ollama/model/renderers/qwen3coder_test.go

371 lines
9.8 KiB
Go
Raw Normal View History

add qwen3-coder tool support The format qwen3-coder uses is relatively unique, both in rendering and in parsing. To implement parsing, I wrote a custom parser in similar style to harmony. For the rendering, I found that the logic would be much more difficult to follow in a template, so I introduced the concept of a built-in renderer that uses go code, rather than a template to generate prompts. I set us up for future built-in parsers and renderers by making it so they can be specified in a Modelfile like so: ``` RENDERER "qwen3-coder" PARSER "qwen3-coder" ``` These need to be provided explicitly because the architecture alone is not enough to understand what format the model expects to receive, and what format we expect it to output (e.g., qwen3-coder is `qwen3moe`, which includes other qwen3-family models as well) I haven't converted harmony to be one of these "built-ins" yet, since some of it is in flux with the changes @ParthSareen has been making to move harmony to the runner. It is likely that many other built-ins will need to move to the runner as well, but I'm able to slightly defer that decision since qwen3-coder doesn't have thinking (and therefore doesn't need to be in the runner to make structured outputs work). I expect to unify harmony with this approach very soon. Whether a particular model supports tools or thinking was previously inferred from templates, but without a template we now also use the parser itself to declare what it supports. If we have future models that re-use the same parsing format, but have different capabilities, we'll want to parameterize them and give them different names to be specified as a `PARSER`. Misc changes: - I worked on the renderer by diffing outputs from the reference implementation and ours. To make it easier to do this, I extended <https://github.com/ollama/ollama/pull/11875> to also support returning the prompt via the openai compat layer
2025-09-12 04:40:35 +08:00
package renderers
import (
"testing"
"github.com/google/go-cmp/cmp"
"github.com/ollama/ollama/api"
)
func TestQwen3CoderRenderer(t *testing.T) {
tests := []struct {
name string
msgs []api.Message
tools []api.Tool
expected string
}{
{
name: "basic",
msgs: []api.Message{
{Role: "system", Content: "You are a helpful assistant."},
{Role: "user", Content: "Hello, how are you?"},
},
expected: `<|im_start|>system
You are a helpful assistant.<|im_end|>
<|im_start|>user
Hello, how are you?<|im_end|>
<|im_start|>assistant
`,
},
{
name: "with tools and response",
msgs: []api.Message{
{Role: "system", Content: "You are a helpful assistant with access to tools."},
{Role: "user", Content: "What is the weather like in San Francisco?"},
{
Role: "assistant",
Content: "I'll check the weather in San Francisco for you.",
ToolCalls: []api.ToolCall{
{
Function: api.ToolCallFunction{
Name: "get_weather",
Arguments: map[string]any{
"unit": "fahrenheit",
},
},
},
},
},
{Role: "tool", Content: "{\"location\": \"San Francisco, CA\", \"temperature\": 68, \"condition\": \"partly cloudy\", \"humidity\": 65, \"wind_speed\": 12}", ToolName: "get_weather"},
{Role: "user", Content: "That sounds nice! What about New York?"},
},
tools: []api.Tool{
{Function: api.ToolFunction{
Name: "get_weather",
Description: "Get the current weather in a given location",
Parameters: api.ToolFunctionParameters{
Required: []string{"unit"},
Properties: map[string]api.ToolProperty{
"unit": {Type: api.PropertyType{"string"}, Enum: []any{"celsius", "fahrenheit"}, Description: "The unit of temperature"},
// TODO(drifkin): add multiple params back once we have predictable
// order via some sort of ordered map type (see
// <https://github.com/ollama/ollama/issues/12244>)
/*
"location": {Type: api.PropertyType{"string"}, Description: "The city and state, e.g. San Francisco, CA"},
*/
},
},
}},
},
expected: `<|im_start|>system
You are a helpful assistant with access to tools.
# Tools
You have access to the following functions:
<tools>
<function>
<name>get_weather</name>
<description>Get the current weather in a given location</description>
<parameters>
<parameter>
<name>unit</name>
<type>string</type>
<description>The unit of temperature</description>
<enum>["celsius","fahrenheit"]</enum>
</parameter>
<required>["unit"]</required>
</parameters>
</function>
</tools>
If you choose to call a function ONLY reply in the following format with NO suffix:
<tool_call>
<function=example_function_name>
<parameter=example_parameter_1>
value_1
</parameter>
<parameter=example_parameter_2>
This is the value for the second parameter
that can span
multiple lines
</parameter>
</function>
</tool_call>
<IMPORTANT>
Reminder:
- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags
- Required parameters MUST be specified
- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after
- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls
</IMPORTANT><|im_end|>
<|im_start|>user
What is the weather like in San Francisco?<|im_end|>
<|im_start|>assistant
I'll check the weather in San Francisco for you.
