Initial commit

.gitattributes

@@ -0,0 +1,37 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz 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
+RobertML.png filter=lfs diff=lfs merge=lfs -text
+backup.png filter=lfs diff=lfs merge=lfs -text

newcomer3

@@ -0,0 +1 @@
+Subproject commit 6eded002a0390c4b390c99fdce837c7ab936c7bb

pyproject.toml

@@ -0,0 +1,54 @@
+[build-system]
+requires = ["setuptools >= 75.0"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "flux-schnell-edge-inference"
+description = "An edge-maxxing model submission by RobertML 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",
+  "edge-maxxing-pipelines @ git+https://github.com/womboai/edge-maxxing@7c760ac54f6052803dadb3ade8ebfc9679a94589#subdirectory=pipelines",
+  "gitpython>=3.1.43",
+  "hf_transfer==0.1.8",
+  "torchao==0.6.1",
+  "setuptools>=75.3.0",
+]
+
+[[tool.edge-maxxing.models]]
+repository = "black-forest-labs/FLUX.1-schnell"
+revision = "741f7c3ce8b383c54771c7003378a50191e9efe9"
+exclude = ["transformer"]
+
+[[tool.edge-maxxing.models]]
+repository = "RobertML/FLUX.1-schnell-int8wo"
+revision = "307e0777d92df966a3c0f99f31a6ee8957a9857a"
+
+[[tool.edge-maxxing.models]]
+repository = "city96/t5-v1_1-xxl-encoder-bf16"
+revision = "1b9c856aadb864af93c1dcdc226c2774fa67bc86"
+
+[[tool.edge-maxxing.models]]
+repository = "RobertML/FLUX.1-schnell-vae_e3m2"
+revision = "da0d2cd7815792fb40d084dbd8ed32b63f153d8d"
+
+[[tool.edge-maxxing.models]]
+repository = "madebyollin/taef1"
+revision = "2d552378e58c9c94201075708d7de4e1163b2689"
+
+[[tool.edge-maxxing.models]]
+repository = "manbeast3b/flux.1-schnell-full1"
+revision = "cb1b599b0d712b9aab2c4df3ad27b050a27ec146"
+
+
+
+[project.scripts]
+start_inference = "main:main"
+

