baalajimaestro/ollama
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

Merge https://github.com/ollama/ollama

Browse source 
Signed-off-by: baalajimaestro <baalajimaestro@ptr.moe>
This commit is contained in:
baalajimaestro 2024-09-08 09:26:04 +05:30
commit 76c9dc57fd
Signed by: baalajimaestro
GPG key ID: B5B69626E67EE82A
53 changed files with 868 additions and 25243 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

2
CONTRIBUTING.md
View file

@ -18,7 +18,7 @@ See the [development documentation](./docs/development.md) for instructions on h
* New features: new features (e.g. API fields, environment variables) add surface area to Ollama and make it harder to maintain in the long run as they cannot be removed without potentially breaking users in the future.
* Refactoring: large code improvements are important, but can be harder or take longer to review and merge.
* Documentation: small updates to fill in or dorrect missing documentation is helpful, however large documentation additions can be hard to maintain over time.
* Documentation: small updates to fill in or correct missing documentation is helpful, however large documentation additions can be hard to maintain over time.
### Issues that may not be accepted

27
README.md
View file

@ -295,13 +295,23 @@ See the [API documentation](./docs/api.md) for all endpoints.
- [Olpaka](https://github.com/Otacon/olpaka) (User-friendly Flutter Web App for Ollama)
- [OllamaSpring](https://github.com/CrazyNeil/OllamaSpring) (Ollama Client for macOS)
- [LLocal.in](https://github.com/kartikm7/llocal) (Easy to use Electron Desktop Client for Ollama)
- [AiLama](https://github.com/zeyoyt/ailama) (A Discord User App that allows you to interact with Ollama anywhere in discord )
- [Ollama with Google Mesop](https://github.com/rapidarchitect/ollama_mesop/) (Mesop Chat Client implementation with Ollama)
- [Painting Droid](https://github.com/mateuszmigas/painting-droid) (Painting app with AI integrations)
- [Kerlig AI](https://www.kerlig.com/) (AI writing assistant for macOS)
- [AI Studio](https://github.com/MindWorkAI/AI-Studio)
- [Sidellama](https://github.com/gyopak/sidellama) (browser-based LLM client)
- [LLMStack](https://github.com/trypromptly/LLMStack) (No-code multi-agent framework to build LLM agents and workflows)
- [BoltAI for Mac](https://boltai.com) (AI Chat Client for Mac)
- [Harbor](https://github.com/av/harbor) (Containerized LLM Toolkit with Ollama as default backend)
- [Go-CREW](https://www.jonathanhecl.com/go-crew/) (Powerful Offline RAG in Golang)
- [PartCAD](https://github.com/openvmp/partcad/) (CAD model generation with OpenSCAD and CadQuery)
- [Ollama4j Web UI](https://github.com/ollama4j/ollama4j-web-ui) - Java-based Web UI for Ollama built with Vaadin, Spring Boot and Ollama4j
- [PyOllaMx](https://github.com/kspviswa/pyOllaMx) - macOS application capable of chatting with both Ollama and Apple MLX models.
- [Claude Dev](https://github.com/saoudrizwan/claude-dev) - VSCode extension for multi-file/whole-repo coding
- [Cherry Studio](https://github.com/kangfenmao/cherry-studio) (Desktop client with Ollama support)
- [ConfiChat](https://github.com/1runeberg/confichat) (Lightweight, standalone, multi-platform, and privacy focused LLM chat interface with optional encryption)
- [Archyve](https://github.com/nickthecook/archyve) (RAG-enabling document library)
### Terminal
@ -327,6 +337,9 @@ See the [API documentation](./docs/api.md) for all endpoints.
- [gollama](https://github.com/sammcj/gollama)
- [Ollama eBook Summary](https://github.com/cognitivetech/ollama-ebook-summary/)
### Apple Vision Pro
- [Enchanted](https://github.com/AugustDev/enchanted)
### Database
- [MindsDB](https://github.com/mindsdb/mindsdb/blob/staging/mindsdb/integrations/handlers/ollama_handler/README.md) (Connects Ollama models with nearly 200 data platforms and apps)
@ -335,8 +348,11 @@ See the [API documentation](./docs/api.md) for all endpoints.
### Package managers
- [Pacman](https://archlinux.org/packages/extra/x86_64/ollama/)
- [Gentoo](https://github.com/gentoo/guru/tree/master/app-misc/ollama)
- [Helm Chart](https://artifacthub.io/packages/helm/ollama-helm/ollama)
- [Guix channel](https://codeberg.org/tusharhero/ollama-guix)
- [Nix package](https://search.nixos.org/packages?channel=24.05&show=ollama&from=0&size=50&sort=relevance&type=packages&query=ollama)
- [Flox](https://flox.dev/blog/ollama-part-one)
### Libraries
@ -347,11 +363,12 @@ See the [API documentation](./docs/api.md) for all endpoints.
- [LangChainRust](https://github.com/Abraxas-365/langchain-rust) with [example](https://github.com/Abraxas-365/langchain-rust/blob/main/examples/llm_ollama.rs)
- [LlamaIndex](https://gpt-index.readthedocs.io/en/stable/examples/llm/ollama.html)
- [LiteLLM](https://github.com/BerriAI/litellm)
- [OllamaFarm for Go](https://github.com/presbrey/ollamafarm)
- [OllamaSharp for .NET](https://github.com/awaescher/OllamaSharp)
- [Ollama for Ruby](https://github.com/gbaptista/ollama-ai)
- [Ollama-rs for Rust](https://github.com/pepperoni21/ollama-rs)
- [Ollama-hpp for C++](https://github.com/jmont-dev/ollama-hpp)
- [Ollama4j for Java](https://github.com/amithkoujalgi/ollama4j)
- [Ollama4j for Java](https://github.com/ollama4j/ollama4j)
- [ModelFusion Typescript Library](https://modelfusion.dev/integration/model-provider/ollama)
- [OllamaKit for Swift](https://github.com/kevinhermawan/OllamaKit)
- [Ollama for Dart](https://github.com/breitburg/dart-ollama)
@ -368,11 +385,16 @@ See the [API documentation](./docs/api.md) for all endpoints.
- [Portkey](https://portkey.ai/docs/welcome/integration-guides/ollama)
- [PromptingTools.jl](https://github.com/svilupp/PromptingTools.jl) with an [example](https://svilupp.github.io/PromptingTools.jl/dev/examples/working_with_ollama)
- [LlamaScript](https://github.com/Project-Llama/llamascript)
- [Gollm](https://docs.gollm.co/examples/ollama-example)
- [Ollamaclient for Golang](https://github.com/xyproto/ollamaclient)
- [High-level function abstraction in Go](https://gitlab.com/tozd/go/fun)
- [Ollama PHP](https://github.com/ArdaGnsrn/ollama-php)
### Mobile
- [Enchanted](https://github.com/AugustDev/enchanted)
- [Maid](https://github.com/Mobile-Artificial-Intelligence/maid)
- [ConfiChat](https://github.com/1runeberg/confichat) (Lightweight, standalone, multi-platform, and privacy focused LLM chat interface with optional encryption)
### Extensions & Plugins
@ -397,11 +419,14 @@ See the [API documentation](./docs/api.md) for all endpoints.
- [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 Hugging Face)
- [Page Assist](https://github.com/n4ze3m/page-assist) (Chrome Extension)
- [Plasmoid Ollama Control](https://github.com/imoize/plasmoid-ollamacontrol) (KDE Plasma extension that allows you to quickly manage/control Ollama model)
- [AI Telegram Bot](https://github.com/tusharhero/aitelegrambot) (Telegram bot using Ollama in backend)
- [AI ST Completion](https://github.com/yaroslavyaroslav/OpenAI-sublime-text) (Sublime Text 4 AI assistant plugin with Ollama support)
- [Discord-Ollama Chat Bot](https://github.com/kevinthedang/discord-ollama) (Generalized TypeScript Discord Bot w/ Tuning Documentation)
- [Discord AI chat/moderation bot](https://github.com/rapmd73/Companion) Chat/moderation bot written in python. Uses Ollama to create personalities.
- [Headless Ollama](https://github.com/nischalj10/headless-ollama) (Scripts to automatically install ollama client & models on any OS for apps that depends on ollama server)
- [vnc-lm](https://github.com/jk011ru/vnc-lm) (A containerized Discord bot with support for attachments and web links)
- [LSP-AI](https://github.com/SilasMarvin/lsp-ai) (Open-source language server for AI-powered functionality)
### Supported backends

15
cmd/cmd.go
View file

@ -726,14 +726,17 @@ func ShowHandler(cmd *cobra.Command, args []string) error {
}
func showInfo(resp *api.ShowResponse) {
arch := resp.ModelInfo["general.architecture"].(string)
modelData := [][]string{
{"arch", arch},
{"parameters", resp.Details.ParameterSize},
{"quantization", resp.Details.QuantizationLevel},
{"context length", fmt.Sprintf("%v", resp.ModelInfo[fmt.Sprintf("%s.context_length", arch)].(float64))},
{"embedding length", fmt.Sprintf("%v", resp.ModelInfo[fmt.Sprintf("%s.embedding_length", arch)].(float64))},
}
if resp.ModelInfo != nil {
arch := resp.ModelInfo["general.architecture"].(string)
modelData = append(modelData,
[]string{"arch", arch},
[]string{"context length", fmt.Sprintf("%v", resp.ModelInfo[fmt.Sprintf("%s.context_length", arch)].(float64))},
[]string{"embedding length", fmt.Sprintf("%v", resp.ModelInfo[fmt.Sprintf("%s.embedding_length", arch)].(float64))},
)
}
mainTableData := [][]string{
@ -1418,6 +1421,8 @@ func NewCLI() *cobra.Command {
envVars["OLLAMA_TMPDIR"],
envVars["OLLAMA_FLASH_ATTENTION"],
envVars["OLLAMA_LLM_LIBRARY"],
envVars["OLLAMA_GPU_OVERHEAD"],
envVars["OLLAMA_LOAD_TIMEOUT"],
})
default:
appendEnvDocs(cmd, envs)

16
convert/convert_gemma2.go
View file

@ -34,10 +34,20 @@ func (p *gemma2Model) KV(t *Tokenizer) llm.KV {
}
func (p *gemma2Model) Replacements() []string {
return append(
p.gemmaModel.Replacements(),
return []string{
"model.embed_tokens", "token_embd",
"model.norm", "output_norm",
"model.layers", "blk",
"input_layernorm", "attn_norm",
"self_attn.q_proj", "attn_q",
"self_attn.k_proj", "attn_k",
"self_attn.v_proj", "attn_v",
"self_attn.o_proj", "attn_output",
"mlp.gate_proj", "ffn_gate",
"mlp.down_proj", "ffn_down",
"mlp.up_proj", "ffn_up",
"post_attention_layernorm", "post_attention_norm",
"pre_feedforward_layernorm", "ffn_norm",
"post_feedforward_layernorm", "post_ffw_norm",
)
}
}

70
convert/convert_test.go
View file

@ -15,6 +15,7 @@ import (
"os"
"path/filepath"
"slices"
"strings"
"testing"
"golang.org/x/exp/maps"
@ -22,6 +23,12 @@ import (
"github.com/ollama/ollama/llm"
)
type tensorData struct {
Offsets []int `json:"data_offsets"`
Type string `json:"dtype"`
Shape []int `json:"shape"`
}
func convertFull(t *testing.T, fsys fs.FS) (*os.File, llm.KV, llm.Tensors) {
t.Helper()
@ -96,6 +103,7 @@ func TestConvertModel(t *testing.T) {
"Mistral-7B-Instruct-v0.2",
"Mixtral-8x7B-Instruct-v0.1",
"gemma-2b-it",
"gemma-2-2b-it",
// microsoft/Phi-3-mini-128-instruct@d548c233192db00165d842bf8edff054bb3212f8
"Phi-3-mini-128k-instruct",
"all-MiniLM-L6-v2",
@ -140,6 +148,36 @@ func TestConvertModel(t *testing.T) {
}
}
func TestConvertInvalidTensorNames(t *testing.T) {
f, err := os.CreateTemp(t.TempDir(), "testmodel")
if err != nil {
t.Fatal(err)
}
defer f.Close()
tempDir := t.TempDir()
td := map[string]*tensorData{}
offset := 4096
td["model.layers.0.self_attn.q_proj.weight"] = &tensorData{
Offsets: []int{0, offset},
Type: "F32",
Shape: []int{4096, 4096},
}
td["blk.0.attn_q.weight"] = &tensorData{
Offsets: []int{offset, offset * 2},
Type: "F32",
Shape: []int{4096, 4096},
}
generateSafetensorTestData(t, tempDir, td)
err = ConvertModel(os.DirFS(tempDir), f)
if err == nil || !strings.HasPrefix(err.Error(), "duplicate tensor name") {
t.Errorf("expected error but didn't get one")
}
}
func TestConvertInvalidDatatype(t *testing.T) {
f, err := os.CreateTemp(t.TempDir(), "testmodel")
if err != nil {
@ -148,23 +186,10 @@ func TestConvertInvalidDatatype(t *testing.T) {
defer f.Close()
tempDir := t.TempDir()
generateSafetensorTestData(t, tempDir)
err = ConvertModel(os.DirFS(tempDir), f)
if err == nil || err.Error() != "unsupported safetensors model" {
t.Errorf("expected error but didn't get one")
}
}
func generateSafetensorTestData(t *testing.T, tempDir string) {
type tensorData struct {
Offsets []int `json:"data_offsets"`
Type string `json:"dtype"`
Shape []int `json:"shape"`
}
offset := 4096 * 14336
td := map[string]*tensorData{}
offset := 4096 * 14336
td["model.layers.0.mlp.down_proj.weight"] = &tensorData{
Offsets: []int{0, offset},
Type: "I8",
@ -175,8 +200,16 @@ func generateSafetensorTestData(t *testing.T, tempDir string) {
Type: "U8",
Shape: []int{},
}
generateSafetensorTestData(t, tempDir, td)
data, err := json.Marshal(td)
err = ConvertModel(os.DirFS(tempDir), f)
if err == nil || err.Error() != "unsupported safetensors model" {
t.Errorf("expected error but didn't get one")
}
}
func generateSafetensorTestData(t *testing.T, tempDir string, tensorData map[string]*tensorData) {
data, err := json.Marshal(tensorData)
if err != nil {
t.Fatal(err)
}
@ -322,11 +355,6 @@ func TestConvertAdapter(t *testing.T) {
}
func generateLoraTestData(t *testing.T, tempDir string) {
type tensorData struct {
Offsets []int `json:"data_offsets"`
Type string `json:"dtype"`
Shape []int `json:"shape"`
}
offset := 4096 * 8 * 4
td := map[string]*tensorData{"__metadata__": nil}

9
convert/reader_safetensors.go
View file

@ -49,12 +49,19 @@ func parseSafetensors(fsys fs.FS, replacer *strings.Replacer, ps ...string) ([]T
keys := maps.Keys(headers)
slices.Sort(keys)
names := make(map[string]struct{}, len(keys))
for _, key := range keys {
if value := headers[key]; value.Type != "" {
// bitsandbytes quantized models are unsupported
if len(value.Shape) == 0 {
return nil, errors.New("unsupported safetensors model")
}
ggufName := replacer.Replace(key)
if _, ok := names[ggufName]; ok {
return nil, fmt.Errorf("duplicate tensor name '%s' was found for this model", ggufName)
}
names[ggufName] = struct{}{}
ts = append(ts, safetensor{
fs: fsys,
path: p,
@ -62,7 +69,7 @@ func parseSafetensors(fsys fs.FS, replacer *strings.Replacer, ps ...string) ([]T
offset: safetensorsPad(n, value.Offsets[0]),
size: safetensorsPad(n, value.Offsets[1]) - safetensorsPad(n, value.Offsets[0]),
tensorBase: &tensorBase{
name: replacer.Replace(key),
name: ggufName,
shape: value.Shape,
},
})

