mlx-teacache 0.1.0__tar.gz

mlx_teacache-0.1.0/.gitignore

@@ -0,0 +1,17 @@
+__pycache__/
+*.py[cod]
+*$py.class
+.DS_Store
+.venv/
+.hypothesis/
+.pytest_cache/
+.ruff_cache/
+.mypy_cache/
+dist/
+build/
+*.egg-info/
+.coverage
+htmlcov/
+tests/converted/*.safetensors
+tests/_artifacts/
+src/mlx_teacache/_version.py

mlx_teacache-0.1.0/CHANGELOG.md

@@ -0,0 +1,34 @@
+# Changelog
+
+All notable changes to this project are documented here.
+
+Format follows [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
+Project follows [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
+
+## [Unreleased]
+
+## [0.1.0] — 2026-05-15 (planned)
+
+### Added
+- Initial public release.
+- `apply_teacache(flux, *, rel_l1_thresh=0.20, ...)` for FLUX.1 dev/schnell and FLUX.2 Klein 4b.
+- Context-manager-compatible `TeaCacheHandle` with live `.stats`, `.provenance`, `.restore()`.
+- Built-in polynomial coefficients vendored from ali-vilab/TeaCache (FLUX.1) and derived in-repo (FLUX.2 Klein 4b).
+- Custom-coefficient override path.
+- Auto-disable on FLUX.2 CFG (`guidance > 1.0`); bit-exact vanilla fallback.
+- img2img rejection with `Img2ImgNotSupportedError`.
+- Threshold-zero fast path: at `rel_l1_thresh <= 0` the wrapper skips building cache tensors entirely (cheap no-op).
+- Five-tier test pyramid: shape/dtype unit tests, paired same-process latent parity (bit-exact for FLUX.1, cosine ≥ 0.97 for FLUX.2), image-level SSIM gates on VAE-decoded outputs.
+- Trusted-Publishing release pipeline.
+
+### Calibration notes (2026-05-15)
+- Default `rel_l1_thresh` chosen by visual comparison + SSIM measurement: at 0.25 some text/synthetic prompts changed rendering style; at 0.20 outputs are indistinguishable from vanilla while still skipping ~25% of steps.
+- FLUX.1-dev / 25 steps / M1 Max: measured 1.48× speedup at default threshold, SSIM ≥ 0.80 on a 5-prompt suite.
+- FLUX.1 polynomial coefficients corrected from a transcription error in an earlier revision (`c0..c3` were ~10× too large); now match ali-vilab upstream exactly.
+
+### Known limitations
+- v0.1 supports txt2img only.
+- FLUX.2 Klein variants other than `flux2-klein-4b` are not in v0.1.
+- Distilled-step schedules (FLUX.1 schnell 4-step, Klein 4-step) see no measurable speedup.
+- M3+ users lose mflux's `mx.compile` of `_predict`; net behavior unmeasured in v0.1 (M1 Max benchmarks only).
+- FLUX.2 parity is numerical, not bit-exact: vanilla-compiled vs wrapper-eager Metal kernel dispatch differs by ~1 ULP per element. Cosine similarity ≥ 0.99 measured; image-level SSIM is the user-facing guarantee. CFG-fallback path remains bit-exact.

mlx_teacache-0.1.0/LICENSE

@@ -0,0 +1,201 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file type. We recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright 2026 Denis Ineshin
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

mlx_teacache-0.1.0/NOTICE

@@ -0,0 +1,53 @@
+mlx-teacache
+Copyright 2026 Denis Ineshin
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+------------------------------------------------------------------------------
+Third-party attributions
+------------------------------------------------------------------------------
+
+This product includes software derived from ali-vilab/TeaCache
+(https://github.com/ali-vilab/TeaCache), licensed under Apache License 2.0:
+  Copyright 2024-2025 the TeaCache authors.
+
+  The polynomial coefficients in src/mlx_teacache/coefficients.py for the
+  FLUX architecture (flux1-dev, flux1-schnell) are vendored verbatim from
+  ali-vilab/TeaCache/TeaCache4FLUX/teacache_flux.py.
+
+------------------------------------------------------------------------------
+
+This product includes software derived from filipstrand/mflux
+(https://github.com/filipstrand/mflux), licensed under MIT License:
+
+  MIT License
+
+  Copyright (c) 2024-2026 Filip Strand
+
+  Permission is hereby granted, free of charge, to any person obtaining a copy
+  of this software and associated documentation files (the "Software"), to deal
+  in the Software without restriction, including without limitation the rights
+  to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+  copies of the Software, and to permit persons to whom the Software is
+  furnished to do so, subject to the following conditions:
+
+  The above copyright notice and this permission notice shall be included in all
+  copies or substantial portions of the Software.
+
+  THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+  IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+  FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+  AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+  LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+  OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+  SOFTWARE.
+
+  mlx-teacache/src/mlx_teacache/integrations/mflux/forward.py reimplements the
+  transformer forward passes from mflux's Flux1.Transformer and
+  Flux2Klein.Flux2Transformer for the purpose of inserting TeaCache gating
+  between body and tail. The reimplementation closely follows mflux's
+  internal structure as of 0.17.5.

