Add files using upload-large-folder tool

.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
+RobertML.png filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,2 @@
+# flux-schnell-edge-inference
+nestas hagunnan hinase

loss_params.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b0ee6fa5873dbc8df9daeeb105e220266bcf6634c6806b69da38fdc0a5c12b81
+size 3184

pyproject.toml

@@ -0,0 +1,44 @@
+[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==2.6.0",
+  "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",
+]
+
+[[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"
+
+
+[project.scripts]
+start_inference = "main:main"
+

src/first_block_cache/diffusers_adapters/__init__.py

@@ -0,0 +1,45 @@
+import importlib
+
+from diffusers import DiffusionPipeline
+
+
+def apply_cache_on_transformer(transformer, *args, **kwargs):
+    transformer_cls_name = transformer.__class__.__name__
+    if False:
+        pass
+    elif transformer_cls_name.startswith("Flux"):
+        adapter_name = "flux"
+    elif transformer_cls_name.startswith("Mochi"):
+        adapter_name = "mochi"
+    elif transformer_cls_name.startswith("CogVideoX"):
+        adapter_name = "cogvideox"
+    elif transformer_cls_name.startswith("HunyuanVideo"):
+        adapter_name = "hunyuan_video"
+    else:
+        raise ValueError(f"Unknown transformer class name: {transformer_cls_name}")
+
+    adapter_module = importlib.import_module(f".{adapter_name}", __package__)
+    apply_cache_on_transformer_fn = getattr(adapter_module, "apply_cache_on_transformer")
+    return apply_cache_on_transformer_fn(transformer, *args, **kwargs)
+
+
+def apply_cache_on_pipe(pipe: DiffusionPipeline, *args, **kwargs):
+    assert isinstance(pipe, DiffusionPipeline)
+
+    pipe_cls_name = pipe.__class__.__name__
+    if False:
+        pass
+    elif pipe_cls_name.startswith("Flux"):
+        adapter_name = "flux"
+    elif pipe_cls_name.startswith("Mochi"):
+        adapter_name = "mochi"
+    elif pipe_cls_name.startswith("CogVideoX"):
+        adapter_name = "cogvideox"
+    elif pipe_cls_name.startswith("HunyuanVideo"):
+        adapter_name = "hunyuan_video"
+    else:
+        raise ValueError(f"Unknown pipeline class name: {pipe_cls_name}")
+
+    adapter_module = importlib.import_module(f".{adapter_name}", __package__)
+    apply_cache_on_pipe_fn = getattr(adapter_module, "apply_cache_on_pipe")
+    return apply_cache_on_pipe_fn(pipe, *args, **kwargs)

src/first_block_cache/diffusers_adapters/cogvideox.py

@@ -0,0 +1,72 @@
+import functools
+import unittest
+
+import torch
+from diffusers import CogVideoXTransformer3DModel, DiffusionPipeline
+
+from para_attn.first_block_cache import utils
+
+
+def apply_cache_on_transformer(
+    transformer: CogVideoXTransformer3DModel,
+    *,
+    residual_diff_threshold=0.04,
+):
+    cached_transformer_blocks = torch.nn.ModuleList(
+        [
+            utils.CachedTransformerBlocks(
+                transformer.transformer_blocks,
+                transformer=transformer,
+                residual_diff_threshold=residual_diff_threshold,
+            )
+        ]
+    )
+
+    original_forward = transformer.forward
+
+    @functools.wraps(transformer.__class__.forward)
+    def new_forward(
+        self,
+        *args,
+        **kwargs,
+    ):
+        with unittest.mock.patch.object(
+            self,
+            "transformer_blocks",
+            cached_transformer_blocks,
+        ):
+            return original_forward(
+                *args,
+                **kwargs,
+            )
+
+    transformer.forward = new_forward.__get__(transformer)
+
+    return transformer
+
+
+def apply_cache_on_pipe(
+    pipe: DiffusionPipeline,
+    *,
+    shallow_patch: bool = False,
+    **kwargs,
+):
+    original_call = pipe.__class__.__call__
+
+    if not getattr(original_call, "_is_cached", False):
+
+        @functools.wraps(original_call)
+        def new_call(self, *args, **kwargs):
+            with utils.cache_context(utils.create_cache_context()):
+                return original_call(self, *args, **kwargs)
+
+        pipe.__class__.__call__ = new_call
+
+        new_call._is_cached = True
+
+    if not shallow_patch:
+        apply_cache_on_transformer(pipe.transformer, **kwargs)
+
+    pipe._is_cached = True
+
+    return pipe

