Use of this source code is governed by a BSD-style
-// license that can be found in the LICENSE file.
-
-// Code originally from https://go-review.googlesource.com/c/crypto/+/218620
-
-// TODO: replace with upstream once the above change is merged and released.
-
-package format
-
-import (
-	"crypto"
-	"crypto/ed25519"
-	"crypto/rand"
-	"encoding/binary"
-	"encoding/pem"
-	"fmt"
-
-	"golang.org/x/crypto/ssh"
-)
-
-const privateKeyAuthMagic = "openssh-key-v1\x00"
-
-type openSSHEncryptedPrivateKey struct {
-	CipherName string
-	KDFName    string
-	KDFOptions string
-	KeysCount  uint32
-	PubKey     []byte
-	KeyBlocks  []byte
-}
-
-type openSSHPrivateKey struct {
-	Check1  uint32
-	Check2  uint32
-	Keytype string
-	Rest    []byte `ssh:"rest"`
-}
-
-type openSSHEd25519PrivateKey struct {
-	Pub     []byte
-	Priv    []byte
-	Comment string
-	Pad     []byte `ssh:"rest"`
-}
-
-func OpenSSHPrivateKey(key crypto.PrivateKey, comment string) (*pem.Block, error) {
-	var check uint32
-	if err := binary.Read(rand.Reader, binary.BigEndian, &check); err != nil {
-		return nil, err
-	}
-
-	var pk1 openSSHPrivateKey
-	pk1.Check1 = check
-	pk1.Check2 = check
-
-	var w openSSHEncryptedPrivateKey
-	w.KeysCount = 1
-
-	if k, ok := key.(*ed25519.PrivateKey); ok {
-		key = *k
-	}
-
-	switch k := key.(type) {
-	case ed25519.PrivateKey:
-		pub, priv := k[32:], k
-		key := openSSHEd25519PrivateKey{
-			Pub:     pub,
-			Priv:    priv,
-			Comment: comment,
-		}
-
-		pk1.Keytype = ssh.KeyAlgoED25519
-		pk1.Rest = ssh.Marshal(key)
-
-		w.PubKey = ssh.Marshal(struct {
-			KeyType string
-			Pub     []byte
-		}{
-			ssh.KeyAlgoED25519, pub,
-		})
-	default:
-		return nil, fmt.Errorf("ssh: unknown key type %T", k)
-	}
-
-	w.KeyBlocks = openSSHPadding(ssh.Marshal(pk1), 8)
-
-	w.CipherName, w.KDFName, w.KDFOptions = "none", "none", ""
-
-	return &pem.Block{
-		Type:  "OPENSSH PRIVATE KEY",
-		Bytes: append([]byte(privateKeyAuthMagic), ssh.Marshal(w)...),
-	}, nil
-}
-
-func openSSHPadding(block []byte, blocksize int) []byte {
-	for i, j := 0, len(block); (j+i)%blocksize != 0; i++ {
-		block = append(block, byte(i+1))
-	}
-
-	return block
-}

go.mod

@@ -3,8 +3,10 @@ module github.com/jmorganca/ollama
 go 1.21
 
 require (
+	github.com/containerd/console v1.0.3
 	github.com/emirpasic/gods v1.18.1
 	github.com/gin-gonic/gin v1.9.1
+	github.com/google/uuid v1.0.0
 	github.com/olekukonko/tablewriter v0.0.5
 	github.com/spf13/cobra v1.7.0
 	github.com/stretchr/testify v1.8.4
@@ -12,26 +14,12 @@ require (
 )
 
 require (
-	github.com/containerd/console v1.0.3 // indirect
-	github.com/cratonica/2goarray v0.0.0-20190331194516-514510793eaa // indirect
 	github.com/davecgh/go-spew v1.1.1 // indirect
-	github.com/getlantern/context v0.0.0-20190109183933-c447772a6520 // indirect
-	github.com/getlantern/errors v0.0.0-20190325191628-abdb3e3e36f7 // indirect
-	github.com/getlantern/golog v0.0.0-20190830074920-4ef2e798c2d7 // indirect
-	github.com/getlantern/hex v0.0.0-20190417191902-c6586a6fe0b7 // indirect
-	github.com/getlantern/hidden v0.0.0-20190325191715-f02dbb02be55 // indirect
-	github.com/getlantern/ops v0.0.0-20190325191751-d70cb0d6f85f // indirect
-	github.com/getlantern/systray v1.2.2 // indirect
-	github.com/go-stack/stack v1.8.0 // indirect
-	github.com/google/uuid v1.0.0 // indirect
 	github.com/mattn/go-runewidth v0.0.14 // indirect
-	github.com/oxtoacart/bpool v0.0.0-20190530202638-03653db5a59c // indirect
-	github.com/pborman/uuid v1.2.1 // indirect
 	github.com/pmezard/go-difflib v1.0.0 // indirect
 	github.com/rivo/uniseg v0.2.0 // indirect
 )
 
-
 require (
 	github.com/bytedance/sonic v1.9.1 // indirect
 	github.com/chenzhuoyu/base64x v0.0.0-20221115062448-fe3a3abad311 // indirect

