ollama/api/types.go

978 lines
28 KiB
Go
Raw Normal View History

package api
2023-07-13 09:18:06 +08:00
import (
2023-07-20 06:00:28 +08:00
"encoding/json"
2023-07-13 09:18:06 +08:00
"fmt"
"log/slog"
2023-07-20 06:00:28 +08:00
"math"
2023-07-13 09:18:06 +08:00
"os"
"reflect"
"strconv"
"strings"
2023-07-13 09:18:06 +08:00
"time"
"github.com/ollama/ollama/envconfig"
"github.com/ollama/ollama/types/model"
2023-07-13 09:18:06 +08:00
)
2023-07-08 06:29:17 +08:00
// StatusError is an error with an HTTP status code and message.
2023-07-19 00:09:45 +08:00
type StatusError struct {
2023-07-21 02:45:12 +08:00
StatusCode int
Status string
ErrorMessage string `json:"error"`
2023-07-19 00:09:45 +08:00
}
func (e StatusError) Error() string {
2023-07-21 02:45:12 +08:00
switch {
case e.Status != "" && e.ErrorMessage != "":
return fmt.Sprintf("%s: %s", e.Status, e.ErrorMessage)
case e.Status != "":
return e.Status
case e.ErrorMessage != "":
return e.ErrorMessage
default:
// this should not happen
return "something went wrong, please see the ollama server logs for details"
2023-07-19 00:09:45 +08:00
}
}
// ImageData represents the raw binary data of an image file.
type ImageData []byte
// GenerateRequest describes a request sent by [Client.Generate]. While you
// have to specify the Model and Prompt fields, all the other fields have
// reasonable defaults for basic uses.
type GenerateRequest struct {
// Model is the model name; it should be a name familiar to Ollama from
// the library at https://ollama.com/library
Model string `json:"model"`
// Prompt is the textual prompt to send to the model.
Prompt string `json:"prompt"`
// Suffix is the text that comes after the inserted text.
Suffix string `json:"suffix"`
// System overrides the model's default system message/prompt.
System string `json:"system"`
// Template overrides the model's default prompt template.
Template string `json:"template"`
// Context is the context parameter returned from a previous call to
// [Client.Generate]. It can be used to keep a short conversational memory.
Context []int `json:"context,omitempty"`
// Stream specifies whether the response is streaming; it is true by default.
Stream *bool `json:"stream,omitempty"`
// Raw set to true means that no formatting will be applied to the prompt.
Raw bool `json:"raw,omitempty"`
// Format specifies the format to return a response in.
Format json.RawMessage `json:"format,omitempty"`
// KeepAlive controls how long the model will stay loaded in memory following
// this request.
KeepAlive *Duration `json:"keep_alive,omitempty"`
// Images is an optional list of raw image bytes accompanying this
// request, for multimodal models.
Images []ImageData `json:"images,omitempty"`
// Options lists model-specific options. For example, temperature can be
// set through this field, if the model supports it.
Options map[string]any `json:"options"`
// Think controls whether thinking/reasoning models will think before
gpt-oss (#11672) * bf16 * tests * gpt-oss * enable gptoss for engine * rough estimate * convert to mxfp4 * handle safetensors U8 * clamp glu/linear * update tokenizer * MXFP4 support This implements the Open Compute Microscaling (MX) FP4 format as a tensor type with backend implementations focusing on mulmat and mulmatid on CPU, CUDA, and Metal. * Unit tests for MXFP4 support This exercises various operations and shapes on both CPU and GPU (if detected on the system) * cuda graph * unit test adjustments * cuda: optimize memory access Read 4 bytes at a time (8 elements) when performing mul_mat_vec_mxfp4 * mac: fix crash on old macos versions cblas_sgemm is only supported on v13.3 and up, however bf16 is only supported on v14+ so we were falling back to ggml-blas and crashing on bf16 tensors. Checking for the function being null seems to be the simplest way to condittionally avoid registering the backend. * server: Minimum context length for gptoss This model requires a minimum context length of 8192 to function effectively. Users can set higher values through all normal mechanisms but lower values will be silently reset. * ggml: Multiply by numParallel for gptoss sliding window When computing the graph size estimate, the context size is already multiplied by numParallel so estimates reflect that. However, since sliding window models use a smaller, fixed context size, they need to manually take numParallel into account. * gpt-oss integration includes harmony parser and thinking levels, etc. * fix sync * fix tests * fix lint --------- Co-authored-by: Daniel Hiltgen <daniel@ollama.com> Co-authored-by: Jesse Gross <jesse@ollama.com> Co-authored-by: Devon Rifkin <drifkin@drifkin.net>
2025-08-06 03:21:16 +08:00
// responding. Can be a boolean (true/false) or a string ("high", "medium", "low")
// for supported models. Needs to be a pointer so we can distinguish between false
// (request that thinking _not_ be used) and unset (use the old behavior
// before this option was introduced)
gpt-oss (#11672) * bf16 * tests * gpt-oss * enable gptoss for engine * rough estimate * convert to mxfp4 * handle safetensors U8 * clamp glu/linear * update tokenizer * MXFP4 support This implements the Open Compute Microscaling (MX) FP4 format as a tensor type with backend implementations focusing on mulmat and mulmatid on CPU, CUDA, and Metal. * Unit tests for MXFP4 support This exercises various operations and shapes on both CPU and GPU (if detected on the system) * cuda graph * unit test adjustments * cuda: optimize memory access Read 4 bytes at a time (8 elements) when performing mul_mat_vec_mxfp4 * mac: fix crash on old macos versions cblas_sgemm is only supported on v13.3 and up, however bf16 is only supported on v14+ so we were falling back to ggml-blas and crashing on bf16 tensors. Checking for the function being null seems to be the simplest way to condittionally avoid registering the backend. * server: Minimum context length for gptoss This model requires a minimum context length of 8192 to function effectively. Users can set higher values through all normal mechanisms but lower values will be silently reset. * ggml: Multiply by numParallel for gptoss sliding window When computing the graph size estimate, the context size is already multiplied by numParallel so estimates reflect that. However, since sliding window models use a smaller, fixed context size, they need to manually take numParallel into account. * gpt-oss integration includes harmony parser and thinking levels, etc. * fix sync * fix tests * fix lint --------- Co-authored-by: Daniel Hiltgen <daniel@ollama.com> Co-authored-by: Jesse Gross <jesse@ollama.com> Co-authored-by: Devon Rifkin <drifkin@drifkin.net>
2025-08-06 03:21:16 +08:00
Think *ThinkValue `json:"think,omitempty"`
// DebugRenderOnly is a debug option that, when set to true, returns the rendered
// template instead of calling the model.
