ollama/harmony/harmonyparser.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

556 lines
17 KiB
Go

package harmony
import (
"encoding/json"
"fmt"
"log/slog"
"strings"
"unicode"
"github.com/ollama/ollama/api"
"github.com/ollama/ollama/logutil"
)
type harmonyParserState int
const (
harmonyParserState_LookingForMessageStart harmonyParserState = iota
harmonyParserState_ParsingHeader
harmonyParserState_ParsingContent
)
func (s harmonyParserState) String() string {
switch s {
// we're looking for the message start tag
case harmonyParserState_LookingForMessageStart:
return "LookingForMessageStart"
case harmonyParserState_ParsingHeader:
return "ParsingHeader"
case harmonyParserState_ParsingContent:
return "ParsingContent"
default:
return "Unknown"
}
}
type HarmonyParser struct {
state harmonyParserState
MessageStartTag string
MessageEndTag string
HeaderEndTag string
acc strings.Builder
lifetimeAcc strings.Builder
}
type HarmonyEvent interface {
isHarmonyEvent()
}
type HarmonyEventMessageStart struct{}
func (HarmonyEventMessageStart) isHarmonyEvent() {}
type HarmonyEventHeaderComplete struct {
Header HarmonyHeader
}
func (HarmonyEventHeaderComplete) isHarmonyEvent() {}
type HarmonyEventContentEmitted struct {
Content string
}
func (HarmonyEventContentEmitted) isHarmonyEvent() {}
type HarmonyEventMessageEnd struct{}
func (HarmonyEventMessageEnd) isHarmonyEvent() {}
type HarmonyHeader struct {
Role string
Channel string
Recipient string
}
func (s *HarmonyParser) AddImplicitStart() {
s.acc.WriteString("<|start|>assistant")
}
func (s *HarmonyParser) AddImplicitStartOrPrefill(lastMessage *api.Message) {
if lastMessage != nil && lastMessage.Role == "assistant" {
// handle prefilling conditions
if lastMessage.Content != "" {
s.acc.WriteString("<|start|>assistant<|channel|>final<|message|>")
return
} else if lastMessage.Thinking != "" {
s.acc.WriteString("<|start|>assistant<|channel|>analysis<|message|>")
return
}
}
s.AddImplicitStart()
}
func (s *HarmonyParser) AddContent(content string) []HarmonyEvent {
s.lifetimeAcc.WriteString(content)
s.acc.WriteString(content)
var events []HarmonyEvent
keepLooping := true
// we loop because we might pass through multiple parsing states in a single
// call to addContent, and we want to make sure callers don't have to wait for
// data that's already unambiguous
for keepLooping {
var newEvents []HarmonyEvent
newEvents, keepLooping = eat(s)
events = append(events, newEvents...)
}
return events
}
// the additional bool return is true iff we should continue eating
func eat(s *HarmonyParser) ([]HarmonyEvent, bool) {
switch s.state {
case harmonyParserState_LookingForMessageStart:
// does the acc contain the message start tag?
if strings.Contains(s.acc.String(), s.MessageStartTag) {
// split the acc into the message start tag and the rest
split := strings.SplitN(s.acc.String(), s.MessageStartTag, 2)
before := split[0]
if before != "" {
slog.Warn("harmony parser: found message start tag in the middle of the content", "content", s.acc.String())
}
after := split[1]
s.acc.Reset()
s.acc.WriteString(after)
s.state = harmonyParserState_ParsingHeader
return []HarmonyEvent{HarmonyEventMessageStart{}}, true
}
// no match, so we keep accumulating
return nil, false
case harmonyParserState_ParsingHeader:
if strings.Contains(s.acc.String(), s.HeaderEndTag) {
split := strings.SplitN(s.acc.String(), s.HeaderEndTag, 2)
header := split[0]
after := split[1]
s.acc.Reset()
s.acc.WriteString(after)
s.state = harmonyParserState_ParsingContent
return []HarmonyEvent{HarmonyEventHeaderComplete{Header: s.parseHeader(header)}}, true
}
return nil, false
case harmonyParserState_ParsingContent:
if strings.Contains(s.acc.String(), s.MessageEndTag) {
// if we already have the message end tag, we can emit the content up to it
split := strings.SplitN(s.acc.String(), s.MessageEndTag, 2)
content := split[0]
after := split[1]
s.acc.Reset()
s.acc.WriteString(after)
s.state = harmonyParserState_LookingForMessageStart
events := []HarmonyEvent{}
if content != "" {
events = append(events, HarmonyEventContentEmitted{Content: content})
