Upload folder using huggingface_hub

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+sample.png filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -0,0 +1,9 @@
+**/.cache
+**/__pycache__
+**/*.egg-info
+*.safetensors
+**/.venv
+.venv
+.git
+*.png
+*.jpeg

README.md

@@ -0,0 +1,19 @@
+# flux-schnell-edge-inference
+
+This holds the baseline for the FLUX Schnel NVIDIA GeForce RTX 4090 contest, which can be forked freely and optimized
+
+Some recommendations are as follows:
+- Installing dependencies should be done in `pyproject.toml`, including git dependencies
+- HuggingFace models should be specified in the `models` array in the `pyproject.toml` file, and will be downloaded before benchmarking
+    - The pipeline does **not** have internet access so all dependencies and models must be included in the `pyproject.toml`
+    - Compiled models should be hosted on HuggingFace and included in the `models` array in the `pyproject.toml` (rather than compiling during loading). Loading time matters far more than file sizes
+- Avoid changing `src/main.py`, as that includes mostly protocol logic. Most changes should be in `models` and `src/pipeline.py`
+- Ensure the entire repository (excluding dependencies and HuggingFace models) is under 16MB
+
+For testing, you need a docker container with pytorch and ubuntu 22.04.
+You can download your listed dependencies with `uv`, installed with:
+```bash
+pipx ensurepath
+pipx install uv
+```
+You can then relock with `uv lock`, and then run with `uv run start_inference`

pyproject.toml

@@ -0,0 +1,30 @@
+[build-system]
+requires = ["setuptools >= 75.0"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "flux-schnell-edge-inference"
+description = "An edge-maxxing model submission for the 4090 Flux contest"
+requires-python = ">=3.10,<3.13"
+version = "8"
+dependencies = [
+  "diffusers==0.31.0",
+  "transformers==4.46.2",
+  "accelerate==1.1.0",
+  "omegaconf==2.3.0",
+  "torch==2.5.1",
+  "protobuf==5.28.3",
+  "sentencepiece==0.2.0",
+  "hf_transfer==0.1.8",
+  "edge-maxxing-pipelines @ git+https://github.com/womboai/edge-maxxing@7c760ac54f6052803dadb3ade8ebfc9679a94589#subdirectory=pipelines",
+  "torchao>=0.6.1",
+  "ipython>=8.29.0",
+  "setuptools >= 75.0"
+]
+
+[tool.edge-maxxing]
+repository = "jokerbit/flux.1-schnell-Robert-int8wo"
+revision = "5ef0012f11a863e5111ec56540302a023bc8587b"
+
+[project.scripts]
+start_inference = "main:main"

src/main.py

@@ -0,0 +1,50 @@
+from io import BytesIO
+from multiprocessing.connection import Listener
+from os import chmod, remove
+from os.path import abspath, exists
+from pathlib import Path
+
+from PIL.JpegImagePlugin import JpegImageFile
+from pipelines.models import TextToImageRequest
+
+from pipeline import load_pipeline, infer
+
+SOCKET = abspath(Path(__file__).parent.parent / "inferences.sock")
+
+
+def main():
+    print(f"Loading pipeline")
+    pipeline = load_pipeline()
+
+    print(f"Pipeline loaded, creating socket at '{SOCKET}'")
+
+    if exists(SOCKET):
+        remove(SOCKET)
+
+    with Listener(SOCKET) as listener:
+        chmod(SOCKET, 0o777)
+
+        print(f"Awaiting connections")
+        with listener.accept() as connection:
+            print(f"Connected")
+
+            while True:
+                try:
+                    request = TextToImageRequest.model_validate_json(connection.recv_bytes().decode("utf-8"))
+                except EOFError:
+                    print(f"Inference socket exiting")
+
+                    return
+
+                image = infer(request, pipeline)
+
+                data = BytesIO()
+                image.save(data, format=JpegImageFile.format)
+
+                packet = data.getvalue()
+
+                connection.send_bytes(packet)
+
+
+if __name__ == '__main__':
+    main()

