Compare commits

...

2 Commits

Author SHA1 Message Date
jmorganca
38ed7c7a4f no parser yet 2025-10-13 14:38:10 -07:00
jmorganca
9ff8e5a64d wip 2025-10-13 13:01:54 -07:00
5 changed files with 572 additions and 0 deletions

219
model/parsers/glm46.go Normal file
View File

@@ -0,0 +1,219 @@
package parsers
import (
"context"
"encoding/json"
"log/slog"
"strings"
"unicode"
"github.com/ollama/ollama/api"
"github.com/ollama/ollama/logutil"
)
const (
glm46CollectingContent glm46ParserState = iota
CollectingThinkingContent
CollectingToolContent
)
const (
thinkingCloseTag = "</think>"
)
// TODO(gguo): add a field for isThinking
type GLM46Parser struct {
state qwenParserState
buffer strings.Builder
tools []api.Tool
}
func (p *GLM46Parser) HasToolSupport() bool {
return true
}
// TODO(gguo): changes this to reference an objects param
func (p *GLM46Parser) HasThinkingSupport() bool {
return true
}
func (p *GLM46Parser) Init(tools []api.Tool, lastMessage *api.Message) []api.Tool {
p.tools = tools
// p.state = p.initialState()
return tools
}
type glm46EventThinkingContent struct {
content string
}
func (glm46EventThinkingContent) isGLM46Event() {}
func (p *GLM46Parser) Add(s string, done bool) (content string, thinking string, calls []api.ToolCall, err error) {
p.buffer.WriteString(s)
events := p.parseEvents()
var toolCalls []api.ToolCall
var sb strings.Builder
for _, event := range events {
switch event := event.(type) {
case glm46EventRawToolCall:
toolCall, err := parseJSONToolCall(event, p.tools)
if err != nil {
slog.Warn("qwen tool call parsing failed", "error", err)
return "", "", nil, err
}
toolCalls = append(toolCalls, toolCall)
case glm46EventThinkingContent:
sb.WriteString(event.content)
case glm46EventContent:
// TODO(drifkin): if the same turn contains multiple interleaved content
// events, we naively append them together here.
sb.WriteString(event.content)
}
}
return sb.String(), "", toolCalls, nil
}
func (p *GLM46Parser) parseEvents() []glm46Event {
var all []glm46Event
keepLooping := true
for keepLooping {
var events []glm46Event
events, keepLooping = p.eat()
if len(events) > 0 {
all = append(all, events...)
}
}
if len(all) > 0 {
slog.Log(context.TODO(), logutil.LevelTrace, "qwen events parsed", "events", all, "state", p.state, "buffer", p.buffer.String())
}
return all
}
func emitContentBeforeTag(p *GLM46Parser, events []glm46Event, tag string) []glm46Event {
split := strings.SplitN(p.buffer.String(), tag, 2)
before := split[0]
before = strings.TrimRightFunc(before, unicode.IsSpace)
if len(before) > 0 {
events = append(events, glm46EventContent{content: before})
}
after := split[1]
p.buffer.Reset()
p.buffer.WriteString(after)
return events
}
func (p *GLM46Parser) eat() ([]glm46Event, bool) {
var events []glm46Event
switch p.state {
case glm46CollectingContent:
if strings.Contains(p.buffer.String(), toolOpenTag) {
events = emitContentBeforeTag(p, events, toolOpenTag)
p.state = glm46CollectingToolContent
return events, true
} else if overlapLen := overlap(p.buffer.String(), toolOpenTag); overlapLen > 0 {
beforePartialTag := p.buffer.String()[:len(p.buffer.String())-overlapLen]
trailingWhitespaceLen := trailingWhitespaceLen(beforePartialTag)
ambiguousStart := len(beforePartialTag) - trailingWhitespaceLen
unambiguous := p.buffer.String()[:ambiguousStart]
ambiguous := p.buffer.String()[ambiguousStart:]
p.buffer.Reset()
p.buffer.WriteString(ambiguous)
if len(unambiguous) > 0 { // why does qwen3coder not have this here
events = append(events, glm46EventContent{content: unambiguous})
}
return events, false
} else {
whitespaceLen := trailingWhitespaceLen(p.buffer.String())
ambiguousStart := len(p.buffer.String()) - whitespaceLen
unambiguous := p.buffer.String()[:ambiguousStart]
ambiguous := p.buffer.String()[ambiguousStart:]
p.buffer.Reset()
p.buffer.WriteString(ambiguous)
if len(unambiguous) > 0 {
events = append(events, glm46EventContent{content: unambiguous})
}
return events, false
}
case CollectingToolContent:
if strings.Contains(p.buffer.String(), glm46ToolCloseTag) {
split := strings.SplitN(p.buffer.String(), toolCloseTag, 2)
before := split[0]
if len(before) == 0 {
slog.Warn("qwen tool call closing tag found but no content before it")
}
after := strings.TrimLeftFunc(split[1], unicode.IsSpace)
events = append(events, glm46EventRawToolCall{raw: before})
p.buffer.Reset()
p.buffer.WriteString(after)
p.state = glm46CollectingContent
return events, true
} else {
return events, false
}
case glm46CollectingThinkingContent: // so we want to hip the unambiguous stuff
if strings.Contains(p.buffer.String(), thinkingCloseTag) {
split := strings.SplitN(p.buffer.String(), thinkingCloseTag, 2)
before := split[0]
if len(before) == 0 {
slog.Warn("qwen tool call closing tag found but no content before it")
}
after := strings.TrimLeftFunc(split[1], unicode.IsSpace)
if len(before) > 0 {
events = append(events, glm46EventThinkingContent{content: before})
}
p.buffer.Reset()
p.buffer.WriteString(after)
p.state = glm46CollectingContent
return events, true
} else if overlapLen := overlap(p.buffer.String(), thinkingCloseTag); overlapLen > 0 { // we see part of a close thinking tag
beforePartialTag := p.buffer.String()[:len(p.buffer.String())-overlapLen]
trailingWhitespaceLen := trailingWhitespaceLen(beforePartialTag)
ambiguousStart := len(beforePartialTag) - trailingWhitespaceLen
unambiguous := p.buffer.String()[:ambiguousStart]
ambiguous := p.buffer.String()[ambiguousStart:]
p.buffer.Reset()
p.buffer.WriteString(ambiguous)
if len(unambiguous) > 0 {
events = append(events, glm46EventThinkingContent{content: unambiguous})
}
return events, false
} else {
whitespaceLen := trailingWhitespaceLen(p.buffer.String())
ambiguousStart := len(p.buffer.String()) - whitespaceLen
unambiguous := p.buffer.String()[:ambiguousStart]
ambiguous := p.buffer.String()[ambiguousStart:]
p.buffer.Reset()
p.buffer.WriteString(ambiguous)
if len(unambiguous) > 0 {
events = append(events, glm46EventThinkingContent{content: unambiguous})
}
return events, false
}
default:
panic("unreachable")
}
}
func parseJSONToolCall(raw glm46EventRawToolCall, tools []api.Tool) (api.ToolCall, error) {
var toolCallFunction api.ToolCallFunction
if err := json.Unmarshal([]byte(raw.raw), &toolCallFunction); err != nil {
return api.ToolCall{}, err
}
toolCall := api.ToolCall{}
toolCall.Function = toolCallFunction
return toolCall, nil
}