<tool_call>
<function=get_weather>
<parameter=unit>
fahrenheit
</parameter>
</function>
</tool_call><|im_end|>
<|im_start|>user
<tool_response>
{"location": "San Francisco, CA", "temperature": 68, "condition": "partly cloudy", "humidity": 65, "wind_speed": 12}
</tool_response>
<|im_end|>
<|im_start|>user
That sounds nice! What about New York?<|im_end|>
<|im_start|>assistant
`,
},
{
name: "parallel tool calls",
msgs: []api.Message{
{Role: "system", Content: "You are a helpful assistant with access to tools."},
{Role: "user", Content: "call double(1) and triple(2)"},
{Role: "assistant", Content: "I'll call double(1) and triple(2) for you.", ToolCalls: []api.ToolCall{
{Function: api.ToolCallFunction{Name: "double", Arguments: map[string]any{"number": "1"}}},
{Function: api.ToolCallFunction{Name: "triple", Arguments: map[string]any{"number": "2"}}},
}},
{Role: "tool", Content: "{\"number\": 2}", ToolName: "double"},
{Role: "tool", Content: "{\"number\": 6}", ToolName: "triple"},
},
tools: []api.Tool{
{Function: api.ToolFunction{Name: "double", Description: "Double a number", Parameters: api.ToolFunctionParameters{Properties: map[string]api.ToolProperty{
"number": {Type: api.PropertyType{"string"}, Description: "The number to double"},
}}}},
{Function: api.ToolFunction{Name: "triple", Description: "Triple a number", Parameters: api.ToolFunctionParameters{Properties: map[string]api.ToolProperty{
"number": {Type: api.PropertyType{"string"}, Description: "The number to triple"},
}}}},
},
expected: `<|im_start|>system
You are a helpful assistant with access to tools.
# Tools
You have access to the following functions:
<tools>
<function>
<name>double</name>
<description>Double a number</description>
<parameters>
<parameter>
<name>number</name>
<type>string</type>
<description>The number to double</description>
</parameter>
</parameters>
</function>
<function>
<name>triple</name>
<description>Triple a number</description>
<parameters>
<parameter>
<name>number</name>
<type>string</type>
<description>The number to triple</description>
</parameter>
</parameters>
</function>
</tools>
If you choose to call a function ONLY reply in the following format with NO suffix:
<tool_call>
<function=example_function_name>
<parameter=example_parameter_1>
value_1
</parameter>
<parameter=example_parameter_2>
This is the value for the second parameter
that can span
multiple lines
</parameter>
</function>
</tool_call>
<IMPORTANT>
Reminder:
- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags
- Required parameters MUST be specified
- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after
- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls
</IMPORTANT><|im_end|>
<|im_start|>user
call double(1) and triple(2)<|im_end|>
<|im_start|>assistant
I'll call double(1) and triple(2) for you.
<tool_call>
<function=double>
<parameter=number>
1
</parameter>
</function>
</tool_call>
<tool_call>
<function=triple>
<parameter=number>
2
</parameter>
</function>
</tool_call><|im_end|>
<|im_start|>user
<tool_response>
{"number": 2}
</tool_response>
<tool_response>
{"number": 6}
</tool_response>
<|im_end|>
<|im_start|>assistant
`,
},
{
name: "prefill",
msgs: []api.Message{
{Role: "system", Content: "You are a helpful assistant."},
{Role: "user", Content: "Tell me something interesting."},
{Role: "assistant", Content: "I'll tell you something interesting about cats"},
},
expected: `<|im_start|>system
You are a helpful assistant.<|im_end|>
<|im_start|>user
Tell me something interesting.<|im_end|>
<|im_start|>assistant
I'll tell you something interesting about cats`,
},
{
name: "complex tool call arguments should remain json encoded",
msgs: []api.Message{
{Role: "user", Content: "call tool"},
{Role: "assistant", ToolCalls: []api.ToolCall{
{Function: api.ToolCallFunction{
Name: "echo",
Arguments: map[string]any{
"payload": map[string]any{"foo": "bar"},
},
}},
}},
{Role: "tool", Content: "{\"payload\": {\"foo\": \"bar\"}}", ToolName: "echo"},
},
expected: `<|im_start|>user
call tool<|im_end|>
<|im_start|>assistant
<tool_call>
<function=echo>
<parameter=payload>
{"foo":"bar"}
</parameter>
</function>
</tool_call><|im_end|>
<|im_start|>user
<tool_response>
{"payload": {"foo": "bar"}}
</tool_response>
<|im_end|>
<|im_start|>assistant
`,
},
}
for _, tt := range tests {
t.Run(tt.name, func(t *testing.T) {
rendered, err := Qwen3CoderRenderer(tt.msgs, tt.tools, nil)
if err != nil {
t.Fatal(err)
}
if diff := cmp.Diff(rendered, tt.expected); diff != "" {
t.Errorf("mismatch (-got +want):\n%s", diff)
}
})
}
}
func TestFormatToolCallArgument(t *testing.T) {
tests := []struct {
name string
arg any
expected string
}{
{
name: "string",
arg: "foo",
// notice no quotes around the string
expected: "foo",
},
{
name: "map",
arg: map[string]any{"foo": "bar"},
expected: "{\"foo\":\"bar\"}",
},
{
name: "number",
arg: 1,
expected: "1",
},
{
name: "boolean",
arg: true,
expected: "true",
},
}
for _, tt := range tests {
t.Run(tt.name, func(t *testing.T) {
got := formatToolCallArgument(tt.arg)
if got != tt.expected {
t.Errorf("formatToolCallArgument(%v) = %v, want %v", tt.arg, got, tt.expected)
}
})
}
}
func TestQwen3ToolDefinitionTypes(t *testing.T) {
tests := []struct {
name string
propertyType api.PropertyType
expected string
}{
{
name: "simple",
propertyType: api.PropertyType{"string"},
expected: "string",
},
{
name: "multiple",
propertyType: api.PropertyType{"string", "number"},
expected: "[\"string\",\"number\"]",
},
{
name: "empty",
propertyType: api.PropertyType{},
expected: "[]",
},
}
for _, tt := range tests {
t.Run(tt.name, func(t *testing.T) {
got := formatToolDefinitionType(tt.propertyType)
if got != tt.expected {
t.Errorf("formatToolDefinitionType() = %v, want %v", got, tt.expected)
}
})
}
}