312
convert/testdata/gemma-2-2b-it.json vendored Normal file
View file

@ -0,0 +1,312 @@
{
"general.architecture": "gemma2",
"general.file_type": "1",
"general.quantization_version": "2",
"gemma2.block_count": "26",
"gemma2.context_length": "8192",
"gemma2.embedding_length": "2304",
"gemma2.feed_forward_length": "9216",
"gemma2.attention.head_count": "8",
"gemma2.attention.head_count_kv": "4",
"gemma2.attention.key_length": "256",
"gemma2.attention.value_length": "256",
"gemma2.attention.layer_norm_rms_epsilon": "1e-06",
"tokenizer.ggml.model": "llama",
"tokenizer.ggml.add_bos_token": "true",
"tokenizer.ggml.add_eos_token": "false",
"tokenizer.ggml.bos_token_id": "2",
"tokenizer.ggml.eos_token_id": "1",
"tokenizer.ggml.padding_token_id": "0",
"tokenizer.ggml.unknown_token_id": "3",
"tokenizer.ggml.scores": "0872465d173867d755d3ee728f882b9dc2057a0bfd596fe1e3d131522f1250d8",
"tokenizer.ggml.token_type": "8d40143b3477df77beea4139420335ede458bf5e14102f01b0170197b55da8d8",
"tokenizer.ggml.tokens": "c6e66de1841f04de8b8d236d461ab720a4c9b9b5414dc293a09c6e10eab45fda",
"token_embd.weight": "64a9d30707e659e2e673656d71f5aef7a9fb9fd83bb9a77558dfc5abbe218a05",
"blk.0.attn_k.weight": "d8b4437c5edb3cddf6af9987038e1bb2b191c4f0fce0e160d2abace717f5d5d7",
"blk.0.attn_norm.weight": "1eb73e3f7aa8e502f6ca31cd19efbb8e4fd9a89692e13e48ac8205545a7fa7e8",
"blk.0.attn_output.weight": "39e7b78e57d356a22dd89ce1c4d7163b970712ba756545e1703f97866cd2192e",
"blk.0.attn_q.weight": "795058e23b6109febd9d55c89e1eebe6af0714ec8c56fd86a160876a6135ffe8",
"blk.0.attn_v.weight": "0cd6e583d1887c020472e961bbb113fe5a0d23ae2f1c2c876fc366cdb7692b52",
"blk.0.ffn_down.weight": "51eb4d962189e945a84e94e0dc1aad3f8f90cc1a11e18029670afcd0ea0acb1b",
"blk.0.ffn_gate.weight": "9811a29b8ad48432925897ab21dfcb13c5cbd372aeccbbefca9b7866883b4ce3",
"blk.0.ffn_norm.weight": "92cbf4652ef503c1de5b10f2be00b3fcf00100980cb3baa8f3013a8d8bf3d851",
"blk.0.ffn_up.weight": "af87de21746879483ed1b374cdd76b19ba11ca2b6dbb1beba98efdf3be3e8077",
"blk.0.post_attention_norm.weight": "32e135f1f258ffe407018899e39af1725d59d66d60022b9a21575ba160e0357a",
"blk.0.post_ffw_norm.weight": "ba286f5ac11b07fbc986173708c66f1920427be5a6d108af38fa0a837c1c8eb6",
"blk.1.attn_k.weight": "51584435552051f7fade76beca582b3f7190cf7fc07adcf527c2774d4b1c3901",
"blk.1.attn_norm.weight": "6833104c7fbf35a7e799ae56c262b97fffa14789642aee14381b25acd21ed80a",
"blk.1.attn_output.weight": "14c39481369087bf292ac9a3ab2ef166f9fe376a9f90c246653213ef264febdc",
"blk.1.attn_q.weight": "443f64ae2229f857c69d6bebb7800b685786cb77884c3ae19d4286aeed081325",
"blk.1.attn_v.weight": "0df482de2038f1e4c8a7733ac0ddb69ad90759dab5968b942af0155588de4c4a",
"blk.1.ffn_down.weight": "66f30763a8bbbcaea609a0087ed75fadb5e771c06378dd2cea94cf17e492e8cf",
"blk.1.ffn_gate.weight": "a7151bff00a545fa18b2c92dcd2a14572ccf9beb957a6c494f1374e8ebe174c9",
"blk.1.ffn_norm.weight": "e197d71ea11b5276bc0167d2663b88089b3ff42b47ba91e85f6c5d95f6306435",
"blk.1.ffn_up.weight": "57c182e0b14cccd1350d388f0c616991702e74281db54637451b70f4ccc24f9b",
"blk.1.post_attention_norm.weight": "3c56f837168d784c2d8bac247c130bdca6610c095c8da4558c536ccad7605609",
"blk.1.post_ffw_norm.weight": "d2a51d320fd01069dd7ccaa7082f16a7faeb671885607d7900b10a89c354d0fa",
"blk.2.attn_k.weight": "bc103c818192de7ce36caaf89dc117be4df13fb902e0bd9a23c64edace5df9b6",
"blk.2.attn_norm.weight": "0f2503aa126083a5d6ac72481be1ef66c6014705b573682b35bd864e4749a3d5",
"blk.2.attn_output.weight": "05fcd4a1226e482f91803a266f72caca887a93e63c2d2ba5611ab3c68d38743a",
"blk.2.attn_q.weight": "6a10b5c2fd423d1e4c4fd60fa8c154a0159b6b2501ea79cae2ef19f45a674e5e",
"blk.2.attn_v.weight": "3cf891945a1f8ae7cc908a5c6b729ff5b70f4436c5ffdbf245cc0ed4cc19cd1b",
"blk.2.ffn_down.weight": "ea204fd04e0d2fc728a9861a459216bbfec629c152004ba625f52cd8837bd51e",
"blk.2.ffn_gate.weight": "3a3518729f1b8b64a82b8792f33987db5418fdb094be0263c68f146a5c38de54",
"blk.2.ffn_norm.weight": "754ede678b725de41a34b82f0edf7688b5c065be7c0d46df6f7ad9430d986884",
"blk.2.ffn_up.weight": "ffdcb88439f5828ffbd9fc844b03ff91637b790b9838097258cc3ae75935720c",
"blk.2.post_attention_norm.weight": "4b3f53b7ba26e8c36b2dfda3b7e5fc4b1065257cefdea235fc7df9af130ac2fd",
"blk.2.post_ffw_norm.weight": "e550369e26b8485e2b54ad34b34bc98af5494287dcc513c2c39cf1eaa5b89d07",
"blk.3.attn_k.weight": "89f24ea450e37d9e95757651a83205c085d81b354ee9489dd6310a391d8409f3",
"blk.3.attn_norm.weight": "24e2ea662b7cb822b4ca5cd61bc17f2709f406d990ec3b4a0dac1cc112db45cf",
"blk.3.attn_output.weight": "ac4dad69473c6e3fac56669212cadd8c34ecc5973d945972e974d94805334967",
"blk.3.attn_q.weight": "b6a9c9a7d4722b9096631c65de62228dfddca6e26edfe6af7fce01e116ef0f4c",
"blk.3.attn_v.weight": "f272a960a40093942309bc342a379984cbacec2d7bc64428db3f64e6b1887ed4",
"blk.3.ffn_down.weight": "c0188ba50d8228805982029c277fc0e87aa57473b8363037c648f6d006ff828a",
"blk.3.ffn_gate.weight": "a04aec1561ee6c0fbb18c3db49dc62fb533619cf697fd548cbf2279761aaec3b",
"blk.3.ffn_norm.weight": "bc053837d44087ec05eb5d9458357b2a5be787789b19cdbbdc694b57697f99a6",
"blk.3.ffn_up.weight": "b3ce8b274f20796d3b1a7c08ba27a919066f9de89a782faa544c4a8d6bea1382",
"blk.3.post_attention_norm.weight": "9c922dee7a7df5667289e2788e60170238239cee2dfdbbd9e435763f9f416718",
"blk.3.post_ffw_norm.weight": "b682544ac953ad2e0b49027ed8916f2e9d1aba5d1587bb4127ac703570c7a03a",
"blk.4.attn_k.weight": "143b0cbb4b787b95c2b6212374410e32173ccef2adb914908a2f89a7916de512",
"blk.4.attn_norm.weight": "5668f60491b780273745192662d02c9a92a4f692b29d16aa0bbc7413fec4f85b",
"blk.4.attn_output.weight": "b9f2bdb68be1e0cf66dd19f8fa2afb105910ad2ef394864cb32cea8f8944e0d5",
"blk.4.attn_q.weight": "ddcf1343dafbc2dfcd0b8741225af22fe4b54b2becce29240bd01c34265d126c",
"blk.4.attn_v.weight": "6dc7074366e7ed52d9f48c594dcc85bef738e096276cb99d28228c89eecc5b9c",
"blk.4.ffn_down.weight": "30334ffc59ce343cf2a1b973174acb7722823463adc07e19a99bd0f404bc9906",
"blk.4.ffn_gate.weight": "890f7c8af208d63b28db52c4b8c16c2288a382d87ff5a6a6d6b0a5b3bf27e6cd",
"blk.4.ffn_norm.weight": "ff0316cc7847221eb86a90c1ab441d4ee61553d410c66414a7755021b3b12448",
"blk.4.ffn_up.weight": "6af97d113f91564c636734f215e25ee602d48eb045458f300b3ec7582be0f41d",
"blk.4.post_attention_norm.weight": "69438f231e105e68216b078bdeb35a7cdc8b12c4e2845e18ecf4c8d361d6a321",
"blk.4.post_ffw_norm.weight": "0fd535da78bcf2b32c95b05b2b83dc49817393765be90d8cc1ed3d56f47b68ec",
"blk.5.attn_k.weight": "0166eb3c6d20dcf3d3c169e94caa8dee057535bb525e29f698fb6f8844f18a6c",
"blk.5.attn_norm.weight": "a7808f27f164023d5cde2be00fc23cac6c71aa0ddeb60bc23e12411b80087672",
"blk.5.attn_output.weight": "8b65b2027a0842b68c5308f91d6a31de9599d794157d77df8418b19f9e0d9334",
"blk.5.attn_q.weight": "966bc626ef2c2394d872087a41c126bb1b67d1d5f6de920204ef5e5b16c34003",
"blk.5.attn_v.weight": "9a362aef3f4437fbf0ef6e1ba785f3329c3db2960f93fe36547d2795e9c254ea",
"blk.5.ffn_down.weight": "63e53541d34197720c06f297aa8142ac6b6eec002c7987b296f26e8b1400f931",
"blk.5.ffn_gate.weight": "d9591fdd32f783e0fc26e20d5d587ee8971ac8ae2e4c818c6eac1c125c7c7f37",
"blk.5.ffn_norm.weight": "677334cc60ecce3a7f4ab3acda15d359353d7358872f614ad8914e3780e9fc6e",
"blk.5.ffn_up.weight": "a63764110e1c655ffbd55af0669b2dfe4cc29d0e198d33a8e5426461b08a85f7",
"blk.5.post_attention_norm.weight": "c55499f859b2c0a7f5cabceaae47309a5ad38bc29d0f4a8db81f1357023162a9",
"blk.5.post_ffw_norm.weight": "82752754665f842418f3e302cb5f43d1e0504dcd124c4b8ddb77018b2c793837",
"blk.6.attn_k.weight": "e20a5f0d6c807273c8d491439566b428497ac02097cf0aa55e33748c28e14be6",
"blk.6.attn_norm.weight": "2c6ba42fd3c73d72073ced03a32dd28d70a89ed9bbbc8fea1ba03a7ade951e6c",
"blk.6.attn_output.weight": "4de7c5c2f4a133a266e17ed8c14c52959466b54cc7ab9e19f789a33b4850f284",
"blk.6.attn_q.weight": "56462d921800e6b8cd2213fef04c4ff16d728905cb2f4c58e966d0a053a3b0ae",
"blk.6.attn_v.weight": "b758dcbff769d6240c2245ede1dbc62c4170a67c77458e866312589220fe29af",
"blk.6.ffn_down.weight": "582247fb3c2bf687cbe9413fe18d18ad47bef4b65df7d78905e10335c6134764",
"blk.6.ffn_gate.weight": "3035444d5286aefb7a6d04e55bc27e1fac7cf895cd5be02319a431b8e047b4ae",
"blk.6.ffn_norm.weight": "e582d24c66e01b96faa20ce6adfda3d8583b11e809bff89969927398175e369a",
"blk.6.ffn_up.weight": "6f4b7bbfedeacf61a4866ae0616c4ba6c9e856662e8f00ae6aaec7f52c53e7b4",
"blk.6.post_attention_norm.weight": "8fe51b50bd677d21586aecab0b565c4bf9fa68ad50bfe366f45e8fea3c657ca8",
"blk.6.post_ffw_norm.weight": "81ba3cb4c2bf5c546b86855b7a885d3fafededc67eb3a35cd3598b03c9e26e65",
"blk.7.attn_k.weight": "2e044179cdcae0946708c86bfea7aa0391e1f7e2a09b33fca035d384cc3ca758",
"blk.7.attn_norm.weight": "94b48c546b046803c60e75a3acb17a356b710735989938021b565f68df9b4985",
"blk.7.attn_output.weight": "65709b4ad7a581f4d75793d39d4032a359f6bcc0c3835205242a0b99e5b66824",
"blk.7.attn_q.weight": "8ded993c95d1f7caf201ceb6fa035cd6ed6d351b50b999fa9355dfee9486cb5b",
"blk.7.attn_v.weight": "c92d5e2d2d48397542bc03bea25bf39154075e66c5bb1ead85188505aa04ae91",
"blk.7.ffn_down.weight": "e8ba8fb57208805ef1dc23cd7c86e9a2d1fb7c52c3940d292cd5bb2eb24b3fac",
"blk.7.ffn_gate.weight": "f0f06d6a2e06c5ac252083bc61d05c814e6289d3f4e4a87d2f06918254c02c36",
"blk.7.ffn_norm.weight": "ebf8ef775f72624148e09d68a4332187a7a5020c521fe0623da1cd3485ad33e0",
"blk.7.ffn_up.weight": "a554adc4fc7122c247c77670e169916ba1794c787b5be30a2b36705138f1f746",
"blk.7.post_attention_norm.weight": "3aa6bc21d85c3a0c12b964e82b12feaedfdd13130c3cd2229228e24e0967ebdf",
"blk.7.post_ffw_norm.weight": "508bc7b19ee8ff08f0007c890133a462fc57c7e72b16ee8f6dd64def264ef876",
"blk.8.attn_k.weight": "363c8e74056642fe9e7c2f3f9769d57319cd3fa0a6022810189ab8d894322885",
"blk.8.attn_norm.weight": "685b49a1f1acb169f4df0bdd8e3de6943f3033cebad14b898a72000595610d92",
"blk.8.attn_output.weight": "7bde571e4efef1c6a6143f0526721dfb59e0a0ea0e1a3616a322b2eb937efa48",
"blk.8.attn_q.weight": "fc993dbc1074c28a0e1d85e5ab2f4ea6a9c6c1affe7ee56027000a275daed9b6",
"blk.8.attn_v.weight": "281e8791d3aef9b3864f1cb054da0ae0c2fef4ce0a58b1bad8bc136b2fa0f62b",
"blk.8.ffn_down.weight": "b1164a2578a7f87ed99c2bbc76c5dfbbbc6a1a803605391acc3f320fc989ffd7",
"blk.8.ffn_gate.weight": "6b39a3b3aaaa79aee61416b54d62160b9258042650e61c6b47bc77c2dd17daf3",
"blk.8.ffn_norm.weight": "17ea1362c72da27f12bc936500492035bdef3fd8f940cb12b57f37d42ba8ecb1",
"blk.8.ffn_up.weight": "bc3a7c47afc440d2bdf8fbe9ddf2c9220467472c60c8b4ded8c0f181470ec96c",
"blk.8.post_attention_norm.weight": "5c506204e00411ef9c8b4134d40eedcc19fffe68dd0af7d7cc49dcabf2dfac7e",
"blk.8.post_ffw_norm.weight": "002faec235c3678864e2901eed275ce4e9dc229164a91c9cd4c965142ba62305",
"blk.9.attn_k.weight": "0bab39d8c237f1b6d0010db40467142625a9e6f2e0e4c49a56c12b41e4e0b1fa",
"blk.9.attn_norm.weight": "de5f38e873b17f07aa7598831b89cc1cae2c9bc3eb2e042ee9af059d2563e84e",
"blk.9.attn_output.weight": "8a8184702c25a62df9ff309c0c7badc8587208523b2be3e8fa90ce7080573e6f",
"blk.9.attn_q.weight": "7c961b2431b09ddf95377acd07201cb91bf13d9cd3ae0f2c25c7d6a0358d9f50",
"blk.9.attn_v.weight": "e22d240cb4743067033e659cbf210ebe2ebbab3e1dea6ccbe5eaa982382ca038",
"blk.9.ffn_down.weight": "a426f81210f03d6ad53277416e1fdcdf37d8065e4817613edaf6c67a343426be",
"blk.9.ffn_gate.weight": "a82eba825cb77b8e64f85ff99ede2fc71bc9b01751eeb17e9e6c246ee12ea62e",
"blk.9.ffn_norm.weight": "1a97f9b1302a3a326d534c5c3fed2db6db0ae45fd0edd381a3e4fc1c75d81030",
"blk.9.ffn_up.weight": "5f20bac2bbf03bb42adb92fbf99561651e1edda57e0b61935ac7f6c08c0ed7cb",
"blk.9.post_attention_norm.weight": "9f9866d13988e1946b1e1c80d9374a92a6e3be33748f8eaed3e126d1e1a4c796",
"blk.9.post_ffw_norm.weight": "a6896dbf698db4dbbe5dbf12417d4fd80e9cad0c539c858892ec0aa5b046bb58",
"blk.10.attn_k.weight": "ca8446e5d21ecd4e6a70dca8d321be480be4fba94d70cba065205436feb44270",
"blk.10.attn_norm.weight": "4f41fe290e8f21f63b82151b6cce94bf7318d121468816b0c58af0ff7c1658ab",
"blk.10.attn_output.weight": "c626d2e9681c5c941bbde43dddfae1a8d4986bf2be4470857bc8e8bd7f869044",
"blk.10.attn_q.weight": "1e61b210a13a429977325cf15d781ab77d604cfa862f4270329cbd94237d5835",
"blk.10.attn_v.weight": "8ff8d3e3f058ec3b35ada1057f2ed59c06494d0e0be6a8dc3ff9edf9f0e1a115",
"blk.10.ffn_down.weight": "bcebc04219f8081a5f483e58103c0ddbbbc631a0a54fd6dd9d55778e041f70ee",
"blk.10.ffn_gate.weight": "7a23a1e620ef871384ddf9611ccdcfb893fbf013cc203ac8e72f745420f1eea0",
"blk.10.ffn_norm.weight": "e3a375e43c349a1c6c66c22328e513cc1af3137fe839e43dc8e9be2f65914fd7",
"blk.10.ffn_up.weight": "5d182e7c94369194fca5f19cbbe668a999911e57f3d363bc7fb6088428700cb9",
"blk.10.post_attention_norm.weight": "b841c6308296e8984f3c5f549c6e3a242f4b3e19141e1f54cc08de9c46759c09",
"blk.10.post_ffw_norm.weight": "9d66fa05b5c940208f634f5053d809094c99a2a10a1d1e8847c8281fbd99fb49",
"blk.11.attn_k.weight": "14adf24ebb2bb17b336ca81cec3e690fd854782f4440ca6c66cc1d7e7bf1c850",
"blk.11.attn_norm.weight": "2d2213f311f50414702b5b34f22aafb9d9a0b6787243e7578562583dc40ad195",
"blk.11.attn_output.weight": "de1f14cc2a7fff00cf11b229f0576999205f17b9536e97abc9d6de3cc79a7884",
"blk.11.attn_q.weight": "2bcc5c147524003109ece0be08b89ac8b25baa71416ffa76573c6c052ffc6eea",
"blk.11.attn_v.weight": "2e6ab8573070c22dc1e0d7aebe4d52123226dacf7822dcce06fadbb38fb036a4",
"blk.11.ffn_down.weight": "1b86902f4e36868421e5228b9445051f8290b292df22a6d1af836dcecc1f25c3",
"blk.11.ffn_gate.weight": "e756e8081bd0a16aea4a9ef5076ad102113524f7a3d50a3a77aaa7f7938b63e8",
"blk.11.ffn_norm.weight": "6913887267be227cf9d1991a3dd8db2e7e74bb9b5fbdfcb9ac954fd7d7b95b3b",
"blk.11.ffn_up.weight": "619a3ac0609ebdf42c3fb2b6e4b1db48df79e6dd8418d7ab8f1bbff13d8a6a50",
"blk.11.post_attention_norm.weight": "e4b4ba92cef7b6a78407e8ab1b0307d47dac6c3df7b6817e28038317ff662d7e",
"blk.11.post_ffw_norm.weight": "40aceeec58cb855f0c158c9cc217168fcd5d0e735567d587217b1d78df17bc5f",
"blk.12.attn_k.weight": "c54c5a4d4892522022d1aa2204cfc624f0b4042caa536e678967316293fe5cb1",
"blk.12.attn_norm.weight": "7cd2ef58298569ffdf244d9b390f3917245276c8206e5780af5f96d8c0bbb446",
"blk.12.attn_output.weight": "85495ef9cc8b3deb21f741bde463ff6493acae2be51f02ecdeef952cbdec3375",
"blk.12.attn_q.weight": "d19383f83fd119bfb8c0280c9515705c11d8e7d502019fcf8f49efeef0d106d0",
"blk.12.attn_v.weight": "869ac669ba49531d9128892a0e27cef15de508ff40cdf80cc1681dde50d09204",
"blk.12.ffn_down.weight": "578f39f8f9fc2f09138afc884a952d7cc3a9a31de4216acd10e88e19e0b75f8c",