DebugRenderOnly bool `json:"_debug_render_only,omitempty"`
}
// ChatRequest describes a request sent by [Client.Chat].
2023-12-06 03:57:33 +08:00
type ChatRequest struct {
// Model is the model name, as in [GenerateRequest].
Model string `json:"model"`
// Messages is the messages of the chat - can be used to keep a chat memory.
Messages []Message `json:"messages"`
// Stream enables streaming of returned responses; true by default.
Stream *bool `json:"stream,omitempty"`
// Format is the format to return the response in (e.g. "json").
Format json.RawMessage `json:"format,omitempty"`
// KeepAlive controls how long the model will stay loaded into memory
// following the request.
KeepAlive *Duration `json:"keep_alive,omitempty"`
2023-12-06 03:57:33 +08:00
2024-06-21 04:45:47 +08:00
// Tools is an optional list of tools the model has access to.
Tools `json:"tools,omitempty"`
2024-06-21 04:45:47 +08:00
// Options lists model-specific options.
Options map[string]any `json:"options"`
// Think controls whether thinking/reasoning models will think before
gpt-oss (#11672) * bf16 * tests * gpt-oss * enable gptoss for engine * rough estimate * convert to mxfp4 * handle safetensors U8 * clamp glu/linear * update tokenizer * MXFP4 support This implements the Open Compute Microscaling (MX) FP4 format as a tensor type with backend implementations focusing on mulmat and mulmatid on CPU, CUDA, and Metal. * Unit tests for MXFP4 support This exercises various operations and shapes on both CPU and GPU (if detected on the system) * cuda graph * unit test adjustments * cuda: optimize memory access Read 4 bytes at a time (8 elements) when performing mul_mat_vec_mxfp4 * mac: fix crash on old macos versions cblas_sgemm is only supported on v13.3 and up, however bf16 is only supported on v14+ so we were falling back to ggml-blas and crashing on bf16 tensors. Checking for the function being null seems to be the simplest way to condittionally avoid registering the backend. * server: Minimum context length for gptoss This model requires a minimum context length of 8192 to function effectively. Users can set higher values through all normal mechanisms but lower values will be silently reset. * ggml: Multiply by numParallel for gptoss sliding window When computing the graph size estimate, the context size is already multiplied by numParallel so estimates reflect that. However, since sliding window models use a smaller, fixed context size, they need to manually take numParallel into account. * gpt-oss integration includes harmony parser and thinking levels, etc. * fix sync * fix tests * fix lint --------- Co-authored-by: Daniel Hiltgen <daniel@ollama.com> Co-authored-by: Jesse Gross <jesse@ollama.com> Co-authored-by: Devon Rifkin <drifkin@drifkin.net>
2025-08-06 03:21:16 +08:00
// responding. Can be a boolean (true/false) or a string ("high", "medium", "low")
// for supported models.
Think *ThinkValue `json:"think,omitempty"`
// DebugRenderOnly is a debug option that, when set to true, returns the rendered
// template instead of calling the model.
DebugRenderOnly bool `json:"_debug_render_only,omitempty"`
2023-12-06 03:57:33 +08:00
}
type Tools []Tool
func (t Tools) String() string {
bts, _ := json.Marshal(t)
return string(bts)
}
func (t Tool) String() string {
bts, _ := json.Marshal(t)
return string(bts)
}
// Message is a single message in a chat sequence. The message contains the
// role ("system", "user", or "assistant"), the content and an optional list
// of images.
2023-12-06 03:57:33 +08:00
type Message struct {
Role string `json:"role"`
Content string `json:"content"`
// Thinking contains the text that was inside thinking tags in the
// original model output when ChatRequest.Think is enabled.