}
events = append(events, HarmonyEventMessageEnd{})
return events, true
} else if overlapLen := overlap(s.acc.String(), s.MessageEndTag); overlapLen > 0 {
// if our suffix contains the start of the message end tag, we can emit
// the content up to the start of the message end tag
content := s.acc.String()[:len(s.acc.String())-overlapLen]
remaining := s.acc.String()[len(s.acc.String())-overlapLen:]
s.acc.Reset()
s.acc.WriteString(remaining)
// emit the content we know isn't part of the message end tag, and keep
// accumulating to disambiguate the rest
if content == "" {
return nil, false
}
return []HarmonyEvent{HarmonyEventContentEmitted{Content: content}}, false
} else {
// no end tag, so it's still normal content that we can immediately emit
content := s.acc.String()
if content == "" {
return nil, false
}
s.acc.Reset()
return []HarmonyEvent{HarmonyEventContentEmitted{Content: content}}, false
}
}
return nil, false
}
func (s *HarmonyParser) parseHeader(raw string) HarmonyHeader {
harmonyHeader := HarmonyHeader{}
// if `<|constrain|>` is present, ensure it has a space before it so it gets
// parsed as a separate token, even if the model didn't include the space
if strings.Contains(raw, "<|constrain|>") {
raw = strings.Replace(raw, "<|constrain|>", " <|constrain|>", 1)
raw = strings.TrimSpace(raw)
}
// look for the optional channel tag, which is `<|channel|>` followed by the
// channel name, all without any whitespace
channelIndex := strings.Index(raw, "<|channel|>")
if channelIndex != -1 {
before := raw[:channelIndex]
after := raw[channelIndex+len("<|channel|>"):]
// the channel name is `after` all the way up to the first (if any) whitespace character
idx := strings.IndexFunc(after, func(r rune) bool {
return unicode.IsSpace(r)
})
if idx == -1 {
idx = len(after)
}
harmonyHeader.Channel = after[:idx]
after = after[idx:]
// now we remove the channel tag from the raw string to further process
raw = before + after
raw = strings.TrimSpace(raw)
}
// split the header into whitespace-separated tokens
tokens := strings.Fields(raw)
// the first token is treated as the role
if len(tokens) == 0 {
slog.Error("harmony parser: missing role in header", "header", raw)
return harmonyHeader
}
role := tokens[0]
tokens = tokens[1:]
// special case: if role starts with to= then it's a tool call
if strings.HasPrefix(role, "to=") {
harmonyHeader.Recipient = role[3:]
harmonyHeader.Role = "tool"
} else {
harmonyHeader.Role = role
}
// the recipient (if any) can be specified before or after the channel tag, so
// we check it at the end once we've already parsed the channel and role
if harmonyHeader.Recipient == "" && len(tokens) > 0 && strings.HasPrefix(tokens[0], "to=") {
harmonyHeader.Recipient = tokens[0][3:]
}
return harmonyHeader
}
// longest overlap between suffix of s and prefix of delim
func overlap(s, delim string) int {
max := min(len(delim), len(s))
for i := max; i > 0; i-- {
if strings.HasSuffix(s, delim[:i]) {
return i
}
}
return 0
}
// harmonyMessageState represents the current state of message processing
type harmonyMessageState int
const (
harmonyMessageState_Normal harmonyMessageState = iota
harmonyMessageState_Thinking
harmonyMessageState_ToolCalling
)
// HarmonyMessageHandler processes harmony events and accumulates content appropriately.
// This is a higher level interface that maps harmony concepts into ollama concepts
type HarmonyMessageHandler struct {
state harmonyMessageState
HarmonyParser *HarmonyParser
FunctionNameMap *FunctionNameMap
toolAccumulator *HarmonyToolCallAccumulator
convertedTools map[string]struct{}
}
// NewHarmonyMessageHandler creates a new message handler
func NewHarmonyMessageHandler() *HarmonyMessageHandler {
return &HarmonyMessageHandler{
state: harmonyMessageState_Normal,
HarmonyParser: &HarmonyParser{
MessageStartTag: "<|start|>",
MessageEndTag: "<|end|>",
HeaderEndTag: "<|message|>",
},
FunctionNameMap: NewFunctionNameMap(),
convertedTools: make(map[string]struct{}),
}
}
// AddContent processes the content and returns the content, thinking, and tool content.