"blk.12.ffn_gate.weight": "e29a0186bc6c4a0720246306e922d3a83f777dadcf4ac80bad468287031cc8b5",
"blk.12.ffn_norm.weight": "e1ee95c6584b5cb57fcf1db8ce2bcc03aff91eb389238c094a61c00dde93d1f2",
"blk.12.ffn_up.weight": "2a826f06d7cdfb3edc6ae250ff44363ef77a2a9cdf96313e23a331b99ebfa17d",
"blk.12.post_attention_norm.weight": "4bafc7699b948d5cbc0d3e09b418b06c6abc4651a61ada9609d9a2f21c7e5607",
"blk.12.post_ffw_norm.weight": "bbb8c34a7176bb1a49f9fe2bacca0bd26b673d52c0835b2e90fa11f2962f077f",
"blk.13.attn_k.weight": "ffeefccfe8255d1b694382012ff4134eee5fec9d9491c8d0ff0a13832d1a37e8",
"blk.13.attn_norm.weight": "35713726529e3887c4135a88e86e8a4d7270ba5b9f2d1ab462622fbf40a7cdce",
"blk.13.attn_output.weight": "0d60b7c5cd71190a9ef4b873b0f516be15447c32d83914db2794b14592b0b460",
"blk.13.attn_q.weight": "8296069e65bef794cefc61257fc65789b3cb22955e30f3df129205e5041b2222",
"blk.13.attn_v.weight": "ca0f4ab9d16a748fc643a5c0c7a19826a811bf2a4e7316a8c935d4bf0ce8abc6",
"blk.13.ffn_down.weight": "d5514e0c8e7b3ed1cbcc1605eb5be1733b6ab3514cf8a0508fc72f7d05ed8bcb",
"blk.13.ffn_gate.weight": "8108e517a82e08a3aefbbd267bfa50a1668f92a76273280ce8a6bc1f6dd61521",
"blk.13.ffn_norm.weight": "5fcb6132d2134bf1f835b904a99820fa501dbc57d2224129f7098bf3cabc1d36",
"blk.13.ffn_up.weight": "6d744b7cd390a3cae3aa350dd379b81246acd056a2259996b6aaadece8465ccc",
"blk.13.post_attention_norm.weight": "e08b14698912509790e9575b8676971fbb0a4d82d719367e3756c0d0c4ab8cc0",
"blk.13.post_ffw_norm.weight": "2b196e4450fc5f1e7367b2cf7fe33a15fe919fbcdd861d11002346f16e980535",
"blk.14.attn_k.weight": "120e5f48d7268dfd9ab5f4bc9cc57a7cec63ea9635f56b80d435eb22936e9483",
"blk.14.attn_norm.weight": "146367bcce4db72cc894419a2e0145a6f533507dd68e4739c10ee480308c401f",
"blk.14.attn_output.weight": "720fa0165e756876c5cb6ad9e2780dd910390933f3f8849e5add5da04266650b",
"blk.14.attn_q.weight": "f5183466f56219ca1aca52d8b82c2d966a4198fea40fdd6b39f4d8b06ca2a6dd",
"blk.14.attn_v.weight": "24f8ea3d5512cd37c43c8329cb0da0c90d1895aef763ac2dcee3fe5157ec50a2",
"blk.14.ffn_down.weight": "e29960965b384ae5ab3d898a4dbaa8fddd28fa0e477ac28bcac49dec12a5ac67",
"blk.14.ffn_gate.weight": "6d0d6a74bfe9692e8f8eedff0fc34fc4fa1c8687794f35f2e2b033ab2d7510b8",
"blk.14.ffn_norm.weight": "f7036c1a9a71e046c9d2af16e9218fda5dbb0f7241ab44747abed1f0f9d602ca",
"blk.14.ffn_up.weight": "7d69ea1424007ffc9c12247dd0308c616e93ac02a59ec341cfa48f92d6ce3b10",
"blk.14.post_attention_norm.weight": "65b9712834d9445d4236bec362f3fb795c20d60c541b3dc6dbb7914d9b493e41",
"blk.14.post_ffw_norm.weight": "9c6a8da2e4e437d5cfdf3b9097e9f8b64bf07946a048badec20f4d374613f38f",
"blk.15.attn_k.weight": "864bc618303a0e4ee67fb1d5e751de61e936cd51e96669dd86f8cd08f2305045",
"blk.15.attn_norm.weight": "f9f4187da6eeadc2fc5921d8fe669741697d16c13d71e4aaeb73b82f50dc577e",
"blk.15.attn_output.weight": "ce2419a0b097036b2a31f2f4ad731d5814bcc2ef4c511786e24471e5eefd273b",
"blk.15.attn_q.weight": "9539db5a970d11ebe99722d1e13fcd635e250033630811efe583d2f97778e4a9",
"blk.15.attn_v.weight": "1c834b48ccd88adaeabb7d8bcb6be0bcd6d5ac1354ce88fc28f19a1a96b81ab3",
"blk.15.ffn_down.weight": "bc1f97a65dde6fa2c1e5397afb612266944b343f2eaa868b635ddd25829f8a42",
"blk.15.ffn_gate.weight": "1b14529d57056b79037f6cb5008132e62cc35992353b38dda59572274623103b",
"blk.15.ffn_norm.weight": "9af77458de9ee55c66f93865759f9c2c398557f94f3fa8fa6af30543d7339cde",
"blk.15.ffn_up.weight": "41d524a26b61a9595816b4fd53cf57ef50a702e4ef32933ff6136dca9136a267",
"blk.15.post_attention_norm.weight": "c60a03cd0e63a7db5c80015e58e9b97ba2208caa19f66a6fef5c4447eca900ce",
"blk.15.post_ffw_norm.weight": "34f7f9f96769215bbc3d17084df091864aef96a6645b7d0b3b7d9bd92f1a4b0b",
"blk.16.attn_k.weight": "7e27240d9f3a8c6cf0f4a980113d43234f514eadc3e3e1792b86efb29ffb1a6d",
"blk.16.attn_norm.weight": "af798acc0899282a30448edec48223b3e8efda177090273e612d8eca5e377301",
"blk.16.attn_output.weight": "79df39a3709d3d53e84146291e0944a7a653d06705293d9ccb5648dceadb432c",
"blk.16.attn_q.weight": "db58a1c3b83ad294804e5fd7321005719e200659173466df5a52a182b80b7165",
"blk.16.attn_v.weight": "2af6d48cbaeb225b5c1a704f76abd89c8ab1521417695b112b4dcc2cbd39b74d",
"blk.16.ffn_down.weight": "fc1c813eb5e7da3d6194569d6cb21602fc6eff2dc8e1b0eb753f2d5df148189c",
"blk.16.ffn_gate.weight": "7a80bcbc42464bd55df4814a6edbd7b5c153e0428323bbe49de55e2d2add33e7",
"blk.16.ffn_norm.weight": "2041685ee926d30f3f2ae4ec35b5688f1cd834167a6359a7d4057eac804c58b2",
"blk.16.ffn_up.weight": "8da4b718973ac1d43b928829bc45e062fd101984d6c98dd825bd7c5d08ebfbe3",
"blk.16.post_attention_norm.weight": "975c48fe680a6167438a106140a8872eee7765191f152d80e3b8ddf47693e095",
"blk.16.post_ffw_norm.weight": "4de2d4d483acfe4fc77860ea929025df2f4e15c10729413f36a18c94eaa6d689",
"blk.17.attn_k.weight": "f937e61f0af8c4cd98ee742648eb60e02e579683e21d421071295a3b70aebaad",
"blk.17.attn_norm.weight": "c3270583ed28b7e423f5b170c59113234f258169b93a867d9274f4c10b7cb115",
"blk.17.attn_output.weight": "b8c1150e81e685e539a5dcf2c19047a24eba2b281fabe166674b1d71ef4612ea",
"blk.17.attn_q.weight": "c255100ae2011e7dc7e3bf3bc3ccd96d859fbb98581cae993d7b82c1ba8e8b39",
"blk.17.attn_v.weight": "5830bb0a555984c6485348067f70b5d22ae337c011aa9248dac2ff4c95944551",
"blk.17.ffn_down.weight": "8ff9a7cccaa3776434a9d895aae4fb5c36c736bf2ec98784226b4c234940fbb0",
"blk.17.ffn_gate.weight": "1b52876739712831c272911533da206f407b46034a1a4ae8a88c1f96b6bd5747",
"blk.17.ffn_norm.weight": "d0e16ba5e87c91b545334e022058c7d03849665c3b1a6298771b656531366b66",
"blk.17.ffn_up.weight": "4dd6211d01dbebbe21052708eddc242b082a58b5f18ed16479e17987c1d3432e",
"blk.17.post_attention_norm.weight": "6f49c775c7417dade77ba8268a0f8441c1e5ec28b5d7e4dc5ed07a04d04600c8",
"blk.17.post_ffw_norm.weight": "b91a0bb2e6679e9c9be06ad323adae441d00a3d673efb19d7c4954be2aa84b27",
"blk.18.attn_k.weight": "22b565ace1b4da8b33865a58625be1d90beea9891f29686a69fa9cf7c93217db",
"blk.18.attn_norm.weight": "3e0160d7063c8753de65d2356a66648e47d921efdc5c917efb8209892120f8db",
"blk.18.attn_output.weight": "e3180f0bb4ca90b31e9b08158db38e332de62dfbaefe34aa94cc316409331e09",
"blk.18.attn_q.weight": "f3a5a83614c3ba7ea41cdd5b1b0819a241ee2a951a381ce4a9e001d3f700ed8f",
"blk.18.attn_v.weight": "f3350a5984fb951fc738adcf78147e6d812ff1c576670c460cafc99c253c1654",
"blk.18.ffn_down.weight": "9e9d09b13a33525e14bdaee6efc65c551ac7cf7680e534b940ab122a3a7c1ac9",
"blk.18.ffn_gate.weight": "ebaec8b4b578a2e8d815baac12f1675c208f80c68074d5a18288a2e1a60680ee",
"blk.18.ffn_norm.weight": "33e7687c53a242f2f8dc7093a491c97b18d4a5a8c14d183f02bd586a770f05aa",
"blk.18.ffn_up.weight": "78a1816662378ce56cc870e705174492781897b3afd2d4d97a51f10f2f2987c1",
"blk.18.post_attention_norm.weight": "a58dde3f12df3e94cbc27d87c8ea86f89af8a388a506446ff6758f05399b05fc",
"blk.18.post_ffw_norm.weight": "cebf90cc143577d483cca27b032dfd82031ee59bdf17c0e2cf60a0a3ad5bf996",
"blk.19.attn_k.weight": "4683375d0599ac9e2232196aae1e90af13a14cae26e865465de5c8e257bb2055",
"blk.19.attn_norm.weight": "f3eba936bfb1814bbcb0a1d62739eb66daac839df8c9c836fe0e94860df88525",
"blk.19.attn_output.weight": "51c0f01d38a9dcfe9bdbc4643576fab164c1d9e4b7168b7695c0ee55e6965667",
"blk.19.attn_q.weight": "28d15b69b8416f2e7ddc88fe381cb1e2ef2ad705fb1c268139ba96498cc74848",
"blk.19.attn_v.weight": "6860f1cd720638e63a981fa2c0b4db900129826bcb9823c9ddf9fb8b1b9f3383",
"blk.19.ffn_down.weight": "bc7f2d7827ee01c2dd41401c7b3b1700ad3a4ff620e8bb734f92630d342dcc7f",
"blk.19.ffn_gate.weight": "54d03ef69ba373fc410fbca8f1e34a565d58e4296d9a035ff7e48340b9c848e7",
"blk.19.ffn_norm.weight": "9178fc796a340ee6e8128ca74c0cb6203d1adbed6927af4e5ac7863da57affc7",
"blk.19.ffn_up.weight": "a77bd708026c6e83ad5c79c223278e74621bcf74a9641c7818d96b595daaad20",
"blk.19.post_attention_norm.weight": "ae94aa26f4c411bf9496a6fd4a6df64ee589ee1ae9a04b531d45acc95721e582",
"blk.19.post_ffw_norm.weight": "9ad210700edeef12133bdcff04bf1c7f62b49f6f4a9ba483c7cdc59857c24a5c",
"blk.20.attn_k.weight": "e35bce1e9f4a7a09ef34721f57ea38cfca68c272f52d923fe50af8308f66cfaa",
"blk.20.attn_norm.weight": "644800f6926fd34f233795c4dec1151a295d2138ca8cac33e3e48167d26f8b41",
"blk.20.attn_output.weight": "8d3758cd236471741e1ad66c0710cb79077dc8c7a3a292d35bc551c0c5abe627",
"blk.20.attn_q.weight": "c333b1f0f6f956b5d73891df10b1a0321e55fc31c40d623a24e1f52caa6a998b",
"blk.20.attn_v.weight": "8562b418d0c4868a050fb19fa3fcaf50a8cf1c669f537d666c80c7b3a04714e1",
"blk.20.ffn_down.weight": "97efb608ac44cc804198faec3ee66eafe56ced6b7ca5359700c6f1df75b7205e",
"blk.20.ffn_gate.weight": "5c61151d86f28415c73c73d90ec088c646cbe5c1640197caf58eb501ba7db293",
"blk.20.ffn_norm.weight": "24bbe0a701afd4bbeea65b3edde712b3cbb2281043bbc43dbf250582453116ed",
"blk.20.ffn_up.weight": "e170cf68e249566aa99eb6f6b265679bf9a5a6b76830ba24e7e130c2515910c4",
"blk.20.post_attention_norm.weight": "e092d751cfe20dbf2d348358f3b38397bd83e4ed94d6bbaa6bbaddcd902b2ac4",
"blk.20.post_ffw_norm.weight": "219a18a47dcba76e669e4322223a5a9227bd3db1de3fbd3d3cfb22e54a783c5a",
"blk.21.attn_k.weight": "c3a095ebddb42c63824f1c98da65263dc88e4d790a26aa1632840b44f5cc7cb1",
"blk.21.attn_norm.weight": "ef8bbaded5fbc45ad9cf3985ae02174524e7090fe6362811124f942ef643bec7",
"blk.21.attn_output.weight": "668f018aba72baac6252aa3ad58569ddd55ab751a0dd8d7bcc9fb9b6efb4bf53",
"blk.21.attn_q.weight": "e759c65663089f3bbbd51847934c185e680c82f1249065d5d487da638e519e6d",
"blk.21.attn_v.weight": "2ff57762686cf9ba1f5a6be76503454b97556ce67f4ac98254bd0562231197ba",
"blk.21.ffn_down.weight": "3fd106556fb721b1c28ae3f4026bc83eb1b08ed910f2ba5f466c6b5f327d91cb",
"blk.21.ffn_gate.weight": "338022d882f4b6619e8054a6fb909696fa3eef3013cf69b65c3cacdfc5b9e42c",
"blk.21.ffn_norm.weight": "1e77660c23a3f9653ee721a863d1960f773d87437cabc4dc0a6e17ee3d4e5e44",
"blk.21.ffn_up.weight": "7d31b20fbc2e6eba8f350f170069dc36f0cb12f68fbc4206ec5022a74085ebcb",
"blk.21.post_attention_norm.weight": "9638bae8d8bdcd7ed68da282979cd84a07c41ff9cabcaea94ebc846a1803db23",
"blk.21.post_ffw_norm.weight": "d622ef11115fe0cbe04b727d5a3b6371e7f39bf08c8d5eb9bc6da52e3f3cfb9d",
"blk.22.attn_k.weight": "5c321cb29deffbe57de200dd206a62005f1e80acb86c4fd2349dd44c8d3594fd",
"blk.22.attn_norm.weight": "198d949705d7170a331d75889d8c7500c3635254dac2cc6aa4dc35d556584536",
"blk.22.attn_output.weight": "19805cd5d7025b457e5d41d70db8b3fd63c2dd0e4a94d3ef1704d50ef4e749e8",
"blk.22.attn_q.weight": "177836cd583fc87405975ddc21ebfebdaa090a0363799664c72caa3da851ae2c",
"blk.22.attn_v.weight": "fea255692483e30d0108f9e4e250eb3ed7dbda8d83f499b06519b8c223ae6096",
"blk.22.ffn_down.weight": "00cb8939f03e5817d6d412de8cf2c923c9568d5493e382cec7faf5718fb034eb",
"blk.22.ffn_gate.weight": "b0591065b91281b2fbd8a9567f3568d40479f680e1f0a29e27ae213f37642489",
"blk.22.ffn_norm.weight": "96b5c5d0737c2ceb8fc869f54adb9e5f46e28cb7b177c40f49fa926b923c00f8",
"blk.22.ffn_up.weight": "81f472185b24344ab0594ea8246cc6e200e0dc1cab4943e74fbe4ca19d5a9701",
"blk.22.post_attention_norm.weight": "27fa9aa6260aa3071e0391e1a1d49322dcb6e8072315b8a9b7064087108dbd06",
"blk.22.post_ffw_norm.weight": "f37e1dcd7f643d9545675ffe9dc527a11eba86eb204989c2f44f636b266d896a",
"blk.23.attn_k.weight": "5d82f36658a56c3f94d0bb2d61f65509c966fa6568f81812e0d3e338b380ef8c",
"blk.23.attn_norm.weight": "b7983f88d9cad88bc88a528923e6da592ad20e699965b223ebc10840fe1f4fec",
"blk.23.attn_output.weight": "59f97f80f430d71606aab0158a195aed29ccd3405e6c0a5c41c809be8eb01898",
"blk.23.attn_q.weight": "53ac4789fe958919cc02ea4222bcd64c0ea1b4baa54304bff46635bdf42f7490",
"blk.23.attn_v.weight": "ec8abe09b9e84dbb52c7a068094657c6d3c62fe551ba8d7c3a3f23da622e9756",
"blk.23.ffn_down.weight": "3cf547eccb1b82aa64f208cee9682d7f558ca84e0aead7d9d3d1420d90f3d992",
"blk.23.ffn_gate.weight": "366aa2486d911ba81eb519119e13807deacf7e9908bc1975a2a63e00d6b10124",
"blk.23.ffn_norm.weight": "6d1d4a4af34bb7dc090ac87d6457d398c3e0fb68bd2e2b60b099dc318b6cfac3",
"blk.23.ffn_up.weight": "53f76692e253f5d2420b3f200c731b9f3b7a83e379920b4a067c729b4674aa4d",
"blk.23.post_attention_norm.weight": "7c952fa0efa76b3f048c8c4c9e8dcb5e3724d231327eda6423a34d3f3d3367de",
"blk.23.post_ffw_norm.weight": "7ab188cfe61f0a91b40309a0ab6bfa99f19d0ff2a37b6ac10e5f0c7f44eb5270",
"blk.24.attn_k.weight": "225798792f9bfdd10eff0505ebe61e0aad0209c17b431f6044ee7968ffe8c198",
"blk.24.attn_norm.weight": "635e3c1ebf5219bbebfc40ef164bc32d2b726ef595a94da64ac524ae878e2915",
"blk.24.attn_output.weight": "482f5bb2db8d9ed22b253d9a3296333b239efe698e5992e5d77e7e12dc2a5cf5",
"blk.24.attn_q.weight": "43805bbccddb65d58fffc4be9b5c374d4e1df1395ec1e1ffb4bcff03e98d5adb",
"blk.24.attn_v.weight": "fa741af54b4a3b1775d32f59134756090c5df2e7345a12a2d8db94fe289667a7",
"blk.24.ffn_down.weight": "83c6351e3162626b276f524a57836144625c2556dbe321b57cbd8fd486a68fab",
"blk.24.ffn_gate.weight": "fbe66be0d84d12cea5176cc7eaef64382ffc7324cd9d6266a3342dc43442f2ac",
"blk.24.ffn_norm.weight": "77c1445a8639ad24938bdf0280233eea2362d47391421833dfa72ec756dfc1e8",
"blk.24.ffn_up.weight": "78235ac729ee23c1cf1ae543751e3af32776d8808cee6e529c2a625a1f027654",
"blk.24.post_attention_norm.weight": "161f71b6d07628d43e4ae51a4c9088ec6ca2db123a17986a14505d83fdd04dad",
"blk.24.post_ffw_norm.weight": "cf1ba692aa683368b02ac413e69b2521b98c69a5274eacbb54165b53bf38a8b2",
"blk.25.attn_k.weight": "057a56bd8c8d2b41608d1f71faa3052902152ddf85e47669ad950c1c3e77c33f",
"blk.25.attn_norm.weight": "b7179fe02c334da556ddcf6c1b502245639a728c4cbba8b552d8e1df4565ee9d",
"blk.25.attn_output.weight": "4fed8b05b08a0ff75ffd022701bbeb52f17b23d09332a1ddcba737244bd0d3b0",
"blk.25.attn_q.weight": "c52e99f5d38bf7538d6106a0bbf38ac6dc6296bca9a3f849afa384ea67b4af01",
"blk.25.attn_v.weight": "c49c23d8e1cfa6a8eb971eb69942204890c6d7d830dc8774c84b108a80598912",
"blk.25.ffn_down.weight": "c08d4dc8412b19fdc870c164b83c341b236ec6fe7bb4a9bcfe0dc100faa20286",
"blk.25.ffn_gate.weight": "1a4cb3f36735d59181721471452807903006539e5e1b5ceb4f72d1d7ae134127",
"blk.25.ffn_norm.weight": "8fd6bd0dcec5198761525a36992a57c9ec5e9da60a22092839a84ae8c4e87f26",
"blk.25.ffn_up.weight": "3a00f39bdd5f31dc5e3b281d2002e1ac4f2475d49a0ac1d7720a25b377dcd04a",
"blk.25.post_attention_norm.weight": "e5f31a648612c859b6d21c9ee426e87a86cb1973dfdd86276c767371d9cef5ad",
"blk.25.post_ffw_norm.weight": "553c3bd774922c99c2384380a142d019881d30dbf0fe3bf9430dabfb3f6cbd33",
"output_norm.weight": "49445c4585ab0a8135717a0bdb1cda4a062a030177d0119561d91542aec5744b"
}