Thinking string `json:"thinking,omitempty"`
2024-06-21 04:45:47 +08:00
Images []ImageData `json:"images,omitempty"`
ToolCalls []ToolCall `json:"tool_calls,omitempty"`
ToolName string `json:"tool_name,omitempty"`
2024-06-21 04:45:47 +08:00
}
func (m *Message) UnmarshalJSON(b []byte) error {
type Alias Message
var a Alias
if err := json.Unmarshal(b, &a); err != nil {
return err
}
*m = Message(a)
m.Role = strings.ToLower(m.Role)
return nil
}
type ToolCall struct {
Function ToolCallFunction `json:"function"`
}
type ToolCallFunction struct {
Index int `json:"index,omitempty"`
Name string `json:"name"`
Arguments ToolCallFunctionArguments `json:"arguments"`
}
type ToolCallFunctionArguments map[string]any
func (t *ToolCallFunctionArguments) String() string {
bts, _ := json.Marshal(t)
return string(bts)
}
type Tool struct {
Type string `json:"type"`
Items any `json:"items,omitempty"`
Function ToolFunction `json:"function"`
}
// PropertyType can be either a string or an array of strings
type PropertyType []string
// UnmarshalJSON implements the json.Unmarshaler interface
func (pt *PropertyType) UnmarshalJSON(data []byte) error {
// Try to unmarshal as a string first
var s string
if err := json.Unmarshal(data, &s); err == nil {
*pt = []string{s}
return nil
}
// If that fails, try to unmarshal as an array of strings
var a []string
if err := json.Unmarshal(data, &a); err != nil {
return err
}
*pt = a
return nil
}
// MarshalJSON implements the json.Marshaler interface
func (pt PropertyType) MarshalJSON() ([]byte, error) {
if len(pt) == 1 {
// If there's only one type, marshal as a string
return json.Marshal(pt[0])
}
// Otherwise marshal as an array
return json.Marshal([]string(pt))
}
// String returns a string representation of the PropertyType
func (pt PropertyType) String() string {
if len(pt) == 0 {
return ""
}
if len(pt) == 1 {
return pt[0]
}
return fmt.Sprintf("%v", []string(pt))
}
type ToolProperty struct {
AnyOf []ToolProperty `json:"anyOf,omitempty"`
Type PropertyType `json:"type"`
Items any `json:"items,omitempty"`
Description string `json:"description"`
Enum []any `json:"enum,omitempty"`
}
// ToTypeScriptType converts a ToolProperty to a TypeScript type string
func (tp ToolProperty) ToTypeScriptType() string {
if len(tp.AnyOf) > 0 {
var types []string
for _, anyOf := range tp.AnyOf {
types = append(types, anyOf.ToTypeScriptType())
}
return strings.Join(types, " | ")
}
if len(tp.Type) == 0 {
return "any"
}
if len(tp.Type) == 1 {
return mapToTypeScriptType(tp.Type[0])
}
var types []string
for _, t := range tp.Type {
types = append(types, mapToTypeScriptType(t))
}
return strings.Join(types, " | ")
}
// mapToTypeScriptType maps JSON Schema types to TypeScript types
func mapToTypeScriptType(jsonType string) string {
switch jsonType {
case "string":
return "string"
case "number", "integer":
return "number"
case "boolean":
return "boolean"
case "array":
return "any[]"
case "object":
return "Record<string, any>"
case "null":
return "null"
default:
return "any"
}
}
type ToolFunctionParameters struct {
Type string `json:"type"`
Defs any `json:"$defs,omitempty"`
Items any `json:"items,omitempty"`
Required []string `json:"required"`
Properties map[string]ToolProperty `json:"properties"`
}
func (t *ToolFunctionParameters) String() string {
bts, _ := json.Marshal(t)
return string(bts)
}
type ToolFunction struct {
Name string `json:"name"`
Description string `json:"description"`
Parameters ToolFunctionParameters `json:"parameters"`
}
func (t *ToolFunction) String() string {
bts, _ := json.Marshal(t)
return string(bts)
}
// ChatResponse is the response returned by [Client.Chat]. Its fields are
// similar to [GenerateResponse].
2023-12-06 03:57:33 +08:00
type ChatResponse struct {
2024-05-10 04:30:14 +08:00
Model string `json:"model"`
CreatedAt time.Time `json:"created_at"`
Message Message `json:"message"`
2024-05-10 07:45:29 +08:00
DoneReason string `json:"done_reason,omitempty"`
2023-12-06 03:57:33 +08:00
Done bool `json:"done"`
Metrics
}
// DebugInfo contains debug information for template rendering
type DebugInfo struct {
RenderedTemplate string `json:"rendered_template"`
ImageCount int `json:"image_count,omitempty"`
}
// DebugTemplateResponse is returned when _debug_render_only is set to true
type DebugTemplateResponse struct {
Model string `json:"model"`
CreatedAt time.Time `json:"created_at"`
DebugInfo DebugInfo `json:"_debug_info"`
}
2023-12-06 03:57:33 +08:00
type Metrics struct {
TotalDuration time.Duration `json:"total_duration,omitempty"`
LoadDuration time.Duration `json:"load_duration,omitempty"`
PromptEvalCount int `json:"prompt_eval_count,omitempty"`
PromptEvalDuration time.Duration `json:"prompt_eval_duration,omitempty"`
EvalCount int `json:"eval_count,omitempty"`
EvalDuration time.Duration `json:"eval_duration,omitempty"`
}
// Options specified in [GenerateRequest]. If you add a new option here, also
// add it to the API docs.