mlx_teacache-0.1.0/PKG-INFO

@@ -0,0 +1,172 @@
+Metadata-Version: 2.4
+Name: mlx-teacache
+Version: 0.1.0
+Summary: TeaCache step-skipping for FLUX diffusion on Apple Silicon, in pure MLX
+Project-URL: Homepage, https://github.com/IonDen/mlx-teacache
+Project-URL: Issues, https://github.com/IonDen/mlx-teacache/issues
+Author-email: Denis Ineshin <denis.ineshin@gmail.com>
+License-Expression: Apache-2.0
+License-File: LICENSE
+License-File: NOTICE
+Keywords: apple-silicon,diffusion,flux,mlx,teacache
+Classifier: Development Status :: 3 - Alpha
+Classifier: Intended Audience :: Developers
+Classifier: License :: OSI Approved :: Apache Software License
+Classifier: Operating System :: MacOS :: MacOS X
+Classifier: Programming Language :: Python :: 3 :: Only
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Programming Language :: Python :: 3.13
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Requires-Python: >=3.11
+Requires-Dist: mlx>=0.31
+Requires-Dist: numpy
+Provides-Extra: mflux
+Requires-Dist: mflux<0.18,>=0.17; extra == 'mflux'
+Description-Content-Type: text/markdown
+
+# mlx-teacache
+
+[![PyPI version](https://img.shields.io/pypi/v/mlx-teacache.svg)](https://pypi.org/project/mlx-teacache/)
+[![Python versions](https://img.shields.io/pypi/pyversions/mlx-teacache.svg)](https://pypi.org/project/mlx-teacache/)
+[![License: Apache 2.0](https://img.shields.io/pypi/l/mlx-teacache.svg)](https://github.com/IonDen/mlx-teacache/blob/main/LICENSE)
+[![CI](https://github.com/IonDen/mlx-teacache/actions/workflows/ci.yml/badge.svg)](https://github.com/IonDen/mlx-teacache/actions/workflows/ci.yml)
+
+**TeaCache step-skipping for FLUX diffusion on Apple Silicon, in pure MLX.**
+
+`mlx-teacache` is the first MLX port of TeaCache — a training-free inference optimization that skips ~20-50% of denoising steps in FLUX-family diffusion models by predicting which steps contribute little to the final image. Measured ~1.48× wall-clock speedup at the default threshold on FLUX.1-dev / 25 steps with visually-equivalent output (SSIM ≥ 0.80 on a 5-prompt suite).
+
+## What it does
+
+Diffusion models run the same big transformer 20-50 times in a loop. Between consecutive steps the output changes very little; TeaCache uses a tiny polynomial fit to predict which steps can reuse the previous step's output. On M1 Max FLUX.1-dev @ 25 steps the default threshold (`rel_l1_thresh=0.20`) skips 6 of 25 steps for a ~1.48× speedup.
+
+```python
+from mflux.models.flux2.variants.txt2img.flux2_klein import Flux2Klein
+from mflux.models.common.config.model_config import ModelConfig
+from mlx_teacache import apply_teacache
+
+flux = Flux2Klein(quantize=4, model_config=ModelConfig.flux2_klein_4b())
+with apply_teacache(flux):  # default rel_l1_thresh=0.20
+    flux.generate_image(prompt="a red apple", seed=42, num_inference_steps=25)
+```
+
+## Install
+
+```bash
+pip install "mlx-teacache[mflux]"
+# or with uv:
+uv add "mlx-teacache[mflux]"
+```
+
+Requires Python ≥ 3.11 and Apple Silicon. The `[mflux]` extra pulls in `mflux>=0.17,<0.18`.
+
+```bash
+pip install "mlx-teacache==0.1.0[mflux]"  # pin for reproducibility