src/ghanta.py

@@ -0,0 +1,74 @@
+import torch
+from typing import Tuple, Callable
+def hacer_nada(x: torch.Tensor, modo: str = None):
+    return x
+def brujeria_mps(entrada, dim, indice):
+    if entrada.shape[-1] == 1:
+        return torch.gather(entrada.unsqueeze(-1), dim - 1 if dim < 0 else dim, indice.unsqueeze(-1)).squeeze(-1)
+    else:
+        return torch.gather(entrada, dim, indice)
+def emparejamiento_suave_aleatorio_2d(
+    metrica: torch.Tensor,
+    ancho: int,
+    alto: int,
+    paso_x: int,
+    paso_y: int,
+    radio: int,
+    sin_aleatoriedad: bool = False,
+    generador: torch.Generator = None
+) -> Tuple[Callable, Callable]:
+    lote, num_nodos, _ = metrica.shape
+    if radio <= 0:
+        return hacer_nada, hacer_nada
+    recopilar = brujeria_mps if metrica.device.type == "mps" else torch.gather
+    with torch.no_grad():
+        alto_paso_y, ancho_paso_x = alto // paso_y, ancho // paso_x
+        if sin_aleatoriedad:
+            indice_aleatorio = torch.zeros(alto_paso_y, ancho_paso_x, 1, device=metrica.device, dtype=torch.int64)
+        else:
+            indice_aleatorio = torch.randint(paso_y * paso_x, size=(alto_paso_y, ancho_paso_x, 1), device=generador.device, generator=generador).to(metrica.device)
+        vista_buffer_indice = torch.zeros(alto_paso_y, ancho_paso_x, paso_y * paso_x, device=metrica.device, dtype=torch.int64)
+        vista_buffer_indice.scatter_(dim=2, index=indice_aleatorio, src=-torch.ones_like(indice_aleatorio, dtype=indice_aleatorio.dtype))
+        vista_buffer_indice = vista_buffer_indice.view(alto_paso_y, ancho_paso_x, paso_y, paso_x).transpose(1, 2).reshape(alto_paso_y * paso_y, ancho_paso_x * paso_x)
+        if (alto_paso_y * paso_y) < alto or (ancho_paso_x * paso_x) < ancho:
+            buffer_indice = torch.zeros(alto, ancho, device=metrica.device, dtype=torch.int64)
+            buffer_indice[:(alto_paso_y * paso_y), :(ancho_paso_x * paso_x)] = vista_buffer_indice
+        else:
+            buffer_indice = vista_buffer_indice
+        indice_aleatorio = buffer_indice.reshape(1, -1, 1).argsort(dim=1)
+        del buffer_indice, vista_buffer_indice
+        num_destino = alto_paso_y * ancho_paso_x
+        indices_a = indice_aleatorio[:, num_destino:, :]
+        indices_b = indice_aleatorio[:, :num_destino, :]
+        def dividir(x):
+            canales = x.shape[-1]
+            origen = recopilar(x, dim=1, index=indices_a.expand(lote, num_nodos - num_destino, canales))
+            destino = recopilar(x, dim=1, index=indices_b.expand(lote, num_destino, canales))
+            return origen, destino
+        metrica = metrica / metrica.norm(dim=-1, keepdim=True)
+        a, b = dividir(metrica)
+        puntuaciones = a @ b.transpose(-1, -2)
+        radio = min(a.shape[1], radio)
+        nodo_max, nodo_indice = puntuaciones.max(dim=-1)
+        indice_borde = nodo_max.argsort(dim=-1, descending=True)[..., None]
+        indice_no_emparejado = indice_borde[..., radio:, :]
+        indice_origen = indice_borde[..., :radio, :]
+        indice_destino = recopilar(nodo_indice[..., None], dim=-2, index=indice_origen)
+    def fusionar(x: torch.Tensor, modo="mean") -> torch.Tensor:
+        origen, destino = dividir(x)
+        n, t1, c = origen.shape
+        no_emparejado = recopilar(origen, dim=-2, index=indice_no_emparejado.expand(n, t1 - radio, c))
+        origen = recopilar(origen, dim=-2, index=indice_origen.expand(n, radio, c))
+        destino = destino.scatter_reduce(-2, indice_destino.expand(n, radio, c), origen, reduce=modo)
+        return torch.cat([no_emparejado, destino], dim=1)
+    def desfusionar(x: torch.Tensor) -> torch.Tensor:
+        longitud_no_emparejado = indice_no_emparejado.shape[1]
+        no_emparejado, destino = x[..., :longitud_no_emparejado, :], x[..., longitud_no_emparejado:, :]
+        _, _, c = no_emparejado.shape
+        origen = recopilar(destino, dim=-2, index=indice_destino.expand(lote, radio, c))
+        salida = torch.zeros(lote, num_nodos, c, device=x.device, dtype=x.dtype)
+        salida.scatter_(dim=-2, index=indices_b.expand(lote, num_destino, c), src=destino)
+        salida.scatter_(dim=-2, index=recopilar(indices_a.expand(lote, indices_a.shape[1], 1), dim=1, index=indice_no_emparejado).expand(lote, longitud_no_emparejado, c), src=no_emparejado)
+        salida.scatter_(dim=-2, index=recopilar(indices_a.expand(lote, indices_a.shape[1], 1), dim=1, index=indice_origen).expand(lote, radio, c), src=origen)
+        return salida
+    return fusionar, desfusionar

src/main.py

@@ -0,0 +1,55 @@
+import atexit
+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 git import Repo
+import torch
+
+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 at_exit():
+    torch.cuda.empty_cache()
+
+
+def main():
+    atexit.register(at_exit)
+
+    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")
+            generator = torch.Generator("cuda")
+            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, generator.manual_seed(request.seed))
+                data = BytesIO()
+                image.save(data, format=JpegImageFile.format)
+
+                packet = data.getvalue()
+
+                connection.send_bytes(packet )
+
+
+if __name__ == '__main__':
+    main()