go.sum

@@ -7,8 +7,6 @@ github.com/chenzhuoyu/base64x v0.0.0-20221115062448-fe3a3abad311/go.mod h1:b583j
 github.com/containerd/console v1.0.3 h1:lIr7SlA5PxZyMV30bDW0MGbiOPXwc63yRuCP0ARubLw=
 github.com/containerd/console v1.0.3/go.mod h1:7LqA/THxQ86k76b8c/EMSiaJ3h1eZkMkXar0TQ1gf3U=
 github.com/cpuguy83/go-md2man/v2 v2.0.2/go.mod h1:tgQtvFlXSQOSOSIRvRPT7W67SCa46tRHOmNcaadrF8o=
-github.com/cratonica/2goarray v0.0.0-20190331194516-514510793eaa h1:Wg+722vs7a2zQH5lR9QWYsVbplKeffaQFIs5FTdfNNo=
-github.com/cratonica/2goarray v0.0.0-20190331194516-514510793eaa/go.mod h1:6Arca19mRx58CA7OWEd7Wu1NpC1rd3uDnNs6s1pj/DI=
 github.com/creack/pty v1.1.9/go.mod h1:oKZEueFk5CKHvIhNR5MUki03XCEU+Q6VDXinZuGJ33E=
 github.com/davecgh/go-spew v1.1.0/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=
 github.com/davecgh/go-spew v1.1.1 h1:vj9j/u1bqnvCEfJOwUhtlOARqs3+rkHYY13jYWTU97c=
@@ -17,20 +15,6 @@ github.com/emirpasic/gods v1.18.1 h1:FXtiHYKDGKCW2KzwZKx0iC0PQmdlorYgdFG9jPXJ1Bc
 github.com/emirpasic/gods v1.18.1/go.mod h1:8tpGGwCnJ5H4r6BWwaV6OrWmMoPhUl5jm/FMNAnJvWQ=
 github.com/gabriel-vasile/mimetype v1.4.2 h1:w5qFW6JKBz9Y393Y4q372O9A7cUSequkh1Q7OhCmWKU=
 github.com/gabriel-vasile/mimetype v1.4.2/go.mod h1:zApsH/mKG4w07erKIaJPFiX0Tsq9BFQgN3qGY5GnNgA=
-github.com/getlantern/context v0.0.0-20190109183933-c447772a6520 h1:NRUJuo3v3WGC/g5YiyF790gut6oQr5f3FBI88Wv0dx4=
-github.com/getlantern/context v0.0.0-20190109183933-c447772a6520/go.mod h1:L+mq6/vvYHKjCX2oez0CgEAJmbq1fbb/oNJIWQkBybY=
-github.com/getlantern/errors v0.0.0-20190325191628-abdb3e3e36f7 h1:6uJ+sZ/e03gkbqZ0kUG6mfKoqDb4XMAzMIwlajq19So=
-github.com/getlantern/errors v0.0.0-20190325191628-abdb3e3e36f7/go.mod h1:l+xpFBrCtDLpK9qNjxs+cHU6+BAdlBaxHqikB6Lku3A=
-github.com/getlantern/golog v0.0.0-20190830074920-4ef2e798c2d7 h1:guBYzEaLz0Vfc/jv0czrr2z7qyzTOGC9hiQ0VC+hKjk=
-github.com/getlantern/golog v0.0.0-20190830074920-4ef2e798c2d7/go.mod h1:zx/1xUUeYPy3Pcmet8OSXLbF47l+3y6hIPpyLWoR9oc=
-github.com/getlantern/hex v0.0.0-20190417191902-c6586a6fe0b7 h1:micT5vkcr9tOVk1FiH8SWKID8ultN44Z+yzd2y/Vyb0=
-github.com/getlantern/hex v0.0.0-20190417191902-c6586a6fe0b7/go.mod h1:dD3CgOrwlzca8ed61CsZouQS5h5jIzkK9ZWrTcf0s+o=
-github.com/getlantern/hidden v0.0.0-20190325191715-f02dbb02be55 h1:XYzSdCbkzOC0FDNrgJqGRo8PCMFOBFL9py72DRs7bmc=
-github.com/getlantern/hidden v0.0.0-20190325191715-f02dbb02be55/go.mod h1:6mmzY2kW1TOOrVy+r41Za2MxXM+hhqTtY3oBKd2AgFA=
-github.com/getlantern/ops v0.0.0-20190325191751-d70cb0d6f85f h1:wrYrQttPS8FHIRSlsrcuKazukx/xqO/PpLZzZXsF+EA=
-github.com/getlantern/ops v0.0.0-20190325191751-d70cb0d6f85f/go.mod h1:D5ao98qkA6pxftxoqzibIBBrLSUli+kYnJqrgBf9cIA=
-github.com/getlantern/systray v1.2.2 h1:dCEHtfmvkJG7HZ8lS/sLklTH4RKUcIsKrAD9sThoEBE=
-github.com/getlantern/systray v1.2.2/go.mod h1:pXFOI1wwqwYXEhLPm9ZGjS2u/vVELeIgNMY5HvhHhcE=
 github.com/gin-contrib/cors v1.4.0 h1:oJ6gwtUl3lqV0WEIwM/LxPF1QZ5qe2lGWdY2+bz7y0g=
 github.com/gin-contrib/cors v1.4.0/go.mod h1:bs9pNM0x/UsmHPBWT2xZz9ROh8xYjYkiURUfmBoMlcs=
 github.com/gin-contrib/sse v0.1.0 h1:Y/yl/+YNO8GZSjAhjMsSuLt29uWRFHdHYUb5lYOV9qE=
@@ -40,6 +24,7 @@ github.com/gin-gonic/gin v1.9.1 h1:4idEAncQnU5cB7BeOkPtxjfCSye0AAm1R0RVIqJ+Jmg=
 github.com/gin-gonic/gin v1.9.1/go.mod h1:hPrL7YrpYKXt5YId3A/Tnip5kqbEAP+KLuI3SUcPTeU=
 github.com/go-playground/assert/v2 v2.0.1/go.mod h1:VDjEfimB/XKnb+ZQfWdccd7VUvScMdVu0Titje2rxJ4=
 github.com/go-playground/assert/v2 v2.2.0 h1:JvknZsQTYeFEAhQwI4qEt9cyV5ONwRHC+lYKSsYSR8s=
+github.com/go-playground/assert/v2 v2.2.0/go.mod h1:VDjEfimB/XKnb+ZQfWdccd7VUvScMdVu0Titje2rxJ4=
 github.com/go-playground/locales v0.14.0/go.mod h1:sawfccIbzZTqEDETgFXqTho0QybSa7l++s0DH+LDiLs=
 github.com/go-playground/locales v0.14.1 h1:EWaQ/wswjilfKLTECiXz7Rh+3BjFhfDFKv/oXslEjJA=
 github.com/go-playground/locales v0.14.1/go.mod h1:hxrqLVvrK65+Rwrd5Fc6F2O76J/NuW9t0sjnWqG1slY=
@@ -49,8 +34,6 @@ github.com/go-playground/universal-translator v0.18.1/go.mod h1:xekY+UJKNuX9WP91
 github.com/go-playground/validator/v10 v10.10.0/go.mod h1:74x4gJWsvQexRdW8Pn3dXSGrTK4nAUsbPlLADvpJkos=
 github.com/go-playground/validator/v10 v10.14.0 h1:vgvQWe3XCz3gIeFDm/HnTIbj6UGmg/+t63MyGU2n5js=
 github.com/go-playground/validator/v10 v10.14.0/go.mod h1:9iXMNT7sEkjXb0I+enO7QXmzG6QCsPWY4zveKFVRSyU=
-github.com/go-stack/stack v1.8.0 h1:5SgMzNM5HxrEjV0ww2lTmX6E2Izsfxas4+YHWRs3Lsk=
-github.com/go-stack/stack v1.8.0/go.mod h1:v0f6uXyyMGvRgIKkXu+yp6POWl0qKG85gN/melR3HDY=
 github.com/goccy/go-json v0.9.7/go.mod h1:6MelG93GURQebXPDq3khkgXZkazVtN9CRI+MGFi0w8I=
 github.com/goccy/go-json v0.10.2 h1:CrxCmQqYDkv1z7lO7Wbh2HN93uovUHgrECaO5ZrCXAU=
 github.com/goccy/go-json v0.10.2/go.mod h1:6MelG93GURQebXPDq3khkgXZkazVtN9CRI+MGFi0w8I=
@@ -79,8 +62,6 @@ github.com/kr/text v0.2.0/go.mod h1:eLer722TekiGuMkidMxC/pM04lWEeraHUUmBw8l2grE=
 github.com/leodido/go-urn v1.2.1/go.mod h1:zt4jvISO2HfUBqxjfIshjdMTYS56ZS/qv49ictyFfxY=