+```
+
+## Quick start
+
+FLUX.1 dev:
+
+```python
+from mflux.models.flux.variants.txt2img.flux import Flux1
+from mlx_teacache import apply_teacache
+
+flux = Flux1.from_name("dev", quantize=4)
+with apply_teacache(flux) as handle:  # default rel_l1_thresh=0.20
+    flux.generate_image(prompt="...", seed=42, num_inference_steps=25, guidance=3.5)
+    print(f"Speedup: {handle.stats.speedup_estimate:.2f}×")
+```
+
+## Threshold guide
+
+Measured on M1 Max 32GB, FLUX.1-dev @ 25 steps, bf16, `seed=42`, `guidance=3.5`, red-apple prompt:
+
+| `rel_l1_thresh` | Skipped steps | Speedup | SSIM vs vanilla | Recommended use |
+|---|---|---|---|---|
+| 0.10 | 0 / 25 | 1.07× | 1.0000 | Cache never engages |
+| 0.15 | 0 / 25 | 1.13× | 1.0000 | Cache never engages |
+| **0.20 (default)** | **6 / 25** | **1.48×** | **≥ 0.80 (5-prompt suite)** | **Visually-lossless sweet spot** |
+| 0.25 | 11 / 25 | 1.96× | 0.57-0.93 | Visible style changes on text/synthetic prompts |
+
+The 0.20 default was chosen after side-by-side visual comparison: at 0.25 a text prompt that vanilla renders as neon tubes can come out as dot-matrix; at 0.20 the output is indistinguishable from vanilla while still skipping ~25% of steps. SSIM is a conservative metric on high-frequency-detail prompts (text, synthetic patterns).
+
+## Supported models
+
+| Variant id | mflux class + config | Coefficient source |
+|---|---|---|
+| `flux1-dev` | `Flux1(model_config=ModelConfig.dev())` | upstream ali-vilab/TeaCache |
+| `flux1-schnell` | `Flux1(model_config=ModelConfig.schnell())` | upstream (shared with dev) |
+| `flux2-klein-4b` | `Flux2Klein(model_config=ModelConfig.flux2_klein_4b())` | in-repo (see `docs/calibration.md`) |
+
+## Combining with mlx-taef
+
+```python
+from mlx_taef.integrations.mflux import LivePreviewCallback
+from mlx_teacache import apply_teacache
+
+preview = LivePreviewCallback(variant="taef2", every=5, save_to="preview.png",
+                              latent_height=32, latent_width=32)
+flux.callbacks.register(preview)
+
+with apply_teacache(flux):  # default rel_l1_thresh=0.20
+    flux.generate_image(prompt="...", seed=42, num_inference_steps=25)
+```
+
+## Inspecting stats
+
+```python
+handle = apply_teacache(flux)
+flux.generate_image(prompt="apple", seed=42, num_inference_steps=25)
+print(handle.stats.computed_count, handle.stats.skipped_count)
+print(handle.stats.last_generation.decisions[5])  # per-step record
+```
+
+## Custom coefficients
+
+```python
+custom = [...]  # length 5, all finite
+apply_teacache(flux, coefficients=custom)
+```
+
+## How it works
+
+TeaCache observes that in diffusion denoising, consecutive transformer outputs change very little between most pairs of adjacent steps. The expensive transformer body (all the joint + single attention blocks) produces a residual that's added to the input — and that residual stays roughly stable for stretches of the denoising trajectory.
+
