Files
ollama/model/model.go
Daniel Hiltgen 96b202d34b Add support for gemma4 (#15214)
* bench: add prompt calibration, context size flag, and NumCtx reporting

Add --num-ctx flag to set context size, and report NumCtx in model info
header. Calibrate tokens-per-word ratio during warmup using actual
tokenization metrics from the model, replacing the fixed 1.3 heuristic.
This produces more accurate prompt token counts for --prompt-tokens.

Also add fetchContextLength() to query running model context via /api/ps.

* integration: improve vision test robustness and add thinking tests

Add skipIfNoVisionOverride() to skip vision tests when OLLAMA_TEST_MODEL
is set to a non-vision model. Add Think:false to context exhaustion test
to prevent thinking models from using all context before the test can
measure it. Add third test image (ollama homepage) and replace OCR test
with ImageDescription test using it. Relax match strings for broader
model compatibility. Add TestThinkingEnabled and TestThinkingSuppressed
to verify thinking output and channel tag handling.

* gemma4: add Gemma 4 GGML model support

Add full Gemma 4 model family support (E2B, E4B, 26B MoE, 31B Dense)
for the GGML backend including text, vision, converter, parser, and
renderer.

Text model features:
- Sliding window + full attention with per-layer patterns
- KV sharing across layers with donor map
- Per-layer embeddings (PLE) with learned projections
- MoE routing with RMSNorm + learned scale
- Proportional RoPE with freq_factors for global attention
- Final logit softcapping

Vision model features:
- SigLIP vision encoder with 2D RoPE
- ClippableLinear with input/output clamping via packed v.clamp_data
- Adaptive average pooling with nMerge kernel
- Multi-modal projection with unweighted RMSNorm

Converter:
- Safetensors to GGUF with vision tensor renaming
- Fused MoE gate_up_proj splitting
- Vision patch embedding reshape (HF to Conv2D layout)
- Packed clamp data tensor for ClippableLinear bounds
- Proportional RoPE freq_factors generation

Also includes:
- BackendGet() on ml.Tensor for reading weight tensor data
- Q6_K CUDA get_rows kernel support
- MoE-aware ffn_down quantization layer counting
- Gemma4 parser with tool calling and thinking support
- Gemma4 renderer with structured tool format
- Architecture-based auto-detection of renderer/parser/stop tokens
- Integration test gemma4 model list additions

* gemma4: add audio support with USM conformer encoder

Add audio encoding for Gemma 4 using the USM conformer architecture:
- Converter: audio tensor mapping, SSCP/conformer/embedder name replacements,
  softplus repacker for per_dim_scale, F32 enforcement for conv weights
- GGML backend: Conv1DDW and PadExt tensor ops
- Audio encoder: SSCP Conv2D, 12 conformer blocks (FFW + block-local
  attention with relative position embeddings + LightConv1d + FFW),
  output projection, audio-to-text embedding projector
- Audio preprocessing: WAV decode, mel spectrogram, FFT (pure Go)
- Model wiring: WAV detection, audio token handling, unified PostTokenize

Correctly transcribes "why is the sky blue" from test audio.

* integration: add gemma4 audio tests including OpenAI API coverage

Test audio transcription and response via the Ollama native API, plus
two new tests exercising the OpenAI-compatible endpoints:
- /v1/audio/transcriptions (multipart form upload)
- /v1/chat/completions with input_audio content type

All tests use capability checks and skip models without audio support.

* gemma4: add OpenAI audio API support and capability detection

- Add CapabilityAudio and detect from audio.block_count in GGUF
- Add /v1/audio/transcriptions endpoint with TranscriptionMiddleware
- Add input_audio content type support in /v1/chat/completions
- Add TranscriptionRequest/Response types in openai package

* gemma4: add audio input support for run command

- /audio toggle in interactive mode for voice chat
- Platform-specific microphone recording (AVFoundation on macOS,
  PulseAudio/ALSA on Linux, WASAPI on Windows)
- Space to start/stop recording, automatic chunking for long audio

* gemma4: add transcribe command (ollama transcribe MODEL)