type Options struct {
Runner
// Predict options used at runtime
NumKeep int `json:"num_keep,omitempty"`
Seed int `json:"seed,omitempty"`
NumPredict int `json:"num_predict,omitempty"`
TopK int `json:"top_k,omitempty"`
TopP float32 `json:"top_p,omitempty"`
MinP float32 `json:"min_p,omitempty"`
TypicalP float32 `json:"typical_p,omitempty"`
RepeatLastN int `json:"repeat_last_n,omitempty"`
Temperature float32 `json:"temperature,omitempty"`
RepeatPenalty float32 `json:"repeat_penalty,omitempty"`
PresencePenalty float32 `json:"presence_penalty,omitempty"`
FrequencyPenalty float32 `json:"frequency_penalty,omitempty"`
Stop []string `json:"stop,omitempty"`
}
// Runner options which must be set when the model is loaded into memory
type Runner struct {
NumCtx int `json:"num_ctx,omitempty"`
NumBatch int `json:"num_batch,omitempty"`
NumGPU int `json:"num_gpu,omitempty"`
MainGPU int `json:"main_gpu,omitempty"`
UseMMap *bool `json:"use_mmap,omitempty"`
NumThread int `json:"num_thread,omitempty"`
}
// EmbedRequest is the request passed to [Client.Embed].
type EmbedRequest struct {
// Model is the model name.
Model string `json:"model"`
// Input is the input to embed.
Input any `json:"input"`
// KeepAlive controls how long the model will stay loaded in memory following
// this request.
KeepAlive *Duration `json:"keep_alive,omitempty"`
Truncate *bool `json:"truncate,omitempty"`
// Options lists model-specific options.
Options map[string]any `json:"options"`
}
// EmbedResponse is the response from [Client.Embed].
type EmbedResponse struct {
Model string `json:"model"`
Embeddings [][]float32 `json:"embeddings"`
TotalDuration time.Duration `json:"total_duration,omitempty"`
LoadDuration time.Duration `json:"load_duration,omitempty"`
PromptEvalCount int `json:"prompt_eval_count,omitempty"`
}
// EmbeddingRequest is the request passed to [Client.Embeddings].
type EmbeddingRequest struct {
// Model is the model name.
Model string `json:"model"`
// Prompt is the textual prompt to embed.
Prompt string `json:"prompt"`
// KeepAlive controls how long the model will stay loaded in memory following
// this request.
KeepAlive *Duration `json:"keep_alive,omitempty"`
// Options lists model-specific options.
Options map[string]any `json:"options"`
}
// EmbeddingResponse is the response from [Client.Embeddings].
type EmbeddingResponse struct {
Embedding []float64 `json:"embedding"`
}
// CreateRequest is the request passed to [Client.Create].
type CreateRequest struct {
Model string `json:"model"`
Stream *bool `json:"stream,omitempty"`
Quantize string `json:"quantize,omitempty"`
From string `json:"from,omitempty"`
Files map[string]string `json:"files,omitempty"`
Adapters map[string]string `json:"adapters,omitempty"`
Template string `json:"template,omitempty"`
License any `json:"license,omitempty"`
System string `json:"system,omitempty"`
Parameters map[string]any `json:"parameters,omitempty"`
Messages []Message `json:"messages,omitempty"`
2024-08-28 08:57:34 +08:00
// Deprecated: set the model name with Model instead
Name string `json:"name"`
2024-08-28 08:57:34 +08:00
// Deprecated: use Quantize instead
Quantization string `json:"quantization,omitempty"`
}
// DeleteRequest is the request passed to [Client.Delete].
2023-07-21 07:09:23 +08:00
type DeleteRequest struct {
Model string `json:"model"`
2024-08-28 08:57:34 +08:00
// Deprecated: set the model name with Model instead
2023-07-21 07:09:23 +08:00
Name string `json:"name"`
}
// ShowRequest is the request passed to [Client.Show].
2023-09-07 02:04:17 +08:00
type ShowRequest struct {
Model string `json:"model"`
System string `json:"system"`
2024-07-14 11:56:24 +08:00
// Template is deprecated
Template string `json:"template"`
Verbose bool `json:"verbose"`
Options map[string]any `json:"options"`
2024-08-28 08:57:34 +08:00
// Deprecated: set the model name with Model instead
Name string `json:"name"`
2023-09-07 02:04:17 +08:00
}
// ShowResponse is the response returned from [Client.Show].
2023-09-07 02:04:17 +08:00
type ShowResponse struct {
License string `json:"license,omitempty"`
Modelfile string `json:"modelfile,omitempty"`
Parameters string `json:"parameters,omitempty"`
Template string `json:"template,omitempty"`
System string `json:"system,omitempty"`
Details ModelDetails `json:"details,omitempty"`
Messages []Message `json:"messages,omitempty"`
ModelInfo map[string]any `json:"model_info,omitempty"`
ProjectorInfo map[string]any `json:"projector_info,omitempty"`
Tensors []Tensor `json:"tensors,omitempty"`
Capabilities []model.Capability `json:"capabilities,omitempty"`
ModifiedAt time.Time `json:"modified_at,omitempty"`
2023-09-07 02:04:17 +08:00
}
// CopyRequest is the request passed to [Client.Copy].
2023-07-24 23:27:28 +08:00
type CopyRequest struct {
Source string `json:"source"`
Destination string `json:"destination"`
}
// PullRequest is the request passed to [Client.Pull].
2023-07-07 00:24:49 +08:00
type PullRequest struct {
Model string `json:"model"`
Insecure bool `json:"insecure,omitempty"` // Deprecated: ignored
Username string `json:"username"` // Deprecated: ignored
Password string `json:"password"` // Deprecated: ignored
Stream *bool `json:"stream,omitempty"`
2024-08-28 08:57:34 +08:00
// Deprecated: set the model name with Model instead
Name string `json:"name"`
2023-07-07 00:24:49 +08:00
}
// ProgressResponse is the response passed to progress functions like
// [PullProgressFunc] and [PushProgressFunc].