 github.com/leodido/go-urn v1.2.4 h1:XlAE/cm/ms7TE/VMVoduSpNBoyc2dOxHs5MZSwAN63Q=
 github.com/leodido/go-urn v1.2.4/go.mod h1:7ZrI8mTSeBSHl/UaRyKQW1qZeMgak41ANeCNaVckg+4=
-github.com/lxn/walk v0.0.0-20210112085537-c389da54e794/go.mod h1:E23UucZGqpuUANJooIbHWCufXvOcT6E7Stq81gU+CSQ=
-github.com/lxn/win v0.0.0-20210218163916-a377121e959e/go.mod h1:KxxjdtRkfNoYDCUP5ryK7XJJNTnpC8atvtmTheChOtk=
 github.com/mattn/go-isatty v0.0.14/go.mod h1:7GGIvUiUoEMVVmxf/4nioHXj79iQHKdU27kJ6hsGG94=
 github.com/mattn/go-isatty v0.0.19 h1:JITubQf0MOLdlGRuRq+jtsDlekdYPia9ZFsB8h/APPA=
 github.com/mattn/go-isatty v0.0.19/go.mod h1:W+V8PltTTMOvKvAeJH7IuucS94S2C6jfK/D7dTCTo3Y=
@@ -94,12 +75,8 @@ github.com/modern-go/reflect2 v1.0.2 h1:xBagoLtFs94CBntxluKeaWgTMpvLxC4ur3nMaC9G
 github.com/modern-go/reflect2 v1.0.2/go.mod h1:yWuevngMOJpCy52FWWMvUC8ws7m/LJsjYzDa0/r8luk=
 github.com/olekukonko/tablewriter v0.0.5 h1:P2Ga83D34wi1o9J6Wh1mRuqd4mF/x/lgBS7N7AbDhec=
 github.com/olekukonko/tablewriter v0.0.5/go.mod h1:hPp6KlRPjbx+hW8ykQs1w3UBbZlj6HuIJcUGPhkA7kY=
-github.com/oxtoacart/bpool v0.0.0-20190530202638-03653db5a59c h1:rp5dCmg/yLR3mgFuSOe4oEnDDmGLROTvMragMUXpTQw=
-github.com/oxtoacart/bpool v0.0.0-20190530202638-03653db5a59c/go.mod h1:X07ZCGwUbLaax7L0S3Tw4hpejzu63ZrrQiUe6W0hcy0=
 github.com/pbnjay/memory v0.0.0-20210728143218-7b4eea64cf58 h1:onHthvaw9LFnH4t2DcNVpwGmV9E1BkGknEliJkfwQj0=
 github.com/pbnjay/memory v0.0.0-20210728143218-7b4eea64cf58/go.mod h1:DXv8WO4yhMYhSNPKjeNKa5WY9YCIEBRbNzFFPJbWO6Y=
-github.com/pborman/uuid v1.2.1 h1:+ZZIw58t/ozdjRaXh/3awHfmWRbzYxJoAdNJxe/3pvw=
-github.com/pborman/uuid v1.2.1/go.mod h1:X/NO0urCmaxf9VXbdlT7C2Yzkj2IKimNn4k+gtPdI/k=
 github.com/pelletier/go-toml/v2 v2.0.1/go.mod h1:r9LEWfGN8R5k0VXJ+0BkIe7MYkRdwZOjgMj2KwnJFUo=
 github.com/pelletier/go-toml/v2 v2.0.8 h1:0ctb6s9mE31h0/lhu+J6OPmVeDxJn+kYnJc2jZR9tGQ=
 github.com/pelletier/go-toml/v2 v2.0.8/go.mod h1:vuYfssBdrU2XDZ9bYydBu6t+6a6PYNcZljzZR9VXg+4=
@@ -112,7 +89,6 @@ github.com/rogpeppe/go-internal v1.6.1/go.mod h1:xXDCJY+GAPziupqXw64V24skbSoqbTE
 github.com/rogpeppe/go-internal v1.8.0 h1:FCbCCtXNOY3UtUuHUYaghJg4y7Fd14rXifAYUAtL9R8=
 github.com/rogpeppe/go-internal v1.8.0/go.mod h1:WmiCO8CzOY8rg0OYDC4/i/2WRWAB6poM+XZ2dLUbcbE=
 github.com/russross/blackfriday/v2 v2.1.0/go.mod h1:+Rmxgy9KzJVeS9/2gXHxylqXiyQDYRxCVz55jmeOWTM=
-github.com/skratchdot/open-golang v0.0.0-20200116055534-eef842397966/go.mod h1:sUM3LWHvSMaG192sy56D9F7CNvL7jUJVXoqM1QKLnog=
 github.com/spf13/cobra v1.7.0 h1:hyqWnYt1ZQShIddO5kBpj3vu05/++x6tJ6dg8EC572I=
 github.com/spf13/cobra v1.7.0/go.mod h1:uLxZILRyS/50WlhOIKD7W6V5bgeIt+4sICxh6uRMrb0=
 github.com/spf13/pflag v1.0.5 h1:iy+VFUOCP1a+8yFto/drg2CJ5u0yRoB7fZw3DKv/JXA=
@@ -149,14 +125,12 @@ golang.org/x/net v0.17.0 h1:pVaXccu2ozPjCXewfr1S7xza/zcXTity9cCdXQYSjIM=
 golang.org/x/net v0.17.0/go.mod h1:NxSsAGuq816PNPmqtQdLE42eU2Fs7NoRIZrHJAlaCOE=
 golang.org/x/sync v0.3.0 h1:ftCYgMx6zT/asHUrPw8BLLscYtGznsLAnjq5RH9P66E=
 golang.org/x/sync v0.3.0/go.mod h1:FU7BRWz2tNW+3quACPkgCx/L+uEAv1htQ0V83Z9Rj+Y=
-golang.org/x/sys v0.0.0-20201018230417-eeed37f84f13/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs=
 golang.org/x/sys v0.0.0-20201119102817-f84b799fce68/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs=
 golang.org/x/sys v0.0.0-20210124154548-22da62e12c0c/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs=
 golang.org/x/sys v0.0.0-20210615035016-665e8c7367d1/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
 golang.org/x/sys v0.0.0-20210630005230-0f9fa26af87c/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
 golang.org/x/sys v0.0.0-20210806184541-e5e7981a1069/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
 golang.org/x/sys v0.0.0-20220704084225-05e143d24a9e/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
-golang.org/x/sys v0.1.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
 golang.org/x/sys v0.6.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
 golang.org/x/sys v0.13.0 h1:Af8nKPmuFypiUBjVoU9V20FiaFXOcuZI21p0ycVYYGE=
 golang.org/x/sys v0.13.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
@@ -173,7 +147,6 @@ google.golang.org/protobuf v1.26.0-rc.1/go.mod h1:jlhhOSvTdKEhbULTjvd4ARK9grFBp0
 google.golang.org/protobuf v1.28.0/go.mod h1:HV8QOd/L58Z+nl8r43ehVNZIU/HEI6OcFqwMG9pJV4I=
 google.golang.org/protobuf v1.30.0 h1:kPPoIgf3TsEvrm0PFe15JQ+570QVxYzEvvHqChK+cng=
 google.golang.org/protobuf v1.30.0/go.mod h1:HV8QOd/L58Z+nl8r43ehVNZIU/HEI6OcFqwMG9pJV4I=
-gopkg.in/Knetic/govaluate.v3 v3.0.0/go.mod h1:csKLBORsPbafmSCGTEh3U7Ozmsuq8ZSIlKk1bcqph0E=
 gopkg.in/check.v1 v0.0.0-20161208181325-20d25e280405/go.mod h1:Co6ibVJAznAaIkqp8huTwlJQCZ016jof/cbN4VW5Yz0=
 gopkg.in/check.v1 v1.0.0-20180628173108-788fd7840127/go.mod h1:Co6ibVJAznAaIkqp8huTwlJQCZ016jof/cbN4VW5Yz0=
 gopkg.in/check.v1 v1.0.0-20201130134442-10cb98267c6c h1:Hei/4ADfdWqJk1ZMxUNpqntNwaWcugrBjAiHlqqRiVk=