- Interactive mode with readline prompt and slash commands
- Non-interactive mode for piped audio or record-until-Ctrl+C
- Chunked streaming transcription for long recordings
- Word-wrapped output matching run command style

* gemma4: add parser, renderer, and integration test plumbing

* gemma4: fix renderer to emit BOS token

* gemma4: add OpenAI audio transcription API and input_audio support

* gemma4: update converter for new weight drop naming

* gemma4: add per_expert_scale to MoE router and fix moe_intermediate_size config

* gemma4: rewrite renderer to match HF Jinja2 template exactly

Fix 8 bugs found by building 55 reference tests verified against the
HF Jinja2 chat template (VERIFY_JINJA2=1 shells out to Python):

- Tool responses use separate <|turn>tool turns (not inline tags)
- Tool calls emitted before content in assistant messages
- Thinking content stripped from assistant history (strip_thinking)
- User, tool, and system content trimmed (template does | trim)
- Empty system message still emits system turn (check role, not content)
- Nested object properties rendered recursively with required field
- Array items specification rendered for array-type properties
- OBJECT/ARRAY type-specific rendering comma logic matches template

Also adds Required field to api.ToolProperty for nested object schemas,
replaces old gemma4_test.go with comprehensive gemma4_reference_test.go,
and commits the Jinja2 template as testdata for verification.

* gemma4: fix MoE fused gate_up split and multiline tool-call arg parsing

- Text MoE: split `ffn_gate_up_exps` into contiguous `[gate|up]` halves instead of stride-2 slices.
- Parser: escape control characters in `<|"|>...<|"|>` string literals when converting tool-call args to JSON.
- Fixes warnings like `invalid character '\n' in string literal` for multiline tool arguments.
- Add Gemma4 parser regressions for multiline tool-call args and `gemma4ArgsToJSON`.

* cmd: simplify audio input to dropped file attachments

* gemma4: use full SWA memory for better cache reuse

* gemma4: initialize clamps after backend load

* convert: align gemma4 audio tensor renames with llama.cpp

* Remove redundant comments in gemma4 vision model

* Format Gemma4 MoE block field alignment

* use 4096 kvcache.NewSWAMemCache

* convert: support new Gemma4 audio_tower tensor naming (#15221)