src/pipeline.py

@@ -0,0 +1,164 @@
+from diffusers import FluxPipeline, AutoencoderKL, AutoencoderTiny
+from diffusers.image_processor import VaeImageProcessor
+from diffusers.schedulers import FlowMatchEulerDiscreteScheduler
+from huggingface_hub.constants import HF_HUB_CACHE
+from transformers import T5EncoderModel, T5TokenizerFast, CLIPTokenizer, CLIPTextModel
+import torch
+import torch._dynamo
+import gc
+from PIL import Image as img
+from PIL.Image import Image
+from pipelines.models import TextToImageRequest
+from torch import Generator
+import time
+from diffusers import DiffusionPipeline
+from torchao.quantization import quantize_, int8_weight_only, fpx_weight_only
+
+import torch
+import math
+from typing import Type, Dict, Any, Tuple, Callable, Optional, Union
+import ghanta 
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.loaders import FromOriginalModelMixin, PeftAdapterMixin
+from diffusers.models.attention import FeedForward
+from diffusers.models.attention_processor import (
+    Attention,
+    AttentionProcessor,
+    FluxAttnProcessor2_0,
+    FusedFluxAttnProcessor2_0,
+)
+from diffusers.models.modeling_utils import ModelMixin
+from diffusers.models.normalization import AdaLayerNormContinuous, AdaLayerNormZero, AdaLayerNormZeroSingle
+from diffusers.utils import USE_PEFT_BACKEND, is_torch_version, logging, scale_lora_layers, unscale_lora_layers
+from diffusers.utils.import_utils import is_torch_npu_available
+from diffusers.utils.torch_utils import maybe_allow_in_graph
+from diffusers.models.embeddings import CombinedTimestepGuidanceTextProjEmbeddings, CombinedTimestepTextProjEmbeddings, FluxPosEmbed
+from diffusers.models.modeling_outputs import Transformer2DModelOutput
+
+import os
+os.environ['PYTORCH_CUDA_ALLOC_CONF']="expandable_segments:True"
+os.environ["TOKENIZERS_PARALLELISM"] = "True"
+torch._dynamo.config.suppress_errors = True
+
+class BasicQuantization:
+    def __init__(self, bits=1):
+        self.bits = bits
+        self.qmin = -(2**(bits-1))
+        self.qmax = 2**(bits-1) - 1
+
+    def quantize_tensor(self, tensor):
+        scale = (tensor.max() - tensor.min()) / (self.qmax - self.qmin)
+        zero_point = self.qmin - torch.round(tensor.min() / scale)
+        qtensor = torch.round(tensor / scale + zero_point)
+        qtensor = torch.clamp(qtensor, self.qmin, self.qmax)
+        return (qtensor - zero_point) * scale, scale, zero_point
+
+class ModelQuantization:
+    def __init__(self, model, bits=7):
+        self.model = model
+        self.quant = BasicQuantization(bits)
+
+    def quantize_model(self):
+        for name, module in self.model.named_modules():
+            if isinstance(module, torch.nn.Linear):
+                if hasattr(module, 'weightML'):
+                    quantized_weight, _, _ = self.quant.quantize_tensor(module.weight)
+                    module.weight = torch.nn.Parameter(quantized_weight)
+                if hasattr(module, 'bias') and module.bias is not None:
+                    quantized_bias, _, _ = self.quant.quantize_tensor(module.bias)
+                    module.bias = torch.nn.Parameter(quantized_bias)
+
+
+def inicializar_generador(dispositivo: torch.device, respaldo: torch.Generator = None):
+    if dispositivo.type == "cpu":
+        return torch.Generator(device="cpu").set_state(torch.get_rng_state())
+    elif dispositivo.type == "cuda":
+        return torch.Generator(device=dispositivo).set_state(torch.cuda.get_rng_state())
+    else:
+        if respaldo is None:
+            return inicializar_generador(torch.device("cpu"))
+        else:
+            return respaldo
+
+def calcular_fusion(x: torch.Tensor, info_tome: Dict[str, Any]) -> Tuple[Callable, ...]:
+    alto_original, ancho_original = info_tome["size"]
+    tokens_originales = alto_original * ancho_original
+    submuestreo = int(math.ceil(math.sqrt(tokens_originales // x.shape[1])))
+    argumentos = info_tome["args"]
+    if submuestreo <= argumentos["down"]:
+        ancho = int(math.ceil(ancho_original / submuestreo))
+        alto = int(math.ceil(alto_original / submuestreo))
+        radio = int(x.shape[1] * argumentos["ratio"])
+
+        if argumentos["generator"] is None:
+            argumentos["generator"] = inicializar_generador(x.device)
+        elif argumentos["generator"].device != x.device:
+            argumentos["generator"] = inicializar_generador(x.device, respaldo=argumentos["generator"])
+
+        usar_aleatoriedad = argumentos["rando"]
+        fusion, desfusion = ghanta.emparejamiento_suave_aleatorio_2d(
+            x, ancho, alto, argumentos["sx"], argumentos["sy"], radio,
+            sin_aleatoriedad=not usar_aleatoriedad, generador=argumentos["generator"]
+        )
+    else:
+        fusion, desfusion = (hacer_nada, hacer_nada)
+    fusion_a, desfusion_a = (fusion, desfusion) if argumentos["m1"] else (hacer_nada, hacer_nada)
+    fusion_c, desfusion_c = (fusion, desfusion) if argumentos["m2"] else (hacer_nada, hacer_nada)
+    fusion_m, desfusion_m = (fusion, desfusion) if argumentos["m3"] else (hacer_nada, hacer_nada)
+    return fusion_a, fusion_c, fusion_m, desfusion_a, desfusion_c, desfusion_m
+
+from diffusers import FluxPipeline, FluxTransformer2DModel
+Pipeline = None
+torch.backends.cuda.matmul.allow_tf32 = True
+torch.backends.cudnn.enabled = True
+torch.backends.cudnn.benchmark = True
+
+ckpt_id = "black-forest-labs/FLUX.1-schnell"
+ckpt_revision = "741f7c3ce8b383c54771c7003378a50191e9efe9" 
+
+TinyVAE = "madebyollin/taef1"
+TinyVAE_REV = "2d552378e58c9c94201075708d7de4e1163b2689"
+
+os.environ['PYTORCH_CUDA_ALLOC_CONF']="expandable_segments:True"
+os.environ["TOKENIZERS_PARALLELISM"] = "True"
+torch._dynamo.config.suppress_errors = True
+
+def empty_cache():
+    gc.collect()
+    torch.cuda.empty_cache()
+    torch.cuda.reset_max_memory_allocated()
+    torch.cuda.reset_peak_memory_stats()
+
+
+def load_pipeline() -> Pipeline:
+    path = os.path.join(HF_HUB_CACHE, "models--manbeast3b--flux.1-schnell-full1/snapshots/cb1b599b0d712b9aab2c4df3ad27b050a27ec146/transformer")
+    transformer = FluxTransformer2DModel.from_pretrained(path, torch_dtype=torch.bfloat16, use_safetensors=False)
+    vae = AutoencoderTiny.from_pretrained(
+            TinyVAE, 
+            revision=TinyVAE_REV, 
+            local_files_only=True,
+            torch_dtype=torch.bfloat16)
+    pipeline = FluxPipeline.from_pretrained(ckpt_id, revision=ckpt_revision, transformer=transformer, vae=vae, local_files_only=True, torch_dtype=torch.bfloat16,)
+    pipeline.to("cuda")
+    # quantize_(pipeline.vae, int8_weight_only())
+    pipeline.transformer = torch.compile(pipeline.transformer, mode="max-autotune", fullgraph=True)
+    pipeline.vae = torch.compile(pipeline.vae, mode="reduce-overhead", fullgraph=True)
+    for _ in range(3):
+        pipeline(prompt="insensible, timbale, pothery, electrovital, actinogram, taxis, intracerebellar, centrodesmus", width=1024, height=1024, guidance_scale=0.0, num_inference_steps=4, max_sequence_length=256)
+    return pipeline
+
+
+sample = None
+@torch.no_grad()
+def infer(request: TextToImageRequest, pipeline: Pipeline, generator: Generator) -> Image:
+    global sample
+    if not sample:
+        sample=1
+        empty_cache()
+    image=pipeline(request.prompt,generator=generator, guidance_scale=0.0, num_inference_steps=4, max_sequence_length=256, height=request.height, width=request.width, output_type="pil").images[0]
+    return image