gpu/gpu_darwin.go

@@ -1,12 +1,14 @@
 //go:build darwin
 
 package gpu
-
+/*
+#cgo CFLAGS: -x objective-c
+#cgo LDFLAGS: -framework Foundation -framework CoreGraphics -framework Metal
+#include "gpu_info_darwin.h"
+*/
 import "C"
 import (
 	"runtime"
-
-	"github.com/pbnjay/memory"
 )
 
 // CheckVRAM returns the free VRAM in bytes on Linux machines with NVIDIA GPUs
@@ -15,19 +17,8 @@ func CheckVRAM() (int64, error) {
 		// gpu not supported, this may not be metal
 		return 0, nil
 	}
-
-	// on macOS, there's already buffer for available vram (see below) so just return the total
-	systemMemory := int64(memory.TotalMemory())
-
-	// macOS limits how much memory is available to the GPU based on the amount of system memory
-	// TODO: handle case where iogpu.wired_limit_mb is set to a higher value
-	if systemMemory <= 36*1024*1024*1024 {
-		systemMemory = systemMemory * 2 / 3
-	} else {
-		systemMemory = systemMemory * 3 / 4
-	}
-
-	return systemMemory, nil
+	recommendedMaxVRAM := int64(C.getRecommendedMaxVRAM())
+	return recommendedMaxVRAM, nil
 }
 
 func GetGPUInfo() GpuInfo {

gpu/gpu_info_cuda.c

@@ -156,7 +156,7 @@ void cuda_check_vram(cuda_handle_t h, mem_info_t *resp) {
     }
 
     LOG(h.verbose, "[%d] CUDA totalMem %ld\n", i, memInfo.total);
-    LOG(h.verbose, "[%d] CUDA usedMem %ld\n", i, memInfo.free);
+    LOG(h.verbose, "[%d] CUDA usedMem %ld\n", i, memInfo.used);
 
     resp->total += memInfo.total;
     resp->free += memInfo.free;

gpu/gpu_info_darwin.h

@@ -0,0 +1,3 @@
+#import <Metal/Metal.h>
+#include <stdint.h>
+uint64_t getRecommendedMaxVRAM();

gpu/gpu_info_darwin.m

@@ -0,0 +1,11 @@
+//go:build darwin
+#include "gpu_info_darwin.h"
+
+uint64_t getRecommendedMaxVRAM()
+{
+	id<MTLDevice> device = MTLCreateSystemDefaultDevice();
+	uint64_t result = device.recommendedMaxWorkingSetSize;
+	CFRelease(device);
+	return result;
+}
+

llm/dyn_ext_server.go

@@ -106,7 +106,12 @@ func newDynExtServer(library, model string, adapters, projectors []string, opts
 	sparams.memory_f16 = C.bool(opts.F16KV)
 	sparams.use_mlock = C.bool(opts.UseMLock)
 	sparams.use_mmap = C.bool(opts.UseMMap)
-	sparams.numa = C.bool(opts.UseNUMA)
+
+	if opts.UseNUMA {
+		sparams.numa = C.int(1)
+	} else {
+		sparams.numa = C.int(0)
+	}
 
 	sparams.lora_adapters = nil
 	for i := 0; i < len(adapters); i++ {
@@ -194,7 +199,6 @@ func (llm *dynExtServer) Predict(ctx context.Context, predict PredictOpts, fn fu
 		request["grammar"] = jsonGrammar
 	}
 
-	var whitespace int
 	retryDelay := 100 * time.Microsecond
 	for retries := 0; retries < maxRetries; retries++ {
 		if retries > 0 {
@@ -253,24 +257,6 @@ func (llm *dynExtServer) Predict(ctx context.Context, predict PredictOpts, fn fu
 					break out
 				}
 
-				// detect if p.Content is entirely whitespace
-				if predict.Format == "json" && strings.TrimSpace(p.Content) == "" {
-					whitespace++
-
-					// if we get 100 consecutive whitespace responses, cancel
-					if whitespace > 100 {
-						slog.Debug("cancelling due to excessive whitespace")
-						C.dyn_llama_server_completion_cancel(llm.s, resp.id, &resp)
-						if resp.id < 0 {
-							return extServerResponseToErr(resp)
-						}
-
-						return nil
-					}
-				} else {
-					whitespace = 0
-				}
-
 				if p.Content != "" {
 					fn(PredictResult{
 						Content: p.Content,

llm/ext_server/ext_server.cpp

@@ -80,7 +80,7 @@ void llama_server_init(ext_server_params *sparams, ext_server_resp_t *err) {
     params.main_gpu = sparams->main_gpu;
     params.use_mlock = sparams->use_mlock;
     params.use_mmap = sparams->use_mmap;
-    params.numa = sparams->numa;
+    params.numa = (ggml_numa_strategy)sparams->numa;
     params.embedding = sparams->embedding;
     if (sparams->model != NULL) {
       params.model = sparams->model;
@@ -111,7 +111,8 @@ void llama_server_init(ext_server_params *sparams, ext_server_resp_t *err) {
     }
 #endif
 
-    llama_backend_init(params.numa);
+    llama_backend_init();
+    llama_numa_init(params.numa);
 