Co-authored-by: jmorganca <jmorganca@gmail.com>

* fix integration test defaults for audio

* review comments and lint fixes

* remove unused audio/video files

---------

Co-authored-by: jmorganca <jmorganca@gmail.com>
2026-04-02 11:33:33 -07:00

355 lines
9.1 KiB
Go

package model
import (
"errors"
"fmt"
_ "image/jpeg"
_ "image/png"
"log/slog"
"os"
"reflect"
"strconv"
"strings"
_ "golang.org/x/image/bmp"
_ "golang.org/x/image/tiff"
_ "golang.org/x/image/webp"
"github.com/ollama/ollama/fs"
fsggml "github.com/ollama/ollama/fs/ggml"
"github.com/ollama/ollama/kvcache"
"github.com/ollama/ollama/logutil"
"github.com/ollama/ollama/ml"
_ "github.com/ollama/ollama/ml/backend"
"github.com/ollama/ollama/ml/nn/pooling"
"github.com/ollama/ollama/model/input"
"github.com/ollama/ollama/tokenizer"
)
var (
ErrNoVisionModel = errors.New("this model is missing data required for image input")
ErrUnsupportedModel = errors.New("model not supported")
ErrUnsupportedTokenizer = errors.New("tokenizer not supported")
)
// Model implements a specific model architecture, defining the forward pass and any model-specific configuration
type Model interface {
Forward(ml.Context, input.Batch) (ml.Tensor, error)
Backend() ml.Backend
Config() config
}
// Validator is an optional interface that models can implement to perform
// validation after tensors have been loaded. If validation fails, model
// loading will fail with the returned error.
type Validator interface {
Validate() error
}
// PostLoader is an optional interface that models can implement to run
// initialization steps after backend weights have been loaded.
type PostLoader interface {
PostLoad() error
}
// MultimodalProcessor must be implemented by multimodal models.
type MultimodalProcessor interface {
// EncodeMultimodal processes a single input (such as an image) and
// generates an output (typically an embedding) that can be used by the model.
//
// The return value is one or more tensors, each with optional model-specific
// opaque metadata. Typically, the tensors might be views into an embedding
// with each view representing a chunk of data that can be processed independently
// in different batches.
//
// The result may be cached by the runner.
EncodeMultimodal(ml.Context, []byte) ([]input.Multimodal, error)
// PostTokenize is called after tokenization to allow the model to edit the
// input stream to correctly arrange multimodal elements.
//
// The input is a slice of tokens with the results of EncodeMultimodal interleaved
// in the order that the user provided them. Each element of the slice will be
// either a single token or single multimodal object.
//
// The model must ensure that inputs are stored according to how they will be
// processed and stored in the cache. For example, Llava-style models should insert
// placeholder tokens equal to the feature size of the corresponding image with
// the image itself attached to and split across these tokens. When Forward is called
// a partial subset of these tokens may be submitted according to the batch size.
//
// This function is also responsible for updating MultimodalHash for any Multimodal
// that is modified to ensure that there is a unique hash value that accurately
// represents the contents.
PostTokenize([]*input.Input) ([]*input.Input, error)
}
// Base implements the common fields and methods for all models
type Base struct {
b ml.Backend
config
}
type config struct {
Cache kvcache.Cache
}
// Backend returns the underlying backend that will run the model
func (m *Base) Backend() ml.Backend {
return m.b
}
func (m *Base) Config() config {
return m.config
}
var models = make(map[string]func(fs.Config) (Model, error))
// Register registers a model constructor for the given architecture
func Register(name string, f func(fs.Config) (Model, error)) {
if _, ok := models[name]; ok {
panic("model: model already registered")
}
models[name] = f
}
// New initializes a new model instance with the provided configuration based on the metadata in the model file
func New(modelPath string, params ml.BackendParams) (Model, error) {
b, err := ml.NewBackend(modelPath, params)
if err != nil {
return nil, err
}
m, err := modelForArch(b.Config())
if err != nil {
return nil, err
}
base := Base{b: b, config: m.Config()}
v := reflect.ValueOf(m)
v.Elem().Set(populateFields(base, v.Elem()))
if validator, ok := m.(Validator); ok {
if err := validator.Validate(); err != nil {
return nil, err
}
}
return m, nil
}