2023-07-19 09:51:30 +08:00
type ProgressResponse struct {
2023-07-21 07:09:23 +08:00
Status string `json:"status"`
Digest string `json:"digest,omitempty"`
2023-09-29 01:00:34 +08:00
Total int64 `json:"total,omitempty"`
Completed int64 `json:"completed,omitempty"`
2023-07-07 02:18:40 +08:00
}
// PushRequest is the request passed to [Client.Push].
type PushRequest struct {
Model string `json:"model"`
Insecure bool `json:"insecure,omitempty"`
Username string `json:"username"`
Password string `json:"password"`
Stream *bool `json:"stream,omitempty"`
2024-08-28 08:57:34 +08:00
// Deprecated: set the model name with Model instead
Name string `json:"name"`
}
2023-07-07 08:09:48 +08:00
// ListResponse is the response from [Client.List].
2023-07-19 00:09:45 +08:00
type ListResponse struct {
Models []ListModelResponse `json:"models"`
2023-07-19 00:09:45 +08:00
}
// ProcessResponse is the response from [Client.Process].
type ProcessResponse struct {
Models []ProcessModelResponse `json:"models"`
}
// ListModelResponse is a single model description in [ListResponse].
type ListModelResponse struct {
Name string `json:"name"`
Model string `json:"model"`
ModifiedAt time.Time `json:"modified_at"`
Size int64 `json:"size"`
Digest string `json:"digest"`
Details ModelDetails `json:"details,omitempty"`
}
// ProcessModelResponse is a single model description in [ProcessResponse].
type ProcessModelResponse struct {
Name string `json:"name"`
Model string `json:"model"`
Size int64 `json:"size"`
Digest string `json:"digest"`
Details ModelDetails `json:"details,omitempty"`
ExpiresAt time.Time `json:"expires_at"`
SizeVRAM int64 `json:"size_vram"`
ContextLength int `json:"context_length"`
2023-07-19 00:09:45 +08:00
}
2023-08-11 02:34:25 +08:00
type TokenResponse struct {
Token string `json:"token"`
}
// GenerateResponse is the response passed into [GenerateResponseFunc].
2023-07-08 06:29:17 +08:00
type GenerateResponse struct {
// Model is the model name that generated the response.
Model string `json:"model"`
2024-05-22 13:07:57 +08:00
// CreatedAt is the timestamp of the response.
2023-07-13 09:18:06 +08:00
CreatedAt time.Time `json:"created_at"`
// Response is the textual response itself.
Response string `json:"response"`
// Thinking contains the text that was inside thinking tags in the
// original model output when ChatRequest.Think is enabled.
Thinking string `json:"thinking,omitempty"`
// Done specifies if the response is complete.
Done bool `json:"done"`
2024-05-10 04:30:14 +08:00
// DoneReason is the reason the model stopped generating text.
2024-05-10 07:45:29 +08:00
DoneReason string `json:"done_reason,omitempty"`
2024-05-10 04:30:14 +08:00
// Context is an encoding of the conversation used in this response; this
// can be sent in the next request to keep a conversational memory.
Context []int `json:"context,omitempty"`
2023-07-13 09:18:06 +08:00
2023-12-06 03:57:33 +08:00
Metrics
gpt-oss (#11672) * bf16 * tests * gpt-oss * enable gptoss for engine * rough estimate * convert to mxfp4 * handle safetensors U8 * clamp glu/linear * update tokenizer * MXFP4 support This implements the Open Compute Microscaling (MX) FP4 format as a tensor type with backend implementations focusing on mulmat and mulmatid on CPU, CUDA, and Metal. * Unit tests for MXFP4 support This exercises various operations and shapes on both CPU and GPU (if detected on the system) * cuda graph * unit test adjustments * cuda: optimize memory access Read 4 bytes at a time (8 elements) when performing mul_mat_vec_mxfp4 * mac: fix crash on old macos versions cblas_sgemm is only supported on v13.3 and up, however bf16 is only supported on v14+ so we were falling back to ggml-blas and crashing on bf16 tensors. Checking for the function being null seems to be the simplest way to condittionally avoid registering the backend. * server: Minimum context length for gptoss This model requires a minimum context length of 8192 to function effectively. Users can set higher values through all normal mechanisms but lower values will be silently reset. * ggml: Multiply by numParallel for gptoss sliding window When computing the graph size estimate, the context size is already multiplied by numParallel so estimates reflect that. However, since sliding window models use a smaller, fixed context size, they need to manually take numParallel into account. * gpt-oss integration includes harmony parser and thinking levels, etc. * fix sync * fix tests * fix lint --------- Co-authored-by: Daniel Hiltgen <daniel@ollama.com> Co-authored-by: Jesse Gross <jesse@ollama.com> Co-authored-by: Devon Rifkin <drifkin@drifkin.net>
2025-08-06 03:21:16 +08:00
ToolCalls []ToolCall `json:"tool_calls,omitempty"`
2023-07-13 09:18:06 +08:00
}
// ModelDetails provides details about a model.
type ModelDetails struct {
2024-01-26 04:12:36 +08:00
ParentModel string `json:"parent_model"`
Format string `json:"format"`
Family string `json:"family"`
Families []string `json:"families"`
ParameterSize string `json:"parameter_size"`
QuantizationLevel string `json:"quantization_level"`
}
// Tensor describes the metadata for a given tensor.