     // load the model
     if (!llama->load_model(params)) {
@@ -145,9 +146,9 @@ void llama_server_start() {
       llama->queue_tasks.on_new_task(std::bind(
         &llama_server_context::process_single_task, llama, std::placeholders::_1));
       llama->queue_tasks.on_finish_multitask(std::bind(
-          &llama_server_context::on_finish_multitask, llama, std::placeholders::_1));
-      llama->queue_tasks.on_all_tasks_finished(std::bind(
-          &llama_server_context::run_on_all_tasks_finished, llama));
+        &llama_server_context::on_finish_multitask, llama, std::placeholders::_1));
+      llama->queue_tasks.on_run_slots(std::bind(
+        &llama_server_context::update_slots, llama));
       llama->queue_results.on_multitask_update(std::bind(
           &llama_server_queue::update_multitask,
           &llama->queue_tasks,
@@ -208,7 +209,6 @@ void llama_server_completion(const char *json_req, ext_server_resp_t *resp) {
 void llama_server_completion_next_result(const int task_id,
                                          ext_server_task_result_t *resp) {
   assert(llama != NULL && resp != NULL);
-  std::string msg;
   resp->id = -1;
   resp->stop = false;
   resp->error = false;

llm/ext_server/ext_server.h

@@ -41,7 +41,7 @@ typedef struct ext_server_params {
   int32_t main_gpu;      // the GPU that is used for scratch and small tensors
   bool use_mlock;        // force system to keep model in RAM
   bool use_mmap;         // use mmap if possible
-  bool numa;             // attempt optimizations that help on some NUMA systems
+  int numa;              // attempt optimizations that help on some NUMA systems
   bool embedding;        // get only sentence embedding
   ext_server_lora_adapter_t *lora_adapters;
   char *mmproj;

llm/generate/gen_common.sh

@@ -101,7 +101,7 @@ compress_libs() {
     pids=""
     rm -rf ${BUILD_DIR}/lib/*.${LIB_EXT}*.gz
     for lib in ${BUILD_DIR}/lib/*.${LIB_EXT}* ; do
-        gzip --best -f ${lib} &
+        gzip -n --best -f ${lib} &
         pids+=" $!"
     done
     echo 

llm/generate/gen_windows.ps1

@@ -154,8 +154,9 @@ apply_patches
 # -DLLAMA_AVX2 -- 2013 Intel Haswell & 2015 AMD Excavator / 2017 AMD Zen
 # -DLLAMA_FMA (FMA3) -- 2013 Intel Haswell & 2012 AMD Piledriver
 
-$script:commonCpuDefs = @("-DCMAKE_POSITION_INDEPENDENT_CODE=on", "-DLLAMA_NATIVE=off")
+$script:commonCpuDefs = @("-DCMAKE_POSITION_INDEPENDENT_CODE=on")
 
+init_vars
 $script:cmakeDefs = $script:commonCpuDefs + @("-DLLAMA_AVX=off", "-DLLAMA_AVX2=off", "-DLLAMA_AVX512=off", "-DLLAMA_FMA=off", "-DLLAMA_F16C=off") + $script:cmakeDefs
 $script:buildDir="${script:llamacppDir}/build/windows/${script:ARCH}/cpu"
 write-host "Building LCD CPU"
@@ -164,6 +165,7 @@ install
 sign
 compress_libs
 
+init_vars
 $script:cmakeDefs = $script:commonCpuDefs + @("-DLLAMA_AVX=on", "-DLLAMA_AVX2=off", "-DLLAMA_AVX512=off", "-DLLAMA_FMA=off", "-DLLAMA_F16C=off") + $script:cmakeDefs
 $script:buildDir="${script:llamacppDir}/build/windows/${script:ARCH}/cpu_avx"
 write-host "Building AVX CPU"
@@ -172,6 +174,7 @@ install
 sign
 compress_libs
 
+init_vars
 $script:cmakeDefs = $script:commonCpuDefs + @("-DLLAMA_AVX=on", "-DLLAMA_AVX2=on", "-DLLAMA_AVX512=off", "-DLLAMA_FMA=on", "-DLLAMA_F16C=on") + $script:cmakeDefs
 $script:buildDir="${script:llamacppDir}/build/windows/${script:ARCH}/cpu_avx2"
 write-host "Building AVX2 CPU"

llm/ggml.go

@@ -31,6 +31,11 @@ const (
 	fileTypeQ5_K_S
 	fileTypeQ5_K_M
 	fileTypeQ6_K
+	fileTypeIQ2_XXS
+	fileTypeIQ2_XS
+	fileTypeQ2_K_S
+	fileTypeQ3_K_XS
+	fileTypeIQ3_XXS
 )
 
 func fileType(fileType uint32) string {
@@ -69,6 +74,16 @@ func fileType(fileType uint32) string {
 		return "Q5_K_M"
 	case fileTypeQ6_K:
 		return "Q6_K"
+	case fileTypeIQ2_XXS:
+		return "IQ2_XXS"
+	case fileTypeIQ2_XS:
+		return "IQ2_XS"
+	case fileTypeQ2_K_S:
+		return "Q2_K_S"
+	case fileTypeQ3_K_XS:
+		return "Q3_K_XS"
+	case fileTypeIQ3_XXS:
+		return "IQ3_XXS"
 	default:
 		return "unknown"
 	}

llm/gguf.go

@@ -115,6 +115,14 @@ func (t tensor) typeSize() uint64 {
 		return 2 + 2 + 12 + blockSize/8 + blockSize/2
 	case 14: // Q6_K
 		return blockSize/2 + blockSize/4 + blockSize/16 + 2
+	case 15: // Q8_K
+		return 2 + blockSize + 2*blockSize/16
+	case 16: // IQ2_XXS
+		return 2 + 2*blockSize/8
+	case 17: // IQ2_XS
+		return 2 + 2*blockSize/8 + blockSize/32
+	case 18: // IQ3_XXS
+		return 2 + 3*blockSize/8
 	default:
 		return 0
 	}

llm/patches/01-cache.diff

@@ -1,9 +1,9 @@
 diff --git a/examples/server/server.cpp b/examples/server/server.cpp
-index d86d7e04..2694e92e 100644
+index 2b2f4a0f..afac49af 100644
 --- a/examples/server/server.cpp
 +++ b/examples/server/server.cpp
-@@ -901,13 +901,15 @@ struct llama_server_context
-                 slot.sent_count += result.text_to_send.size();
+@@ -997,13 +997,15 @@ struct llama_server_context
+                 slot.n_sent_text += result.text_to_send.size();
                  // add the token to slot queue and cache
              }
 -            slot.add_token_string(result);

llm/patches/02-cudaleaks.diff

@@ -1,30 +1,29 @@
 diff --git a/examples/server/server.cpp b/examples/server/server.cpp
-index 3102762c..568ac1d0 100644
+index 2b2f4a0f..25857bdd 100644
 --- a/examples/server/server.cpp
 +++ b/examples/server/server.cpp
-@@ -307,6 +307,10 @@ struct llama_client_slot
-     }
- };
+@@ -31,6 +31,10 @@
+ #include <atomic>
+ #include <signal.h>
  
