flash attn

main.py

@@ -3,6 +3,7 @@ import os
 
 import torch
 import torch.optim as optim
+from torch.cuda.amp import autocast
 from torchvision.utils import make_grid
 from tqdm import tqdm
 
@@ -17,8 +18,8 @@ train_dataset, test_dataset = get_dataset()
 
 device = "cuda:1"
 
-batch_size = 8 * 4
+batch_size = 8 * 4 * 2
-accumulation_steps = 4
+accumulation_steps = 2
 total_epoch = 500
 
 steps_per_epoch = len(train_dataset) // batch_size
@@ -28,9 +29,11 @@ warmup_steps = int(0.05 * total_steps)
 grad_norm = 1.0
 
 
-model = UTransformer.from_pretrained_backbone(
+model = (
-    "facebook/dinov3-vitl16-pretrain-sat493m"
+    UTransformer.from_pretrained_backbone("facebook/dinov3-vitl16-pretrain-sat493m")
-).to(device)
+    .to(device)
+    .bfloat16()
+)
 rf = RF(model, "icfm", "lpips_mse")
 optimizer = optim.AdamW(model.parameters(), lr=3e-4)
 
@@ -81,6 +84,7 @@ for epoch in range(start_epoch, total_epoch):
         cloud = batch["cloud"].to(device)
         gt = batch["gt"].to(device)
 
+        with autocast(dtype=torch.bfloat16):
             loss, blsct, loss_list = rf.forward(gt, cloud)
             loss = loss / accumulation_steps
             loss.backward()

pyproject.toml

@@ -22,4 +22,11 @@ dependencies = [
     "tqdm>=4.67.1",
     "transformers>=4.56.2",
     "wandb[media]>=0.22.0",
+    "flash-attn"
 ]
+
+[tool.uv.extra-build-dependencies]
+flash-attn = [{ requirement = "torch", match-runtime = true }]
+
+[tool.uv.extra-build-variables]
+flash-attn = { FLASH_ATTENTION_SKIP_CUDA_BUILD = "TRUE" }

src/model/dino.py

@@ -5,6 +5,7 @@ import numpy as np
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
+from flash_attn import flash_attn_func
 from transformers.models.dinov3_vit.configuration_dinov3_vit import DINOv3ViTConfig
 
 
@@ -194,53 +195,54 @@ class DINOv3ViTAttention(nn.Module):
     def forward(
         self,
         hidden_states: torch.Tensor,
-        attention_mask: Optional[torch.Tensor] = None,
+        other: torch.Tensor | None = None,
+        attention_mask: Optional[torch.Tensor] = None,  # wont work rn
         position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
     ) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
-        assert position_embeddings is not None
+        # assert position_embeddings is not None
 
         batch_size, patches, _ = hidden_states.size()
 
         query_states = self.q_proj(hidden_states)
-        key_states = self.k_proj(hidden_states)
+        key_states = self.k_proj(hidden_states) if other is None else self.k_proj(other)
-        value_states = self.v_proj(hidden_states)
+        value_states = (
+            self.v_proj(hidden_states) if other is None else self.v_proj(other)
+        )
 
         query_states = query_states.view(
-            batch_size, patches, self.num_heads, self.head_dim
+            batch_size, -1, self.num_heads, self.head_dim
         ).transpose(1, 2)
         key_states = key_states.view(
-            batch_size, patches, self.num_heads, self.head_dim
+            batch_size, -1, self.num_heads, self.head_dim
         ).transpose(1, 2)
         value_states = value_states.view(
-            batch_size, patches, self.num_heads, self.head_dim
+            batch_size, -1, self.num_heads, self.head_dim
         ).transpose(1, 2)
 
+        if position_embeddings is not None:
             cos, sin = position_embeddings
             query_states, key_states = apply_rotary_pos_emb(
                 query_states, key_states, cos, sin
             )
 
-        attn_weights = (
+        query_states = query_states.transpose(1, 2)
-            torch.matmul(query_states, key_states.transpose(-1, -2)) * self.scaling
+        key_states = key_states.transpose(1, 2)
-        )
+        value_states = value_states.transpose(1, 2)
-        attn_weights = F.softmax(attn_weights, dim=-1, dtype=torch.float32).to(
+        dropout_p = self.dropout if self.training else 0.0
-            query_states.dtype
+
+        attn_output = flash_attn_func(
+            query_states,
+            key_states,
+            value_states,
+            dropout_p=dropout_p,
+            softmax_scale=None,  # Will use default 1/sqrt(headdim)
+            causal=False,
         )
 
-        if self.training:
+        attn_output = attn_output.reshape(batch_size, patches, -1).contiguous()  # type: ignore
-            attn_weights = F.dropout(
-                attn_weights, p=self.dropout, training=self.training
-            )
-
-        if attention_mask is not None:
-            attn_weights = attn_weights * attention_mask
-
-        attn_output = torch.matmul(attn_weights, value_states)
-        attn_output = attn_output.transpose(1, 2).contiguous()
-        attn_output = attn_output.reshape(batch_size, patches, -1).contiguous()
         attn_output = self.o_proj(attn_output)
 