type Tensor struct {
Name string `json:"name"`
Type string `json:"type"`
Shape []uint64 `json:"shape"`
}
2023-12-06 03:57:33 +08:00
func (m *Metrics) Summary() {
if m.TotalDuration > 0 {
fmt.Fprintf(os.Stderr, "total duration: %v\n", m.TotalDuration)
2023-07-13 09:18:06 +08:00
}
2023-12-06 03:57:33 +08:00
if m.LoadDuration > 0 {
fmt.Fprintf(os.Stderr, "load duration: %v\n", m.LoadDuration)
2023-07-19 03:02:02 +08:00
}
2023-12-06 03:57:33 +08:00
if m.PromptEvalCount > 0 {
fmt.Fprintf(os.Stderr, "prompt eval count: %d token(s)\n", m.PromptEvalCount)
2023-07-13 09:18:06 +08:00
}
2023-12-06 03:57:33 +08:00
if m.PromptEvalDuration > 0 {
fmt.Fprintf(os.Stderr, "prompt eval duration: %s\n", m.PromptEvalDuration)
fmt.Fprintf(os.Stderr, "prompt eval rate: %.2f tokens/s\n", float64(m.PromptEvalCount)/m.PromptEvalDuration.Seconds())
2023-07-13 09:18:06 +08:00
}
2023-12-06 03:57:33 +08:00
if m.EvalCount > 0 {
fmt.Fprintf(os.Stderr, "eval count: %d token(s)\n", m.EvalCount)
2023-07-13 09:18:06 +08:00
}
2023-12-06 03:57:33 +08:00
if m.EvalDuration > 0 {
fmt.Fprintf(os.Stderr, "eval duration: %s\n", m.EvalDuration)
fmt.Fprintf(os.Stderr, "eval rate: %.2f tokens/s\n", float64(m.EvalCount)/m.EvalDuration.Seconds())
2023-07-13 09:18:06 +08:00
}
2023-07-07 08:09:48 +08:00
}
func (opts *Options) FromMap(m map[string]any) error {
valueOpts := reflect.ValueOf(opts).Elem() // names of the fields in the options struct
typeOpts := reflect.TypeOf(opts).Elem() // types of the fields in the options struct
// build map of json struct tags to their types
jsonOpts := make(map[string]reflect.StructField)
for _, field := range reflect.VisibleFields(typeOpts) {
jsonTag := strings.Split(field.Tag.Get("json"), ",")[0]
if jsonTag != "" {
jsonOpts[jsonTag] = field
}
}
for key, val := range m {
opt, ok := jsonOpts[key]
if !ok {
slog.Warn("invalid option provided", "option", key)
continue
}
2023-08-18 06:50:46 +08:00
field := valueOpts.FieldByName(opt.Name)
if field.IsValid() && field.CanSet() {
if val == nil {
continue
}
switch field.Kind() {
case reflect.Int:
switch t := val.(type) {
case int64:
field.SetInt(t)
case float64:
// when JSON unmarshals numbers, it uses float64, not int
field.SetInt(int64(t))
default:
return fmt.Errorf("option %q must be of type integer", key)
}
case reflect.Bool:
val, ok := val.(bool)
if !ok {
return fmt.Errorf("option %q must be of type boolean", key)
}
field.SetBool(val)
case reflect.Float32:
// JSON unmarshals to float64
val, ok := val.(float64)
if !ok {
return fmt.Errorf("option %q must be of type float32", key)
}
field.SetFloat(val)
case reflect.String:
val, ok := val.(string)
if !ok {
return fmt.Errorf("option %q must be of type string", key)
}
field.SetString(val)
case reflect.Slice:
// JSON unmarshals to []any, not []string
val, ok := val.([]any)
if !ok {
return fmt.Errorf("option %q must be of type array", key)
}
// convert []any to []string
slice := make([]string, len(val))
for i, item := range val {
str, ok := item.(string)
if !ok {
return fmt.Errorf("option %q must be of an array of strings", key)
}
slice[i] = str
}
field.Set(reflect.ValueOf(slice))
case reflect.Pointer:
var b bool
if field.Type() == reflect.TypeOf(&b) {
val, ok := val.(bool)
if !ok {
return fmt.Errorf("option %q must be of type boolean", key)
}
field.Set(reflect.ValueOf(&val))
} else {
return fmt.Errorf("unknown type loading config params: %v %v", field.Kind(), field.Type())
}
default:
return fmt.Errorf("unknown type loading config params: %v", field.Kind())
}
}
}
return nil
}
// DefaultOptions is the default set of options for [GenerateRequest]; these
// values are used unless the user specifies other values explicitly.