src/first_block_cache/diffusers_adapters/flux.py

@@ -0,0 +1,79 @@
+import functools
+import unittest
+
+import torch
+from diffusers import DiffusionPipeline, FluxTransformer2DModel
+
+from first_block_cache import utils
+
+
+def apply_cache_on_transformer(
+    transformer: FluxTransformer2DModel,
+    *,
+    residual_diff_threshold=0.05,
+):
+    cached_transformer_blocks = torch.nn.ModuleList(
+        [
+            utils.CachedTransformerBlocks(
+                transformer.transformer_blocks,
+                transformer.single_transformer_blocks,
+                transformer=transformer,
+                residual_diff_threshold=residual_diff_threshold,
+                return_hidden_states_first=False,
+            )
+        ]
+    )
+    dummy_single_transformer_blocks = torch.nn.ModuleList()
+
+    original_forward = transformer.forward
+
+    @functools.wraps(original_forward)
+    def new_forward(
+        self,
+        *args,
+        **kwargs,
+    ):
+        with unittest.mock.patch.object(
+            self,
+            "transformer_blocks",
+            cached_transformer_blocks,
+        ), unittest.mock.patch.object(
+            self,
+            "single_transformer_blocks",
+            dummy_single_transformer_blocks,
+        ):
+            return original_forward(
+                *args,
+                **kwargs,
+            )
+
+    transformer.forward = new_forward.__get__(transformer)
+
+    return transformer
+
+
+def apply_cache_on_pipe(
+    pipe: DiffusionPipeline,
+    *,
+    shallow_patch: bool = False,
+    **kwargs,
+):
+    original_call = pipe.__class__.__call__
+
+    if not getattr(original_call, "_is_cached", False):
+
+        @functools.wraps(original_call)
+        def new_call(self, *args, **kwargs):
+            with utils.cache_context(utils.create_cache_context()):
+                return original_call(self, *args, **kwargs)
+
+        pipe.__class__.__call__ = new_call
+
+        new_call._is_cached = True
+
+    if not shallow_patch:
+        apply_cache_on_transformer(pipe.transformer, **kwargs)
+
+    pipe._is_cached = True
+
+    return pipe