-        return attn_output, attn_weights
+        return attn_output
 
 
 class DINOv3ViTLayerScale(nn.Module):
@@ -350,7 +352,7 @@ class DINOv3ViTLayer(nn.Module):
 
         residual = hidden_states
         hidden_states = self.norm1(hidden_states)
-        hidden_states, _ = self.attention(
+        hidden_states = self.attention(
             hidden_states,
             attention_mask=attention_mask,
             position_embeddings=position_embeddings,

src/model/utransformer.py

@@ -8,6 +8,7 @@ from torch import nn
 from transformers.models.dinov3_vit.configuration_dinov3_vit import DINOv3ViTConfig
 
 from src.model.dino import (
+    DINOv3ViTAttention,
     DINOv3ViTEmbeddings,
     DINOv3ViTLayer,
     DINOv3ViTLayerScale,
@@ -81,12 +82,7 @@ class DinoConditionedLayer(DINOv3ViTLayer):
         self.is_encoder = is_encoder
 
         self.norm_cond = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
-        self.cond = nn.MultiheadAttention(
+        self.cond = DINOv3ViTAttention(config)
-            config.hidden_size,
-            config.num_attention_heads,
-            config.drop_path_rate,
-            batch_first=True,
-        )
         self.layer_scale_cond = DINOv3ViTLayerScale(config)
 
         # no init zeros!
@@ -114,7 +110,7 @@ class DinoConditionedLayer(DINOv3ViTLayer):
 
         residual = hidden_states
         hidden_states = self.norm1(hidden_states)
-        hidden_states, _ = self.attention(
+        hidden_states = self.attention(
             hidden_states,
             attention_mask=attention_mask,
             position_embeddings=position_embeddings,
@@ -125,8 +121,9 @@ class DinoConditionedLayer(DINOv3ViTLayer):
         if do_condition:
             residual = hidden_states
             hidden_states = self.norm_cond(hidden_states)
-            hidden_states, _ = self.cond(
+            hidden_states = self.cond(
-                hidden_states, conditioning_input, conditioning_input
+                hidden_states,
+                conditioning_input,
             )
             hidden_states = self.layer_scale_cond(hidden_states)
             hidden_states = self.drop_path(hidden_states) + residual
@@ -188,9 +185,7 @@ class DinoV3ViTDecoder(nn.Module):
         self.patch_size = config.patch_size
 
         self.projection = nn.Linear(
-            config.hidden_size,
+            config.hidden_size, config.num_channels * (self.patch_size**2), bias=True
-            config.num_channels * (self.patch_size**2),
-            bias=True,
         )
         self.pixel_shuffle = nn.PixelShuffle(self.patch_size)
 
@@ -209,11 +204,8 @@ class DinoV3ViTDecoder(nn.Module):
         w_grid = image_size[1] // p
 
         assert x.shape[1] == h_grid * w_grid
-
         x = self.projection(x)
-
         x = x.reshape(batch_size, h_grid, w_grid, -1).permute(0, 3, 1, 2)
-
         x = self.pixel_shuffle(x)
 
         return x

uv.lock

@@ -297,6 +297,7 @@ source = { virtual = "." }
 dependencies = [
     { name = "datasets" },
     { name = "einops" },
+    { name = "flash-attn" },
     { name = "lpips" },
     { name = "pyright" },
     { name = "python-lsp-server" },
@@ -318,6 +319,7 @@ dependencies = [
 requires-dist = [
     { name = "datasets", specifier = ">=4.1.1" },
     { name = "einops", specifier = ">=0.8.1" },
+    { name = "flash-attn" },
     { name = "lpips", specifier = ">=0.1.4" },
     { name = "pyright", specifier = ">=1.1.405" },
     { name = "python-lsp-server", specifier = ">=1.13.1" },
@@ -531,6 +533,16 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/42/14/42b2651a2f46b022ccd948bca9f2d5af0fd8929c4eec235b8d6d844fbe67/filelock-3.19.1-py3-none-any.whl", hash = "sha256:d38e30481def20772f5baf097c122c3babc4fcdb7e14e57049eb9d88c6dc017d", size = 15988, upload-time = "2025-08-14T16:56:01.633Z" },
 ]
 
+[[package]]
+name = "flash-attn"
+version = "2.8.3"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "einops" },
+    { name = "torch" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/3b/b2/8d76c41ad7974ee264754709c22963447f7f8134613fd9ce80984ed0dab7/flash_attn-2.8.3.tar.gz", hash = "sha256:1e71dd64a9e0280e0447b8a0c2541bad4bf6ac65bdeaa2f90e51a9e57de0370d", size = 8447812, upload-time = "2025-08-15T08:28:12.911Z" }
+
 [[package]]
 name = "fonttools"
 version = "4.60.1"