Local ckpt support for python mlx

.github/workflows/lint.yml

@@ -14,8 +14,16 @@ jobs:
       - name: Checkout code
         uses: actions/checkout@v3
 
+      - name: Cache Homebrew
+        uses: actions/cache@v3
+        with:
+          path: /home/linuxbrew/.linuxbrew
+          key: ${{ runner.os }}-homebrew-${{ hashFiles('**/Brewfile.lock.json') }}
+          restore-keys: |
+            ${{ runner.os }}-homebrew-
+
       - name: Set up Homebrew
-        id: set-up-homebrew
+        if: steps.cache-homebrew.outputs.cache-hit != 'true'
         uses: Homebrew/actions/setup-homebrew@master
 
       - name: Setup environment

README.md

@@ -74,6 +74,7 @@ Some notable optional arguments:
 - For image-to-image, use `--image-path` (path to input image) and `--denoise` (value between 0. and 1.)
 - T5 text embeddings, use `--t5`
 - For different resolutions, use `--height` and `--width`
+- For using a local checkpoint, use `--local-ckpt </path/to/ckpt.safetensors>` (e.g. `~/models/stable-diffusion-3-medium/sd3_medium.safetensors`).
 
 Please refer to the help menu for all available arguments: `diffusionkit-cli -h`.
 

python/src/mlx/__init__.py

@@ -48,7 +48,9 @@ def __init__(
         model_size: str = "2b",
         low_memory_mode: bool = True,
         a16: bool = False,
+        local_ckpt=None,
     ):
+        model_io.LOCAl_SD3_CKPT = local_ckpt
         self.dtype = mx.float16 if w16 else mx.float32
         self.activation_dtype = mx.float16 if a16 else mx.float32
         self.use_t5 = use_t5

python/src/mlx/model_io.py

@@ -62,6 +62,8 @@
     "8b": 192,
 }
 
+LOCAl_SD3_CKPT = None
+
 
 def mmdit_state_dict_adjustments(state_dict, prefix=""):
     # Remove prefix
@@ -453,7 +455,8 @@ def load_mmdit(
     model = MMDiT(config)
 
     mmdit_weights = _MMDIT[key][model_key]
-    weights = mx.load(hf_hub_download(key, mmdit_weights))
+    mmdit_weights_ckpt = LOCAl_SD3_CKPT or hf_hub_download(key, mmdit_weights)
+    weights = mx.load(mmdit_weights_ckpt)
     weights = mmdit_state_dict_adjustments(weights, prefix="model.diffusion_model.")
     weights = {k: v.astype(dtype) for k, v in weights.items()}
     model.update(tree_unflatten(tree_flatten(weights)))
@@ -548,7 +551,8 @@ def load_vae_decoder(
 
     dtype = mx.float16 if float16 else mx.float32
     vae_weights = _MMDIT[key][model_key]
-    weights = mx.load(hf_hub_download(key, vae_weights))
+    vae_weights_ckpt = LOCAl_SD3_CKPT or hf_hub_download(key, vae_weights)
+    weights = mx.load(vae_weights_ckpt)
     weights = vae_decoder_state_dict_adjustments(
         weights, prefix="first_stage_model.decoder."
     )
@@ -575,7 +579,8 @@ def load_vae_encoder(
 
     dtype = mx.float16 if float16 else mx.float32
     vae_weights = _MMDIT[key][model_key]
-    weights = mx.load(hf_hub_download(key, vae_weights))
+    vae_weights_ckpt = LOCAl_SD3_CKPT or hf_hub_download(key, vae_weights)
+    weights = mx.load(vae_weights_ckpt)
     weights = vae_encoder_state_dict_adjustments(
         weights, prefix="first_stage_model.encoder."
     )

python/src/mlx/scripts/generate_images.py

@@ -99,6 +99,12 @@ def cli():
         default=0.0,
         help="Denoising factor when an input image is provided. (between 0.0 and 1.0)",
     )
+    parser.add_argument(
+        "--local-ckpt",
+        default=None,
+        type=str,
+        help="Path to the local mmdit checkpoint.",
+    )
     args = parser.parse_args()
 
     if args.benchmark_mode:
@@ -119,6 +125,7 @@ def cli():
         model_size=args.model_size,
         low_memory_mode=args.low_memory_mode,
         a16=args.a16,
+        local_ckpt=args.local_ckpt,
     )
 
     # Ensure that models are read in memory if needed