+TeaCache trains a tiny polynomial that predicts how much the output will change given how much the input has changed (measured as relative L1 distance of the modulated block-0 input). When the predicted accumulated change since the last actual compute step is small, TeaCache reuses the cached residual instead of running the full transformer body — only the cheap prelude (embeddings) and tail (norm + projection) still run.
+
+mlx-teacache implements this for mflux on Apple Silicon. We replace `flux.transformer` with a per-instance proxy for FLUX.1, and `flux._predict` with an instance-level closure for FLUX.2 (to keep gating live even when `mx.compile` is normally used). The polynomial coefficients for FLUX.1 are vendored from upstream; the FLUX.2 Klein 4b coefficients are derived in-repo via `scripts/calibrate.py`. See the TeaCache paper at https://liewfeng.github.io/TeaCache/ and `docs/superpowers/spikes/2026-05-14-mlx-teacache-phase-0-spike.md` for the Apple-Silicon-specific design notes.
+
+## Benchmarks
+
+M1 Max 32GB, default threshold (`rel_l1_thresh=0.20`):
+
+| Model | Steps | Vanilla | With TeaCache | Speedup | SSIM (PR-gate prompt) |
+|---|---|---|---|---|---|
+| FLUX.1-dev @ 512² | 25 | ~5:18 | ~3:35 | 1.48× | ≥ 0.90 |
+| FLUX.2 Klein 4b @ 512² | 8 | ~37s | ~30s | ~1.2× | ≥ 0.85 |
+
+FLUX.2's smaller speedup at 8 steps reflects the short non-distilled denoising trajectory at Klein's recommended step count. Longer schedules (`num_inference_steps>=15`) give larger gains.
+
+## Limitations
+
+- **v0.1 is txt2img only.** img2img inputs raise `Img2ImgNotSupportedError`. v0.2 planned.
+- **FLUX.2 CFG (`guidance > 1.0`) auto-falls-back to vanilla mflux.** Output is bit-exact. v0.2 will add per-branch caching.
+- **Distilled schedules see no speedup.** FLUX.1 schnell 4-step and Klein 4-step defaults have too few non-forced steps. Use `num_inference_steps >= 10` for benefit.
+- **Klein variants other than `flux2-klein-4b` are not supported in v0.1.** 9b and base configs planned for v0.2.
+- **M3+ users lose mflux's `mx.compile` of `_predict`.** v0.1 provides a manual benchmark recipe (`docs/m3-plus-tradeoff.md`) and does not claim a measured M3+ speedup.
+- **FLUX.2 parity is numerical, not bit-exact.** Because the wrapper replaces a function mflux wraps in `mx.compile`, vanilla-compiled vs wrapper-eager differ by ~1 ULP per element from Metal kernel-dispatch noise (compounds across steps; cosine similarity stays ≥ 0.99). The CFG-fallback path remains bit-exact. End-to-end image quality (SSIM ≥ 0.85 on Klein 4b) is the user-facing guarantee.
+- **mflux pin is strict (`>=0.17,<0.18`).** Bumping requires a deliberate release.
+- **Parent-level `flux.parameters()` may miss transformer parameters while patched.** Use `flux.transformer.parameters()` directly, or `handle.restore()` first.
+
+## Contributing
+
+Open an issue at https://github.com/IonDen/mlx-teacache/issues.
+
+## License + acknowledgements
+
+Apache-2.0. See `LICENSE` and `NOTICE`.
+
+- [ali-vilab/TeaCache](https://github.com/ali-vilab/TeaCache) — upstream method and FLUX.1 coefficients.
+- [filipstrand/mflux](https://github.com/filipstrand/mflux) — MLX FLUX runner this library integrates with.
+- [Apple ML Explore](https://github.com/ml-explore/mlx) — MLX.

