TRT decoder

flux/math.py

@@ -1,5 +1,5 @@
 import torch
-from einops import rearrange
+#from einops import rearrange
 from torch import Tensor
 from torch.nn.attention import SDPBackend, sdpa_kernel
 
@@ -10,19 +10,28 @@ def attention(q: Tensor, k: Tensor, v: Tensor, pe: Tensor) -> Tensor:
     # Only enable flash attention backend
     with sdpa_kernel(SDPBackend.FLASH_ATTENTION):
         x = torch.nn.functional.scaled_dot_product_attention(q, k, v)
-    x = rearrange(x, "B H L D -> B L (H D)")
-
+    # x = rearrange(x, "B H L D -> B L (H D)")
+    x = x.transpose(1, 2).contiguous().view(x.size(0), x.size(2), -1)
     return x
 
 
 def rope(pos: Tensor, dim: int, theta: int) -> Tensor:
     assert dim % 2 == 0
-    scale = torch.arange(0, dim, 2, dtype=torch.float64, device=pos.device) / dim
+    # f64 is problematic
+    # https://github.com/pytorch/TensorRT/blob/v2.4.0/py/torch_tensorrt/dynamo/conversion/converter_utils.py#L380
+    scale = torch.arange(0, dim, 2, dtype=torch.float32, device=pos.device) / dim
+    # scale = torch.arange(0, dim, 2, dtype=torch.float64, device=pos.device) / dim
     omega = 1.0 / (theta**scale)
-    out = torch.einsum("...n,d->...nd", pos, omega)
-    out = torch.stack([torch.cos(out), -torch.sin(out), torch.sin(out), torch.cos(out)], dim=-1)
-    out = rearrange(out, "b n d (i j) -> b n d i j", i=2, j=2)
-    return out.float()
+    # out = torch.einsum("...n,d->...nd", pos, omega)
+    out = pos.unsqueeze(-1) * omega
+    # out = torch.stack([torch.cos(out), -torch.sin(out), torch.sin(out), torch.cos(out)], dim=-1)
+    cos_out = torch.cos(out)
+    sin_out = torch.sin(out)
+    out = torch.stack([cos_out, -sin_out, sin_out, cos_out], dim=-1)
+    # out = rearrange(out, "b n d (i j) -> b n d i j", i=2, j=2)
+    # Reshaping the tensor to (..., n, d, 2, 2)
+    out = out.view(*out.shape[:-1], 2, 2)
+    return out # .float()
 
 
 def apply_rope(xq: Tensor, xk: Tensor, freqs_cis: Tensor) -> tuple[Tensor, Tensor]:

flux/model.py

@@ -81,7 +81,7 @@ def __init__(self, params: FluxParams):
 
     def forward(
         self,
-        img: Tensor,
+        img: Tensor, # (bs, dynamic, 64)
         img_ids: Tensor,
         txt: Tensor,
         txt_ids: Tensor,
@@ -93,7 +93,7 @@ def forward(
             raise ValueError("Input img and txt tensors must have 3 dimensions.")
 
         # running on sequences img
-        img = self.img_in(img)
+        img = self.img_in(img) # (bs, dynamic, hidden_size)
         vec = self.time_in(timestep_embedding(timesteps, 256))
         if self.params.guidance_embed:
             if guidance is None:

flux/modules/autoencoder.py

@@ -105,6 +105,16 @@ def forward(self, x: Tensor):
         x = self.conv(x)
         return x
 
+class DownBlock(nn.Module):
+    def __init__(self, block: list, downsample: nn.Module) -> None:
+        super().__init__()
+        # we're doing this instead of a flat nn.Sequential to preserve the keys "block" "downsample"
+        self.block = nn.Sequential(*block)
+        self.downsample = downsample
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.downsample(self.block(x))
+
 
 class Encoder(nn.Module):
     def __init__(
@@ -128,23 +138,25 @@ def __init__(
         curr_res = resolution
         in_ch_mult = (1,) + tuple(ch_mult)
         self.in_ch_mult = in_ch_mult
-        self.down = nn.ModuleList()
+        down_layers = []
         block_in = self.ch
+        # ideally, this would all append to a single flat nn.Sequential
+        # we cannot do this due to the existing state dict keys
         for i_level in range(self.num_resolutions):
-            block = nn.ModuleList()
-            attn = nn.ModuleList()
             block_in = ch * in_ch_mult[i_level]
             block_out = ch * ch_mult[i_level]
+            block_layers = []
             for _ in range(self.num_res_blocks):
-                block.append(ResnetBlock(in_channels=block_in, out_channels=block_out))
-                block_in = block_out
-            down = nn.Module()
-            down.block = block
-            down.attn = attn
+                block_layers.append(ResnetBlock(in_channels=block_in, out_channels=block_out))
+                block_in = block_out # ?
+            # originally this provided for attn layers, but those are never actually created
             if i_level != self.num_resolutions - 1:
-                down.downsample = Downsample(block_in)
+                downsample = Downsample(block_in)
                 curr_res = curr_res // 2
-            self.down.append(down)
+            else:
+                downsample = nn.Identity()
+            down_layers.append(DownBlock(block_layers, downsample))
+        self.down = nn.Sequential(*down_layers)
 
