ollama/template/template.go
Daniel Hiltgen 9db4bdbad6
runner: Remove CGO engines, use llama-server exclusively for GGML models (#16031)
* broad lint fixes to sidestep CI scope glitch

* runner: Remove CGO engines, use llama-server exclusively for GGML models

Remove the vendored GGML and llama.cpp backend, CGO runner, Go model
implementations, and sample.  llama-server (built from upstream llama.cpp via
FetchContent) is now the sole inference engine for GGUF-based models.
(Safetensor based models continue to run on the new MLX engine.)  This allows
us to more rapidly pick up new capabilities and fixes from llama.cpp as they
come out.

On windows this now requires recent AMD driver versions to support ROCm v7 as
llama.cpp currently does not support building against v6.

* llama/compat: load Ollama-format GGUFs in llama-server

Squashed from upstream/jmorganca/llama-compat on 2026-04-29.
Source tip: 0c33775d37.

Original source commits:
- 25223160d llama/compat: add in-memory shim so llama-server can load Ollama-format GGUFs
- 7449b539a llm,server: route Ollama-format gemma3 blobs through llama/compat
- 436f2e2b1 llama/compat: make patch-apply idempotent
- 8c2c9d4c8 llama/compat: extend gemma3 handler to cover 1B and 270M blobs
- 021389f7b llama/compat: shrink clip.cpp injection from 18 lines to 1
- 61b367ec2 llama/compat: shrink patch to pure call-site hooks (34 -> 20 lines)
- 36049361c llama/compat: simplify shim (gemma3-tested)
- 8fa664865 llama/compat: add qwen35moe text handler
- db0c74530 llama/compat: add qwen35moe vision (clip) support
- 2a388da77 llama/compat: split shared infra into a util TU
- 9a69a17dc llama/compat: document non-public API dependencies
- d0f38a915 llama/compat: add gpt-oss and lfm2 handlers
- 086071822 llama/compat: add mistral3 text handler (vision TODO)
- 63bde9ff7 llama/compat: add mistral3 vision (clip) support
- 3a57b89d5 llama/compat: apply LLaMA RoPE permute to mistral3 vision Q/K
- 99cb87439 llama/compat: add qwen35, gemma4, deepseek-ocr handlers
- 2c7850dba llama/compat: add nemotron_h_moe handler (latent FFN + MTP skip)
- 9e3b54225 llama/compat: add llama4 text + clip handlers
- 034fee349 llama/compat: add gemma4 clip handler (gemma4v projector)
- 9945c5a93 server: remove dhiltgen/* compat redirect table
- 5d4539101 llama/compat: rewrite gemma4 tokenizer model to BPE
- 7e0765327 llama/compat: add glm-ocr text handler + text-loader load-op hook
- f1bd1a25a llama/compat: add glm-ocr clip handler (glm4v projector)
- 4b5cf3420 llama/compat: collapse text-loader hook back to one new patch line
- eb4ecf4fc llama/compat: extend gemma4 clip handler to gemma4a (audio)
- a23a5e76f llama/compat: fix gemma4a per-block norm tensor mapping
- cd2dcaff4 llama/compat: add embeddinggemma handler
- 1ce8a6b26 llama/compat: add qwen3-vl + qwen2.5-vl handlers
- fd98ffa1e llama/compat: add gemma3n + glm4moelite handlers
- cc7bdf0bc llama/compat: handle null buft in maybe_load_tensor
- 0c33775d3 llama/compat: disable mmap when load_op transforms text-side tensors

* refine implementation

* ci: fix windows MLX build

* ci: fix windows llama-server build

* ci: fix windows rocm build

* ci: windows mlx tuning

Shorten long-tail on build, and get OllamaSetup.exe back under 2g limit

* ci: fix windows dependencies

* win: fix dependency gathering

* disable openmp

* win: arm64 cross-compile build

also DRY out CI steps

* scheduler improvements

* ci: improvements from #15982

* win: favor ninja for faster developer builds

* win: fix build

* win: fix arm64 cross-compile

* win: avoid spaces in compiler path

* misc discovery fixes, and bos handling

* lint fixes

* win: fix arm cross-compile build/CI bugs

* llama.cpp update

* win: handle multiple CRT dirs

* vulkan: add windows iGPU detection

* fix creation bugs for patched models, other refactoring work

* tune batch size for better performance

* ci and lint fixes

* fix repeat_last_n bug

* build: revamp build for better developer UX

* amd, sampler, qwen3next fixes

* version bump

* fix mlx build

* revamp GPU discovery

Scanning the output of llama-server is turning out to be too error prone across
llama.cpp updates, so this switches to a thin dynamic library load against the
bundled GGML libraries so more details can be gathered from the API.

