fix(diffusers): sd_embed is not always available (#8602)

Seems sd_embed doesn't play well with MPS and L4T. Making it optional

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

backend/python/diffusers/backend.py

@@ -40,7 +40,21 @@ from compel import Compel, ReturnedEmbeddingsType
 from optimum.quanto import freeze, qfloat8, quantize
 from transformers import T5EncoderModel
 from safetensors.torch import load_file
-from sd_embed.embedding_funcs import get_weighted_text_embeddings_sd15, get_weighted_text_embeddings_sdxl, get_weighted_text_embeddings_sd3, get_weighted_text_embeddings_flux1
+# Try to import sd_embed - it might not always be available
+try:
+    from sd_embed.embedding_funcs import (
+        get_weighted_text_embeddings_sd15,
+        get_weighted_text_embeddings_sdxl,
+        get_weighted_text_embeddings_sd3,
+        get_weighted_text_embeddings_flux1,
+    )
+    SD_EMBED_AVAILABLE = True
+except ImportError:
+    get_weighted_text_embeddings_sd15 = None
+    get_weighted_text_embeddings_sdxl = None
+    get_weighted_text_embeddings_sd3 = None
+    get_weighted_text_embeddings_flux1 = None
+    SD_EMBED_AVAILABLE = False
 
 # Import LTX-2 specific utilities
 from diffusers.pipelines.ltx2.export_utils import encode_video as ltx2_encode_video
@@ -49,6 +63,9 @@ from diffusers import LTX2VideoTransformer3DModel, GGUFQuantizationConfig
 _ONE_DAY_IN_SECONDS = 60 * 60 * 24
 COMPEL = os.environ.get("COMPEL", "0") == "1"
 SD_EMBED = os.environ.get("SD_EMBED", "0") == "1"
+# Warn if SD_EMBED is enabled but the module is not available
+if SD_EMBED and not SD_EMBED_AVAILABLE:
+    print("WARNING: SD_EMBED is enabled but sd_embed module is not available. Falling back to standard prompt processing.", file=sys.stderr)
 XPU = os.environ.get("XPU", "0") == "1"
 CLIPSKIP = os.environ.get("CLIPSKIP", "1") == "1"
 SAFETENSORS = os.environ.get("SAFETENSORS", "1") == "1"
@@ -743,7 +760,7 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
                 guidance_scale=self.cfg_scale,
                 **kwargs
             ).images[0]
-        elif SD_EMBED:
+        elif SD_EMBED and SD_EMBED_AVAILABLE:
             if self.PipelineType == "StableDiffusionPipeline":
                 (
                     kwargs["prompt_embeds"],

backend/python/diffusers/requirements-l4t12.txt

@@ -4,7 +4,6 @@ git+https://github.com/huggingface/diffusers
 transformers
 accelerate
 compel
-git+https://github.com/xhinker/sd_embed
 peft
 optimum-quanto
 numpy<2

backend/python/diffusers/requirements-l4t13.txt

@@ -4,7 +4,6 @@ git+https://github.com/huggingface/diffusers
 transformers
 accelerate
 compel
-git+https://github.com/xhinker/sd_embed
 peft
 optimum-quanto
 numpy<2

backend/python/diffusers/requirements-mps.txt

@@ -5,7 +5,6 @@ opencv-python
 transformers
 accelerate
 compel
-git+https://github.com/xhinker/sd_embed
 peft
 sentencepiece
 optimum-quanto