src/pipeline.py

@@ -0,0 +1,165 @@
+import os
+from diffusers import FluxPipeline, AutoencoderKL, FluxTransformer2DModel
+from diffusers.image_processor import VaeImageProcessor
+from transformers import T5EncoderModel, T5TokenizerFast, CLIPTokenizer, CLIPTextModel, CLIPTextConfig, T5Config
+import torch
+import gc
+from PIL.Image import Image
+from pipelines.models import TextToImageRequest
+from torch import Generator
+from torchao.quantization import quantize_, int8_weight_only
+from time import perf_counter
+
+
+HOME = os.environ["HOME"]
+QUANTIZED_MODEL = ["text_encoder_2"]
+os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:False,garbage_collection_threshold:0.01"
+FLUX_CHECKPOINT = "jokerbit/flux.1-schnell-Robert-int8wo"
+FLUX_CACHE = os.path.join(HOME, ".cache/huggingface/hub/models--jokerbit--flux.1-schnell-Robert-int8wo/snapshots/5ef0012f11a863e5111ec56540302a023bc8587b")
+torch.backends.cudnn.benchmark = True
+torch.backends.cuda.matmul.allow_tf32 = True
+torch.cuda.set_per_process_memory_fraction(0.99)
+
+QUANT_CONFIG = int8_weight_only()
+DTYPE = torch.bfloat16
+NUM_STEPS = 4
+PROMPT = 'martyr, semiconformity, peregrination, quip, twineless, emotionless, tawa, depickle'
+
+
+def empty_cache():
+    gc.collect()
+    torch.cuda.empty_cache()
+    torch.cuda.reset_max_memory_allocated()
+    torch.cuda.reset_peak_memory_stats()
+
+
+def quantize(pipe, config):
+    if "text_encoder" in QUANTIZED_MODEL:
+        quantize_(pipe.text_encoder, config)
+    if "text_encoder_2" in QUANTIZED_MODEL:
+        quantize_(pipe.text_encoder_2, config)
+    if "transformer" in QUANTIZED_MODEL:
+        quantize_(pipe.transformer, config, device="cuda")
+    if "vae" in QUANTIZED_MODEL:
+        quantize_(pipe.vae, config)
+    return pipe
+
+
+def load_pipeline() -> FluxPipeline:
+    empty_cache()
+    transformer = FluxTransformer2DModel.from_pretrained(os.path.join(FLUX_CACHE, "transformer"), use_safetensors=False, torch_dtype=DTYPE)
+    pipe = FluxPipeline.from_pretrained(FLUX_CHECKPOINT,
+                                        transformer=transformer,
+                                        torch_dtype=DTYPE) 
+    pipe.vae.enable_tiling()
+    pipe.vae.enable_slicing()
+    quantize(pipe, QUANT_CONFIG)
+    pipe.to("cuda")
+    request = TextToImageRequest(prompt=PROMPT, height=1024, width=1024, seed=666)
+    infer(request, pipe)
+    # pipe.enable_model_cpu_offload()
+    return pipe
+
+
+def encode_prompt(_pipeline, prompt: str):
+    pipeline = FluxPipeline.from_pipe(
+            _pipeline,
+            transformer=None,
+            vae=None,
+            ).to("cuda")
+    with torch.no_grad():
+        outputs = pipeline.encode_prompt(
+                prompt=prompt,
+                prompt_2=None,
+                max_sequence_length=256)
+    del pipeline 
+    empty_cache()
+    return outputs
+
+
+def infer_latents(_pipeline, prompt_embeds, pooled_prompt_embeds, width: int | None, height: int | None, seed: int | None):
+    pipeline = FluxPipeline.from_pipe(
+            _pipeline,
+            text_encoder=None,
+            text_encoder_2=None,
+            tokenizer=None,
+            tokenizer_2=None,
+            vae=None,
+            ).to("cuda")
+
+    if seed is None:
+        generator = None
+    else:
+        generator = Generator(pipeline.device).manual_seed(seed)
+    outputs = pipeline(
+            prompt_embeds=prompt_embeds,
+            pooled_prompt_embeds=pooled_prompt_embeds,
+            num_inference_steps=4,
+            guidance_scale=0.0,
+            width=width,
+            height=height,
+            generator=generator,
+            output_type="latent",
+            ).images
+    del pipeline
+    empty_cache()
+    return outputs
+
+
+def decode_latents(vae, latents, width, height):
+    vae.to("cuda")
+    vae_scale_factor = 2 ** (len(vae.config.block_out_channels))
+    width = width or 64 * vae_scale_factor
+    height = height or 64 * vae_scale_factor
+    image_processor = VaeImageProcessor(vae_scale_factor=vae_scale_factor)
+    with torch.no_grad():
+        latents = FluxPipeline._unpack_latents(latents, height, width, vae_scale_factor)
+        latents = (latents / vae.config.scaling_factor) + vae.config.shift_factor
+        image = vae.decode(latents, return_dict=False)[0] 
+    return image_processor.postprocess(image, output_type="pil")[0]
+
+
+# def infer(request: TextToImageRequest, _pipeline: FluxPipeline) -> Image: 
+#     prompt_embeds, pooled_prompt_embeds, text_ids = encode_prompt(_pipeline, request.prompt) 
+#     latents = infer_latents(_pipeline, prompt_embeds, pooled_prompt_embeds, request.width, request.height, request.seed)
+#     del prompt_embeds
+#     del pooled_prompt_embeds
+#     del text_ids
+#     # _pipeline.transformer.to("cpu")
+#     image = decode_latents(_pipeline.vae, latents, request.width, request.height)
+#     torch.cuda.reset_peak_memory_stats()
+#     return image
+
+
+def infer(request: TextToImageRequest, _pipeline: FluxPipeline) -> Image:
+    if request.seed is None:
+        generator = None
+    else:
+        generator = Generator(device="cuda").manual_seed(request.seed)
+    torch.cuda.reset_peak_memory_stats()
+    image = _pipeline(prompt=request.prompt,
+            width=request.width,
+            height=request.height,
+            guidance_scale=0.0,
+            generator=generator,
+            output_type="pil",                          
+            max_sequence_length=256,
+            num_inference_steps=NUM_STEPS).images[0]
+    return image
+
+if __name__ == "__main__":
+    request = TextToImageRequest(prompt=PROMPT, 
+            height=None,
+            width=None,
+            seed=666)
+    start_time = perf_counter()
+    pipe_ = load_pipeline()
+    stop_time = perf_counter()
+    print(f"Pipeline is loaded in {stop_time - start_time}s")
+    for _ in range(4):
+        start_time = perf_counter()
+        infer(request, pipe_)
+        stop_time = perf_counter()
+        print(f"Request in {stop_time - start_time}s")
+
+    # pipe("cat holding a womboai sign", num_inference_steps=4, guidance_scale=0, generator=torch.Generator(pipe.device).manual_seed(666)).images[0].save("sample.png")