 +#ifdef GGML_USE_CUBLAS
 +extern "C" GGML_CALL void ggml_free_cublas(void);
 +#endif
 +
- struct llama_server_context
- {
-     llama_model *model = nullptr;
-@@ -353,6 +357,10 @@ struct llama_server_context
+ using json = nlohmann::json;
+ 
+ struct server_params {
+@@ -363,6 +367,9 @@ struct llama_server_context
              llama_free_model(model);
              model = nullptr;
          }
 +#ifdef GGML_USE_CUBLAS
 +        ggml_free_cublas();
 +#endif
-+
      }
  
      bool load_model(const gpt_params &params_)
-@@ -3093,6 +3101,7 @@ int main(int argc, char **argv)
+@@ -3494,6 +3501,7 @@ int main(int argc, char **argv)
      sigemptyset (&sigint_action.sa_mask);
      sigint_action.sa_flags = 0;
      sigaction(SIGINT, &sigint_action, NULL);
@@ -32,16 +31,11 @@ index 3102762c..568ac1d0 100644
  #elif defined (_WIN32)
      auto console_ctrl_handler = +[](DWORD ctrl_type) -> BOOL {
          return (ctrl_type == CTRL_C_EVENT) ? (signal_handler(SIGINT), true) : false;
-@@ -3106,3 +3115,4 @@ int main(int argc, char **argv)
-     llama_backend_free();
-     return 0;
- }
-+
 diff --git a/ggml-cuda.cu b/ggml-cuda.cu
-index 96976f24..3543920e 100644
+index 0c6501e9..75c12723 100644
 --- a/ggml-cuda.cu
 +++ b/ggml-cuda.cu
-@@ -39,6 +39,7 @@
+@@ -43,6 +43,7 @@
  #define __shfl_xor_sync(mask, var, laneMask, width) __shfl_xor(var, laneMask, width)
  #define cublasComputeType_t hipblasDatatype_t //deprecated, new hipblasComputeType_t not in 5.6
  #define cublasCreate hipblasCreate
@@ -49,30 +43,30 @@ index 96976f24..3543920e 100644
  #define cublasGemmEx hipblasGemmEx
  #define cublasGemmBatchedEx hipblasGemmBatchedEx
  #define cublasGemmStridedBatchedEx hipblasGemmStridedBatchedEx
-@@ -7928,10 +7929,11 @@ GGML_CALL bool ggml_cublas_loaded(void) {
+@@ -8694,10 +8695,10 @@ GGML_CALL bool ggml_cublas_loaded(void) {
      return g_cublas_loaded;
  }
  
-+static bool g_cublas_initialized = false;
-+
- GGML_CALL void ggml_init_cublas() {
+-GGML_CALL void ggml_init_cublas() {
 -    static bool initialized = false;
++static bool g_cublas_initialized = false;
  
 -    if (!initialized) {
++GGML_CALL void ggml_init_cublas() {
 +    if (!g_cublas_initialized) {
  
  #ifdef __HIP_PLATFORM_AMD__
          // Workaround for a rocBLAS bug when using multiple graphics cards:
-@@ -7941,7 +7943,7 @@ GGML_CALL void ggml_init_cublas() {
+@@ -8707,7 +8708,7 @@ GGML_CALL void ggml_init_cublas() {
  #endif
  
          if (cudaGetDeviceCount(&g_device_count) != cudaSuccess) {
 -            initialized = true;
 +            g_cublas_initialized = true;
              g_cublas_loaded = false;
+             fprintf(stderr, "%s: no " GGML_CUDA_NAME " devices found, " GGML_CUDA_NAME " will be disabled\n", __func__);
              return;
-         }
-@@ -8011,7 +8013,7 @@ GGML_CALL void ggml_init_cublas() {
+@@ -8778,7 +8779,7 @@ GGML_CALL void ggml_init_cublas() {
          // configure logging to stdout
          // CUBLAS_CHECK(cublasLoggerConfigure(1, 1, 0, nullptr));
  
@@ -81,7 +75,7 @@ index 96976f24..3543920e 100644
          g_cublas_loaded = true;
      }
  }
-@@ -11528,3 +11530,17 @@ GGML_CALL int ggml_backend_cuda_reg_devices() {
+@@ -12345,3 +12346,22 @@ GGML_CALL int ggml_backend_cuda_reg_devices() {
      }
      return device_count;
  }
@@ -89,28 +83,32 @@ index 96976f24..3543920e 100644
 +extern "C" GGML_CALL void ggml_free_cublas(void);
 +GGML_CALL void ggml_free_cublas(void) {
 +    for (int id = 0; id < g_device_count; ++id) {
-+#if !defined(GGML_USE_HIPBLAS)
-+        CU_CHECK(cuMemUnmap(g_cuda_pool_addr[id], g_cuda_pool_size[id]));
-+        g_cuda_pool_size[id] = 0;
-+        g_cuda_pool_addr[id] = 0;
++#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
++        if (g_device_caps[id].vmm) {
++            CU_CHECK(cuMemUnmap(g_cuda_pool_addr[id], g_cuda_pool_size[id]));
++            g_cuda_pool_size[id] = 0;
++            g_cuda_pool_addr[id] = 0;
++        }
 +#endif
++        // TODO: free legacy non-vmm memory
++        // destroy cublas handle
 +        CUBLAS_CHECK(cublasDestroy(g_cublas_handles[id]));
 +        g_cublas_handles[id] = nullptr;
 +    }
++
 +    g_cublas_initialized = false;
 +}
-\ No newline at end of file
 diff --git a/ggml-cuda.h b/ggml-cuda.h
-index b1ebd61d..b4c80c2c 100644
+index b1ebd61d..6dd58ddf 100644
 --- a/ggml-cuda.h
 +++ b/ggml-cuda.h
-@@ -20,6 +20,9 @@ extern "C" {
- // Always success. To check if CUDA is actually loaded, use `ggml_cublas_loaded`.
- GGML_API GGML_CALL void   ggml_init_cublas(void);
+@@ -23,6 +23,9 @@ GGML_API GGML_CALL void   ggml_init_cublas(void);
+ // Returns `true` if there are available CUDA devices and cublas loads successfully; otherwise, it returns `false`.
+ GGML_API GGML_CALL bool   ggml_cublas_loaded(void);
  
 +// Release CUDA resources
 +GGML_API GGML_CALL void   ggml_free_cublas(void);
 +
- // Returns `true` if there are available CUDA devices and cublas loads successfully; otherwise, it returns `false`.
- GGML_API GGML_CALL bool   ggml_cublas_loaded(void);
+ GGML_API GGML_CALL void * ggml_cuda_host_malloc(size_t size);
+ GGML_API GGML_CALL void   ggml_cuda_host_free(void * ptr);
  