View File

@@ -21,6 +21,9 @@ func ParserForName(name string) Parser {
case "qwen3-coder":
parser := &Qwen3CoderParser{}
return parser
case "glm-4.6":
parser := &GLM46Parser{}
return parser
case "passthrough":
return &PassthroughParser{}
case "harmony":

View File

@@ -0,0 +1,239 @@
package renderers
import (
"testing"
"github.com/google/go-cmp/cmp"
"github.com/ollama/ollama/api"
)
func TestGLM46Renderer(t *testing.T) {
tests := []struct {
name string
messages []api.Message
tools []api.Tool
thinkValue *api.ThinkValue
expected string
}{
{
name: "basic",
messages: []api.Message{
{Role: "user", Content: "Hello, how are you?"},
},
expected: `[gMASK]<sop><|user|>
Hello, how are you?<|assistant|>`,
},
{
name: "basic with system message",
messages: []api.Message{
{Role: "system", Content: "You are a helpful assistant."},
{Role: "user", Content: "Hello, how are you?"},
},
expected: `[gMASK]<sop><|system|>
You are a helpful assistant.<|user|>
Hello, how are you?<|assistant|>`,
},
{
name: "basic with user assistant user",
messages: []api.Message{
{Role: "user", Content: "What is the capital of France?"},
{Role: "assistant", Thinking: "Let me analyze the request...", Content: "The capital of France is Paris."},
{Role: "user", Content: "Fantastic!"},
},
expected: `[gMASK]<sop><|user|>
What is the capital of France?<|assistant|>
The capital of France is Paris.<|user|>
Fantastic!<|assistant|>`,
},
{
name: "tools",
messages: []api.Message{
{Role: "system", Content: "You are a helpful assistant with access to tools."},
{Role: "user", Content: "What is the weather like in Tokyo?"},
},
tools: []api.Tool{
{
Type: "function",
Function: api.ToolFunction{
Name: "get_weather",
Description: "Get the current weather in a given location",
Parameters: api.ToolFunctionParameters{
Type: "object",
Required: []string{"location"},
Properties: map[string]api.ToolProperty{
"location": {
Type: api.PropertyType{"string"},
Description: "The city and state, e.g. San Francisco, CA",
},
"unit": {
Type: api.PropertyType{"string"},
Enum: []any{"celsius", "fahrenheit"},
},
},
},
},
},
},
expected: `[gMASK]<sop><|system|>
# Tools
You may call one or more functions to assist with the user query.
You are provided with function signatures within <tools></tools> XML tags:
<tools>
{"type":"function","function":{"name":"get_weather","description":"Get the current weather in a given location","parameters":{"type":"object","required":["location"],"properties":{"location":{"type":"string","description":"The city and state, e.g. San Francisco, CA"},"unit":{"type":"string","description":"","enum":["celsius","fahrenheit"]}}}}}
</tools>
For each function call, output the function name and arguments within the following XML format:
<tool_call>{function-name}
<arg_key>{arg-key-1}</arg_key>
<arg_value>{arg-value-1}</arg_value>
<arg_key>{arg-key-2}</arg_key>
<arg_value>{arg-value-2}</arg_value>
...
</tool_call><|system|>
You are a helpful assistant with access to tools.<|user|>
What is the weather like in Tokyo?<|assistant|>`,
},
{
name: "tool calls",
messages: []api.Message{
{Role: "system", Content: "You are a helpful assistant with access to tools."},
{Role: "user", Content: "What is the weather like in Tokyo?"},
{
Role: "assistant",
ToolCalls: []api.ToolCall{
{
Function: api.ToolCallFunction{
Name: "get_weather",
Arguments: api.ToolCallFunctionArguments{
"location": "Tokyo, Japan",
"unit": "celsius",
},
},
},
{
Function: api.ToolCallFunction{
Name: "get_weather",