2
docs/faq.md
View file

@ -194,6 +194,8 @@ Refer to the section [above](#how-do-i-configure-ollama-server) for how to set e
If a different directory needs to be used, set the environment variable `OLLAMA_MODELS` to the chosen directory.
> Note: on Linux using the standard installer, the `ollama` user needs read and write access to the specified directory. To assign the directory to the `ollama` user run `sudo chown -R ollama:ollama <directory>`.
Refer to the section [above](#how-do-i-configure-ollama-server) for how to set environment variables on your platform.
## How can I use Ollama in Visual Studio Code?

2
docs/gpu.md
View file

@ -10,7 +10,7 @@ Check your compute compatibility to see if your card is supported:
| 9.0 | NVIDIA | `H100` |
| 8.9 | GeForce RTX 40xx | `RTX 4090` `RTX 4080 SUPER` `RTX 4080` `RTX 4070 Ti SUPER` `RTX 4070 Ti` `RTX 4070 SUPER` `RTX 4070` `RTX 4060 Ti` `RTX 4060` |
| | NVIDIA Professional | `L4` `L40` `RTX 6000` |
| 8.6 | GeForce RTX 30xx | `RTX 3090 Ti` `RTX 3090` `RTX 3080 Ti` `RTX 3080` `RTX 3070 Ti` `RTX 3070` `RTX 3060 Ti` `RTX 3060` |
| 8.6 | GeForce RTX 30xx | `RTX 3090 Ti` `RTX 3090` `RTX 3080 Ti` `RTX 3080` `RTX 3070 Ti` `RTX 3070` `RTX 3060 Ti` `RTX 3060` `RTX 3050 Ti` `RTX 3050` |
| | NVIDIA Professional | `A40` `RTX A6000` `RTX A5000` `RTX A4000` `RTX A3000` `RTX A2000` `A10` `A16` `A2` |
| 8.0 | NVIDIA | `A100` `A30` |
| 7.5 | GeForce GTX/RTX | `GTX 1650 Ti` `TITAN RTX` `RTX 2080 Ti` `RTX 2080` `RTX 2070` `RTX 2060` |

112
docs/linux.md
View file

@ -1,44 +1,59 @@
# Ollama on Linux
# Linux
## Install
Install Ollama running this one-liner:
To install Ollama, run the following command:
>
```bash
```shell
curl -fsSL https://ollama.com/install.sh | sh
```
## AMD Radeon GPU support
While AMD has contributed the `amdgpu` driver upstream to the official linux
kernel source, the version is older and may not support all ROCm features. We
recommend you install the latest driver from
https://www.amd.com/en/support/linux-drivers for best support of your Radeon
GPU.
## Manual install
### Download `ollama`
Download and extract the package:
Download and extract the Linux package:
```bash
curl -fsSL https://ollama.com/download/ollama-linux-amd64.tgz | sudo tar zx -C /usr
```shell
curl -L https://ollama.com/download/ollama-linux-amd64.tgz -o ollama-linux-amd64.tgz
sudo tar -C /usr -xzf ollama-linux-amd64.tgz
```
If you have an AMD GPU, also download and extract the ROCm package into the same location
```bash
curl -fsSL https://ollama.com/download/ollama-linux-amd64-rocm.tgz | sudo tar zx -C /usr
Start Ollama:
```shell
ollama serve
```
In another terminal, verify that Ollama is running:
```shell
ollama -v
```
### AMD GPU install
If you have an AMD GPU, also download and extract the additional ROCm package:
```shell
curl -L https://ollama.com/download/ollama-linux-amd64-rocm.tgz -o ollama-linux-amd64-rocm.tgz
sudo tar -C /usr -xzf ollama-linux-amd64-rocm.tgz
```
### ARM64 install
Download and extract the ARM64-specific package:
```shell
curl -L https://ollama.com/download/ollama-linux-arm64.tgz -o ollama-linux-arm64.tgz
sudo tar -C /usr -xzf ollama-linux-arm64.tgz
```
### Adding Ollama as a startup service (recommended)
Create a user for Ollama:
Create a user and group for Ollama:
```bash
sudo useradd -r -s /bin/false -m -d /usr/share/ollama ollama
```shell
sudo useradd -r -s /bin/false -U -m -d /usr/share/ollama ollama
sudo usermod -a -G ollama $(whoami)
```
Create a service file in `/etc/systemd/system/ollama.service`:
@ -54,6 +69,7 @@ User=ollama
Group=ollama
Restart=always
RestartSec=3
Environment="PATH=$PATH"
[Install]
WantedBy=default.target
@ -61,46 +77,54 @@ WantedBy=default.target
Then start the service:
```bash
```shell
sudo systemctl daemon-reload
sudo systemctl enable ollama
```
### Install CUDA drivers (optional – for Nvidia GPUs)
### Install CUDA drivers (optional)
[Download and install](https://developer.nvidia.com/cuda-downloads) CUDA.
Verify that the drivers are installed by running the following command, which should print details about your GPU:
```bash
```shell
nvidia-smi
```
### Install ROCm (optional - for Radeon GPUs)
[Download and Install](https://rocm.docs.amd.com/projects/install-on-linux/en/latest/tutorial/quick-start.html)
### Install AMD ROCm drivers (optional)
Make sure to install ROCm v6
[Download and Install](https://rocm.docs.amd.com/projects/install-on-linux/en/latest/tutorial/quick-start.html) ROCm v6.
### Start Ollama
Start Ollama using `systemd`:
Start Ollama and verify it is running:
```bash
```shell
sudo systemctl start ollama
sudo systemctl status ollama
```
## Update
> [!NOTE]
> While AMD has contributed the `amdgpu` driver upstream to the official linux
> kernel source, the version is older and may not support all ROCm features. We
> recommend you install the latest driver from
> https://www.amd.com/en/support/linux-drivers for best support of your Radeon
> GPU.
Update ollama by running the install script again:
## Updating
```bash
Update Ollama by running the install script again:
```shell
curl -fsSL https://ollama.com/install.sh | sh
```
Or by downloading the ollama binary:
Or by re-downloading Ollama:
```bash
curl -fsSL https://ollama.com/download/ollama-linux-amd64.tgz | sudo tar zx -C /usr
```shell
curl -L https://ollama.com/download/ollama-linux-amd64.tgz -o ollama-linux-amd64.tgz
sudo tar -C /usr -xzf ollama-linux-amd64.tgz
```
## Installing specific versions
@ -109,15 +133,15 @@ Use `OLLAMA_VERSION` environment variable with the install script to install a s
For example:
```
curl -fsSL https://ollama.com/install.sh | OLLAMA_VERSION=0.1.32 sh
```shell
curl -fsSL https://ollama.com/install.sh | OLLAMA_VERSION=0.3.9 sh
```
## Viewing logs
To view logs of Ollama running as a startup service, run:
```bash
```shell
journalctl -e -u ollama
```
@ -125,7 +149,7 @@ journalctl -e -u ollama
Remove the ollama service:
```bash
```shell
sudo systemctl stop ollama
sudo systemctl disable ollama
sudo rm /etc/systemd/system/ollama.service
@ -133,13 +157,13 @@ sudo rm /etc/systemd/system/ollama.service
Remove the ollama binary from your bin directory (either `/usr/local/bin`, `/usr/bin`, or `/bin`):
```bash
```shell
sudo rm $(which ollama)
```
Remove the downloaded models and Ollama service user and group:
```bash
```shell
sudo rm -r /usr/share/ollama
sudo userdel ollama
sudo groupdel ollama