* version bump

* missing file

* ci: fix cache miss on rocm build

* refine vulkan dep handling

* fix ps reporting bug on full GPU load

* improve cmake wiring for customized local builds

* version bump

* docker build arg cleanup

* improve windows exit error logs

* fix community gemma4 support and ci flakes

* fix mlx unit test

* tighten up ps logic to avoid double counting fit log lines

* version bump

* fix ps view for full gpu layer offload

* add MTP wiring for llama-server and create with GGUFs

* pick best template by capabilities

* version bump

* ci: harden apt repos

* remove unused cpu core discovery

* adjust batch default logic to reduce OOMs

* support larger tool calls

* fix audio support, template show

* qwen35 mtp patch support

* flesh out dtypes

* rocm deps

* version bump

* lint fix

* block broken gfx1150 on windows

* fix qwen3.5 moe mtp tensors in patch

* mmproj oom fallback and vulkan on by default

* qwen MTP compat fix

* version bump

* ci: fix WoA cross-compile

* ci: workaround ui tool in cross-compile

* version bump

* win: enable OpenMP for CPU builds

* build: improve developer UX

* ci: windows path workaround for CPU build

* win: fix WoA dependencies

* win: fix large offset reads for mmproj patched loads

* version bump

* fix vulkan dup detection

* add OLLAMA_IGPU_ENABLE and largely disable iGPUs by default

* opt-in MTP, win large offset, integraton fixes

* fix unit test scheduler interaction hang

* fix multi-gpu filtering

* version bump

* review comments

* fix thinking level

* fix linux rocm ordering and granite 3.3 template

* version bump

* ci fix - non-shallow MLX checkout

* bypass linux sysfs unit test on windows

---------

Co-authored-by: jmorganca <jmorganca@gmail.com>
2026-05-29 13:35:47 -07:00

651 lines
15 KiB
Go

package template
import (
"bytes"
"embed"
"encoding/json"
"errors"
"io"
"maps"
"math"
"slices"
"strings"
"sync"
"text/template"
"text/template/parse"
"time"
"github.com/agnivade/levenshtein"
"github.com/ollama/ollama/api"
)
//go:embed index.json
var indexBytes []byte
//go:embed *.gotmpl
//go:embed *.json
var templatesFS embed.FS
var templatesOnce = sync.OnceValues(func() ([]*named, error) {
var templates []*named
if err := json.Unmarshal(indexBytes, &templates); err != nil {
return nil, err
}
for _, t := range templates {
bts, err := templatesFS.ReadFile(t.Name + ".gotmpl")
if err != nil {
return nil, err
}
// normalize line endings
t.Bytes = bytes.ReplaceAll(bts, []byte("\r\n"), []byte("\n"))
params, err := templatesFS.ReadFile(t.Name + ".json")
if err != nil {
continue
}
if err := json.Unmarshal(params, &t.Parameters); err != nil {
return nil, err
}
}
return templates, nil
})
type named struct {
Name string `json:"name"`
Template string `json:"template"`
Bytes []byte
Parameters *struct {
Stop []string `json:"stop"`
}
}
func (t named) Reader() io.Reader {
return bytes.NewReader(t.Bytes)
}
func Named(s string) (*named, error) {
templates, err := templatesOnce()
if err != nil {
return nil, err
}
var template *named
score := math.MaxInt
for _, t := range templates {
if s := levenshtein.ComputeDistance(s, t.Template); s < score {
score = s
template = t
}
}
if score < 100 {
return template, nil
}
return nil, errors.New("no matching template found")
}
var DefaultTemplate, _ = Parse("{{ .Prompt }}")
type Template struct {
*template.Template
raw string
}
// response is a template node that can be added to templates that don't already have one
var response = parse.ActionNode{
NodeType: parse.NodeAction,
Pipe: &parse.PipeNode{
NodeType: parse.NodePipe,
Cmds: []*parse.CommandNode{
{
NodeType: parse.NodeCommand,
Args: []parse.Node{
&parse.FieldNode{
NodeType: parse.NodeField,
Ident: []string{"Response"},
},
},
},
},
},
}
var funcs = template.FuncMap{
"json": func(v any) string {
b, _ := json.Marshal(v)
return string(b)
},
"currentDate": func(args ...string) string {