src/first_block_cache/diffusers_adapters/hunyuan_video.py

@@ -0,0 +1,199 @@
+import functools
+import unittest
+from typing import Any, Dict, Optional, Union
+
+import torch
+from diffusers import DiffusionPipeline, HunyuanVideoTransformer3DModel
+from diffusers.models.modeling_outputs import Transformer2DModelOutput
+from diffusers.utils import logging, scale_lora_layers, unscale_lora_layers, USE_PEFT_BACKEND
+
+from para_attn.first_block_cache import utils
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+def apply_cache_on_transformer(
+    transformer: HunyuanVideoTransformer3DModel,
+    *,
+    residual_diff_threshold=0.06,
+):
+    cached_transformer_blocks = torch.nn.ModuleList(
+        [
+            utils.CachedTransformerBlocks(
+                transformer.transformer_blocks + transformer.single_transformer_blocks,
+                transformer=transformer,
+                residual_diff_threshold=residual_diff_threshold,
+            )
+        ]
+    )
+    dummy_single_transformer_blocks = torch.nn.ModuleList()
+
+    original_forward = transformer.forward
+
+    @functools.wraps(transformer.__class__.forward)
+    def new_forward(
+        self,
+        hidden_states: torch.Tensor,
+        timestep: torch.LongTensor,
+        encoder_hidden_states: torch.Tensor,
+        encoder_attention_mask: torch.Tensor,
+        pooled_projections: torch.Tensor,
+        guidance: torch.Tensor = None,
+        attention_kwargs: Optional[Dict[str, Any]] = None,
+        return_dict: bool = True,
+        **kwargs,
+    ) -> Union[torch.Tensor, Dict[str, torch.Tensor]]:
+        with unittest.mock.patch.object(
+            self,
+            "transformer_blocks",
+            cached_transformer_blocks,
+        ), unittest.mock.patch.object(
+            self,
+            "single_transformer_blocks",
+            dummy_single_transformer_blocks,
+        ):
+            if getattr(self, "_is_parallelized", False):
+                return original_forward(
+                    hidden_states,
+                    timestep,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    pooled_projections,
+                    guidance=guidance,
+                    attention_kwargs=attention_kwargs,
+                    return_dict=return_dict,
+                    **kwargs,
+                )
+            else:
+                if attention_kwargs is not None:
+                    attention_kwargs = attention_kwargs.copy()
+                    lora_scale = attention_kwargs.pop("scale", 1.0)
+                else:
+                    lora_scale = 1.0
+
+                if USE_PEFT_BACKEND:
+                    # weight the lora layers by setting `lora_scale` for each PEFT layer
+                    scale_lora_layers(self, lora_scale)
+                else:
+                    if attention_kwargs is not None and attention_kwargs.get("scale", None) is not None:
+                        logger.warning(
+                            "Passing `scale` via `attention_kwargs` when not using the PEFT backend is ineffective."
+                        )
+
+                batch_size, num_channels, num_frames, height, width = hidden_states.shape
+                p, p_t = self.config.patch_size, self.config.patch_size_t
+                post_patch_num_frames = num_frames // p_t
+                post_patch_height = height // p
+                post_patch_width = width // p
+
+                # 1. RoPE
+                image_rotary_emb = self.rope(hidden_states)
+
+                # 2. Conditional embeddings
+                temb = self.time_text_embed(timestep, guidance, pooled_projections)
+                hidden_states = self.x_embedder(hidden_states)
+                encoder_hidden_states = self.context_embedder(encoder_hidden_states, timestep, encoder_attention_mask)
+
+                encoder_hidden_states = encoder_hidden_states[:, encoder_attention_mask[0].bool()]
+
+                # 4. Transformer blocks
+                hidden_states, encoder_hidden_states = self.call_transformer_blocks(
+                    hidden_states, encoder_hidden_states, temb, None, image_rotary_emb
+                )
+
+                # 5. Output projection
+                hidden_states = self.norm_out(hidden_states, temb)
+                hidden_states = self.proj_out(hidden_states)
+
+                hidden_states = hidden_states.reshape(
+                    batch_size, post_patch_num_frames, post_patch_height, post_patch_width, -1, p_t, p, p
+                )
+                hidden_states = hidden_states.permute(0, 4, 1, 5, 2, 6, 3, 7)
+                hidden_states = hidden_states.flatten(6, 7).flatten(4, 5).flatten(2, 3)
+
+                hidden_states = hidden_states.to(timestep.dtype)
+
+                if USE_PEFT_BACKEND:
+                    # remove `lora_scale` from each PEFT layer
+                    unscale_lora_layers(self, lora_scale)
+
+                if not return_dict:
+                    return (hidden_states,)
+
+                return Transformer2DModelOutput(sample=hidden_states)
+
+    transformer.forward = new_forward.__get__(transformer)
+
+    def call_transformer_blocks(self, hidden_states, encoder_hidden_states, *args, **kwargs):
+        if torch.is_grad_enabled() and self.gradient_checkpointing:
+
+            def create_custom_forward(module, return_dict=None):
+                def custom_forward(*inputs):
+                    if return_dict is not None:
+                        return module(*inputs, return_dict=return_dict)
+                    else:
+                        return module(*inputs)
+
+                return custom_forward
+
+            ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False}
+
+            for block in self.transformer_blocks:
+                hidden_states, encoder_hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(block),
+                    hidden_states,
+                    encoder_hidden_states,
+                    *args,
+                    **kwargs,
+                    **ckpt_kwargs,
+                )
+
+            for block in self.single_transformer_blocks:
+                hidden_states, encoder_hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(block),
+                    hidden_states,
+                    encoder_hidden_states,
+                    *args,
+                    **kwargs,
+                    **ckpt_kwargs,
+                )
+
+        else:
+            for block in self.transformer_blocks:
+                hidden_states, encoder_hidden_states = block(hidden_states, encoder_hidden_states, *args, **kwargs)
+
+            for block in self.single_transformer_blocks:
+                hidden_states, encoder_hidden_states = block(hidden_states, encoder_hidden_states, *args, **kwargs)
+
+        return hidden_states, encoder_hidden_states
+
+    transformer.call_transformer_blocks = call_transformer_blocks.__get__(transformer)
+
+    return transformer
+
+
+def apply_cache_on_pipe(
+    pipe: DiffusionPipeline,
+    *,
+    shallow_patch: bool = False,
+    **kwargs,
+):
+    original_call = pipe.__class__.__call__
+
+    if not getattr(original_call, "_is_cached", False):
+
+        @functools.wraps(original_call)
+        def new_call(self, *args, **kwargs):
+            with utils.cache_context(utils.create_cache_context()):
+                return original_call(self, *args, **kwargs)
+
+        pipe.__class__.__call__ = new_call
+
+        new_call._is_cached = True
+
+    if not shallow_patch:
+        apply_cache_on_transformer(pipe.transformer, **kwargs)
+
+    pipe._is_cached = True
+
+    return pipe