6
docs/modelfile.md
View file

@ -128,10 +128,10 @@ Currently supported model architectures:
#### Build from a GGUF file
```modelfile
FROM ./ollama-model.bin
FROM ./ollama-model.gguf
```
The GGUF bin file location should be specified as an absolute path or relative to the `Modelfile` location.
The GGUF file location should be specified as an absolute path or relative to the `Modelfile` location.
### PARAMETER
@ -208,7 +208,7 @@ Currently supported Safetensor adapters:
#### GGUF adapter
```modelfile
ADAPTER ./ollama-lora.bin
ADAPTER ./ollama-lora.gguf
```
### LICENSE

3
docs/windows.md
View file

@ -48,6 +48,9 @@ the explorer window by hitting `<cmd>+R` and type in:
- `explorer %HOMEPATH%\.ollama` contains models and configuration
- `explorer %TEMP%` contains temporary executable files in one or more `ollama*` directories
## Uninstall
The Ollama Windows installer registers an Uninstaller application. Under `Add or remove programs` in Windows Settings, you can uninstall Ollama.
## Standalone CLI

49
envconfig/config.go
View file

@ -30,9 +30,7 @@ func Host() *url.URL {
defaultPort = "443"
}
// trim trailing slashes
hostport = strings.TrimRight(hostport, "/")
hostport, path, _ := strings.Cut(hostport, "/")
host, port, err := net.SplitHostPort(hostport)
if err != nil {
host, port = "127.0.0.1", defaultPort
@ -45,15 +43,13 @@ func Host() *url.URL {
if n, err := strconv.ParseInt(port, 10, 32); err != nil || n > 65535 || n < 0 {
slog.Warn("invalid port, using default", "port", port, "default", defaultPort)
return &url.URL{
Scheme: scheme,
Host: net.JoinHostPort(host, defaultPort),
}
port = defaultPort
}
return &url.URL{
Scheme: scheme,
Host: net.JoinHostPort(host, port),
Path: path,
}
}
@ -116,6 +112,26 @@ func KeepAlive() (keepAlive time.Duration) {
return keepAlive
}
// LoadTimeout returns the duration for stall detection during model loads. LoadTimeout can be configured via the OLLAMA_LOAD_TIMEOUT environment variable.
// Zero or Negative values are treated as infinite.
// Default is 5 minutes.
func LoadTimeout() (loadTimeout time.Duration) {
loadTimeout = 5 * time.Minute
if s := Var("OLLAMA_LOAD_TIMEOUT"); s != "" {
if d, err := time.ParseDuration(s); err == nil {
loadTimeout = d
} else if n, err := strconv.ParseInt(s, 10, 64); err == nil {
loadTimeout = time.Duration(n) * time.Second
}
}
if loadTimeout <= 0 {
return time.Duration(math.MaxInt64)
}
return loadTimeout
}
func Bool(k string) func() bool {
return func() bool {
if s := Var(k); s != "" {
@ -235,6 +251,23 @@ var (
MaxVRAM = Uint("OLLAMA_MAX_VRAM", 0)
)
func Uint64(key string, defaultValue uint64) func() uint64 {
return func() uint64 {
if s := Var(key); s != "" {
if n, err := strconv.ParseUint(s, 10, 64); err != nil {
slog.Warn("invalid environment variable, using default", "key", key, "value", s, "default", defaultValue)
} else {
return n
}
}
return defaultValue
}
}
// Set aside VRAM per GPU
var GpuOverhead = Uint64("OLLAMA_GPU_OVERHEAD", 0)
type EnvVar struct {
Name string
Value any
@ -245,9 +278,11 @@ func AsMap() map[string]EnvVar {
ret := map[string]EnvVar{
"OLLAMA_DEBUG": {"OLLAMA_DEBUG", Debug(), "Show additional debug information (e.g. OLLAMA_DEBUG=1)"},
"OLLAMA_FLASH_ATTENTION": {"OLLAMA_FLASH_ATTENTION", FlashAttention(), "Enabled flash attention"},
"OLLAMA_GPU_OVERHEAD": {"OLLAMA_GPU_OVERHEAD", GpuOverhead(), "Reserve a portion of VRAM per GPU (bytes)"},
"OLLAMA_HOST": {"OLLAMA_HOST", Host(), "IP Address for the ollama server (default 127.0.0.1:11434)"},
"OLLAMA_KEEP_ALIVE": {"OLLAMA_KEEP_ALIVE", KeepAlive(), "The duration that models stay loaded in memory (default \"5m\")"},
"OLLAMA_LLM_LIBRARY": {"OLLAMA_LLM_LIBRARY", LLMLibrary(), "Set LLM library to bypass autodetection"},
"OLLAMA_LOAD_TIMEOUT": {"OLLAMA_LOAD_TIMEOUT", LoadTimeout(), "How long to allow model loads to stall before giving up (default \"5m\")"},
"OLLAMA_MAX_LOADED_MODELS": {"OLLAMA_MAX_LOADED_MODELS", MaxRunners(), "Maximum number of loaded models per GPU"},
"OLLAMA_MAX_QUEUE": {"OLLAMA_MAX_QUEUE", MaxQueue(), "Maximum number of queued requests"},
"OLLAMA_MODELS": {"OLLAMA_MODELS", Models(), "The path to the models directory"},

81
envconfig/config_test.go
View file

@ -13,34 +13,35 @@ func TestHost(t *testing.T) {
value string
expect string
}{
"empty": {"", "127.0.0.1:11434"},
"only address": {"1.2.3.4", "1.2.3.4:11434"},
"only port": {":1234", ":1234"},
"address and port": {"1.2.3.4:1234", "1.2.3.4:1234"},
"hostname": {"example.com", "example.com:11434"},
"hostname and port": {"example.com:1234", "example.com:1234"},
"zero port": {":0", ":0"},
"too large port": {":66000", ":11434"},
"too small port": {":-1", ":11434"},
"ipv6 localhost": {"[::1]", "[::1]:11434"},
"ipv6 world open": {"[::]", "[::]:11434"},
"ipv6 no brackets": {"::1", "[::1]:11434"},
"ipv6 + port": {"[::1]:1337", "[::1]:1337"},
"extra space": {" 1.2.3.4 ", "1.2.3.4:11434"},
"extra quotes": {"\"1.2.3.4\"", "1.2.3.4:11434"},
"extra space+quotes": {" \" 1.2.3.4 \" ", "1.2.3.4:11434"},
"extra single quotes": {"'1.2.3.4'", "1.2.3.4:11434"},
"http": {"http://1.2.3.4", "1.2.3.4:80"},
"http port": {"http://1.2.3.4:4321", "1.2.3.4:4321"},
"https": {"https://1.2.3.4", "1.2.3.4:443"},
"https port": {"https://1.2.3.4:4321", "1.2.3.4:4321"},
"empty": {"", "http://127.0.0.1:11434"},
"only address": {"1.2.3.4", "http://1.2.3.4:11434"},
"only port": {":1234", "http://:1234"},
"address and port": {"1.2.3.4:1234", "http://1.2.3.4:1234"},
"hostname": {"example.com", "http://example.com:11434"},
"hostname and port": {"example.com:1234", "http://example.com:1234"},
"zero port": {":0", "http://:0"},
"too large port": {":66000", "http://:11434"},
"too small port": {":-1", "http://:11434"},
"ipv6 localhost": {"[::1]", "http://[::1]:11434"},
"ipv6 world open": {"[::]", "http://[::]:11434"},
"ipv6 no brackets": {"::1", "http://[::1]:11434"},
"ipv6 + port": {"[::1]:1337", "http://[::1]:1337"},
"extra space": {" 1.2.3.4 ", "http://1.2.3.4:11434"},
"extra quotes": {"\"1.2.3.4\"", "http://1.2.3.4:11434"},
"extra space+quotes": {" \" 1.2.3.4 \" ", "http://1.2.3.4:11434"},
"extra single quotes": {"'1.2.3.4'", "http://1.2.3.4:11434"},
"http": {"http://1.2.3.4", "http://1.2.3.4:80"},
"http port": {"http://1.2.3.4:4321", "http://1.2.3.4:4321"},
"https": {"https://1.2.3.4", "https://1.2.3.4:443"},
"https port": {"https://1.2.3.4:4321", "https://1.2.3.4:4321"},
"proxy path": {"https://example.com/ollama", "https://example.com:443/ollama"},
}
for name, tt := range cases {
t.Run(name, func(t *testing.T) {
t.Setenv("OLLAMA_HOST", tt.value)
if host := Host(); host.Host != tt.expect {
t.Errorf("%s: expected %s, got %s", name, tt.expect, host.Host)
if host := Host(); host.String() != tt.expect {
t.Errorf("%s: expected %s, got %s", name, tt.expect, host.String())
}
})
}
@ -214,6 +215,40 @@ func TestKeepAlive(t *testing.T) {
}
}
func TestLoadTimeout(t *testing.T) {
defaultTimeout := 5 * time.Minute
cases := map[string]time.Duration{
"": defaultTimeout,
"1s": time.Second,
"1m": time.Minute,
"1h": time.Hour,
"5m0s": defaultTimeout,
"1h2m3s": 1*time.Hour + 2*time.Minute + 3*time.Second,
"0": time.Duration(math.MaxInt64),
"60": 60 * time.Second,
"120": 2 * time.Minute,
"3600": time.Hour,
"-0": time.Duration(math.MaxInt64),
"-1": time.Duration(math.MaxInt64),
"-1m": time.Duration(math.MaxInt64),
// invalid values
" ": defaultTimeout,
"???": defaultTimeout,
"1d": defaultTimeout,
"1y": defaultTimeout,
"1w": defaultTimeout,
}
for tt, expect := range cases {
t.Run(tt, func(t *testing.T) {
t.Setenv("OLLAMA_LOAD_TIMEOUT", tt)
if actual := LoadTimeout(); actual != expect {
t.Errorf("%s: expected %s, got %s", tt, expect, actual)
}
})
}
}
func TestVar(t *testing.T) {
cases := map[string]string{
"value": "value",

2
gpu/cuda_common.go
View file

@ -57,7 +57,7 @@ func cudaVariant(gpuInfo CudaGPUInfo) string {
}
}
if gpuInfo.computeMajor < 6 || gpuInfo.DriverMajor < 12 {
if gpuInfo.computeMajor < 6 || gpuInfo.DriverMajor < 12 || (gpuInfo.DriverMajor == 12 && gpuInfo.DriverMinor == 0) {
return "v11"
}
return "v12"

2
llm/ext_server/CMakeLists.txt vendored
View file

@ -2,7 +2,7 @@ set(TARGET ollama_llama_server)
option(LLAMA_SERVER_VERBOSE "Build verbose logging option for Server" ON)
set(LLAMA_SERVER_LDFLAGS $ENV{LLAMA_SERVER_LDFLAGS})
include_directories(${CMAKE_CURRENT_SOURCE_DIR})
add_executable(${TARGET} server.cpp utils.hpp json.hpp httplib.h)
add_executable(${TARGET} server.cpp utils.hpp httplib.h)
install(TARGETS ${TARGET} RUNTIME)
target_compile_definitions(${TARGET} PRIVATE
SERVER_VERBOSE=$<BOOL:${LLAMA_SERVER_VERBOSE}>

24596
llm/ext_server/json.hpp vendored
View file

File diff suppressed because it is too large Load diff

25
llm/ext_server/server.cpp vendored
View file

@ -262,7 +262,7 @@ struct server_slot {
char buffer[512];
double t_token = t_prompt_processing / n_prompt_tokens_processed;
double n_tokens_second = 1e3 / t_prompt_processing * n_prompt_tokens_processed;
sprintf(buffer, "prompt eval time = %10.2f ms / %5d tokens (%8.2f ms per token, %8.2f tokens per second)",
snprintf(buffer, sizeof(buffer), "prompt eval time = %10.2f ms / %5d tokens (%8.2f ms per token, %8.2f tokens per second)",
t_prompt_processing, n_prompt_tokens_processed,
t_token, n_tokens_second);
LOG_DEBUG(buffer, {
@ -276,7 +276,7 @@ struct server_slot {
t_token = t_token_generation / n_decoded;
n_tokens_second = 1e3 / t_token_generation * n_decoded;
sprintf(buffer, "generation eval time = %10.2f ms / %5d runs (%8.2f ms per token, %8.2f tokens per second)",
snprintf(buffer, sizeof(buffer), "generation eval time = %10.2f ms / %5d runs (%8.2f ms per token, %8.2f tokens per second)",
t_token_generation, n_decoded,
t_token, n_tokens_second);
LOG_DEBUG(buffer, {
@ -288,7 +288,7 @@ struct server_slot {
{"n_tokens_second", n_tokens_second},
});
sprintf(buffer, " total time = %10.2f ms", t_prompt_processing + t_token_generation);
snprintf(buffer, sizeof(buffer), " total time = %10.2f ms", t_prompt_processing + t_token_generation);
LOG_DEBUG(buffer, {
{"slot_id", id},
{"task_id", task_id},
@ -425,7 +425,7 @@ struct llama_server_context
n_ctx = llama_n_ctx(ctx);
add_bos_token = llama_should_add_bos_token(model);
add_bos_token = llama_add_bos_token(model);
return true;
}
@ -1031,7 +1031,7 @@ struct llama_server_context
continue;
}
if (!llava_image_embed_make_with_clip_img(clp_ctx, params.n_threads, img.img_data, &img.image_embedding, &img.image_tokens)) {
if (!llava_image_embed_make_with_clip_img(clp_ctx, params.cpuparams.n_threads, img.img_data, &img.image_embedding, &img.image_tokens)) {
LOG_TEE("Error processing the given image");
return false;
}
@ -2014,7 +2014,7 @@ static void server_print_usage(const char *argv0, const gpt_params &params,
printf("options:\n");
printf(" -h, --help show this help message and exit\n");
printf(" -v, --verbose verbose output (default: %s)\n", server_verbose ? "enabled" : "disabled");
printf(" -t N, --threads N number of threads to use during computation (default: %d)\n", params.n_threads);
printf(" -t N, --threads N number of threads to use during computation (default: %d)\n", params.cpuparams.n_threads);
printf(" -tb N, --threads-batch N number of threads to use during batch and prompt processing (default: same as --threads)\n");
printf(" --threads-http N number of threads in the http server pool to process requests (default: max(hardware concurrency - 1, --parallel N + 2))\n");
printf(" -c N, --ctx-size N size of the prompt context (default: %d)\n", params.n_ctx);
@ -2287,7 +2287,7 @@ static void server_params_parse(int argc, char **argv, server_params &sparams, g
invalid_param = true;
break;
}
params.n_threads = std::stoi(argv[i]);
params.cpuparams.n_threads = std::stoi(argv[i]);
}
else if (arg == "--grp-attn-n" || arg == "-gan")
{
@ -2315,7 +2315,7 @@ static void server_params_parse(int argc, char **argv, server_params &sparams, g
invalid_param = true;
break;
}
params.n_threads_batch = std::stoi(argv[i]);
params.cpuparams_batch.n_threads = std::stoi(argv[i]);
}
else if (arg == "--threads-http")
{
@ -2626,6 +2626,11 @@ static void server_params_parse(int argc, char **argv, server_params &sparams, g
params.kv_overrides.back().key[0] = 0;
}
postprocess_cpu_params(params.cpuparams, nullptr);
postprocess_cpu_params(params.cpuparams_batch, &params.cpuparams);
postprocess_cpu_params(params.draft_cpuparams, &params.cpuparams);
postprocess_cpu_params(params.draft_cpuparams_batch, &params.cpuparams_batch);
if (invalid_param)
{
fprintf(stderr, "error: invalid parameter for argument: %s\n", arg.c_str());
@ -2775,8 +2780,8 @@ int main(int argc, char **argv) {
{"commit", LLAMA_COMMIT}});
LOG_INFO("system info", {
{"n_threads", params.n_threads},
{"n_threads_batch", params.n_threads_batch},
{"n_threads", params.cpuparams.n_threads},
{"n_threads_batch", params.cpuparams_batch.n_threads},
{"total_threads", std::thread::hardware_concurrency()},
{"system_info", llama_print_system_info()},
});