// content and thinking are already fully parsed, but tool content still needs to be passed to the tool parser
func (h *HarmonyMessageHandler) AddContent(content string, toolParser *HarmonyToolCallAccumulator) (string, string, string) {
contentSb := strings.Builder{}
thinkingSb := strings.Builder{}
toolContentSb := strings.Builder{}
events := h.HarmonyParser.AddContent(content)
for _, event := range events {
switch event := event.(type) {
case HarmonyEventHeaderComplete:
logutil.Trace("harmony event header complete", "header", event.Header)
switch event.Header.Channel {
case "analysis":
if event.Header.Recipient != "" {
h.state = harmonyMessageState_ToolCalling
// event.Header.Recipient is the tool name, something like
// "browser.search" for a built-in, or "functions.calc" for a
// custom one
toolParser.SetToolName(event.Header.Recipient)
} else {
h.state = harmonyMessageState_Thinking
}
case "commentary":
if event.Header.Recipient != "" {
h.state = harmonyMessageState_ToolCalling
toolParser.SetToolName(event.Header.Recipient)
} else {
h.state = harmonyMessageState_Normal
}
case "final":
h.state = harmonyMessageState_Normal
}
case HarmonyEventContentEmitted:
logutil.Trace("harmony event content", "content", event.Content, "state", h.state)
if h.state == harmonyMessageState_Normal {
contentSb.WriteString(event.Content)
} else if h.state == harmonyMessageState_Thinking {
thinkingSb.WriteString(event.Content)
} else if h.state == harmonyMessageState_ToolCalling {
toolContentSb.WriteString(event.Content)
}
case HarmonyEventMessageEnd:
h.state = harmonyMessageState_Normal
}
}
return contentSb.String(), thinkingSb.String(), toolContentSb.String()
}
func (h *HarmonyMessageHandler) CreateToolParser() *HarmonyToolCallAccumulator {
return &HarmonyToolCallAccumulator{
state: harmonyToolCallState_Normal,
currentToolName: nil,
}
}
type harmonyToolCallState int
const (
harmonyToolCallState_Normal harmonyToolCallState = iota
harmonyToolCallState_ToolCalling
)
type HarmonyToolCallAccumulator struct {
state harmonyToolCallState
acc strings.Builder
currentToolName *string
}
func (a *HarmonyToolCallAccumulator) SetToolName(toolName string) {
a.currentToolName = &toolName
}
func (a *HarmonyToolCallAccumulator) Add(content string) {
a.acc.WriteString(content)
}
func (a *HarmonyToolCallAccumulator) Drain() (*string, string) {
str := a.acc.String()
a.state = harmonyToolCallState_Normal
a.acc.Reset()
return a.currentToolName, str
}
func (a *HarmonyToolCallAccumulator) Content() string {
return a.acc.String()
}
// FunctionNameMap maps a user-specified function name to a valid function
// name for harmony (which look like TypeScript identifiers). This is needed to
// transform user-specified function names, which might contain characters that
// are not allowed in TypeScript identifiers
type FunctionNameMap struct {
userToHarmony map[string]string
harmonyToUser map[string]string
}
func NewFunctionNameMap() *FunctionNameMap {
return &FunctionNameMap{
userToHarmony: make(map[string]string),
harmonyToUser: make(map[string]string),
}
}
// Init initializes the handler with tools, optional last message, and think value
// Implements the Parser interface
func (h *HarmonyMessageHandler) Init(tools []api.Tool, lastMessage *api.Message, thinkValue *api.ThinkValue) []api.Tool {
// Initialize the harmony parser
if h.HarmonyParser == nil {
h.HarmonyParser = &HarmonyParser{
MessageStartTag: "<|start|>",
MessageEndTag: "<|end|>",
HeaderEndTag: "<|message|>",
}
}
// Handle prefill for chat mode
if lastMessage != nil {
h.HarmonyParser.AddImplicitStartOrPrefill(lastMessage)
} else {
h.HarmonyParser.AddImplicitStart()
}
// Initialize tool accumulator
h.toolAccumulator = h.CreateToolParser()
// Process tools and return renamed versions
if len(tools) == 0 {
return tools
}
processedTools := make([]api.Tool, len(tools))
copy(processedTools, tools)
for i, tool := range processedTools {
if tool.Function.Name != "" {
processedTools[i].Function.Name = h.FunctionNameMap.ConvertAndAdd(tool.Function.Name)
h.convertedTools[tool.Function.Name] = struct{}{}
}
}
return processedTools
}
// Add implements the Parser interface - processes streamed content and extracts content, thinking, and tool calls
func (h *HarmonyMessageHandler) Add(s string, done bool) (content string, thinking string, calls []api.ToolCall, err error) {
content, thinking, toolContent := h.AddContent(s, h.toolAccumulator)
if toolContent != "" {
h.toolAccumulator.Add(toolContent)
}
// tool calls always happen one at a time, and always at the end of a message,
// so for simplicity we defer parsing them until we know we're done
if done {
toolName, raw := h.toolAccumulator.Drain()
if toolName != nil {
name := strings.TrimPrefix(*toolName, "functions.")