         # middle
         self.mid = nn.Module()
@@ -158,18 +170,10 @@ def __init__(
 
     def forward(self, x: Tensor) -> Tensor:
         # downsampling
-        hs = [self.conv_in(x)]
-        for i_level in range(self.num_resolutions):
-            for i_block in range(self.num_res_blocks):
-                h = self.down[i_level].block[i_block](hs[-1])
-                if len(self.down[i_level].attn) > 0:
-                    h = self.down[i_level].attn[i_block](h)
-                hs.append(h)
-            if i_level != self.num_resolutions - 1:
-                hs.append(self.down[i_level].downsample(hs[-1]))
+        h = self.conv_in(h)
+        h = self.down(h)
 
         # middle
-        h = hs[-1]
         h = self.mid.block_1(h)
         h = self.mid.attn_1(h)
         h = self.mid.block_2(h)
@@ -179,6 +183,15 @@ def forward(self, x: Tensor) -> Tensor:
         h = self.conv_out(h)
         return h
 
+class UpBlock(nn.Module):
+    def __init__(self, block: list, upsample: nn.Module) -> None:
+        super().__init__()
+        self.block = nn.Sequential(*block)
+        self.upsample = upsample
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.upsample(self.block(x))
+
 
 class Decoder(nn.Module):
     def __init__(
@@ -214,26 +227,37 @@ def __init__(
         self.mid.block_2 = ResnetBlock(in_channels=block_in, out_channels=block_in)
 
         # upsampling
-        self.up = nn.ModuleList()
+        up_blocks = []
+        # 3, 2, 1, 0, descending order
         for i_level in reversed(range(self.num_resolutions)):
-            block = nn.ModuleList()
-            attn = nn.ModuleList()
+            level_blocks = []
             block_out = ch * ch_mult[i_level]
             for _ in range(self.num_res_blocks + 1):
-                block.append(ResnetBlock(in_channels=block_in, out_channels=block_out))
+                level_blocks.append(ResnetBlock(in_channels=block_in, out_channels=block_out))
                 block_in = block_out
-            up = nn.Module()
-            up.block = block
-            up.attn = attn
             if i_level != 0:
-                up.upsample = Upsample(block_in)
+                upsample = Upsample(block_in)
                 curr_res = curr_res * 2
-            self.up.insert(0, up)  # prepend to get consistent order
+            else:
+                upsample = nn.Identity()
+            # 0, 1, 2, 3, ascending order
+            up_blocks.insert(0, UpBlock(level_blocks, upsample)) # prepend to get consistent order
+        self.up = nn.Sequential(*up_blocks)
 
         # end
         self.norm_out = nn.GroupNorm(num_groups=32, num_channels=block_in, eps=1e-6, affine=True)
         self.conv_out = nn.Conv2d(block_in, out_ch, kernel_size=3, stride=1, padding=1)
 
+    # this is a hack to get something like property but only evaluate it once
+    # we're doing it like this so that up_descending isn't in the state_dict keys
+    # without adding anything conditional to the main flow
+    def __getattr__(self, name):
+        if name == "up_descending":
+            self.up_descending = nn.Sequential(*reversed(self.up))
+            Decoder.__getattr__ = nn.Module.__getattr__
+            return self.up_descending
+        return super().__getattr__(name)
+
     def forward(self, z: Tensor) -> Tensor:
         # z to block_in
         h = self.conv_in(z)
@@ -244,13 +268,7 @@ def forward(self, z: Tensor) -> Tensor:
         h = self.mid.block_2(h)
 
         # upsampling
-        for i_level in reversed(range(self.num_resolutions)):
-            for i_block in range(self.num_res_blocks + 1):
-                h = self.up[i_level].block[i_block](h)
-                if len(self.up[i_level].attn) > 0:
-                    h = self.up[i_level].attn[i_block](h)
-            if i_level != 0:
-                h = self.up[i_level].upsample(h)
+        h = self.up_descending(h)
 
         # end
         h = self.norm_out(h)

flux/modules/layers.py

@@ -25,7 +25,7 @@ def forward(self, ids: Tensor) -> Tensor:
         return emb.unsqueeze(1)
 
 
-def timestep_embedding(t: Tensor, dim, max_period=10000, time_factor: float = 1000.0):
+def timestep_embedding(t: Tensor, dim, max_period=10000):
     """
     Create sinusoidal timestep embeddings.
     :param t: a 1-D Tensor of N indices, one per batch element.
@@ -34,6 +34,7 @@ def timestep_embedding(t: Tensor, dim, max_period=10000, time_factor: float = 10
     :param max_period: controls the minimum frequency of the embeddings.
     :return: an (N, D) Tensor of positional embeddings.
     """
+    time_factor = torch.tensor(1000.0)
     t = time_factor * t
     half = dim // 2
     freqs = torch.exp(-math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half).to(

flux/sampling.py

@@ -30,7 +30,9 @@ def get_noise(
     )
 