Arguments: api.ToolCallFunctionArguments{
"location": "Japan",
"unit": "fahrenheit",
},
},
},
},
},
{
Role: "tool",
Content: "{\"temperature\": 22, \"weather\": \"partly cloudy\", \"humidity\": 65}",
ToolName: "get_weather",
},
{
Role: "tool",
Content: "{\"temperature\": 68, \"weather\": \"sunny\", \"humidity\": 75}",
ToolName: "get_weather",
},
{
Role: "assistant",
Content: "The weather in Tokyo is currently partly cloudy with a temperature of 22°C and 65% humidity. It's a pleasant day with moderate temperatures.",
},
},
tools: []api.Tool{
{
Type: "function",
Function: api.ToolFunction{
Name: "get_weather",
Description: "Get the current weather in a given location",
Parameters: api.ToolFunctionParameters{
Type: "object",
Required: []string{"location"},
Properties: map[string]api.ToolProperty{
"location": {
Type: api.PropertyType{"string"},
Description: "The city and state, e.g. San Francisco, CA",
},
"unit": {
Type: api.PropertyType{"string"},
Enum: []any{"celsius", "fahrenheit"},
},
},
},
},
},
},
expected: `[gMASK]<sop><|system|>
# Tools
You may call one or more functions to assist with the user query.
You are provided with function signatures within <tools></tools> XML tags:
<tools>
{"type":"function","function":{"name":"get_weather","description":"Get the current weather in a given location","parameters":{"type":"object","required":["location"],"properties":{"location":{"type":"string","description":"The city and state, e.g. San Francisco, CA"},"unit":{"type":"string","description":"","enum":["celsius","fahrenheit"]}}}}}
</tools>
For each function call, output the function name and arguments within the following XML format:
<tool_call>{function-name}
<arg_key>{arg-key-1}</arg_key>
<arg_value>{arg-value-1}</arg_value>
<arg_key>{arg-key-2}</arg_key>
<arg_value>{arg-value-2}</arg_value>
...
</tool_call><|system|>
You are a helpful assistant with access to tools.<|user|>
What is the weather like in Tokyo?<|assistant|>
<think></think>
<tool_call>get_weather
<arg_key>location</arg_key>
<arg_value>Tokyo, Japan</arg_value>
<arg_key>unit</arg_key>
<arg_value>celsius</arg_value>
</tool_call>
<tool_call>get_weather
<arg_key>location</arg_key>
<arg_value>Japan</arg_value>
<arg_key>unit</arg_key>
<arg_value>fahrenheit</arg_value>
</tool_call><|observation|>
<tool_response>
{"temperature": 22, "weather": "partly cloudy", "humidity": 65}
</tool_response>
<tool_response>
{"temperature": 68, "weather": "sunny", "humidity": 75}
</tool_response><|assistant|>
<think></think>
The weather in Tokyo is currently partly cloudy with a temperature of 22°C and 65% humidity. It's a pleasant day with moderate temperatures.<|assistant|>`,
},
{
name: "think true",
messages: []api.Message{
{Role: "user", Content: "Hello, how are you?"},
},
thinkValue: &api.ThinkValue{Value: true},
expected: `[gMASK]<sop><|user|>
Hello, how are you?<|assistant|>`,
},
{
name: "think false",
messages: []api.Message{
{Role: "user", Content: "Hello, how are you?"},
},
thinkValue: &api.ThinkValue{Value: false},
expected: `[gMASK]<sop><|user|>
Hello, how are you?/nothink<|assistant|>
<think></think>`,
},
}
for _, tt := range tests {
t.Run(tt.name, func(t *testing.T) {
rendered, err := GLM46Renderer(tt.messages, tt.tools, tt.thinkValue)
if err != nil {
t.Fatal(err)
}
if diff := cmp.Diff(rendered, tt.expected); diff != "" {
t.Errorf("mismatch (-got +want):\n%s", diff)
t.Logf("Got:\n%s", rendered)
t.Logf("Expected:\n%s", tt.expected)
}
})
}
}