2
llm/generate/gen_common.sh
View file

@ -87,6 +87,8 @@ apply_patches() {
build() {
cmake -S ${LLAMACPP_DIR} -B ${BUILD_DIR} ${CMAKE_DEFS}
cmake --build ${BUILD_DIR} ${CMAKE_TARGETS} -j8
# remove unnecessary build artifacts
rm -f ${BUILD_DIR}/bin/ggml-common.h ${BUILD_DIR}/bin/ggml-metal.metal
}
compress() {

2
llm/generate/gen_darwin.sh
View file

@ -19,7 +19,7 @@ sign() {
fi
}
COMMON_DARWIN_DEFS="-DBUILD_SHARED_LIBS=off -DCMAKE_OSX_DEPLOYMENT_TARGET=11.3 -DLLAMA_METAL_MACOSX_VERSION_MIN=11.3 -DCMAKE_SYSTEM_NAME=Darwin -DGGML_METAL_EMBED_LIBRARY=on -DGGML_OPENMP=off"
COMMON_DARWIN_DEFS="-DBUILD_SHARED_LIBS=off -DCMAKE_OSX_DEPLOYMENT_TARGET=11.3 -DGGML_METAL_MACOSX_VERSION_MIN=11.3 -DCMAKE_SYSTEM_NAME=Darwin -DGGML_METAL_EMBED_LIBRARY=on -DGGML_OPENMP=off"
case "${GOARCH}" in
"amd64")

6
llm/ggml.go
View file

@ -360,11 +360,13 @@ func (llm GGML) GraphSize(context, batch uint64) (partialOffload, fullOffload ui
switch llm.KV().Architecture() {
case "llama":
fullOffload = 4 * batch * (1 + 4*embedding + context*(1+heads))
fullOffload = max(
4*batch*(1+4*embedding+context*(1+heads)),
4*batch*(embedding+vocab),
)
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+embeddingHeads*headsKV),
4*batch*(embedding+vocab)+embedding*vocab*105/128,
)

2
llm/llama.cpp

@ -1 +1 @@
Subproject commit 1e6f6554aa11fa10160a5fda689e736c3c34169f
Subproject commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177

26
llm/memory.go
View file

@ -7,6 +7,7 @@ import (
"strings"
"github.com/ollama/ollama/api"
"github.com/ollama/ollama/envconfig"
"github.com/ollama/ollama/format"
"github.com/ollama/ollama/gpu"
)
@ -94,6 +95,7 @@ func EstimateGPULayers(gpus []gpu.GpuInfo, ggml *GGML, projectors []string, opts
// Overflow that didn't fit into the GPU
var overflow uint64
overhead := envconfig.GpuOverhead()
availableList := make([]string, len(gpus))
for i, gpu := range gpus {
availableList[i] = format.HumanBytes2(gpu.FreeMemory)
@ -164,8 +166,22 @@ func EstimateGPULayers(gpus []gpu.GpuInfo, ggml *GGML, projectors []string, opts
gzo = gpuZeroOverhead
}
// Only include GPUs that can fit the graph, gpu minimum, the layer buffer and at least more layer
if gpus[i].FreeMemory < gzo+max(graphPartialOffload, graphFullOffload)+gpus[i].MinimumMemory+2*layerSize {
slog.Debug("gpu has too little memory to allocate any layers", "gpu", gpus[i])
if (gpus[i].FreeMemory - overhead) < gzo+max(graphPartialOffload, graphFullOffload)+gpus[i].MinimumMemory+2*layerSize {
slog.Debug("gpu has too little memory to allocate any layers",
"id", gpus[i].ID,
"library", gpus[i].Library,
"variant", gpus[i].Variant,
"compute", gpus[i].Compute,
"driver", fmt.Sprintf("%d.%d", gpus[i].DriverMajor, gpus[i].DriverMinor),
"name", gpus[i].Name,
"total", format.HumanBytes2(gpus[i].TotalMemory),
"available", format.HumanBytes2(gpus[i].FreeMemory),
"minimum_memory", gpus[i].MinimumMemory,
"layer_size", format.HumanBytes2(layerSize),
"gpu_zer_overhead", format.HumanBytes2(gzo),
"partial_offload", format.HumanBytes2(graphPartialOffload),
"full_offload", format.HumanBytes2(graphFullOffload),
)
continue
}
gpusWithSpace = append(gpusWithSpace, gs{i, &gpus[i]})
@ -196,7 +212,7 @@ func EstimateGPULayers(gpus []gpu.GpuInfo, ggml *GGML, projectors []string, opts
for j := len(gpusWithSpace); j > 0; j-- {
g := gpusWithSpace[i%j]
used := gpuAllocations[g.i] + max(graphPartialOffload, graphFullOffload)
if g.g.FreeMemory > used+layerSize {
if (g.g.FreeMemory - overhead) > used+layerSize {
gpuAllocations[g.i] += layerSize
layerCounts[g.i]++
layerCount++
@ -219,7 +235,7 @@ func EstimateGPULayers(gpus []gpu.GpuInfo, ggml *GGML, projectors []string, opts
for j := len(gpusWithSpace); j > 0; j-- {
g := gpusWithSpace[layerCount%j]
used := gpuAllocations[g.i] + max(graphPartialOffload, graphFullOffload)
if g.g.FreeMemory > used+memoryLayerOutput {
if (g.g.FreeMemory - overhead) > used+memoryLayerOutput {
gpuAllocations[g.i] += memoryLayerOutput
layerCounts[g.i]++
layerCount++
@ -306,6 +322,7 @@ func EstimateGPULayers(gpus []gpu.GpuInfo, ggml *GGML, projectors []string, opts
}
func (m MemoryEstimate) log() {
overhead := envconfig.GpuOverhead()
slog.Info(
"offload to "+m.inferenceLibrary,
slog.Group(
@ -323,6 +340,7 @@ func (m MemoryEstimate) log() {
"memory",
// memory available by GPU for offloading
"available", m.availableList,
"gpu_overhead", format.HumanBytes2(overhead),
slog.Group(
"required",
// memory required for full offloading

10
llm/patches/05-default-pretokenizer.diff
View file

@ -1,8 +1,8 @@
diff --git a/src/llama.cpp b/src/llama.cpp
index a207451f..2ddf431d 100644
index 88355971..dd7d41ed 100644
--- a/src/llama.cpp
+++ b/src/llama.cpp
@@ -5347,16 +5347,7 @@ static void llm_load_vocab(
@@ -6083,16 +6083,7 @@ static void llm_load_vocab(
if (vocab.type == LLAMA_VOCAB_TYPE_BPE) {
vocab.tokenizer_add_space_prefix = false;
vocab.tokenizer_clean_spaces = true;
@ -20,9 +20,9 @@ index a207451f..2ddf431d 100644
vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
} else if (
tokenizer_pre == "llama3" ||
@@ -5443,7 +5434,8 @@ static void llm_load_vocab(
tokenizer_pre == "codeshell") {
vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_CODESHELL;
@@ -6188,7 +6179,8 @@ static void llm_load_vocab(
tokenizer_pre == "exaone") {
vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_EXAONE;
} else {
- throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str()));
+ LLAMA_LOG_WARN("%s: missing or unrecognized pre-tokenizer type, using: 'default'\n", __func__);

28
llm/patches/06-embeddings.diff
View file

@ -1,37 +1,36 @@
diff --git a/src/llama.cpp b/src/llama.cpp
index 1fe2b9f7..a43312a7 100644
index 88355971..d7db689b 100644
--- a/src/llama.cpp
+++ b/src/llama.cpp
@@ -13689,7 +13689,7 @@ static size_t llama_output_reserve(llama_context & lctx, size_t n_outputs) {
@@ -15906,7 +15906,7 @@ static size_t llama_output_reserve(llama_context & lctx, size_t n_outputs) {
const auto n_embd = hparams.n_embd;
// TODO: use a per-batch flag for logits presence instead
- const bool has_logits = !cparams.embeddings;
+ const bool has_logits = cparams.causal_attn;
const bool has_embd = lctx.is_encoding || (cparams.embeddings && (cparams.pooling_type == LLAMA_POOLING_TYPE_NONE));
const bool has_embd = cparams.embeddings && (cparams.pooling_type == LLAMA_POOLING_TYPE_NONE);
const size_t logits_size = has_logits ? n_vocab*n_outputs_max : 0;
@@ -13959,17 +13959,25 @@ static int llama_decode_internal(
@@ -16175,20 +16175,23 @@ static int llama_decode_internal(
// no output
res = nullptr;
embd = nullptr;
- } else if (cparams.embeddings) {
- res = nullptr; // do not extract logits for embedding case
- embd = gf->nodes[gf->n_nodes - 1];
- if (strcmp(embd->name, "result_embd_pooled") != 0) {
- embd = gf->nodes[gf->n_nodes - 2];
- res = nullptr; // do not extract logits for embedding case
- embd = nullptr;
+ }
+
+ if (cparams.embeddings) {
+ for (int i = gf->n_nodes - 1; i >= 0; --i) {
for (int i = gf->n_nodes - 1; i >= 0; --i) {
- if (strcmp(gf->nodes[i]->name, "result_embd_pooled") == 0) {
- embd = gf->nodes[i];
+ embd = gf->nodes[i];
+ if (strcmp(embd->name, "result_embd_pooled") == 0) {
+ break;
+ }
break;
}
}
GGML_ASSERT(strcmp(embd->name, "result_embd_pooled") == 0 && "missing embeddings tensor");
- } else {
+ } else {
- GGML_ASSERT(embd != nullptr && "missing embeddings tensor");
} else {
embd = nullptr; // do not extract embeddings when not needed
GGML_ASSERT(strcmp(res->name, "result_output") == 0 && "missing result_output tensor");
}
@ -39,7 +38,6 @@ index 1fe2b9f7..a43312a7 100644
+ if (!cparams.causal_attn) {
+ res = nullptr; // do not extract logits when not needed
+ }
+
// LLAMA_LOG_INFO("graph build time: %.3f ms (%d nodes, %d leafs)\n", (ggml_time_us() - t_start_us)/1000.0, gf->n_nodes, gf->n_leafs);
ggml_backend_sched_alloc_graph(lctx.sched, gf);