// Currently ignoring the format argument, but accepting it for future use
// Default format is YYYY-MM-DD
return time.Now().Format("2006-01-02")
},
"yesterdayDate": func(args ...string) string {
return time.Now().AddDate(0, 0, -1).Format("2006-01-02")
},
"toTypeScriptType": func(v any) string {
if param, ok := v.(api.ToolProperty); ok {
return param.ToTypeScriptType()
}
// Handle pointer case
if param, ok := v.(*api.ToolProperty); ok && param != nil {
return param.ToTypeScriptType()
}
return "any"
},
}
func Parse(s string) (*Template, error) {
tmpl := template.New("").Option("missingkey=zero").Funcs(funcs)
tmpl, err := tmpl.Parse(s)
if err != nil {
return nil, err
}
t := Template{Template: tmpl, raw: s}
vars, err := t.Vars()
if err != nil {
return nil, err
}
if !slices.Contains(vars, "messages") && !slices.Contains(vars, "response") {
// touch up the template and append {{ .Response }}
tmpl.Tree.Root.Nodes = append(tmpl.Tree.Root.Nodes, &response)
}
return &t, nil
}
func (t *Template) String() string {
return t.raw
}
func (t *Template) Vars() ([]string, error) {
var vars []string
for _, tt := range t.Templates() {
for _, n := range tt.Root.Nodes {
v, err := Identifiers(n)
if err != nil {
return vars, err
}
vars = append(vars, v...)
}
}
set := make(map[string]struct{})
for _, n := range vars {
set[strings.ToLower(n)] = struct{}{}
}
return slices.Sorted(maps.Keys(set)), nil
}
func (t *Template) Contains(s string) bool {
return strings.Contains(t.raw, s)
}
type Values struct {
Messages []api.Message
api.Tools
Prompt string
Suffix string
Think bool
// ThinkLevel contains the thinking level if Think is true and a string value was provided
ThinkLevel string
// whether or not the user explicitly set the thinking flag (vs. it being
// implicitly false). Templates can't see whether `Think` is nil
IsThinkSet bool
// forceLegacy is a flag used to test compatibility with legacy templates
forceLegacy bool
}
func (t *Template) Subtree(fn func(parse.Node) bool) *template.Template {
var walk func(parse.Node) parse.Node
walk = func(n parse.Node) parse.Node {
if fn(n) {
return n
}
switch t := n.(type) {
case *parse.ListNode:
for _, c := range t.Nodes {
if n := walk(c); n != nil {
return n
}
}
case *parse.BranchNode:
for _, n := range []*parse.ListNode{t.List, t.ElseList} {
if n != nil {
if n := walk(n); n != nil {
return n
}
}
}
case *parse.IfNode:
return walk(&t.BranchNode)
case *parse.WithNode:
return walk(&t.BranchNode)
case *parse.RangeNode:
return walk(&t.BranchNode)
}
return nil
}
if n := walk(t.Tree.Root); n != nil {
return (&template.Template{
Tree: &parse.Tree{
Root: &parse.ListNode{
Nodes: []parse.Node{n},
},
},
}).Funcs(funcs)
}
return nil
}
func (t *Template) Execute(w io.Writer, v Values) error {
system, messages := collate(v.Messages)
vars, err := t.Vars()
if err != nil {
return err
}
if v.Prompt != "" && v.Suffix != "" {
return t.Template.Execute(w, map[string]any{
"Prompt": v.Prompt,
"Suffix": v.Suffix,
"Response": "",
"Think": v.Think,
"ThinkLevel": v.ThinkLevel,
"IsThinkSet": v.IsThinkSet,
})
} else if !v.forceLegacy && slices.Contains(vars, "messages") {
return t.Template.Execute(w, map[string]any{
"System": system,
"Messages": convertMessagesForTemplate(messages),
"Tools": convertToolsForTemplate(v.Tools),
"Response": "",
"Think": v.Think,
"ThinkLevel": v.ThinkLevel,
"IsThinkSet": v.IsThinkSet,
})
}
system = ""
var b bytes.Buffer
var prompt, response string
for _, m := range messages {
execute := func() error {
if err := t.Template.Execute(&b, map[string]any{
"System": system,
"Prompt": prompt,
"Response": response,
"Think": v.Think,
"ThinkLevel": v.ThinkLevel,
"IsThinkSet": v.IsThinkSet,
}); err != nil {
return err
}
system = ""
prompt = ""
response = ""
return nil
}
switch m.Role {
case "system":
if prompt != "" || response != "" {
if err := execute(); err != nil {
return err
}
}
system = m.Content
case "user":