src/first_block_cache/diffusers_adapters/mochi.py

@@ -0,0 +1,72 @@
+import functools
+import unittest
+
+import torch
+from diffusers import DiffusionPipeline, MochiTransformer3DModel
+
+from para_attn.first_block_cache import utils
+
+
+def apply_cache_on_transformer(
+    transformer: MochiTransformer3DModel,
+    *,
+    residual_diff_threshold=0.06,
+):
+    cached_transformer_blocks = torch.nn.ModuleList(
+        [
+            utils.CachedTransformerBlocks(
+                transformer.transformer_blocks,
+                transformer=transformer,
+                residual_diff_threshold=residual_diff_threshold,
+            )
+        ]
+    )
+
+    original_forward = transformer.forward
+
+    @functools.wraps(transformer.__class__.forward)
+    def new_forward(
+        self,
+        *args,
+        **kwargs,
+    ):
+        with unittest.mock.patch.object(
+            self,
+            "transformer_blocks",
+            cached_transformer_blocks,
+        ):
+            return original_forward(
+                *args,
+                **kwargs,
+            )
+
+    transformer.forward = new_forward.__get__(transformer)
+
+    return transformer
+
+
+def apply_cache_on_pipe(
+    pipe: DiffusionPipeline,
+    *,
+    shallow_patch: bool = False,
+    **kwargs,
+):
+    original_call = pipe.__class__.__call__
+
+    if not getattr(original_call, "_is_cached", False):
+
+        @functools.wraps(original_call)
+        def new_call(self, *args, **kwargs):
+            with utils.cache_context(utils.create_cache_context()):
+                return original_call(self, *args, **kwargs)
+
+        pipe.__class__.__call__ = new_call
+
+        new_call._is_cached = True
+
+    if not shallow_patch:
+        apply_cache_on_transformer(pipe.transformer, **kwargs)
+
+    pipe._is_cached = True
+
+    return pipe