350
llm/patches/09-lora.diff
View file

@ -1,350 +0,0 @@
diff --git a/common/common.cpp b/common/common.cpp
index 2e8374d5..70d0afde 100644
--- a/common/common.cpp
+++ b/common/common.cpp
@@ -2110,9 +2110,21 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
loaded_la.adapter = llama_lora_adapter_init(model, la.path.c_str());
if (loaded_la.adapter == nullptr) {
fprintf(stderr, "%s: error: failed to apply lora adapter '%s'\n", __func__, la.path.c_str());
- llama_free(lctx);
- llama_free_model(model);
- return iparams;
+
+ // if that fails, try loading as ggla for compatibility
+ int err = llama_model_apply_lora_from_file(model,
+ la.path.c_str(),
+ la.scale,
+ nullptr,
+ params.n_threads);
+ if (err != 0) {
+ fprintf(stderr, "%s: error: failed to apply lora adapter\n", __func__);
+ llama_free(lctx);
+ llama_free_model(model);
+ return iparams;
+ } else {
+ break;
+ }
}
iparams.lora_adapters.push_back(loaded_la); // copy to list of loaded adapters
}
diff --git a/include/llama.h b/include/llama.h
index 93fd77ca..b0fb37a6 100644
--- a/include/llama.h
+++ b/include/llama.h
@@ -1160,6 +1160,20 @@ extern "C" {
LLAMA_API void llama_dump_timing_info_yaml(FILE * stream, const struct llama_context * ctx);
+ // Apply a LoRA adapter to a loaded model
+ // path_base_model is the path to a higher quality model to use as a base for
+ // the layers modified by the adapter. Can be NULL to use the current loaded model.
+ // The model needs to be reloaded before applying a new adapter, otherwise the adapter
+ // will be applied on top of the previous one
+ // Returns 0 on success
+ LLAMA_API int32_t llama_model_apply_lora_from_file(
+ const struct llama_model * model,
+ const char * path_lora,
+ float scale,
+ const char * path_base_model,
+ int32_t n_threads);
+
+
#ifdef __cplusplus
}
#endif
diff --git a/src/llama.cpp b/src/llama.cpp
index 80a0dd0f..9d7b0e17 100644
--- a/src/llama.cpp
+++ b/src/llama.cpp
@@ -21880,3 +21880,290 @@ static void llama_log_callback_default(ggml_log_level level, const char * text,
fputs(text, stderr);
fflush(stderr);
}
+
+static int llama_apply_lora_from_file_internal(
+ const struct llama_model & model, const char * path_lora, float scale, const char * path_base_model, int n_threads
+) {
+ LLAMA_LOG_INFO("%s: applying lora adapter from '%s' - please wait ...\n", __func__, path_lora);
+
+ const int64_t t_start_lora_us = ggml_time_us();
+
+ llama_file fin(path_lora, "rb");
+
+ // verify magic and version
+ {
+ uint32_t magic = fin.read_u32();
+ if (magic != LLAMA_FILE_MAGIC_GGLA) {
+ LLAMA_LOG_ERROR("%s: bad file magic\n", __func__);
+ return 1;
+ }
+
+ uint32_t format_version = fin.read_u32();
+ if (format_version != 1) {
+ LLAMA_LOG_ERROR("%s: unsupported file version\n", __func__ );
+ return 1;
+ }
+ }
+
+ int32_t lora_r = fin.read_u32();
+ int32_t lora_alpha = fin.read_u32();
+ float scaling = scale * (float)lora_alpha / (float)lora_r;
+
+ LLAMA_LOG_INFO("%s: r = %d, alpha = %d, scaling = %.2f\n", __func__, lora_r, lora_alpha, scaling);
+
+ // load base model
+ std::unique_ptr<llama_model_loader> ml;
+ if (path_base_model) {
+ LLAMA_LOG_INFO("%s: loading base model from '%s'\n", __func__, path_base_model);
+ ml.reset(new llama_model_loader(path_base_model, /*use_mmap*/ true, /*check_tensors*/ false, /*kv_overrides*/ nullptr));
+ ml->init_mappings(/*prefetch*/ false); // no prefetching
+ }
+
+ struct tensor_meta {
+ std::string name;
+ ggml_type type;
+ int32_t ne[2];
+ size_t offset;
+ };
+ std::map<std::string, tensor_meta> tensor_meta_map;
+
+ // load all tensor meta
+ while (true) {
+ if (fin.tell() == fin.size) {
+ // eof
+ break;
+ }
+
+ int32_t n_dims;
+ int32_t name_len;
+ int32_t ftype;
+
+ fin.read_raw(&n_dims, sizeof(n_dims));
+ fin.read_raw(&name_len, sizeof(name_len));
+ fin.read_raw(&ftype, sizeof(ftype));
+
+ if (n_dims != 1 && n_dims != 2) {
+ LLAMA_LOG_ERROR("%s: unsupported tensor dimension %d\n", __func__, n_dims);
+ return 1;
+ }
+
+ int32_t ne[2] = { 1, 1 };
+ for (int i = 0; i < n_dims; ++i) {
+ fin.read_raw(&ne[i], sizeof(ne[i]));
+ }
+
+ std::string name;
+ {
+ GGML_ASSERT(name_len < GGML_MAX_NAME);
+ char buf[GGML_MAX_NAME];
+ fin.read_raw(buf, name_len);
+ name = std::string(buf, name_len);
+ }
+
+ // check for lora suffix
+ std::string lora_suffix;
+ if (name.length() > 6) {
+ lora_suffix = name.substr(name.length() - 6);
+ }
+ if (lora_suffix != ".loraA" && lora_suffix != ".loraB") {
+ LLAMA_LOG_ERROR("%s: error: '%s' is not a lora tensor\n", __func__, name.c_str());
+ return 1;
+ }
+
+ // tensor type
+ ggml_type wtype;
+ switch (ftype) {
+ case 0: wtype = GGML_TYPE_F32; break;
+ case 1: wtype = GGML_TYPE_F16; break;
+ default:
+ {
+ LLAMA_LOG_ERROR("%s: invalid tensor data type '%d'\n",
+ __func__, ftype);
+ return 1;
+ }
+ }
+
+ // data offset
+ size_t offset = fin.tell();
+ offset = (offset + 31) & -32;
+
+ // skip tensor data
+ fin.seek(offset + ggml_row_size(wtype, ne[0]) * ne[1], SEEK_SET);
+
+ tensor_meta_map.emplace(name, tensor_meta{ name, wtype, { ne[0], ne[1] }, offset });
+ }
+
+ bool warned = false;
+ int n_tensors = 0;
+
+ // apply
+ ggml_backend_t backend_cpu = ggml_backend_cpu_init();
+ if (backend_cpu == nullptr) {
+ LLAMA_LOG_ERROR("%s: error: failed to initialize cpu backend\n", __func__);
+ return 1;
+ }
+ ggml_backend_cpu_set_n_threads(backend_cpu, n_threads);
+
+ std::vector<no_init<uint8_t>> read_buf;
+ for (const auto & it : model.tensors_by_name) {
+ const std::string & base_name = it.first;
+ ggml_tensor * model_t = it.second;
+
+ if (tensor_meta_map.find(base_name + ".loraA") == tensor_meta_map.end() ||
+ tensor_meta_map.find(base_name + ".loraB") == tensor_meta_map.end()) {
+ continue;
+ }
+
+ tensor_meta & metaA = tensor_meta_map.at(base_name + ".loraA");
+ tensor_meta & metaB = tensor_meta_map.at(base_name + ".loraB");
+
+ ggml_init_params lora_init_params = {
+ /* .mem_size */ ggml_tensor_overhead()*128 + ggml_graph_overhead(),
+ /* .mem_buffer */ nullptr,
+ /* .no_alloc */ true,
+ };
+ ggml_context * lora_ctx = ggml_init(lora_init_params);
+ if (lora_ctx == nullptr) {
+ LLAMA_LOG_ERROR("%s: error: failed to initialize lora context\n", __func__);
+ ggml_backend_free(backend_cpu);
+ return 1;
+ }
+
+ // create tensors
+ ggml_tensor * loraA = ggml_new_tensor_2d(lora_ctx, metaA.type, metaA.ne[0], metaA.ne[1]);
+ ggml_tensor * loraB = ggml_new_tensor_2d(lora_ctx, metaB.type, metaB.ne[0], metaB.ne[1]);
+ ggml_set_name(loraA, metaA.name.c_str());
+ ggml_set_name(loraB, metaB.name.c_str());
+
+ ggml_tensor * base_t;
+ if (ml) {
+ if (!ml->get_tensor_meta(base_name.c_str())) {
+ LLAMA_LOG_ERROR("%s: error: tensor '%s' not found in base model\n", __func__, base_name.c_str());
+ return 1;
+ }
+ base_t = ggml_dup_tensor(lora_ctx, ml->get_tensor_meta(base_name.c_str()));
+ } else {
+ base_t = ggml_dup_tensor(lora_ctx, model_t);
+ }
+ ggml_set_name(base_t, base_name.c_str());
+
+ // allocate in backend buffer
+ ggml_backend_buffer_t lora_buf = ggml_backend_alloc_ctx_tensors_from_buft(lora_ctx, ggml_backend_cpu_buffer_type());
+ if (lora_buf == nullptr) {
+ LLAMA_LOG_ERROR("%s: error: failed to allocate lora tensors\n", __func__);
+ return 1;
+ }
+
+ // load tensor data
+ auto load_tensor = [&read_buf, &fin](const tensor_meta & tensor_meta, ggml_tensor * tensor) {
+ read_buf.resize(ggml_nbytes(tensor));
+ fin.seek(tensor_meta.offset, SEEK_SET);
+ fin.read_raw(read_buf.data(), ggml_nbytes(tensor));
+ ggml_backend_tensor_set(tensor, read_buf.data(), 0, read_buf.size());
+ };
+ load_tensor(metaA, loraA);
+ load_tensor(metaB, loraB);
+
+ // load base model tensor data
+ if (ml) {
+ ml->load_data_for(base_t);
+ } else {
+ ggml_backend_tensor_copy(model_t, base_t);
+ }
+
+ if (ggml_is_quantized(base_t->type) && !warned) {
+ LLAMA_LOG_WARN("%s: warning: using a lora adapter with a quantized model may result in poor quality, "
+ "use a f16 or f32 base model with --lora-base\n", __func__);
+ warned = true;
+ }
+
+ if (base_t->ne[0] != loraA->ne[1] || base_t->ne[1] != loraB->ne[1]) {
+ LLAMA_LOG_ERROR("%s: incompatible tensor dimensions (%" PRId64 " and %" PRId64 ");"
+ " are you sure that this adapter is for this model?\n", __func__, base_t->ne[0], loraA->ne[1]);
+ ggml_free(lora_ctx);
+ ggml_backend_buffer_free(lora_buf);
+ ggml_backend_free(backend_cpu);
+ return 1;
+ }
+
+ auto build_lora_graph = [&]() {
+ // w = w + BA*s
+ ggml_tensor * BA = ggml_mul_mat(lora_ctx, loraA, loraB);
+ ggml_set_name(BA, "BA");
+
+ if (scaling != 1.0f) {
+ BA = ggml_scale(lora_ctx, BA, scaling);
+ ggml_set_name(BA, "BA_scaled");
+ }
+
+ ggml_tensor * r;
+ r = ggml_add_inplace(lora_ctx, base_t, BA);
+ ggml_set_name(r, "r_add");
+
+ if (base_t->type != model_t->type) {
+ // convert the result to the model type
+ r = ggml_cast(lora_ctx, r, model_t->type);
+ ggml_set_name(r, "r_cast");
+ }
+
+ return r;
+ };
+
+ ggml_cgraph * gf = ggml_new_graph(lora_ctx);
+ ggml_tensor * r = build_lora_graph();
+ ggml_build_forward_expand(gf, r);
+
+ ggml_backend_buffer_t graph_buf = ggml_backend_alloc_ctx_tensors_from_buft(lora_ctx, ggml_backend_cpu_buffer_type());
+ if (graph_buf == nullptr) {
+ LLAMA_LOG_ERROR("%s: error: failed to allocate graph tensors\n", __func__);
+ ggml_free(lora_ctx);
+ ggml_backend_buffer_free(lora_buf);
+ ggml_backend_free(backend_cpu);
+ return 1;
+ }
+
+ ggml_backend_graph_compute(backend_cpu, gf);
+
+ ggml_backend_tensor_set(model_t, r->data, 0, ggml_nbytes(r));
+
+#if 0
+ // TODO: use scheduler with fallback to CPU for less copies between CPU and GPU
+ //ggml_backend_sched_t sched = ggml_backend_sched_new(backends.data(), backends.size(), GGML_DEFAULT_GRAPH_SIZE);
+
+ // sched compute
+ ggml_build_forward_expand(gf, build_graph());
+ ggml_backend_sched_init_measure(sched, gf);
+
+ // create the graph again, since the previous one was destroyed by the measure
+ ggml_graph_clear(gf);
+ ggml_build_forward_expand(gf, build_graph());
+ ggml_backend_sched_graph_compute(sched, gf);
+ ggml_backend_sched_free(sched);
+#endif
+
+ ggml_backend_buffer_free(lora_buf);
+ ggml_backend_buffer_free(graph_buf);
+ ggml_free(lora_ctx);
+
+ n_tensors++;
+ if (n_tensors % 4 == 0) {
+ LLAMA_LOG_INFO(".");
+ }
+ }
+
+ ggml_backend_free(backend_cpu);
+
+ const int64_t t_lora_us = ggml_time_us() - t_start_lora_us;
+ LLAMA_LOG_INFO(" done (%.2f ms)\n", t_lora_us / 1000.0);
+
+ return 0;
+}
+
+int32_t llama_model_apply_lora_from_file(const struct llama_model * model, const char * path_lora, float scale, const char * path_base_model, int32_t n_threads) {
+ try {
+ return llama_apply_lora_from_file_internal(*model, path_lora, scale, path_base_model, n_threads);
+ } catch (const std::exception & err) {
+ LLAMA_LOG_ERROR("%s: failed to apply lora adapter: %s\n", __func__, err.what());
+ return 1;
+ }
+}
\ No newline at end of file

43
llm/patches/11-phi3-sliding-window.diff
View file

@ -1,43 +0,0 @@
From 6eedae4cf2fcc8015dac79cb3f28f61fcabacab2 Mon Sep 17 00:00:00 2001
From: Michael Yang <mxyng@pm.me>
Date: Wed, 31 Jul 2024 14:57:04 -0700
Subject: [PATCH] phi3 sliding window
---
src/llama.cpp | 6 +++---
1 file changed, 3 insertions(+), 3 deletions(-)
diff --git a/src/llama.cpp b/src/llama.cpp
index a207451f..f2872d4e 100644
--- a/src/llama.cpp
+++ b/src/llama.cpp
@@ -4893,7 +4893,7 @@ static void llm_load_hparams(
} break;
case LLM_ARCH_PHI3:
{
- ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa);
+ ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa, false);
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
switch (hparams.n_layer) {
@@ -10762,7 +10762,7 @@ struct llm_build_context {
struct ggml_tensor * inp_pos = build_inp_pos();
// KQ_mask (mask for 1 head, it will be broadcasted to all heads)
- struct ggml_tensor * KQ_mask_swa = build_inp_KQ_mask_swa();
+ struct ggml_tensor * KQ_mask = hparams.n_swa > 0 ? build_inp_KQ_mask_swa() : build_inp_KQ_mask();
for (int il = 0; il < n_layer; ++il) {
auto residual = inpL;
@@ -10820,7 +10820,7 @@ struct llm_build_context {
cur = llm_build_kv(ctx0, lctx, kv_self, gf,
model.layers[il].wo, model.layers[il].bo,
- Kcur, Vcur, Qcur, KQ_mask_swa, n_tokens, kv_head, n_kv, 1.0f, cb, il);
+ Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f, cb, il);
}
if (il == n_layer - 1) {
--
2.45.2

7
llm/server.go
View file

@ -98,7 +98,7 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
systemTotalMemory = systemMemInfo.TotalMemory
systemFreeMemory = systemMemInfo.FreeMemory
systemSwapFreeMemory = systemMemInfo.FreeSwap
slog.Debug("system memory", "total", format.HumanBytes2(systemTotalMemory), "free", format.HumanBytes2(systemFreeMemory), "free_swap", format.HumanBytes2(systemSwapFreeMemory))
slog.Info("system memory", "total", format.HumanBytes2(systemTotalMemory), "free", format.HumanBytes2(systemFreeMemory), "free_swap", format.HumanBytes2(systemSwapFreeMemory))
}
// If the user wants zero GPU layers, reset the gpu list to be CPU/system ram info
@ -584,8 +584,7 @@ func (s *llmServer) Ping(ctx context.Context) error {
func (s *llmServer) WaitUntilRunning(ctx context.Context) error {
start := time.Now()
stallDuration := 5 * time.Minute // If no progress happens
finalLoadDuration := 5 * time.Minute // After we hit 100%, give the runner more time to come online
stallDuration := envconfig.LoadTimeout() // If no progress happens
stallTimer := time.Now().Add(stallDuration) // give up if we stall
slog.Info("waiting for llama runner to start responding")
@ -637,7 +636,7 @@ func (s *llmServer) WaitUntilRunning(ctx context.Context) error {
stallTimer = time.Now().Add(stallDuration)
} else if !fullyLoaded && int(s.loadProgress*100.0) >= 100 {
slog.Debug("model load completed, waiting for server to become available", "status", status.ToString())
stallTimer = time.Now().Add(finalLoadDuration)
stallTimer = time.Now().Add(stallDuration)
fullyLoaded = true
}
time.Sleep(time.Millisecond * 250)

14
openai/openai.go
View file

@ -79,7 +79,7 @@ type ChatCompletionRequest struct {
Stop any `json:"stop"`
Temperature *float64 `json:"temperature"`
FrequencyPenalty *float64 `json:"frequency_penalty"`
PresencePenalty *float64 `json:"presence_penalty_penalty"`
PresencePenalty *float64 `json:"presence_penalty"`
TopP *float64 `json:"top_p"`
ResponseFormat *ResponseFormat `json:"response_format"`
Tools []api.Tool `json:"tools"`
@ -452,7 +452,7 @@ func fromChatRequest(r ChatCompletionRequest) (*api.ChatRequest, error) {
}
if r.Temperature != nil {
options["temperature"] = *r.Temperature * 2.0
options["temperature"] = *r.Temperature
} else {
options["temperature"] = 1.0
}
@ -462,11 +462,11 @@ func fromChatRequest(r ChatCompletionRequest) (*api.ChatRequest, error) {
}
if r.FrequencyPenalty != nil {
options["frequency_penalty"] = *r.FrequencyPenalty * 2.0
options["frequency_penalty"] = *r.FrequencyPenalty
}
if r.PresencePenalty != nil {
options["presence_penalty"] = *r.PresencePenalty * 2.0
options["presence_penalty"] = *r.PresencePenalty
}
if r.TopP != nil {
@ -513,7 +513,7 @@ func fromCompleteRequest(r CompletionRequest) (api.GenerateRequest, error) {
}
if r.Temperature != nil {
options["temperature"] = *r.Temperature * 2.0
options["temperature"] = *r.Temperature
} else {
options["temperature"] = 1.0
}
@ -522,9 +522,9 @@ func fromCompleteRequest(r CompletionRequest) (api.GenerateRequest, error) {
options["seed"] = *r.Seed
}
options["frequency_penalty"] = r.FrequencyPenalty * 2.0
options["frequency_penalty"] = r.FrequencyPenalty
options["presence_penalty"] = r.PresencePenalty * 2.0
options["presence_penalty"] = r.PresencePenalty
if r.TopP != 0.0 {
options["top_p"] = r.TopP

