From c0efc28968bf52ea13e473109f75a9b966dcf757 Mon Sep 17 00:00:00 2001
From: Ettore Di Giacinto <mudler@localai.io>
Date: Wed, 24 Jun 2026 17:11:34 +0000
Subject: [PATCH] feat(backends): add darwin/metal (MPS) build for trl

Authors backend/python/trl/requirements-mps.txt and wires trl into the
darwin CI matrix and gallery so the MPS training path can be built and
validated on Apple Silicon. The MPS variant installs plain PyPI torch
wheels (MPS-capable on macOS arm64) and the trl training stack; bitsandbytes
is omitted as it is a CUDA-only dependency with poor Apple Silicon support.

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude:opus-4.8 [Claude Code]
---
 .github/backend-matrix.yml              |  3 +++
 backend/index.yaml                      | 11 +++++++++++
 backend/python/trl/requirements-mps.txt | 12 ++++++++++++
 3 files changed, 26 insertions(+)
 create mode 100644 backend/python/trl/requirements-mps.txt

diff --git a/.github/backend-matrix.yml b/.github/backend-matrix.yml
index 593e44cde..729a4d8b3 100644
--- a/.github/backend-matrix.yml
+++ b/.github/backend-matrix.yml
@@ -4974,6 +4974,9 @@ includeDarwin:
   - backend: "kitten-tts"
     tag-suffix: "-metal-darwin-arm64-kitten-tts"
     build-type: "mps"
+  - backend: "trl"
+    tag-suffix: "-metal-darwin-arm64-trl"
+    build-type: "mps"
   - backend: "piper"
     tag-suffix: "-metal-darwin-arm64-piper"
     build-type: "metal"
diff --git a/backend/index.yaml b/backend/index.yaml
index 3f61f7b4e..2e22ed4f3 100644
--- a/backend/index.yaml
+++ b/backend/index.yaml
@@ -5282,6 +5282,7 @@
     nvidia: "cuda12-trl"
     nvidia-cuda-12: "cuda12-trl"
     nvidia-cuda-13: "cuda13-trl"
+    metal: "metal-trl"
 ## TRL backend images
 - !!merge <<: *trl
   name: "cpu-trl"
@@ -5313,6 +5314,16 @@
   uri: "quay.io/go-skynet/local-ai-backends:master-gpu-nvidia-cuda-13-trl"
   mirrors:
     - localai/localai-backends:master-gpu-nvidia-cuda-13-trl
+- !!merge <<: *trl
+  name: "metal-trl"
+  uri: "quay.io/go-skynet/local-ai-backends:latest-metal-darwin-arm64-trl"
+  mirrors:
+    - localai/localai-backends:latest-metal-darwin-arm64-trl
+- !!merge <<: *trl
+  name: "metal-trl-development"
+  uri: "quay.io/go-skynet/local-ai-backends:master-metal-darwin-arm64-trl"
+  mirrors:
+    - localai/localai-backends:master-metal-darwin-arm64-trl
 ## llama.cpp quantization backend
 - &llama-cpp-quantization
   name: "llama-cpp-quantization"
diff --git a/backend/python/trl/requirements-mps.txt b/backend/python/trl/requirements-mps.txt
new file mode 100644
index 000000000..fbdfb6536
--- /dev/null
+++ b/backend/python/trl/requirements-mps.txt
@@ -0,0 +1,12 @@
+torch==2.10.0
+trl
+peft
+datasets>=3.0.0
+transformers>=4.56.2
+accelerate>=1.4.0
+huggingface-hub>=1.3.0
+sentencepiece
+# Note: bitsandbytes is intentionally omitted on MPS. It is only used by the
+# CUDA (cublas) variants for 8-bit/4-bit quantization and has poor support on
+# Apple Silicon. torch here uses the plain PyPI wheels, which ship MPS support
+# on macOS arm64.