llm/patches/02-shutdown.diff

@@ -1,96 +0,0 @@
-diff --git a/examples/server/server.cpp b/examples/server/server.cpp
-index a0b46970..7800c6e7 100644
---- a/examples/server/server.cpp
-+++ b/examples/server/server.cpp
-@@ -28,6 +28,7 @@
- #include <chrono>
- #include <condition_variable>
- #include <atomic>
-+#include <signal.h>
- 
- using json = nlohmann::json;
- 
-@@ -2511,6 +2512,9 @@ static void append_to_generated_text_from_generated_token_probs(llama_server_con
-     }
- }
- 
-+std::function<void(int)> shutdown_handler;
-+inline void signal_handler(int signal) { shutdown_handler(signal); }
-+
- int main(int argc, char **argv)
- {
- #if SERVER_VERBOSE != 1
-@@ -3128,8 +3132,25 @@ int main(int argc, char **argv)
-         std::placeholders::_2,
-         std::placeholders::_3
-     ));
--    llama.queue_tasks.start_loop();
- 
-+    shutdown_handler = [&](int) {
-+        llama.queue_tasks.terminate();
-+    };
-+
-+#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
-+    struct sigaction sigint_action;
-+    sigint_action.sa_handler = signal_handler;
-+    sigemptyset (&sigint_action.sa_mask);
-+    sigint_action.sa_flags = 0;
-+    sigaction(SIGINT, &sigint_action, NULL);
-+#elif defined (_WIN32)
-+    auto console_ctrl_handler = +[](DWORD ctrl_type) -> BOOL {
-+        return (ctrl_type == CTRL_C_EVENT) ? (signal_handler(SIGINT), true) : false;
-+    };
-+    SetConsoleCtrlHandler(reinterpret_cast<PHANDLER_ROUTINE>(console_ctrl_handler), true);
-+#endif
-+    llama.queue_tasks.start_loop();
-+    svr.stop();
-     t.join();
- 
-     llama_backend_free();
-diff --git a/examples/server/utils.hpp b/examples/server/utils.hpp
-index 54854896..0ee670db 100644
---- a/examples/server/utils.hpp
-+++ b/examples/server/utils.hpp
-@@ -220,6 +220,7 @@ inline std::string format_chatml(std::vector<json> messages)
- struct llama_server_queue {
-     int id = 0;
-     std::mutex mutex_tasks;
-+    bool running;
-     // queues
-     std::vector<task_server> queue_tasks;
-     std::vector<task_server> queue_tasks_deferred;
-@@ -278,9 +279,18 @@ struct llama_server_queue {
-         queue_tasks_deferred.clear();
-     }
- 
--    // Start the main loop. This call is blocking
--    [[noreturn]]
-+    // end the start_loop routine
-+    void terminate() {
-+        {
-+            std::unique_lock<std::mutex> lock(mutex_tasks);
-+            running = false;
-+        }
-+        condition_tasks.notify_all();
-+    }
-+
-+    // Start the main loop.
-     void start_loop() {
-+        running = true;
-         while (true) {
-             // new task arrived
-             LOG_VERBOSE("have new task", {});
-@@ -324,8 +334,12 @@ struct llama_server_queue {
-             {
-                 std::unique_lock<std::mutex> lock(mutex_tasks);
-                 if (queue_tasks.empty()) {
-+                    if (!running) {
-+                        LOG_VERBOSE("ending start_loop", {});
-+                        return;
-+                    }
-                     condition_tasks.wait(lock, [&]{
--                        return !queue_tasks.empty();
-+                        return (!queue_tasks.empty() || !running);
-                     });
-                 }
-             }

scripts/build_docker.sh

@@ -5,13 +5,15 @@ set -eu
 export VERSION=${VERSION:-$(git describe --tags --first-parent --abbrev=7 --long --dirty --always | sed -e "s/^v//g")}
 export GOFLAGS="'-ldflags=-w -s \"-X=github.com/jmorganca/ollama/version.Version=$VERSION\" \"-X=github.com/jmorganca/ollama/server.mode=release\"'"
 
+IMAGE_NAME=${IMAGE_NAME:-"ollama/ollama"}
+BUILD_PLATFORM=${BUILD_PLATFORM:-"linux/arm64,linux/amd64"}
 docker build \
     --load \
-    --platform=linux/arm64,linux/amd64 \
+    --platform=${BUILD_PLATFORM} \
     --build-arg=VERSION \
     --build-arg=GOFLAGS \
     -f Dockerfile \
-    -t ollama/ollama:$VERSION \
+    -t ${IMAGE_NAME}:$VERSION \
     .
 
 docker build \
@@ -21,5 +23,12 @@ docker build \
     --build-arg=GOFLAGS \
     --target runtime-rocm \
     -f Dockerfile \
-    -t ollama/ollama:$VERSION-rocm \
+    -t ${IMAGE_NAME}:$VERSION-rocm \
     .
+
+docker tag ${IMAGE_NAME}:$VERSION ${IMAGE_NAME}:latest
+docker tag ${IMAGE_NAME}:$VERSION-rocm ${IMAGE_NAME}:rocm
+
+echo "To release, run:"
+echo "  docker push ${IMAGE_NAME}:$VERSION && docker push ${IMAGE_NAME}:latest"
+echo "  docker push ${IMAGE_NAME}:$VERSION-rocm && docker push ${IMAGE_NAME}:rocm"

scripts/build_windows.ps1

@@ -53,13 +53,14 @@ function buildOllama() {
     write-host "Building ollama CLI"
     & go generate ./...
     if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
-    & go build "-ldflags=-w -s ""-X=github.com/jmorganca/ollama/version.Version=$script:VERSION"" ""-X=github.com/jmorganca/ollama/server.mode=release""" .
+    & go build -ldflags "-s -w -X=github.com/jmorganca/ollama/version.Version=$script:VERSION -X=github.com/jmorganca/ollama/server.mode=release" .
     if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
     if ("${env:KEY_CONTAINER}") {
         & "${script:SignTool}" sign /v /fd sha256 /t http://timestamp.digicert.com /f "${script:OLLAMA_CERT}" `
             /csp "Google Cloud KMS Provider" /kc ${env:KEY_CONTAINER} ollama.exe
         if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
     }
+    New-Item -ItemType Directory -Path .\dist -Force
     cp .\ollama.exe .\dist\ollama-windows-amd64.exe
 }
 
@@ -67,7 +68,7 @@ function buildApp() {
     write-host "Building Ollama App"
     cd "${script:SRC_DIR}\app"
     & windres -l 0 -o ollama.syso ollama.rc
-    & go build "-ldflags=-H windowsgui -w -s ""-X=github.com/jmorganca/ollama/version.Version=$script:VERSION"" ""-X=github.com/jmorganca/ollama/server.mode=release""" .
+    & go build -ldflags "-s -w -H windowsgui -X=github.com/jmorganca/ollama/version.Version=$script:VERSION -X=github.com/jmorganca/ollama/server.mode=release" .
     if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
     if ("${env:KEY_CONTAINER}") {
         & "${script:SignTool}" sign /v /fd sha256 /t http://timestamp.digicert.com /f "${script:OLLAMA_CERT}" `
@@ -129,4 +130,4 @@ try {
 } finally {
     set-location $script:SRC_DIR
     $env:PKG_VERSION=""
-}
+}

scripts/install.sh

@@ -72,7 +72,7 @@ $SUDO install -o0 -g0 -m755 -d $BINDIR
 $SUDO install -o0 -g0 -m755 $TEMP_DIR/ollama $BINDIR/ollama
 
 install_success() { 
-    status 'The Ollama API is now available at 0.0.0.0:11434.'
+    status 'The Ollama API is now available at 127.0.0.1:11434.'
     status 'Install complete. Run "ollama" from the command line.'
 }
 trap install_success EXIT
@@ -88,6 +88,10 @@ configure_systemd() {
         status "Adding ollama user to render group..."
         $SUDO usermod -a -G render ollama
     fi
+    if getent group video >/dev/null 2>&1; then
+        status "Adding ollama user to video group..."
+        $SUDO usermod -a -G video ollama
+    fi
 
     status "Adding current user to ollama group..."
     $SUDO usermod -a -G ollama $(whoami)

server/images.go

@@ -52,6 +52,10 @@ type Model struct {
 	Messages       []Message
 }
 
+func (m *Model) IsEmbedding() bool {
+	return slices.Contains(m.Config.ModelFamilies, "bert") || slices.Contains(m.Config.ModelFamilies, "nomic-bert")
+}
+
 type Message struct {
 	Role    string `json:"role"`
 	Content string `json:"content"`
@@ -1103,18 +1107,7 @@ func makeRequest(ctx context.Context, method string, requestURL *url.URL, header
 		req.ContentLength = contentLength
 	}
 