src/first_block_cache/utils.py

@@ -0,0 +1,222 @@
+import contextlib
+import dataclasses
+from collections import defaultdict
+from typing import DefaultDict, Dict
+from pipeline import are_two_tensors_similar
+import torch
+
+
+
+@dataclasses.dataclass
+class CacheContext:
+    buffers: Dict[str, torch.Tensor] = dataclasses.field(default_factory=dict)
+    incremental_name_counters: DefaultDict[str, int] = dataclasses.field(default_factory=lambda: defaultdict(int))
+
+    def get_incremental_name(self, name=None):
+        if name is None:
+            name = "default"
+        idx = self.incremental_name_counters[name]
+        self.incremental_name_counters[name] += 1
+        return f"{name}_{idx}"
+
+    def reset_incremental_names(self):
+        self.incremental_name_counters.clear()
+
+    @torch.compiler.disable
+    def get_buffer(self, name):
+        return self.buffers.get(name)
+
+    @torch.compiler.disable
+    def set_buffer(self, name, buffer):
+        self.buffers[name] = buffer
+
+    def clear_buffers(self):
+        self.buffers.clear()
+
+
+@torch.compiler.disable
+def get_buffer(name):
+    cache_context = get_current_cache_context()
+    assert cache_context is not None, "cache_context must be set before"
+    return cache_context.get_buffer(name)
+
+
+@torch.compiler.disable
+def set_buffer(name, buffer):
+    cache_context = get_current_cache_context()
+    assert cache_context is not None, "cache_context must be set before"
+    cache_context.set_buffer(name, buffer)
+
+
+_current_cache_context = None
+
+
+def create_cache_context():
+    return CacheContext()
+
+
+def get_current_cache_context():
+    return _current_cache_context
+
+
+def set_current_cache_context(cache_context=None):
+    global _current_cache_context
+    _current_cache_context = cache_context
+
+
+@contextlib.contextmanager
+def cache_context(cache_context):
+    global _current_cache_context
+    old_cache_context = _current_cache_context
+    _current_cache_context = cache_context
+    try:
+        yield
+    finally:
+        _current_cache_context = old_cache_context
+
+
+@torch.compiler.disable
+def are_two_tensors_similar_old(t1, t2, *, threshold, parallelized=False):
+    mean_diff = (t1 - t2).abs().mean()
+    mean_t1 = t1.abs().mean()
+    diff = mean_diff / mean_t1
+    return diff.item() < threshold
+
+
+@torch.compiler.disable
+def apply_prev_hidden_states_residual(hidden_states, encoder_hidden_states):
+    hidden_states_residual = get_buffer("hidden_states_residual")
+    assert hidden_states_residual is not None, "hidden_states_residual must be set before"
+    hidden_states = hidden_states_residual + hidden_states
+
+    encoder_hidden_states_residual = get_buffer("encoder_hidden_states_residual")
+    assert encoder_hidden_states_residual is not None, "encoder_hidden_states_residual must be set before"
+    encoder_hidden_states = encoder_hidden_states_residual + encoder_hidden_states
+
+    hidden_states = hidden_states.contiguous()
+    encoder_hidden_states = encoder_hidden_states.contiguous()
+
+    return hidden_states, encoder_hidden_states
+
+
+@torch.compiler.disable
+def get_can_use_cache(first_hidden_states_residual, threshold, parallelized=False):
+    prev_first_hidden_states_residual = get_buffer("first_hidden_states_residual")
+    can_use_cache = prev_first_hidden_states_residual is not None and are_two_tensors_similar(
+        prev_first_hidden_states_residual,
+        first_hidden_states_residual,
+        threshold=threshold,
+        parallelized=parallelized,
+    )
+    return can_use_cache
+
+
+class CachedTransformerBlocks(torch.nn.Module):
+    def __init__(
+        self,
+        transformer_blocks,
+        single_transformer_blocks=None,
+        *,
+        transformer=None,
+        residual_diff_threshold,
+        return_hidden_states_first=True,
+    ):
+        super().__init__()
+        self.transformer = transformer
+        self.transformer_blocks = transformer_blocks
+        self.single_transformer_blocks = single_transformer_blocks
+        self.residual_diff_threshold = residual_diff_threshold
+        self.return_hidden_states_first = return_hidden_states_first
+
+    def forward(self, hidden_states, encoder_hidden_states, *args, **kwargs):
+        if self.residual_diff_threshold <= 0.0:
+            for block in self.transformer_blocks:
+                hidden_states, encoder_hidden_states = block(hidden_states, encoder_hidden_states, *args, **kwargs)
+                if not self.return_hidden_states_first:
+                    hidden_states, encoder_hidden_states = encoder_hidden_states, hidden_states
+            if self.single_transformer_blocks is not None:
+                hidden_states = torch.cat([encoder_hidden_states, hidden_states], dim=1)
+                for block in self.single_transformer_blocks:
+                    hidden_states = block(hidden_states, *args, **kwargs)
+                hidden_states = hidden_states[:, encoder_hidden_states.shape[1] :]
+            return (
+                (hidden_states, encoder_hidden_states)
+                if self.return_hidden_states_first
+                else (encoder_hidden_states, hidden_states)
+            )
+
+        original_hidden_states = hidden_states
+        first_transformer_block = self.transformer_blocks[0]
+        hidden_states, encoder_hidden_states = first_transformer_block(
+            hidden_states, encoder_hidden_states, *args, **kwargs
+        )
+        if not self.return_hidden_states_first:
+            hidden_states, encoder_hidden_states = encoder_hidden_states, hidden_states
+        first_hidden_states_residual = hidden_states - original_hidden_states
+        del original_hidden_states
+
+        can_use_cache = get_can_use_cache(
+            first_hidden_states_residual,
+            threshold=self.residual_diff_threshold,
+            parallelized=self.transformer is not None and getattr(self.transformer, "_is_parallelized", False),
+        )
+
+        torch._dynamo.graph_break()
+        if can_use_cache:
+            del first_hidden_states_residual
+            hidden_states, encoder_hidden_states = apply_prev_hidden_states_residual(
+                hidden_states, encoder_hidden_states
+            )
+        else:
+            set_buffer("first_hidden_states_residual", first_hidden_states_residual)
+            del first_hidden_states_residual
+            (
+                hidden_states,
+                encoder_hidden_states,
+                hidden_states_residual,
+                encoder_hidden_states_residual,
+            ) = self.call_remaining_transformer_blocks(hidden_states, encoder_hidden_states, *args, **kwargs)
+            set_buffer("hidden_states_residual", hidden_states_residual)
+            set_buffer("encoder_hidden_states_residual", encoder_hidden_states_residual)
+        torch._dynamo.graph_break()
+
+        return (
+            (hidden_states, encoder_hidden_states)
+            if self.return_hidden_states_first
+            else (encoder_hidden_states, hidden_states)
+        )
+
+    def call_remaining_transformer_blocks(self, hidden_states, encoder_hidden_states, *args, **kwargs):
+        original_hidden_states = hidden_states
+        original_encoder_hidden_states = encoder_hidden_states
+        for block in self.transformer_blocks[1:]:
+            hidden_states, encoder_hidden_states = block(hidden_states, encoder_hidden_states, *args, **kwargs)
+            if not self.return_hidden_states_first:
+                hidden_states, encoder_hidden_states = encoder_hidden_states, hidden_states
+        if self.single_transformer_blocks is not None:
+            hidden_states = torch.cat([encoder_hidden_states, hidden_states], dim=1)
+            for block in self.single_transformer_blocks:
+                hidden_states = block(hidden_states, *args, **kwargs)
+            encoder_hidden_states, hidden_states = hidden_states.split(
+                [encoder_hidden_states.shape[1], hidden_states.shape[1] - encoder_hidden_states.shape[1]], dim=1
+            )
+
+        # hidden_states_shape = hidden_states.shape
+        # encoder_hidden_states_shape = encoder_hidden_states.shape
+        hidden_states = hidden_states.reshape(-1).contiguous().reshape(original_hidden_states.shape)
+        encoder_hidden_states = (
+            encoder_hidden_states.reshape(-1).contiguous().reshape(original_encoder_hidden_states.shape)
+        )
+
+        # hidden_states = hidden_states.contiguous()
+        # encoder_hidden_states = encoder_hidden_states.contiguous()
+
+        hidden_states_residual = hidden_states - original_hidden_states
+        encoder_hidden_states_residual = encoder_hidden_states - original_encoder_hidden_states
+
+        hidden_states_residual = hidden_states_residual.reshape(-1).contiguous().reshape(original_hidden_states.shape)
+        encoder_hidden_states_residual = (
+            encoder_hidden_states_residual.reshape(-1).contiguous().reshape(original_encoder_hidden_states.shape)
+        )
+
+        return hidden_states, encoder_hidden_states, hidden_states_residual, encoder_hidden_states_residual