mlx_teacache-0.1.0/README.md

@@ -0,0 +1,145 @@
+# mlx-teacache
+
+[![PyPI version](https://img.shields.io/pypi/v/mlx-teacache.svg)](https://pypi.org/project/mlx-teacache/)
+[![Python versions](https://img.shields.io/pypi/pyversions/mlx-teacache.svg)](https://pypi.org/project/mlx-teacache/)
+[![License: Apache 2.0](https://img.shields.io/pypi/l/mlx-teacache.svg)](https://github.com/IonDen/mlx-teacache/blob/main/LICENSE)
+[![CI](https://github.com/IonDen/mlx-teacache/actions/workflows/ci.yml/badge.svg)](https://github.com/IonDen/mlx-teacache/actions/workflows/ci.yml)
+
+**TeaCache step-skipping for FLUX diffusion on Apple Silicon, in pure MLX.**
+
+`mlx-teacache` is the first MLX port of TeaCache — a training-free inference optimization that skips ~20-50% of denoising steps in FLUX-family diffusion models by predicting which steps contribute little to the final image. Measured ~1.48× wall-clock speedup at the default threshold on FLUX.1-dev / 25 steps with visually-equivalent output (SSIM ≥ 0.80 on a 5-prompt suite).
+
+## What it does
+
+Diffusion models run the same big transformer 20-50 times in a loop. Between consecutive steps the output changes very little; TeaCache uses a tiny polynomial fit to predict which steps can reuse the previous step's output. On M1 Max FLUX.1-dev @ 25 steps the default threshold (`rel_l1_thresh=0.20`) skips 6 of 25 steps for a ~1.48× speedup.
+
+```python
+from mflux.models.flux2.variants.txt2img.flux2_klein import Flux2Klein
+from mflux.models.common.config.model_config import ModelConfig
+from mlx_teacache import apply_teacache
+
+flux = Flux2Klein(quantize=4, model_config=ModelConfig.flux2_klein_4b())
+with apply_teacache(flux):  # default rel_l1_thresh=0.20
+    flux.generate_image(prompt="a red apple", seed=42, num_inference_steps=25)
+```
+
+## Install
+
+```bash
+pip install "mlx-teacache[mflux]"
+# or with uv:
+uv add "mlx-teacache[mflux]"
+```
+
+Requires Python ≥ 3.11 and Apple Silicon. The `[mflux]` extra pulls in `mflux>=0.17,<0.18`.
+
+```bash
+pip install "mlx-teacache==0.1.0[mflux]"  # pin for reproducibility
+```
+
+## Quick start
+
+FLUX.1 dev:
+
+```python
+from mflux.models.flux.variants.txt2img.flux import Flux1
+from mlx_teacache import apply_teacache
+
+flux = Flux1.from_name("dev", quantize=4)
+with apply_teacache(flux) as handle:  # default rel_l1_thresh=0.20
+    flux.generate_image(prompt="...", seed=42, num_inference_steps=25, guidance=3.5)
+    print(f"Speedup: {handle.stats.speedup_estimate:.2f}×")
+```
+
+## Threshold guide
+
+Measured on M1 Max 32GB, FLUX.1-dev @ 25 steps, bf16, `seed=42`, `guidance=3.5`, red-apple prompt:
+
+| `rel_l1_thresh` | Skipped steps | Speedup | SSIM vs vanilla | Recommended use |
+|---|---|---|---|---|
+| 0.10 | 0 / 25 | 1.07× | 1.0000 | Cache never engages |
+| 0.15 | 0 / 25 | 1.13× | 1.0000 | Cache never engages |
+| **0.20 (default)** | **6 / 25** | **1.48×** | **≥ 0.80 (5-prompt suite)** | **Visually-lossless sweet spot** |
+| 0.25 | 11 / 25 | 1.96× | 0.57-0.93 | Visible style changes on text/synthetic prompts |
+
+The 0.20 default was chosen after side-by-side visual comparison: at 0.25 a text prompt that vanilla renders as neon tubes can come out as dot-matrix; at 0.20 the output is indistinguishable from vanilla while still skipping ~25% of steps. SSIM is a conservative metric on high-frequency-detail prompts (text, synthetic patterns).
+
+## Supported models
+
+| Variant id | mflux class + config | Coefficient source |
+|---|---|---|
+| `flux1-dev` | `Flux1(model_config=ModelConfig.dev())` | upstream ali-vilab/TeaCache |
+| `flux1-schnell` | `Flux1(model_config=ModelConfig.schnell())` | upstream (shared with dev) |
+| `flux2-klein-4b` | `Flux2Klein(model_config=ModelConfig.flux2_klein_4b())` | in-repo (see `docs/calibration.md`) |
+
+## Combining with mlx-taef
+
+```python
+from mlx_taef.integrations.mflux import LivePreviewCallback
+from mlx_teacache import apply_teacache
+
+preview = LivePreviewCallback(variant="taef2", every=5, save_to="preview.png",
+                              latent_height=32, latent_width=32)
+flux.callbacks.register(preview)
+
