Implementation:Ollama Ollama Imagegen Manifest Weights
| Knowledge Sources | |
|---|---|
| Domains | Image Generation, Model Loading |
| Last Updated | 2025-02-15 00:00 GMT |
Overview
Provides zero-copy weight loading from Ollama's manifest blob storage using native MLX safetensors mmap.
Description
The weights.go file implements ManifestWeights for loading model tensors from Ollama's content-addressed blob storage. Each tensor is stored as an individual safetensors blob, loaded via native MLX mmap (mlx_load_safetensors) for zero-copy GPU access. The Load method groups tensors by digest to avoid loading the same blob multiple times, handles component prefix stripping (e.g., "text_encoder/model.embed_tokens.weight" becomes "model.embed_tokens.weight"), and detects quantization metadata (quant_type, group_size) from blob headers. It supports three tensor formats: single-tensor blobs (keyed by name or legacy "data"), combined quantized blobs (weight + scale + bias), and packed blobs for expert groups containing multiple tensors.
Usage
Used by all image generation model loaders (Flux2, ZImage, text encoders) to load weights from Ollama's blob storage with zero-copy GPU mapping.
Code Reference
Source Location
- Repository: Ollama
- File: x/imagegen/manifest/weights.go
- Lines: 1-298
Signature
type ManifestWeights struct {
manifest *ModelManifest
component string
tensors map[string]ManifestLayer
cache map[string]*mlx.Array
nativeCache []*mlx.SafetensorsFile
quantType string
groupSize int
}
func LoadWeightsFromManifest(manifest *ModelManifest, component string) (*ManifestWeights, error)
func (mw *ManifestWeights) Load(dtype mlx.Dtype) error
func (mw *ManifestWeights) Get(name string) *mlx.Array
func (mw *ManifestWeights) QuantType() string
func (mw *ManifestWeights) GroupSize() int
Import
import "github.com/ollama/ollama/x/imagegen/manifest"
I/O Contract
Inputs
| Name | Type | Required | Description |
|---|---|---|---|
| manifest | *ModelManifest | Yes | Loaded model manifest |
| component | string | No | Component filter (e.g., "transformer", "" for all) |
| dtype | mlx.Dtype | No | Target dtype for non-quantized tensors (0 = original) |
Outputs
| Name | Type | Description |
|---|---|---|
| *ManifestWeights | *ManifestWeights | Weight loader with zero-copy mmap tensors |
| error | error | Error if no tensors found or blob loading fails |
Usage Examples
manifest, _ := manifest.LoadManifest("flux2-klein:latest")
weights, err := manifest.LoadWeightsFromManifest(manifest, "transformer")
if err != nil {
return err
}
if err := weights.Load(mlx.DtypeBFloat16); err != nil {
return err
}
attnWeight := weights.Get("layers.0.attn.q_proj.weight")