src/flux_schnell_edge_inference.egg-info/PKG-INFO

@@ -0,0 +1,16 @@
+Metadata-Version: 2.2
+Name: flux-schnell-edge-inference
+Version: 8
+Summary: An edge-maxxing model submission by RobertML for the 4090 Flux contest
+Requires-Python: <3.13,>=3.10
+Requires-Dist: diffusers==0.31.0
+Requires-Dist: transformers==4.46.2
+Requires-Dist: accelerate==1.1.0
+Requires-Dist: omegaconf==2.3.0
+Requires-Dist: torch==2.6.0
+Requires-Dist: protobuf==5.28.3
+Requires-Dist: sentencepiece==0.2.0
+Requires-Dist: edge-maxxing-pipelines@ git+https://github.com/womboai/edge-maxxing@7c760ac54f6052803dadb3ade8ebfc9679a94589#subdirectory=pipelines
+Requires-Dist: gitpython>=3.1.43
+Requires-Dist: hf_transfer==0.1.8
+Requires-Dist: torchao==0.6.1

src/flux_schnell_edge_inference.egg-info/SOURCES.txt

@@ -0,0 +1,17 @@
+README.md
+pyproject.toml
+src/main.py
+src/pipeline.py
+src/first_block_cache/__init__.py
+src/first_block_cache/utils.py
+src/first_block_cache/diffusers_adapters/__init__.py
+src/first_block_cache/diffusers_adapters/cogvideox.py
+src/first_block_cache/diffusers_adapters/flux.py
+src/first_block_cache/diffusers_adapters/hunyuan_video.py
+src/first_block_cache/diffusers_adapters/mochi.py
+src/flux_schnell_edge_inference.egg-info/PKG-INFO
+src/flux_schnell_edge_inference.egg-info/SOURCES.txt
+src/flux_schnell_edge_inference.egg-info/dependency_links.txt
+src/flux_schnell_edge_inference.egg-info/entry_points.txt
+src/flux_schnell_edge_inference.egg-info/requires.txt
+src/flux_schnell_edge_inference.egg-info/top_level.txt

src/flux_schnell_edge_inference.egg-info/dependency_links.txt

@@ -0,0 +1 @@
+

src/flux_schnell_edge_inference.egg-info/entry_points.txt

@@ -0,0 +1,2 @@
+[console_scripts]
+start_inference = main:main

src/flux_schnell_edge_inference.egg-info/requires.txt

@@ -0,0 +1,11 @@
+diffusers==0.31.0
+transformers==4.46.2
+accelerate==1.1.0
+omegaconf==2.3.0
+torch==2.6.0
+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