109
model/renderers/gml46.go Normal file
View File

@@ -0,0 +1,109 @@
package renderers
import (
"encoding/json"
"fmt"
"strings"
"github.com/ollama/ollama/api"
)
func GLM46Renderer(messages []api.Message, tools []api.Tool, thinkValue *api.ThinkValue) (string, error) {
var sb strings.Builder
sb.WriteString("[gMASK]<sop>")
var lastUserIndex int
for i, message := range messages {
if message.Role == "user" {
lastUserIndex = i
}
}
if len(tools) > 0 {
sb.WriteString("<|system|>\n")
sb.WriteString("# Tools\n\n")
sb.WriteString("You may call one or more functions to assist with the user query.\n\n")
sb.WriteString("You are provided with function signatures within <tools></tools> XML tags:\n")
sb.WriteString("<tools>\n")
for _, tool := range tools {
d, _ := json.Marshal(tool)
sb.WriteString(string(d) + "\n")
}
sb.WriteString("</tools>\n\n")
sb.WriteString("For each function call, output the function name and arguments within the following XML format:\n")
sb.WriteString("<tool_call>{function-name}\n")
sb.WriteString("<arg_key>{arg-key-1}</arg_key>\n")
sb.WriteString("<arg_value>{arg-value-1}</arg_value>\n")
sb.WriteString("<arg_key>{arg-key-2}</arg_key>\n")
sb.WriteString("<arg_value>{arg-value-2}</arg_value>\n")
sb.WriteString("...\n")
sb.WriteString("</tool_call>")
}
for i, message := range messages {
switch message.Role {
case "user":
sb.WriteString("<|user|>\n")
sb.WriteString(message.Content)
if thinkValue != nil && !thinkValue.Bool() && !strings.HasSuffix(message.Content, "/nothink") {
sb.WriteString("/nothink")
}
case "assistant":
sb.WriteString("<|assistant|>")
if i > lastUserIndex {
if message.Thinking != "" {
sb.WriteString("\n<think>" + message.Thinking + "</think>")
} else {
sb.WriteString("\n<think></think>")
}
}
if message.Content != "" {
sb.WriteString("\n" + message.Content)
}
if len(message.ToolCalls) > 0 {
for _, toolCall := range message.ToolCalls {
sb.WriteString("\n<tool_call>" + toolCall.Function.Name + "\n")
for key, value := range toolCall.Function.Arguments {
sb.WriteString("<arg_key>" + key + "</arg_key>\n")
var valueStr string
if str, ok := value.(string); ok {
valueStr = str
} else {
jsonBytes, err := json.Marshal(value)
if err != nil {
valueStr = fmt.Sprintf("%v", value)
} else {
valueStr = string(jsonBytes)
}
}
sb.WriteString("<arg_value>" + valueStr + "</arg_value>\n")
}
sb.WriteString("</tool_call>")
}
}
case "tool":
if i == 0 || messages[i-1].Role != "tool" {
sb.WriteString("<|observation|>")
}
sb.WriteString("\n<tool_response>\n")
sb.WriteString(message.Content)
sb.WriteString("\n</tool_response>")
case "system":
sb.WriteString("<|system|>\n")
sb.WriteString(message.Content)
}
}
// Add generation prompt
sb.WriteString("<|assistant|>")
fmt.Println("thinkValue", thinkValue, thinkValue.Bool())
if thinkValue != nil && !thinkValue.Bool() {
sb.WriteString("\n<think></think>")
}
return sb.String(), nil
}

View File

@@ -20,6 +20,8 @@ func rendererForName(name string) rendererFunc {
switch name {
case "qwen3-coder":
return Qwen3CoderRenderer
case "glm-4.6":
return GLM46Renderer
default:
return nil
}