if response != "" {
if err := execute(); err != nil {
return err
}
}
prompt = m.Content
case "assistant":
response = m.Content
}
}
var cut bool
nodes := deleteNode(t.Template.Root.Copy(), func(n parse.Node) bool {
if field, ok := n.(*parse.FieldNode); ok && slices.Contains(field.Ident, "Response") {
cut = true
return false
}
return cut
})
tree := parse.Tree{Root: nodes.(*parse.ListNode)}
if err := template.Must(template.New("").AddParseTree("", &tree)).Execute(&b, map[string]any{
"System": system,
"Prompt": prompt,
"Response": response,
"Think": v.Think,
"ThinkLevel": v.ThinkLevel,
"IsThinkSet": v.IsThinkSet,
}); err != nil {
return err
}
_, err = io.Copy(w, &b)
return err
}
// collate messages based on role. consecutive messages of the same role are merged
// into a single message (except for tool messages which preserve individual metadata).
// collate also collects and returns all system messages.
// collate mutates message content adding image tags ([img-%d]) as needed
// todo(parthsareen): revisit for contextual image support
func collate(msgs []api.Message) (string, []*api.Message) {
var system []string
var collated []*api.Message
for i := range msgs {
if msgs[i].Role == "system" {
system = append(system, msgs[i].Content)
}
// merges consecutive messages of the same role into a single message (except for tool messages)
if len(collated) > 0 && collated[len(collated)-1].Role == msgs[i].Role && msgs[i].Role != "tool" {
collated[len(collated)-1].Content += "\n\n" + msgs[i].Content
} else {
collated = append(collated, &msgs[i])
}
}
return strings.Join(system, "\n\n"), collated
}
// templateTools is a slice of templateTool that marshals to JSON.
type templateTools []templateTool
func (t templateTools) String() string {
bts, _ := json.Marshal(t)
return string(bts)
}
// templateArgs is a map type with JSON string output for templates.
type templateArgs map[string]any
func (t templateArgs) String() string {
if t == nil {
return "{}"
}
bts, _ := json.Marshal(t)
return string(bts)
}
// templateProperties is a map type with JSON string output for templates.
type templateProperties map[string]api.ToolProperty
func (t templateProperties) String() string {
if t == nil {
return "{}"
}
bts, _ := json.Marshal(t)
return string(bts)
}
func (t templateTool) String() string {
bts, _ := json.Marshal(t)
return string(bts)
}
// templateTool is a template-compatible representation of api.Tool
// with Properties as a regular map for template ranging.
type templateTool struct {
Type string `json:"type"`
Items any `json:"items,omitempty"`
Function templateToolFunction `json:"function"`
}
type templateToolFunction struct {
Name string `json:"name"`
Description string `json:"description"`
Parameters templateToolFunctionParameters `json:"parameters"`
}
type templateToolFunctionParameters struct {
Type string `json:"type"`
Defs any `json:"$defs,omitempty"`
Items any `json:"items,omitempty"`
Required []string `json:"required,omitempty"`
Properties templateProperties `json:"properties"`
}
// templateToolCall is a template-compatible representation of api.ToolCall
// with Arguments as a regular map for template ranging.
type templateToolCall struct {
ID string
Function templateToolCallFunction
}
type templateToolCallFunction struct {
Index int
Name string
Arguments templateArgs
}
// templateMessage is a template-compatible representation of api.Message
// with ToolCalls converted for template use.
type templateMessage struct {
Role string
Content string
Thinking string
Images []api.ImageData
ToolCalls []templateToolCall
ToolName string
ToolCallID string
}
// convertToolsForTemplate converts Tools to template-compatible format.
func convertToolsForTemplate(tools api.Tools) templateTools {
if tools == nil {
return nil
}
result := make(templateTools, len(tools))
for i, tool := range tools {
result[i] = templateTool{
Type: tool.Type,
Items: tool.Items,