48
openai/openai_test.go
View file

@ -22,7 +22,10 @@ const (
image = `iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAQAAAC1HAwCAAAAC0lEQVR42mNk+A8AAQUBAScY42YAAAAASUVORK5CYII=`
)
var False = false
var (
False = false
True = true
)
func captureRequestMiddleware(capturedRequest any) gin.HandlerFunc {
return func(c *gin.Context) {
@ -70,6 +73,44 @@ func TestChatMiddleware(t *testing.T) {
Stream: &False,
},
},
{
name: "chat handler with options",
body: `{
"model": "test-model",
"messages": [
{"role": "user", "content": "Hello"}
],
"stream": true,
"max_tokens": 999,
"seed": 123,
"stop": ["\n", "stop"],
"temperature": 3.0,
"frequency_penalty": 4.0,
"presence_penalty": 5.0,
"top_p": 6.0,
"response_format": {"type": "json_object"}
}`,
req: api.ChatRequest{
Model: "test-model",
Messages: []api.Message{
{
Role: "user",
Content: "Hello",
},
},
Options: map[string]any{
"num_predict": 999.0, // float because JSON doesn't distinguish between float and int
"seed": 123.0,
"stop": []any{"\n", "stop"},
"temperature": 3.0,
"frequency_penalty": 4.0,
"presence_penalty": 5.0,
"top_p": 6.0,
},
Format: "json",
Stream: &True,
},
},
{
name: "chat handler with image content",
body: `{
@ -186,6 +227,8 @@ func TestChatMiddleware(t *testing.T) {
req, _ := http.NewRequest(http.MethodPost, "/api/chat", strings.NewReader(tc.body))
req.Header.Set("Content-Type", "application/json")
defer func() { capturedRequest = nil }()
resp := httptest.NewRecorder()
router.ServeHTTP(resp, req)
@ -202,7 +245,6 @@ func TestChatMiddleware(t *testing.T) {
if !reflect.DeepEqual(tc.err, errResp) {
t.Fatal("errors did not match")
}
capturedRequest = nil
})
}
}
@ -233,7 +275,7 @@ func TestCompletionsMiddleware(t *testing.T) {
Options: map[string]any{
"frequency_penalty": 0.0,
"presence_penalty": 0.0,
"temperature": 1.6,
"temperature": 0.8,
"top_p": 1.0,
"stop": []any{"\n", "stop"},
},

6
scripts/install.sh
View file

@ -38,7 +38,7 @@ IS_WSL2=false
KERN=$(uname -r)
case "$KERN" in
*icrosoft*WSL2 | *icrosoft*wsl2) IS_WSL2=true;;
*icrosoft) error "Microsoft WSL1 is not currently supported. Please upgrade to WSL2 with 'wsl --set-version <distro> 2'" ;;
*icrosoft) error "Microsoft WSL1 is not currently supported. Please use WSL2 with 'wsl --set-version <distro> 2'" ;;
*) ;;
esac
@ -356,12 +356,12 @@ if ! lsmod | grep -q nvidia || ! lsmod | grep -q nvidia_uvm; then
fi
# make sure the NVIDIA modules are loaded on boot with nvidia-persistenced
if command -v nvidia-persistenced > /dev/null 2>&1; then
if available nvidia-persistenced; then
$SUDO touch /etc/modules-load.d/nvidia.conf
MODULES="nvidia nvidia-uvm"
for MODULE in $MODULES; do
if ! grep -qxF "$MODULE" /etc/modules-load.d/nvidia.conf; then
echo "$MODULE" | sudo tee -a /etc/modules-load.d/nvidia.conf > /dev/null
echo "$MODULE" | $SUDO tee -a /etc/modules-load.d/nvidia.conf > /dev/null
fi
done
fi

3
scripts/rh_linux_deps.sh
View file

@ -30,7 +30,7 @@ if grep -i "centos" /etc/system-release >/dev/null; then
dnf install -y rh-git227-git
ln -s /opt/rh/rh-git227/root/usr/bin/git /usr/local/bin/git
fi
dnf install -y devtoolset-10-gcc devtoolset-10-gcc-c++ pigz
dnf install -y devtoolset-10-gcc devtoolset-10-gcc-c++ pigz findutils
elif grep -i "rocky" /etc/system-release >/dev/null; then
# Temporary workaround until rocky 8 AppStream ships GCC 10.4 (10.3 is incompatible with NVCC)
cat << EOF > /etc/yum.repos.d/Rocky-Vault.repo
@ -45,6 +45,7 @@ EOF
dnf install -y git \
gcc-toolset-10-gcc-10.2.1-8.2.el8 \
gcc-toolset-10-gcc-c++-10.2.1-8.2.el8 \
findutils \
pigz
else
echo "ERROR Unexpected distro"

2
server/download.go
View file

@ -256,7 +256,7 @@ func (b *blobDownload) run(ctx context.Context, requestURL *url.URL, opts *regis
continue
}
defer resp.Body.Close()
if resp.StatusCode != http.StatusTemporaryRedirect {
if resp.StatusCode != http.StatusTemporaryRedirect && resp.StatusCode != http.StatusOK {
return nil, fmt.Errorf("unexpected status code %d", resp.StatusCode)
}
return resp.Location()

4
server/routes_create_test.go
View file

@ -593,9 +593,9 @@ func TestCreateDetectTemplate(t *testing.T) {
checkFileExists(t, filepath.Join(p, "blobs", "*"), []string{
filepath.Join(p, "blobs", "sha256-0d79f567714c62c048378f2107fb332dabee0135d080c302d884317da9433cc5"),
filepath.Join(p, "blobs", "sha256-35360843d0c84fb1506952a131bbef13cd2bb4a541251f22535170c05b56e672"),
filepath.Join(p, "blobs", "sha256-553c4a3f747b3d22a4946875f1cc8ed011c2930d83f864a0c7265f9ec0a20413"),
filepath.Join(p, "blobs", "sha256-c608dc615584cd20d9d830363dabf8a4783ae5d34245c3d8c115edb3bc7b28e4"),
filepath.Join(p, "blobs", "sha256-ea34c57ba5b78b740aafe2aeb74dc6507fc3ad14170b64c26a04fb9e36c88d75"),
filepath.Join(p, "blobs", "sha256-de3959f841e9ef6b4b6255fa41cb9e0a45da89c3066aa72bdd07a4747f848990"),
})
})

3
template/alfred.gotmpl
View file

@ -1 +1,2 @@
{{ if .System }}<start_system>{{ .System }}<end_message>{{ end }}{{ if .Prompt }}<start_user>{{ .Prompt }}<end_message>{{ end }}<start_assistant>{{ .Response }}<end_message>
{{- range .Messages }}<start_{{ .Role }}>{{ .Content }}<end_message>
{{- end }}<start_assistant>

20
template/alpaca.gotmpl
View file

@ -1,8 +1,18 @@
{{ if .System }}{{ .System }}
{{- $system := "" }}
{{- range .Messages }}
{{- if eq .Role "system" }}
{{- if not $system }}{{ $system = .Content }}
{{- else }}{{ $system = printf "%s\n\n%s" $system .Content }}
{{- end }}
{{- else if eq .Role "user" }}
{{- if $system }}{{ $system }}
{{ end }}{{ if .Prompt }}### Instruction:
{{ .Prompt }}
{{ $system = "" }}
{{- end }}### Instruction:
{{ .Content }}
{{ end }}### Response:
{{ .Response }}
{{ else if eq .Role "assistant" }}### Response:
{{ .Content }}
{{ end }}
{{- end }}### Response:

7
template/chatml.gotmpl
View file

@ -1,6 +1,3 @@
{{ if .System }}<|im_start|>system
{{ .System }}<|im_end|>
{{ end }}{{ if .Prompt }}<|im_start|>user
{{ .Prompt }}<|im_end|>
{{- range .Messages }}<|im_start|>{{ .Role }}
{{ .Content }}<|im_end|>
{{ end }}<|im_start|>assistant
{{ .Response }}<|im_end|>

11
template/chatqa.gotmpl
View file

@ -1,6 +1,7 @@
{{ if .System }}System: {{ .System }}
{{ end }}{{ if .Prompt }}User: {{ .Prompt }}
{{ end }}Assistant: {{ .Response }}
{{- range .Messages }}
{{- if eq .Role "system" }}System:
{{- else if eq .Role "user" }}User:
{{- else if eq .Role "assistant" }}Assistant:
{{- end }} {{ .Content }}
{{ end }}Assistant:

16
template/codellama-70b-instruct.gotmpl
View file

@ -1,10 +1,10 @@
{{ if .System }}Source: system
{{ .System }} <step> {{ end }}Source: user
{{ .Prompt }} <step> Source: assistant
{{- if not .Response }}
Destination: user
{{- range .Messages }}Source:
{{- if eq .Role "system" }} system
{{- else if eq .Role "user" }} user
{{- else if eq .Role "assistant" }} assistant
{{- end }}
{{ .Response }} <step>
{{ .Content }} <step> {{ end }}Source: assistant
Destination: user

11
template/falcon-instruct.gotmpl
View file

@ -1,5 +1,8 @@
{{ if .System }}System: {{ .System }}
{{ end }}{{ if .Prompt }}User:
{{ .Prompt }}
{{- range .Messages }}
{{- if eq .Role "system" }}System: {{ .Content }}
{{ continue }}
{{- else if eq .Role "user" }}User:
{{- else if eq .Role "assistant" }}Falcon:
{{- end }}
{{ .Content }}
{{ end }}Falcon:
{{ .Response }}

21
template/gemma-instruct.gotmpl
View file

@ -1,5 +1,16 @@
<start_of_turn>user
{{ if .System }}{{ .System }}
{{ end }}{{ .Prompt }}<end_of_turn>
<start_of_turn>model
{{ .Response }}<end_of_turn>
{{- $system := "" }}
{{- range .Messages }}
{{- if eq .Role "system" }}
{{- if not $system }}{{ $system = .Content }}
{{- else }}{{ $system = printf "%s\n\n%s" $system .Content }}
{{- end }}
{{- continue }}
{{- else if eq .Role "user" }}<start_of_turn>user
{{- if $system }}
{{ $system }}
{{- $system = "" }}
{{- end }}
{{- else if eq .Role "assistant" }}<start_of_turn>model
{{- end }}
{{ .Content }}<end_of_turn>
{{ end }}<start_of_turn>model

13
template/granite-instruct.gotmpl
View file

@ -1,9 +1,8 @@
{{ if .System }}System:
{{ .System }}
{{ end }}{{ if .Prompt }}Question:
{{ .Prompt }}
{{- range .Messages }}
{{- if eq .Role "system" }}System:
{{- else if eq .Role "user" }}Question:
{{- else if eq .Role "assistant" }}Answer:
{{- end }}
{{ .Content }}
{{ end }}Answer:
{{ .Response }}

18
template/llama2-chat.gotmpl
View file

@ -1,6 +1,14 @@
[INST] <<SYS>>
{{- if .System }}
{{ .System }}
{{ end }}<</SYS>>
{{- $system := "" }}[INST] {{ range .Messages }}
{{- if eq .Role "system" }}
{{- if not $system }}{{ $system = .Content }}
{{- else }}{{ $system = printf "%s\n\n%s" $system .Content }}
{{- end }}
{{- else if eq .Role "user" }}<<SYS>>
{{- if $system }}
{{ $system }}
{{ $system = "" }}
{{- end }}<</SYS>>
{{ .Prompt }} [/INST] {{ .Response }}</s><s>
{{ .Content }} [/INST]
{{- else if eq .Role "assistant" }} {{ .Content }}</s><s>[INST] {{ end }}
{{- end }}

8
template/llama3-instruct.gotmpl
View file

@ -1,7 +1,5 @@
{{ if .System }}<|start_header_id|>system<|end_header_id|>
{{- range .Messages }}<|start_header_id|>{{ .Role }}<|end_header_id|>
{{ .System }}<|eot_id|>{{ end }}{{ if .Prompt }}<|start_header_id|>user<|end_header_id|>
{{ .Content }}<|eot_id|>
{{- end }}<|start_header_id|>assistant<|end_header_id|>
{{ .Prompt }}<|eot_id|>{{ end }}<|start_header_id|>assistant<|end_header_id|>
{{ .Response }}<|eot_id|>

19
template/magicoder.gotmpl
View file

@ -1,8 +1,17 @@
{{ if .System }}{{ .System }}
{{- $system := "" }}
{{- range .Messages }}
{{- if eq .Role "system" }}
{{- if not $system }}{{ $system = .Content }}
{{- else }}{{ $system = printf "%s\n\n%s" $system .Content }}
{{- end }}
{{- continue }}
{{- else if eq .Role "user" }}
{{- if $system }}{{ $system }}
{{ end }}{{ if .Prompt }}@@ Instruction
{{ .Prompt }}
{{ $system = "" }}
{{- end }}@@ Instruction
{{- else if eq .Role "assistant" }}@@ Response
{{- end }}
{{ .Content }}
{{ end }}@@ Response
{{ .Response }}

7
template/mistral-instruct.gotmpl
View file

@ -1,3 +1,6 @@
[INST] {{ if .System }}{{ .System }}
[INST] {{ range $index, $_ := .Messages }}
{{- if eq .Role "system" }}{{ .Content }}
{{ end }}{{ .Prompt }}[/INST] {{ .Response }}</s>
{{ else if eq .Role "user" }}{{ .Content }}[/INST]
{{- else if eq .Role "assistant" }} {{ .Content }}</s>[INST] {{ end }}
{{- end }}

7
template/openchat.gotmpl
View file

@ -1 +1,6 @@
{{ if .System }}GPT4 Correct System: {{ .System }}<|end_of_turn|>{{ end }}GPT4 Correct User: {{ .Prompt }}<|end_of_turn|>GPT4 Correct Assistant: {{ .Response }}<|end_of_turn|>
{{- range .Messages }}GPT4 Correct
{{- if eq .Role "system" }} System:
{{- else if eq .Role "user" }} User:
{{- else if eq .Role "assistant" }} Assistant:
{{- end }} {{ .Content }}<|end_of_turn|>
{{- end }}GPT4 Correct Assistant:

7
template/phi-3.gotmpl
View file

@ -1,6 +1,3 @@
{{ if .System }}<|system|>
{{ .System }}<|end|>
{{ end }}{{ if .Prompt }}<|user|>
{{ .Prompt }}<|end|>
{{- range .Messages }}<|{{ .Role }}|>
{{ .Content }}<|end|>
{{ end }}<|assistant|>
{{ .Response }}<|end|>

14
template/solar-instruct.gotmpl
View file

@ -1,9 +1,11 @@
{{ if .System }}### System:
{{ .System }}
{{- range .Messages }}
{{- if eq .Role "system" }}### System:
{{- else if eq .Role "user" }}### User:
{{- else if eq .Role "assistant" }}### Assistant:
{{ .Content }}</s>
{{ end }}{{ if .Prompt }}### User:
{{ .Prompt }}
{{ continue }}
{{- end }}
{{ .Content }}
{{ end }}### Assistant:
{{ .Response }}</s>

20
template/starcoder2-instruct.gotmpl
View file

@ -1,8 +1,18 @@
{{ if .System }}{{ .System }}
{{- $system := "" }}
{{- range .Messages }}
{{- if eq .Role "system" }}
{{- if not $system }}{{ $system = .Content }}
{{- else }}{{ $system = printf "%s\n\n%s" $system .Content }}
{{- end }}
{{- else if eq .Role "user" }}
{{- if $system }}{{ $system }}
{{ end }}{{ if .Prompt }}### Instruction
{{ .Prompt }}
{{ $system = "" }}
{{- end }}### Instruction
{{ .Content }}
{{ end }}### Response
{{ .Response }}<|endoftext|>
{{ else if eq .Role "assistant" }}### Response
{{ .Content }}<|endoftext|>
{{ end }}
{{- end }}### Response

16
template/vicuna.gotmpl
View file

@ -1,4 +1,14 @@
{{ if .System }}{{ .System }}
{{- $system := "" }}
{{- range .Messages }}
{{- if eq .Role "system" }}
{{- if not $system }}{{ $system = .Content }}
{{- else }}{{ $system = printf "%s\n\n%s" $system .Content }}
{{- end }}
{{- else if eq .Role "user" }}
{{- if $system }}{{ $system }}
{{ end }}{{ if .Prompt }}USER: {{ .Prompt }}
{{ end }}ASSISTANT: {{ .Response }}</s>
{{ $system = "" }}
{{- end }}USER: {{ .Content }}
{{ else if eq .Role "assistant" }}ASSISTANT: {{ .Content }}</s>
{{ end }}
{{- end }}ASSISTANT:

7
template/zephyr.gotmpl
View file

@ -1,6 +1,3 @@
{{ if .System }}<|system|>
{{ .System }}</s>
{{ end }}{{ if .Prompt }}<|user|>
{{ .Prompt }}</s>
{{- range .Messages }}<|{{ .Role }}|>
{{ .Content }}</s>
{{ end }}<|assistant|>
{{ .Response }}</s>