-	proxyURL, err := http.ProxyFromEnvironment(req)
-	if err != nil {
-		return nil, err
-	}
-
-	client := http.Client{
-		Transport: &http.Transport{
-			Proxy: http.ProxyURL(proxyURL),
-		},
-	}
-
-	resp, err := client.Do(req)
+	resp, err := http.DefaultClient.Do(req)
 	if err != nil {
 		return nil, err
 	}

server/prompt.go

@@ -121,13 +121,15 @@ func ChatPrompt(tmpl string, messages []api.Message, window int, encode func(str
 				p = prompt{}
 			}
 
-			p.Prompt = msg.Content
-
+			var sb strings.Builder
 			for range msg.Images {
-				p.Prompt += fmt.Sprintf(" [img-%d]", imgId)
+				fmt.Fprintf(&sb, "[img-%d] ", imgId)
 				p.images = append(p.images, imgId)
 				imgId += 1
 			}
+
+			sb.WriteString(msg.Content)
+			p.Prompt = sb.String()
 		case "assistant":
 			if p.Response != "" {
 				prompts = append(prompts, p)

server/prompt_test.go

@@ -155,7 +155,7 @@ func TestChatPrompt(t *testing.T) {
 				{Role: "user", Content: "Hello", Images: []api.ImageData{[]byte("base64")}},
 			},
 			window: 1024,
-			want:   "You are a Wizard. Hello [img-0]",
+			want:   "You are a Wizard. [img-0] Hello",
 		},
 		{
 			name:     "images truncated",
@@ -165,7 +165,7 @@ func TestChatPrompt(t *testing.T) {
 				{Role: "user", Content: "Hello", Images: []api.ImageData{[]byte("img1"), []byte("img2")}},
 			},
 			window: 1024,
-			want:   "You are a Wizard. Hello [img-1]",
+			want:   "You are a Wizard. [img-0] [img-1] Hello",
 		},
 		{
 			name:     "empty list",
@@ -198,7 +198,7 @@ func TestChatPrompt(t *testing.T) {
 			}
 
 			if got != tc.want {
-				t.Errorf("got = %v, want %v", got, tc.want)
+				t.Errorf("got: %q, want: %q", got, tc.want)
 			}
 		})
 	}

server/routes.go

@@ -191,6 +191,11 @@ func GenerateHandler(c *gin.Context) {
 		return
 	}
 
+	if model.IsEmbedding() {
+		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "embedding models do not support generate"})
+		return
+	}
+
 	opts, err := modelOptions(model, req.Options)
 	if err != nil {
 		if errors.Is(err, api.ErrInvalidOpts) {
@@ -245,6 +250,19 @@ func GenerateHandler(c *gin.Context) {
 		slog.Debug("generate handler", "system", req.System)
 
 		var sb strings.Builder
+		for i := range req.Images {
+			fmt.Fprintf(&sb, "[img-%d] ", i)
+		}
+
+		sb.WriteString(req.Prompt)
+
+		p, err := Prompt(req.Template, req.System, sb.String(), "", true)
+		if err != nil {
+			c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+			return
+		}
+
+		sb.Reset()
 		if req.Context != nil {
 			prev, err := loaded.runner.Decode(c.Request.Context(), req.Context)
 			if err != nil {
@@ -255,18 +273,6 @@ func GenerateHandler(c *gin.Context) {
 			sb.WriteString(prev)
 		}
 
-		// write image tags
-		// TODO: limit the number of images to fit in the context similar to the chat endpoint
-		for i := range req.Images {
-			req.Prompt += fmt.Sprintf(" [img-%d]", i)
-		}
-
-		p, err := Prompt(req.Template, req.System, req.Prompt, "", true)
-		if err != nil {
-			c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
-			return
-		}
-
 		sb.WriteString(p)
 
 		prompt = sb.String()
@@ -379,7 +385,7 @@ func GenerateHandler(c *gin.Context) {
 	streamResponse(c, ch)
 }
 
-func EmbeddingHandler(c *gin.Context) {
+func EmbeddingsHandler(c *gin.Context) {
 	loaded.mu.Lock()
 	defer loaded.mu.Unlock()
 
@@ -432,8 +438,9 @@ func EmbeddingHandler(c *gin.Context) {
 		return
 	}
 
-	if !loaded.Options.EmbeddingOnly {
-		c.JSON(http.StatusBadRequest, gin.H{"error": "embedding option must be set to true"})
+	// an empty request loads the model
+	if req.Prompt == "" {
+		c.JSON(http.StatusOK, api.EmbeddingResponse{Embedding: []float64{}})
 		return
 	}
 
@@ -942,7 +949,7 @@ func (s *Server) GenerateRoutes() http.Handler {
 	r.POST("/api/pull", PullModelHandler)
 	r.POST("/api/generate", GenerateHandler)
 	r.POST("/api/chat", ChatHandler)
-	r.POST("/api/embeddings", EmbeddingHandler)
+	r.POST("/api/embeddings", EmbeddingsHandler)
 	r.POST("/api/create", CreateModelHandler)
 	r.POST("/api/push", PushModelHandler)
 	r.POST("/api/copy", CopyModelHandler)
@@ -1143,6 +1150,11 @@ func ChatHandler(c *gin.Context) {
 		return
 	}
 
+	if model.IsEmbedding() {
+		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "embedding models do not support chat"})
+		return
+	}
+
 	opts, err := modelOptions(model, req.Options)
 	if err != nil {
 		if errors.Is(err, api.ErrInvalidOpts) {

server/upload.go

@@ -12,7 +12,6 @@ import (
 	"net/http"
 	"net/url"
 	"os"
-	"strings"
 	"sync"
 	"sync/atomic"
 	"time"
@@ -177,16 +176,14 @@ func (b *blobUpload) Run(ctx context.Context, opts *registryOptions) {
 	requestURL := <-b.nextURL
 
 	// calculate md5 checksum and add it to the commit request
-	var sb strings.Builder
+	md5sum := md5.New()
 	for _, part := range b.Parts {
-		sb.Write(part.Sum(nil))
+		md5sum.Write(part.Sum(nil))
 	}
 
-	md5sum := md5.Sum([]byte(sb.String()))
-
 	values := requestURL.Query()
 	values.Add("digest", b.Digest)
-	values.Add("etag", fmt.Sprintf("%x-%d", md5sum, len(b.Parts)))
+	values.Add("etag", fmt.Sprintf("%x-%d", md5sum.Sum(nil), len(b.Parts)))
 	requestURL.RawQuery = values.Encode()
 
 	headers := make(http.Header)