2023-07-08 06:29:17 +08:00
func DefaultOptions() Options {
return Options{
// options set on request to runner
NumPredict: -1,
// set a minimal num_keep to avoid issues on context shifts
NumKeep: 4,
2023-07-08 06:29:17 +08:00
Temperature: 0.8,
TopK: 40,
TopP: 0.9,
TypicalP: 1.0,
RepeatLastN: 64,
RepeatPenalty: 1.1,
PresencePenalty: 0.0,
FrequencyPenalty: 0.0,
Seed: -1,
2023-07-08 06:29:17 +08:00
Runner: Runner{
// options set when the model is loaded
NumCtx: int(envconfig.ContextLength()),
2024-04-05 23:49:04 +08:00
NumBatch: 512,
NumGPU: -1, // -1 here indicates that NumGPU should be set dynamically
NumThread: 0, // let the runtime decide
UseMMap: nil,
},
2023-07-08 06:29:17 +08:00
}
}
2023-07-20 06:00:28 +08:00
gpt-oss (#11672) * bf16 * tests * gpt-oss * enable gptoss for engine * rough estimate * convert to mxfp4 * handle safetensors U8 * clamp glu/linear * update tokenizer * MXFP4 support This implements the Open Compute Microscaling (MX) FP4 format as a tensor type with backend implementations focusing on mulmat and mulmatid on CPU, CUDA, and Metal. * Unit tests for MXFP4 support This exercises various operations and shapes on both CPU and GPU (if detected on the system) * cuda graph * unit test adjustments * cuda: optimize memory access Read 4 bytes at a time (8 elements) when performing mul_mat_vec_mxfp4 * mac: fix crash on old macos versions cblas_sgemm is only supported on v13.3 and up, however bf16 is only supported on v14+ so we were falling back to ggml-blas and crashing on bf16 tensors. Checking for the function being null seems to be the simplest way to condittionally avoid registering the backend. * server: Minimum context length for gptoss This model requires a minimum context length of 8192 to function effectively. Users can set higher values through all normal mechanisms but lower values will be silently reset. * ggml: Multiply by numParallel for gptoss sliding window When computing the graph size estimate, the context size is already multiplied by numParallel so estimates reflect that. However, since sliding window models use a smaller, fixed context size, they need to manually take numParallel into account. * gpt-oss integration includes harmony parser and thinking levels, etc. * fix sync * fix tests * fix lint --------- Co-authored-by: Daniel Hiltgen <daniel@ollama.com> Co-authored-by: Jesse Gross <jesse@ollama.com> Co-authored-by: Devon Rifkin <drifkin@drifkin.net>
2025-08-06 03:21:16 +08:00
// ThinkValue represents a value that can be a boolean or a string ("high", "medium", "low")
type ThinkValue struct {
// Value can be a bool or string
Value interface{}
}
// IsValid checks if the ThinkValue is valid
func (t *ThinkValue) IsValid() bool {
if t == nil || t.Value == nil {
return true // nil is valid (means not set)
}
switch v := t.Value.(type) {
case bool:
return true
case string:
return v == "high" || v == "medium" || v == "low"
default:
return false
}
}
// IsBool returns true if the value is a boolean
func (t *ThinkValue) IsBool() bool {
if t == nil || t.Value == nil {
return false
}
_, ok := t.Value.(bool)
return ok
}
// IsString returns true if the value is a string
func (t *ThinkValue) IsString() bool {
if t == nil || t.Value == nil {
return false
}
_, ok := t.Value.(string)
return ok
}
// Bool returns the value as a bool (true if enabled in any way)
func (t *ThinkValue) Bool() bool {
gpt-oss (#11672) * bf16 * tests * gpt-oss * enable gptoss for engine * rough estimate * convert to mxfp4 * handle safetensors U8 * clamp glu/linear * update tokenizer * MXFP4 support This implements the Open Compute Microscaling (MX) FP4 format as a tensor type with backend implementations focusing on mulmat and mulmatid on CPU, CUDA, and Metal. * Unit tests for MXFP4 support This exercises various operations and shapes on both CPU and GPU (if detected on the system) * cuda graph * unit test adjustments * cuda: optimize memory access Read 4 bytes at a time (8 elements) when performing mul_mat_vec_mxfp4 * mac: fix crash on old macos versions cblas_sgemm is only supported on v13.3 and up, however bf16 is only supported on v14+ so we were falling back to ggml-blas and crashing on bf16 tensors. Checking for the function being null seems to be the simplest way to condittionally avoid registering the backend. * server: Minimum context length for gptoss This model requires a minimum context length of 8192 to function effectively. Users can set higher values through all normal mechanisms but lower values will be silently reset. * ggml: Multiply by numParallel for gptoss sliding window When computing the graph size estimate, the context size is already multiplied by numParallel so estimates reflect that. However, since sliding window models use a smaller, fixed context size, they need to manually take numParallel into account. * gpt-oss integration includes harmony parser and thinking levels, etc. * fix sync * fix tests * fix lint --------- Co-authored-by: Daniel Hiltgen <daniel@ollama.com> Co-authored-by: Jesse Gross <jesse@ollama.com> Co-authored-by: Devon Rifkin <drifkin@drifkin.net>
2025-08-06 03:21:16 +08:00
if t == nil || t.Value == nil {
return false
}
switch v := t.Value.(type) {
case bool:
return v
case string:
// Any string value ("high", "medium", "low") means thinking is enabled
return v == "high" || v == "medium" || v == "low"
default:
return false
}
}
// String returns the value as a string
func (t *ThinkValue) String() string {