+with apply_teacache(flux):  # default rel_l1_thresh=0.20
+    flux.generate_image(prompt="...", seed=42, num_inference_steps=25)
+```
+
+## Inspecting stats
+
+```python
+handle = apply_teacache(flux)
+flux.generate_image(prompt="apple", seed=42, num_inference_steps=25)
+print(handle.stats.computed_count, handle.stats.skipped_count)
+print(handle.stats.last_generation.decisions[5])  # per-step record
+```
+
+## Custom coefficients
+
+```python
+custom = [...]  # length 5, all finite
+apply_teacache(flux, coefficients=custom)
+```
+
+## How it works
+
+TeaCache observes that in diffusion denoising, consecutive transformer outputs change very little between most pairs of adjacent steps. The expensive transformer body (all the joint + single attention blocks) produces a residual that's added to the input — and that residual stays roughly stable for stretches of the denoising trajectory.
+
+TeaCache trains a tiny polynomial that predicts how much the output will change given how much the input has changed (measured as relative L1 distance of the modulated block-0 input). When the predicted accumulated change since the last actual compute step is small, TeaCache reuses the cached residual instead of running the full transformer body — only the cheap prelude (embeddings) and tail (norm + projection) still run.
+
+mlx-teacache implements this for mflux on Apple Silicon. We replace `flux.transformer` with a per-instance proxy for FLUX.1, and `flux._predict` with an instance-level closure for FLUX.2 (to keep gating live even when `mx.compile` is normally used). The polynomial coefficients for FLUX.1 are vendored from upstream; the FLUX.2 Klein 4b coefficients are derived in-repo via `scripts/calibrate.py`. See the TeaCache paper at https://liewfeng.github.io/TeaCache/ and `docs/superpowers/spikes/2026-05-14-mlx-teacache-phase-0-spike.md` for the Apple-Silicon-specific design notes.
+
+## Benchmarks
+
+M1 Max 32GB, default threshold (`rel_l1_thresh=0.20`):
+
+| Model | Steps | Vanilla | With TeaCache | Speedup | SSIM (PR-gate prompt) |
+|---|---|---|---|---|---|
+| FLUX.1-dev @ 512² | 25 | ~5:18 | ~3:35 | 1.48× | ≥ 0.90 |
+| FLUX.2 Klein 4b @ 512² | 8 | ~37s | ~30s | ~1.2× | ≥ 0.85 |
+
+FLUX.2's smaller speedup at 8 steps reflects the short non-distilled denoising trajectory at Klein's recommended step count. Longer schedules (`num_inference_steps>=15`) give larger gains.
+
+## Limitations
+
+- **v0.1 is txt2img only.** img2img inputs raise `Img2ImgNotSupportedError`. v0.2 planned.
+- **FLUX.2 CFG (`guidance > 1.0`) auto-falls-back to vanilla mflux.** Output is bit-exact. v0.2 will add per-branch caching.
+- **Distilled schedules see no speedup.** FLUX.1 schnell 4-step and Klein 4-step defaults have too few non-forced steps. Use `num_inference_steps >= 10` for benefit.
+- **Klein variants other than `flux2-klein-4b` are not supported in v0.1.** 9b and base configs planned for v0.2.
+- **M3+ users lose mflux's `mx.compile` of `_predict`.** v0.1 provides a manual benchmark recipe (`docs/m3-plus-tradeoff.md`) and does not claim a measured M3+ speedup.
+- **FLUX.2 parity is numerical, not bit-exact.** Because the wrapper replaces a function mflux wraps in `mx.compile`, vanilla-compiled vs wrapper-eager differ by ~1 ULP per element from Metal kernel-dispatch noise (compounds across steps; cosine similarity stays ≥ 0.99). The CFG-fallback path remains bit-exact. End-to-end image quality (SSIM ≥ 0.85 on Klein 4b) is the user-facing guarantee.
+- **mflux pin is strict (`>=0.17,<0.18`).** Bumping requires a deliberate release.
+- **Parent-level `flux.parameters()` may miss transformer parameters while patched.** Use `flux.transformer.parameters()` directly, or `handle.restore()` first.
+
+## Contributing
+
+Open an issue at https://github.com/IonDen/mlx-teacache/issues.
+
+## License + acknowledgements
+
+Apache-2.0. See `LICENSE` and `NOTICE`.
+
+- [ali-vilab/TeaCache](https://github.com/ali-vilab/TeaCache) — upstream method and FLUX.1 coefficients.
+- [filipstrand/mflux](https://github.com/filipstrand/mflux) — MLX FLUX runner this library integrates with.
+- [Apple ML Explore](https://github.com/ml-explore/mlx) — MLX.