centersl/ollama
1
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-03-27 02:58:43 +07:00
Code Issues Packages Projects Releases Wiki Activity

x/imagegen: add naive TeaCache and FP8 quantization support (#13683)

Browse Source 
TeaCache:
- Timestep embedding similarity caching for diffusion models
- Polynomial rescaling with configurable thresholds
- Reduces transformer forward passes by ~30-50%

FP8 quantization:
- Support for FP8 quantized models (8-bit weights with scales)
- QuantizedMatmul on Metal, Dequantize on CUDA
- Client-side quantization via ollama create --quantize fp8

Other bug fixes:
- Fix `/api/show` API for image generation models
- Server properly returns model info (architecture, parameters, quantization)
- Memory allocation optimizations
- CLI improvements for image generation
This commit is contained in:
Jeffrey Morgan
2026-01-12 13:45:22 -08:00
committed by GitHub
parent a937a68317
commit 9667c2282f
26 changed files with 1228 additions and 269 deletions
Download Patch File Download Diff File Expand all files Collapse all files

30
cmd/cmd.go
View File

@@ -100,7 +100,8 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
if filename == "" {
// No Modelfile found - check if current directory is an image gen model
if imagegen.IsTensorModelDir(".") {
return imagegenclient.CreateModel(args[0], ".", p)
quantize, _ := cmd.Flags().GetString("quantize")
return imagegenclient.CreateModel(args[0], ".", quantize, p)
}
reader = strings.NewReader("FROM .\n")
} else {
@@ -464,14 +465,6 @@ func RunHandler(cmd *cobra.Command, args []string) error {
name := args[0]
// Check if this is a known image generation model (skip Show/Pull)
if imagegen.HasTensorLayers(name) {
if opts.Prompt == "" && !interactive {
return errors.New("image generation models require a prompt. Usage: ollama run " + name + " \"your prompt here\"")
}
return imagegen.RunCLI(cmd, name, opts.Prompt, interactive, opts.KeepAlive)
}
info, err := func() (*api.ShowResponse, error) {
showReq := &api.ShowRequest{Name: name}
info, err := client.Show(cmd.Context(), showReq)
@@ -533,6 +526,14 @@ func RunHandler(cmd *cobra.Command, args []string) error {
return generateEmbedding(cmd, name, opts.Prompt, opts.KeepAlive, truncate, dimensions)
}
// Check if this is an image generation model
if slices.Contains(info.Capabilities, model.CapabilityImageGeneration) {
if opts.Prompt == "" && !interactive {
return errors.New("image generation models require a prompt. Usage: ollama run " + name + " \"your prompt here\"")
}
return imagegen.RunCLI(cmd, name, opts.Prompt, interactive, opts.KeepAlive)
}
// Check for experimental flag
isExperimental, _ := cmd.Flags().GetBool("experimental")
yoloMode, _ := cmd.Flags().GetBool("experimental-yolo")
@@ -671,7 +672,11 @@ func PushHandler(cmd *cobra.Command, args []string) error {
bar, ok := bars[resp.Digest]
if !ok {
bar = progress.NewBar(fmt.Sprintf("pushing %s...", resp.Digest[7:19]), resp.Total, resp.Completed)
msg := resp.Status
if msg == "" {
msg = fmt.Sprintf("pushing %s...", resp.Digest[7:19])
}
bar = progress.NewBar(msg, resp.Total, resp.Completed)
bars[resp.Digest] = bar
p.Add(resp.Digest, bar)
}
@@ -837,11 +842,6 @@ func DeleteHandler(cmd *cobra.Command, args []string) error {
}
func ShowHandler(cmd *cobra.Command, args []string) error {
// Check if this is an image generation model
if imagegen.HasTensorLayers(args[0]) {
return imagegen.Show(args[0], os.Stdout)
}
client, err := api.ClientFromEnvironment()
if err != nil {
return err