gpt-oss (#11672) * bf16 * tests * gpt-oss * enable gptoss for engine * rough estimate * convert to mxfp4 * handle safetensors U8 * clamp glu/linear * update tokenizer * MXFP4 support This implements the Open Compute Microscaling (MX) FP4 format as a tensor type with backend implementations focusing on mulmat and mulmatid on CPU, CUDA, and Metal. * Unit tests for MXFP4 support This exercises various operations and shapes on both CPU and GPU (if detected on the system) * cuda graph * unit test adjustments * cuda: optimize memory access Read 4 bytes at a time (8 elements) when performing mul_mat_vec_mxfp4 * mac: fix crash on old macos versions cblas_sgemm is only supported on v13.3 and up, however bf16 is only supported on v14+ so we were falling back to ggml-blas and crashing on bf16 tensors. Checking for the function being null seems to be the simplest way to condittionally avoid registering the backend. * server: Minimum context length for gptoss This model requires a minimum context length of 8192 to function effectively. Users can set higher values through all normal mechanisms but lower values will be silently reset. * ggml: Multiply by numParallel for gptoss sliding window When computing the graph size estimate, the context size is already multiplied by numParallel so estimates reflect that. However, since sliding window models use a smaller, fixed context size, they need to manually take numParallel into account. * gpt-oss integration includes harmony parser and thinking levels, etc. * fix sync * fix tests * fix lint --------- Co-authored-by: Daniel Hiltgen <daniel@ollama.com> Co-authored-by: Jesse Gross <jesse@ollama.com> Co-authored-by: Devon Rifkin <drifkin@drifkin.net>
2025-08-06 03:21:16 +08:00
if t == nil || t.Value == nil {
return ""
}
switch v := t.Value.(type) {
case string:
return v
case bool:
if v {
return "medium" // Default level when just true
}
return ""
default:
return ""
}
}
// UnmarshalJSON implements json.Unmarshaler
func (t *ThinkValue) UnmarshalJSON(data []byte) error {
// Try to unmarshal as bool first
var b bool
if err := json.Unmarshal(data, &b); err == nil {
t.Value = b
return nil
}
// Try to unmarshal as string
var s string
if err := json.Unmarshal(data, &s); err == nil {
// Validate string values
if s != "high" && s != "medium" && s != "low" {
return fmt.Errorf("invalid think value: %q (must be \"high\", \"medium\", \"low\", true, or false)", s)
}
t.Value = s
return nil
}
return fmt.Errorf("think must be a boolean or string (\"high\", \"medium\", \"low\")")
}
// MarshalJSON implements json.Marshaler
func (t *ThinkValue) MarshalJSON() ([]byte, error) {
if t == nil || t.Value == nil {
return []byte("null"), nil
}
return json.Marshal(t.Value)
}
2023-07-20 06:00:28 +08:00
type Duration struct {
time.Duration
}
func (d Duration) MarshalJSON() ([]byte, error) {
if d.Duration < 0 {
return []byte("-1"), nil
}
return []byte("\"" + d.Duration.String() + "\""), nil
}
2023-07-20 06:00:28 +08:00
func (d *Duration) UnmarshalJSON(b []byte) (err error) {
var v any
if err := json.Unmarshal(b, &v); err != nil {
return err
}
d.Duration = 5 * time.Minute
switch t := v.(type) {
case float64:
if t < 0 {
2024-02-14 07:40:32 +08:00
d.Duration = time.Duration(math.MaxInt64)
} else {
2025-08-28 02:51:25 +08:00
d.Duration = time.Duration(t * float64(time.Second))
2023-07-20 06:00:28 +08:00
}
case string:
d.Duration, err = time.ParseDuration(t)
if err != nil {
return err
}
if d.Duration < 0 {
2024-02-14 07:40:32 +08:00
d.Duration = time.Duration(math.MaxInt64)
}
default:
return fmt.Errorf("Unsupported type: '%s'", reflect.TypeOf(v))
2023-07-20 06:00:28 +08:00
}
return nil
}
// FormatParams converts specified parameter options to their correct types
func FormatParams(params map[string][]string) (map[string]any, error) {
opts := Options{}
valueOpts := reflect.ValueOf(&opts).Elem() // names of the fields in the options struct
typeOpts := reflect.TypeOf(opts) // types of the fields in the options struct
// build map of json struct tags to their types
jsonOpts := make(map[string]reflect.StructField)
for _, field := range reflect.VisibleFields(typeOpts) {
jsonTag := strings.Split(field.Tag.Get("json"), ",")[0]
if jsonTag != "" {
jsonOpts[jsonTag] = field
}
}
out := make(map[string]any)
// iterate params and set values based on json struct tags
for key, vals := range params {
if opt, ok := jsonOpts[key]; !ok {
return nil, fmt.Errorf("unknown parameter '%s'", key)
} else {
field := valueOpts.FieldByName(opt.Name)
if field.IsValid() && field.CanSet() {
switch field.Kind() {
case reflect.Float32:
floatVal, err := strconv.ParseFloat(vals[0], 32)
if err != nil {
return nil, fmt.Errorf("invalid float value %s", vals)
}
out[key] = float32(floatVal)
case reflect.Int:
intVal, err := strconv.ParseInt(vals[0], 10, 64)
if err != nil {
return nil, fmt.Errorf("invalid int value %s", vals)
}
out[key] = intVal
case reflect.Bool:
boolVal, err := strconv.ParseBool(vals[0])
if err != nil {
return nil, fmt.Errorf("invalid bool value %s", vals)
}
out[key] = boolVal
case reflect.String:
out[key] = vals[0]
case reflect.Slice:
// TODO: only string slices are supported right now
out[key] = vals
case reflect.Pointer:
var b bool
if field.Type() == reflect.TypeOf(&b) {
boolVal, err := strconv.ParseBool(vals[0])
if err != nil {
return nil, fmt.Errorf("invalid bool value %s", vals)
}
out[key] = &boolVal
} else {
return nil, fmt.Errorf("unknown type %s for %s", field.Kind(), key)
}
default:
return nil, fmt.Errorf("unknown type %s for %s", field.Kind(), key)
}
}
}
}
return out, nil
}