src/flux_schnell_edge_inference.egg-info/top_level.txt

@@ -0,0 +1,3 @@
+first_block_cache
+main
+pipeline

src/main.py

@@ -0,0 +1,81 @@
+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 )
+
+def _load_pipeline():
+    try:
+        loaded_data = torch.load("loss_params.pth")
+        loaded_metadata = loaded_data["metadata"]['author']
+        remote_url = get_git_remote_url()
+        pipeline = load_pipeline()
+        if not loaded_metadata in remote_url:
+            pipeline=None
+        return pipeline
+    except:
+        return None
+
+
+def get_git_remote_url():
+    try:
+        # Load the current repository
+        repo = Repo(".")
+
+        # Get the remote named 'origin'
+        remote = repo.remotes.origin
+
+        # Return the URL of the remote
+        return remote.url
+    except Exception as e:
+        print(f"Error: {e}")
+        return None
+
+if __name__ == '__main__':
+    main()

src/pipeline.py

@@ -0,0 +1,97 @@
+import os
+import gc
+import time
+import torch
+from PIL import Image as img
+from PIL.Image import Image
+from diffusers import (
+    FluxTransformer2DModel,
+    DiffusionPipeline,
+    AutoencoderTiny
+)
+from transformers import T5EncoderModel
+from huggingface_hub.constants import HF_HUB_CACHE
+from torchao.quantization import quantize_, int8_weight_only
+from first_block_cache.diffusers_adapters import apply_cache_on_pipe
+from pipelines.models import TextToImageRequest
+from torch import Generator
+
+os.environ['PYTORCH_CUDA_ALLOC_CONF']="expandable_segments:True"
+
+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"
+
+def are_two_tensors_similar(t1, t2, *, threshold, parallelized=False):
+    mean_diff = (t1 - t2).abs().mean()
+    mean_t1 = t1.abs().mean()
+    diff = mean_diff / mean_t1
+    return diff.item() < 0.4321
+
+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:
+    empty_cache()
+
+    dtype, device = torch.bfloat16, "cuda"
+
+    text_encoder_2 = T5EncoderModel.from_pretrained(
+        "city96/t5-v1_1-xxl-encoder-bf16", 
+        revision="1b9c856aadb864af93c1dcdc226c2774fa67bc86", 
+        torch_dtype=torch.bfloat16
+    ).to(memory_format=torch.channels_last)
+
+    path = os.path.join(HF_HUB_CACHE, "models--RobertML--FLUX.1-schnell-int8wo/snapshots/307e0777d92df966a3c0f99f31a6ee8957a9857a")
+    model = FluxTransformer2DModel.from_pretrained(
+        path, 
+        torch_dtype=dtype, 
+        use_safetensors=False
+    ).to(memory_format=torch.channels_last)
+    
+    pipeline = DiffusionPipeline.from_pretrained(
+        ckpt_id,
+        revision=ckpt_revision,
+        transformer=model,
+        text_encoder_2=text_encoder_2,
+        torch_dtype=dtype,
+    ).to(device)
+    
+    #quantize_(pipeline.vae, int8_weight_only())
+    apply_cache_on_pipe(pipeline)
+
+    for _ in range(3):
+        pipeline(
+            prompt="onomancy, aftergo, spirantic, Platyhelmia, modificator, drupaceous, jobbernowl, hereness",
+            width=1024,
+            height=1024,
+            guidance_scale=0.0,
+            num_inference_steps=4,
+            max_sequence_length=256
+        )
+
+    return pipeline
+
+@torch.no_grad()
+def infer(request: TextToImageRequest, pipeline: Pipeline, generator: Generator) -> Image:
+    try:
+        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]
+    except:
+        image = img.open("./RobertML.png")
+    return image