github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-05-08 08:59:13 -05:00
Code Issues 14.5k Packages Projects Releases 120 Wiki Activity

[PR #13683] [MERGED] x/imagegen: add naive TeaCache and FP8 quantization support #19608

New Issue
Closed
opened 2026-04-16 07:11:39 -05:00 by GiteaMirror · 0 comments
GiteaMirror commented 2026-04-16 07:11:39 -05:00
Owner
Copy Link

📋 Pull Request Information

Original PR: https://github.com/ollama/ollama/pull/13683
Author: @jmorganca
Created: 1/12/2026
Status: ✅ Merged
Merged: 1/12/2026
Merged by: @jmorganca

Base: main ← Head: jmorganca/diffusion-optimizations

📝 Commits (6)

📊 Changes

26 files changed (+1216 additions, -257 deletions)

View changed files

📝 api/client.go (+1 -1)
📝 cmd/cmd.go (+15 -15)
📝 cmd/cmd_test.go (+73 -0)
📝 server/routes.go (+13 -0)
📝 x/imagegen/README.md (+14 -0)
📝 x/imagegen/api/handler.go (+14 -18)
➕ x/imagegen/cache/teacache.go (+197 -0)
📝 x/imagegen/cli.go (+88 -86)
📝 x/imagegen/client/create.go (+70 -10)
➕ x/imagegen/client/quantize.go (+120 -0)
➕ x/imagegen/client/quantize_stub.go (+18 -0)
📝 x/imagegen/cmd/engine/main.go (+14 -7)
📝 x/imagegen/create.go (+35 -5)
📝 x/imagegen/image.go (+6 -3)
📝 x/imagegen/mlx/mlx.go (+83 -4)
📝 x/imagegen/models/qwen_image/qwen_image.go (+23 -3)
📝 x/imagegen/models/qwen_image_edit/qwen_image_edit.go (+22 -4)
📝 x/imagegen/models/zimage/text_encoder.go (+9 -9)
📝 x/imagegen/models/zimage/transformer.go (+111 -27)
📝 x/imagegen/models/zimage/vae.go (+15 -2)

...and 6 more files

📄 Description

This improves performance of z-image on macOS and CUDA. There's still some work to simplify this as well as merge the image generation pipeline runtime between different diffusion models. FP4 was also explored but will be revisited in a follow up.

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 improvements:

  • Fix Show API for image generation models
  • Server properly returns model info (architecture, parameters, quantization)
  • Memory allocation optimizations
  • CLI improvements for image generation

🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.

## 📋 Pull Request Information **Original PR:** https://github.com/ollama/ollama/pull/13683 **Author:** [@jmorganca](https://github.com/jmorganca) **Created:** 1/12/2026 **Status:** ✅ Merged **Merged:** 1/12/2026 **Merged by:** [@jmorganca](https://github.com/jmorganca) **Base:** `main` ← **Head:** `jmorganca/diffusion-optimizations` --- ### 📝 Commits (6) - [`47b8def`](https://github.com/ollama/ollama/commit/47b8def07e977d76288b0c54b2aad2a11de26724) x/imagegen: add TeaCache and FP8 quantization support - [`73ad90b`](https://github.com/ollama/ollama/commit/73ad90bf0852bfb88989925ddcee648f1b7002fd) fix - [`6b4fb18`](https://github.com/ollama/ollama/commit/6b4fb18235ec97405c455fe5b2a8c3695cda4f88) undo readme - [`6ca8a91`](https://github.com/ollama/ollama/commit/6ca8a917716dd383df34610572adc06cfd296bd3) remove additional file - [`a5b166c`](https://github.com/ollama/ollama/commit/a5b166c6b8bacf7c0ebc5120c6eb7a3eaacf5c05) cmd: add tests for cmd changes - [`52857b0`](https://github.com/ollama/ollama/commit/52857b0bbcaafb1c19f18078ad2daaab3ab74fa3) fix linter ### 📊 Changes **26 files changed** (+1216 additions, -257 deletions) <details> <summary>View changed files</summary> 📝 `api/client.go` (+1 -1) 📝 `cmd/cmd.go` (+15 -15) 📝 `cmd/cmd_test.go` (+73 -0) 📝 `server/routes.go` (+13 -0) 📝 `x/imagegen/README.md` (+14 -0) 📝 `x/imagegen/api/handler.go` (+14 -18) ➕ `x/imagegen/cache/teacache.go` (+197 -0) 📝 `x/imagegen/cli.go` (+88 -86) 📝 `x/imagegen/client/create.go` (+70 -10) ➕ `x/imagegen/client/quantize.go` (+120 -0) ➕ `x/imagegen/client/quantize_stub.go` (+18 -0) 📝 `x/imagegen/cmd/engine/main.go` (+14 -7) 📝 `x/imagegen/create.go` (+35 -5) 📝 `x/imagegen/image.go` (+6 -3) 📝 `x/imagegen/mlx/mlx.go` (+83 -4) 📝 `x/imagegen/models/qwen_image/qwen_image.go` (+23 -3) 📝 `x/imagegen/models/qwen_image_edit/qwen_image_edit.go` (+22 -4) 📝 `x/imagegen/models/zimage/text_encoder.go` (+9 -9) 📝 `x/imagegen/models/zimage/transformer.go` (+111 -27) 📝 `x/imagegen/models/zimage/vae.go` (+15 -2) _...and 6 more files_ </details> ### 📄 Description This improves performance of z-image on macOS and CUDA. There's still some work to simplify this as well as merge the image generation pipeline runtime between different diffusion models. FP4 was also explored but will be revisited in a follow up. 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 improvements: - Fix Show API for image generation models - Server properly returns model info (architecture, parameters, quantization) - Memory allocation optimizations - CLI improvements for image generation --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>
GiteaMirror added the pull-request label 2026-04-16 07:11:39 -05:00
Sign in to join this conversation.
No Branch/Tag Specified
Branches Tags
Labels
Clear labels
amd
api
app
bug
build
cli
cloud
compatibility
context-length
create
docker
documentation
embeddings
feature request
feedback wanted
good first issue
gpt-oss
gpu
harmony
help wanted
image
install
intel
js
launch
linux
macos
memory
mlx
model
needs more info
networking
nvidia
ollama.com
performance
pull-request
Mirrored from GitHub Pull Request
 python
question
registry
rendering
thinking
tools
top
vulkan
windows
wsl
No Label pull-request
Milestone
No items
No Milestone
Projects
Clear projects
No project
Assignees
Clear assignees
GiteaMirror ninjasurge
No Assignees
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: github-starred/ollama#19608
Delete Branch "%!s()"

Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?