 
-def prepare(t5: HFEmbedder, clip: HFEmbedder, img: Tensor, prompt: str | list[str]) -> dict[str, Tensor]:
+def prepare(
+    t5: HFEmbedder, clip: HFEmbedder, img: Tensor, prompt: str | list[str]
+) -> dict[str, Tensor]:
     bs, c, h, w = img.shape
     if bs == 1 and not isinstance(prompt, str):
         bs = len(prompt)
@@ -104,7 +106,7 @@ def denoise_single_item(
     vec: Tensor,
     timesteps: list[float],
     guidance: float = 4.0,
-    compile_run: bool = False
+    compile_run: bool = False,
 ):
     img = img.unsqueeze(0)
     img_ids = img_ids.unsqueeze(0)
@@ -127,15 +129,16 @@ def denoise_single_item(
             img_ids=img_ids,
             txt=txt,
             txt_ids=txt_ids,
-            y=vec,
             timesteps=t_vec,
+            y=vec,
             guidance=guidance_vec,
         )
 
         img = img + (t_prev - t_curr) * pred.squeeze(0)
 
     return img, model
 
+
 def denoise(
     model: Flux,
     # model input
@@ -147,7 +150,7 @@ def denoise(
     # sampling parameters
     timesteps: list[float],
     guidance: float = 4.0,
-    compile_run: bool = False
+    compile_run: bool = False,
 ):
     batch_size = img.shape[0]
     output_imgs = []
@@ -162,13 +165,14 @@ def denoise(
             vec[i],
             timesteps,
             guidance,
-            compile_run
+            compile_run,
         )
         compile_run = False
         output_imgs.append(denoised_img)
 
     return torch.cat(output_imgs), model
 
+
 def unpack(x: Tensor, height: int, width: int) -> Tensor:
     return rearrange(
         x,
@@ -177,4 +181,4 @@ def unpack(x: Tensor, height: int, width: int) -> Tensor:
         w=math.ceil(width / 16),
         ph=2,
         pw=2,
-    )
+   )

predict.py

@@ -1,4 +1,7 @@
+import torch_tensorrt
 import os
+#os.environ["TORCH_LOGS"] = "+dynamic"
+#os.environ["TORCH_COMPILE_DEBUG"] = "1"
 import time
 from typing import Any, Dict, Optional
 
@@ -44,6 +47,9 @@
     "https://weights.replicate.delivery/default/falconai/nsfw-image-detection.tar"
 )
 
+DECODER_URL = "https://weights.replicate.delivery/default/official-models/flux/ae/decoder.engine"
+DECODER_PATH = "model-cache/ae/decoder.engine"
+
 # Suppress diffusers nsfw warnings
 logging.getLogger("diffusers").setLevel(logging.CRITICAL)
 logging.getLogger("transformers").setLevel(logging.CRITICAL)
@@ -149,15 +155,37 @@ def base_setup(
 
         device = "cuda"
         max_length = 256 if self.flow_model_name == "flux-schnell" else 512
+        # we still need to load the encoder but it would be better to avoid loading the decoder twice
+        self.ae = load_ae(self.flow_model_name, device="cpu" if self.offload else device)
+        if not os.getenv("COMPILE_ENGINE"):
+            if not os.path.exists(DECODER_PATH):
+                download_weights(DECODER_URL, DECODER_PATH)
+            t = time.time()
+            self.ae.decoder = torch.export.load(DECODER_PATH).module()
+            print(f"loading decoder took {time.time() - t:.3f}s")
+        else:
+            #inputs = [torch.randn([1, 3, 1024, 1024]) # enc/dec
+            t = time.time()
+            inputs = [torch.randn([1, 16, 128, 128], device="cuda")] # dec
+            self.ae.decoder.up_descending # access
+            self.orig_decoder = self.ae.decoder
+            base = {"truncate_long_and_double": True}
+            best = {
+                "num_avg_timing_iters": 2,
+                "use_fast_partitioner": False,
+                "optimization_level": 5,
+            }
+            dec = torch_tensorrt.compile(self.ae.decoder, inputs=inputs, options=base | best)
+            torch_tensorrt.save(dec, "decoder.engine", inputs=inputs)
+            print("compiling and saving decoder took", time.time()-t)
+            self.ae.decoder = dec
+
         self.t5 = load_t5(device, max_length=max_length)
         self.clip = load_clip(device)
         self.flux = load_flow_model(
             self.flow_model_name, device="cpu" if self.offload else device
         )
         self.flux = self.flux.eval()
-        self.ae = load_ae(
-            self.flow_model_name, device="cpu" if self.offload else device
-        )
 
         self.num_steps = 4 if self.flow_model_name == "flux-schnell" else 28
         self.shift = self.flow_model_name != "flux-schnell"