func NewTextProcessor(s string) (tokenizer.Tokenizer, error) {
r, err := os.Open(s)
if err != nil {
return nil, err
}
defer r.Close()
meta, err := fsggml.Decode(r, -1)
if err != nil {
return nil, err
}
m, err := modelForArch(meta.KV())
if err != nil {
return nil, err
}
tp, ok := m.(tokenizer.Tokenizer)
if !ok {
return nil, ErrUnsupportedTokenizer
}
return tp, nil
}
func modelForArch(c fs.Config) (Model, error) {
arch := c.Architecture()
if pooling.Type(c.Uint("pooling_type")) != pooling.TypeNone {
arch = arch + "_embed"
}
f, ok := models[arch]
if !ok {
return nil, ErrUnsupportedModel
}
return f(c)
}
func populateFields(base Base, v reflect.Value, tags ...Tag) reflect.Value {
t := v.Type()
if t.Kind() == reflect.Struct {
allNil := true
for i := range t.NumField() {
tt := t.Field(i).Type
vv := v.Field(i)
if !vv.CanSet() {
continue
}
// make a copy
tagsCopy := tags
if tag := t.Field(i).Tag.Get("gguf"); tag != "" {
tagsCopy = append(tagsCopy, parseTag(tag))
}
if tt == reflect.TypeOf((*Base)(nil)).Elem() {
vv.Set(reflect.ValueOf(base))
} else if tt == reflect.TypeOf((*ml.Tensor)(nil)).Elem() {
var fn func([]Tag, string, string) [][]string
fn = func(tags []Tag, prefix, suffix string) (fullNames [][]string) {
if len(tags) > 0 {
var names []string
if tags[0].name != "" {
for _, n := range append([]string{tags[0].name}, tags[0].alternatives...) {
names = append(names, prefix+n+suffix)
}
}
childNames := fn(tags[1:], tags[0].prefix, tags[0].suffix)
if len(names) == 0 {
// current tag has no name, use child names only
fullNames = append(fullNames, childNames...)
} else if len(childNames) == 0 {
// current tag has names but no children, create branches for each name
for _, name := range names {
fullNames = append(fullNames, []string{name})
}
} else {
// merge each name with each child
for _, name := range names {
for _, childName := range childNames {
fullNames = append(fullNames, append([]string{name}, childName...))
}
}
}
}
return fullNames
}
names := fn(tagsCopy, "", "")
for _, name := range names {
if tensor := base.Backend().Get(strings.Join(name, ".")); tensor != nil {
logutil.Trace("found tensor", "", tensor)
vv.Set(reflect.ValueOf(tensor))
break
}
}
} else if tt.Kind() == reflect.Pointer || tt.Kind() == reflect.Interface {
setPointer(base, vv, tagsCopy)
} else if tt.Kind() == reflect.Slice || tt.Kind() == reflect.Array {
for i := range vv.Len() {
vvv := vv.Index(i)
if vvv.Kind() == reflect.Pointer || vvv.Kind() == reflect.Interface {
setPointer(base, vvv, append(tagsCopy, Tag{name: strconv.Itoa(i)}))
} else {
vvv.Set(populateFields(base, vvv, append(tagsCopy, Tag{name: strconv.Itoa(i)})...))
}
}
}
if !canNil(tt) || !vv.IsNil() {
allNil = false
}
}
if allNil {
return reflect.Zero(t)
}
}
return v
}
func setPointer(base Base, v reflect.Value, tags []Tag) {
vv := v
if v.Kind() == reflect.Interface {
if v.IsNil() {
return
}
vv = vv.Elem()
}
vv = reflect.Indirect(vv)
if v.IsNil() {
vv = reflect.New(v.Type().Elem()).Elem()
}
if f := populateFields(base, vv, tags...); f.CanAddr() {
v.Set(f.Addr())
}
}
type Tag struct {
name,
// prefix and suffix are applied to child tags
prefix,
suffix string
alternatives []string
}
func parseTag(s string) (tag Tag) {
parts := strings.Split(s, ",")
if len(parts) > 0 {
tag.name = parts[0]
for _, part := range parts[1:] {
if value, ok := strings.CutPrefix(part, "alt:"); ok && tag.name == "" {
// elevate alternative to primary if no primary given
tag.name = value
slog.Warn("gguf tag has alt: but no primary name", "tag", s)
} else if ok {
tag.alternatives = append(tag.alternatives, value)
}
if value, ok := strings.CutPrefix(part, "pre:"); ok {
tag.prefix = value
}
if value, ok := strings.CutPrefix(part, "suf:"); ok {
tag.suffix = value
}
}
}
return
}
func canNil(t reflect.Type) bool {
return t.Kind() == reflect.Chan ||
t.Kind() == reflect.Func ||
t.Kind() == reflect.Interface ||
t.Kind() == reflect.Map ||
t.Kind() == reflect.Pointer ||
t.Kind() == reflect.Slice
}
func Forward(ctx ml.Context, m Model, batch input.Batch) (ml.Tensor, error) {
if len(batch.Positions) != len(batch.Sequences) {
return nil, fmt.Errorf("length of positions (%v) must match length of seqs (%v)", len(batch.Positions), len(batch.Sequences))
}
if len(batch.Positions) < 1 {
return nil, errors.New("batch size cannot be less than 1")
}
cache := m.Config().Cache
if cache != nil {
err := cache.StartForward(ctx, batch, false)
if err != nil {
return nil, err
}
}
t, err := m.Forward(ctx, batch)
if err != nil {
return nil, err
}
ctx.Forward(t)
return t, nil
}