Function: templateToolFunction{
Name: tool.Function.Name,
Description: tool.Function.Description,
Parameters: templateToolFunctionParameters{
Type: tool.Function.Parameters.Type,
Defs: tool.Function.Parameters.Defs,
Items: tool.Function.Parameters.Items,
Required: tool.Function.Parameters.Required,
Properties: templateProperties(tool.Function.Parameters.Properties.ToMap()),
},
},
}
}
return result
}
// convertMessagesForTemplate converts Messages to template-compatible format.
func convertMessagesForTemplate(messages []*api.Message) []*templateMessage {
if messages == nil {
return nil
}
result := make([]*templateMessage, len(messages))
for i, msg := range messages {
var toolCalls []templateToolCall
for _, tc := range msg.ToolCalls {
toolCalls = append(toolCalls, templateToolCall{
ID: tc.ID,
Function: templateToolCallFunction{
Index: tc.Function.Index,
Name: tc.Function.Name,
Arguments: templateArgs(tc.Function.Arguments.ToMap()),
},
})
}
result[i] = &templateMessage{
Role: msg.Role,
Content: msg.Content,
Thinking: msg.Thinking,
Images: msg.Images,
ToolCalls: toolCalls,
ToolName: msg.ToolName,
ToolCallID: msg.ToolCallID,
}
}
return result
}
// Identifiers walks the node tree returning any identifiers it finds along the way
func Identifiers(n parse.Node) ([]string, error) {
switch n := n.(type) {
case *parse.ListNode:
var names []string
for _, n := range n.Nodes {
i, err := Identifiers(n)
if err != nil {
return names, err
}
names = append(names, i...)
}
return names, nil
case *parse.TemplateNode:
if n.Pipe == nil {
return nil, errors.New("undefined template specified")
}
return Identifiers(n.Pipe)
case *parse.ActionNode:
if n.Pipe == nil {
return nil, errors.New("undefined action in template")
}
return Identifiers(n.Pipe)
case *parse.BranchNode:
if n.Pipe == nil {
return nil, errors.New("undefined branch")
}
names, err := Identifiers(n.Pipe)
if err != nil {
return names, err
}
for _, n := range []*parse.ListNode{n.List, n.ElseList} {
if n != nil {
i, err := Identifiers(n)
if err != nil {
return names, err
}
names = append(names, i...)
}
}
return names, nil
case *parse.IfNode:
return Identifiers(&n.BranchNode)
case *parse.RangeNode:
return Identifiers(&n.BranchNode)
case *parse.WithNode:
return Identifiers(&n.BranchNode)
case *parse.PipeNode:
var names []string
for _, c := range n.Cmds {
for _, a := range c.Args {
i, err := Identifiers(a)
if err != nil {
return names, err
}
names = append(names, i...)
}
}
return names, nil
case *parse.FieldNode:
return n.Ident, nil
case *parse.VariableNode:
return n.Ident, nil
}
return nil, nil
}
// deleteNode walks the node list and deletes nodes that match the predicate
// this is currently to remove the {{ .Response }} node from templates
func deleteNode(n parse.Node, fn func(parse.Node) bool) parse.Node {
var walk func(n parse.Node) parse.Node
walk = func(n parse.Node) parse.Node {
if fn(n) {
return nil
}
switch t := n.(type) {
case *parse.ListNode:
var nodes []parse.Node
for _, c := range t.Nodes {
if n := walk(c); n != nil {
nodes = append(nodes, n)
}
}
t.Nodes = nodes
return t
case *parse.IfNode:
t.BranchNode = *(walk(&t.BranchNode).(*parse.BranchNode))
case *parse.WithNode:
t.BranchNode = *(walk(&t.BranchNode).(*parse.BranchNode))
case *parse.RangeNode:
t.BranchNode = *(walk(&t.BranchNode).(*parse.BranchNode))
case *parse.BranchNode:
t.List = walk(t.List).(*parse.ListNode)
if t.ElseList != nil {
t.ElseList = walk(t.ElseList).(*parse.ListNode)
}
case *parse.ActionNode:
n := walk(t.Pipe)
if n == nil {
return nil
}
t.Pipe = n.(*parse.PipeNode)
case *parse.PipeNode:
var commands []*parse.CommandNode
for _, c := range t.Cmds {
var args []parse.Node
for _, a := range c.Args {
if n := walk(a); n != nil {
args = append(args, n)
}
}
if len(args) == 0 {
return nil
}
c.Args = args
commands = append(commands, c)
}
if len(commands) == 0 {
return nil
}
t.Cmds = commands
}
return n
}
return walk(n)
}