name = h.FunctionNameMap.OriginalFromConverted(name)
var args api.ToolCallFunctionArguments
if err := json.Unmarshal([]byte(raw), &args); err != nil {
return "", "", nil, fmt.Errorf("error parsing tool call: raw='%s', err=%w", raw, err)
}
calls = append(calls, api.ToolCall{Function: api.ToolCallFunction{Name: name, Arguments: args}})
}
}
return content, thinking, calls, nil
}
// HasToolSupport implements the Parser interface
func (h *HarmonyMessageHandler) HasToolSupport() bool {
return true
}
// HasThinkingSupport implements the Parser interface
func (h *HarmonyMessageHandler) HasThinkingSupport() bool {
return true
}
func (h *HarmonyMessageHandler) PreservedTokens() []string {
// <|call|> is an EOG marker for tool calls. Preserve structural tokens
// used by the parser, but let llama-server stop on the call terminator.
return []string{
"<|start|>",
"<|end|>",
"<|message|>",
"<|channel|>",
"<|constrain|>",
}
}
func (m *FunctionNameMap) ConvertAndAdd(userFunctionName string) string {
harmonyFunctionName := m.deriveName(userFunctionName)
// built-in functions should not be renamed
if userFunctionName == "browser.open" || userFunctionName == "browser.search" || userFunctionName == "browser.find" || userFunctionName == "python" {
harmonyFunctionName = userFunctionName
}
m.userToHarmony[userFunctionName] = harmonyFunctionName
m.harmonyToUser[harmonyFunctionName] = userFunctionName
return harmonyFunctionName
}
// OriginalFromConverted looks up the reverse-mapping of a previously-converted
// user->harmony function name. To unmap reliably, the mapping must exist, as
// the conversion process is not reversible without the appropriate state
func (m *FunctionNameMap) OriginalFromConverted(harmonyFunctionName string) string {
if userFunctionName, ok := m.harmonyToUser[harmonyFunctionName]; ok {
return userFunctionName
}
slog.Warn("harmony parser: no reverse mapping found for function name", "harmonyFunctionName", harmonyFunctionName)
// fallback to the original function name if we can't find a mapping
return harmonyFunctionName
}
// convertToValidChars converts a user-specified function name to a valid
// TypeScript identifier.
//
// Limitations:
//
// - This doesn't restrict reserved TypeScript keywords.
// - We don't perform a real ID_Start/ID_Continue check, and instead use the more
// restrictive unicode.IsLetter/unicode.IsDigit check. Unclear what kind of
// identifiers these models were trained on, so in the end we might want to
// convert unicode-heavy identifiers to their closest ASCII equivalents.
func (m *FunctionNameMap) convertToValidChars(userFunctionName string) string {
mapper := func(r rune) rune {
// first, replace certain characters with underscores
if r == ' ' || r == '-' || r == '.' {
return '_'
}
if unicode.IsLetter(r) || unicode.IsDigit(r) || r == '_' || r == '$' {
return r
}
// finally, remove any other characters
return -1
}
candidate := strings.Map(mapper, userFunctionName)
// set a default name if we end up with nothing left
if candidate == "" {
return "unnamed"
}
// if the candidate starts with a number, prepend an underscore to make it a
// valid identifier
if unicode.IsDigit(rune(candidate[0])) {
candidate = "_" + candidate
}
return candidate
}
func (m *FunctionNameMap) deriveName(userFunctionName string) string {
originalCandidate := m.convertToValidChars(userFunctionName)
candidate := originalCandidate
// Check for dupes, and if so, add a number to the end.
// We start at 2 because if we have dupes and the first is never renamed, it
// makes sense for them to be named, say, `f`, `f_2`, `f_3`
count := 2
for {
if _, exists := m.harmonyToUser[candidate]; !exists {
break
}
candidate = fmt.Sprintf("%s_%d", originalCandidate, count)
count++
}
return candidate
}