feat(diffusers): support flux models (#3129)

* feat(diffusers): support flux models

This adds support for FLUX models. For instance:
https://huggingface.co/black-forest-labs/FLUX.1-dev

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

* feat(diffusers): support FluxTransformer2DModel

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

* Small fixups

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

---------

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

backend/python/diffusers/backend.py

@@ -18,13 +18,13 @@
 import grpc
 
 from diffusers import StableDiffusion3Pipeline, StableDiffusionXLPipeline, StableDiffusionDepth2ImgPipeline, DPMSolverMultistepScheduler, StableDiffusionPipeline, DiffusionPipeline, \
-    EulerAncestralDiscreteScheduler
+    EulerAncestralDiscreteScheduler, FluxPipeline, FluxTransformer2DModel
 from diffusers import StableDiffusionImg2ImgPipeline, AutoPipelineForText2Image, ControlNetModel, StableVideoDiffusionPipeline
 from diffusers.pipelines.stable_diffusion import safety_checker
 from diffusers.utils import load_image, export_to_video
 from compel import Compel, ReturnedEmbeddingsType
-
-from transformers import CLIPTextModel
+from optimum.quanto import freeze, qfloat8, quantize
+from transformers import CLIPTextModel, T5EncoderModel
 from safetensors.torch import load_file
 
 _ONE_DAY_IN_SECONDS = 60 * 60 * 24
@@ -163,6 +163,8 @@ def LoadModel(self, request, context):
             modelFile = request.Model
 
             self.cfg_scale = 7
+            self.PipelineType = request.PipelineType
+
             if request.CFGScale != 0:
                 self.cfg_scale = request.CFGScale
 
@@ -244,6 +246,30 @@ def LoadModel(self, request, context):
                         torch_dtype=torchType,
                         use_safetensors=True,
                         variant=variant)
+            elif request.PipelineType == "FluxPipeline":
+                    self.pipe = FluxPipeline.from_pretrained(
+                        request.Model,
+                        torch_dtype=torch.bfloat16)
+                    if request.LowVRAM:
+                        self.pipe.enable_model_cpu_offload()
+            elif request.PipelineType == "FluxTransformer2DModel":
+                    dtype = torch.bfloat16
+                    # specify from environment or default to "ChuckMcSneed/FLUX.1-dev"
+                    bfl_repo = os.environ.get("BFL_REPO", "ChuckMcSneed/FLUX.1-dev")
+
+                    transformer = FluxTransformer2DModel.from_single_file(modelFile, torch_dtype=dtype)
+                    quantize(transformer, weights=qfloat8)
+                    freeze(transformer)
+                    text_encoder_2 = T5EncoderModel.from_pretrained(bfl_repo, subfolder="text_encoder_2", torch_dtype=dtype)
+                    quantize(text_encoder_2, weights=qfloat8)
+                    freeze(text_encoder_2)
+
+                    self.pipe = FluxPipeline.from_pretrained(bfl_repo, transformer=None, text_encoder_2=None, torch_dtype=dtype)
+                    self.pipe.transformer = transformer
+                    self.pipe.text_encoder_2 = text_encoder_2
+
+                    if request.LowVRAM:
+                        self.pipe.enable_model_cpu_offload()
 
             if CLIPSKIP and request.CLIPSkip != 0:
                 self.clip_skip = request.CLIPSkip
@@ -399,6 +425,13 @@ def GenerateImage(self, request, context):
                 request.seed
             )
 
+        if self.PipelineType == "FluxPipeline":
+            kwargs["max_sequence_length"] = 256
+
+        if self.PipelineType == "FluxTransformer2DModel":
+            kwargs["output_type"] = "pil"
+            kwargs["generator"] = torch.Generator("cpu").manual_seed(0)
+
         if self.img2vid:
             # Load the conditioning image
             image = load_image(request.src)

backend/python/diffusers/requirements-cpu.txt

@@ -5,4 +5,5 @@ accelerate
 compel
 peft
 sentencepiece
-torch
+torch
+optimum-quanto

backend/python/diffusers/requirements-cublas11.txt

@@ -6,4 +6,5 @@ transformers
 accelerate
 compel
 peft
-sentencepiece
+sentencepiece
+optimum-quanto

backend/python/diffusers/requirements-cublas12.txt

@@ -5,4 +5,5 @@ transformers
 accelerate
 compel
 peft
-sentencepiece
+sentencepiece
+optimum-quanto

backend/python/diffusers/requirements-hipblas.txt

@@ -8,3 +8,4 @@ accelerate
 compel
 peft
 sentencepiece
+optimum-quanto

backend/python/diffusers/requirements-intel.txt

@@ -10,4 +10,5 @@ transformers
 accelerate
 compel
 peft
-sentencepiece
+sentencepiece
+optimum-quanto