support bfloat16

README.md

@@ -12,7 +12,6 @@ To generate a 4x4 grid of DALL·E Mega images it takes:
 - 89 sec with a T4 in Colab
 - 48 sec with a P100 in Colab
 - 14 sec with an A100 on Replicate
-- TBD with an H100 (@NVIDIA?)
 
 The flax model and code for converting it to torch can be found [here](https://github.com/kuprel/min-dalle-flax).
 
@@ -30,13 +29,14 @@ Load the model parameters once and reuse the model to generate multiple images.
 from min_dalle import MinDalle
 
 model = MinDalle(
+    models_root='./pretrained',
+    dtype=torch.float32,
     is_mega=True, 
-    is_reusable=True,
-    models_root='./pretrained'
+    is_reusable=True
 )
 ```
 
-The required models will be downloaded to `models_root` if they are not already there.  Once everything has finished initializing, call `generate_image` with some text as many times as you want.  Use a positive `seed` for reproducible results.  Higher values for `log2_supercondition_factor` result in better agreement with the text but a narrower variety of generated images.  Every image token is sampled from the top $k$ most probable tokens.
+The required models will be downloaded to `models_root` if they are not already there.  If you have an Ampere architecture GPU you can set the `dtype=torch.bfloat16` and save GPU memory.  There is still an issue with `dtype=torch.float16` that needs to be sorted out.  Once everything has finished initializing, call `generate_image` with some text as many times as you want.  Use a positive `seed` for reproducible results.  Higher values for `log2_supercondition_factor` result in better agreement with the text but a narrower variety of generated images.  Every image token is sampled from the top-$k$ most probable tokens.
 
 ```python
 image = model.generate_image(

README.rst

@@ -1,16 +1,16 @@
 min(DALL·E)
 ===========
 
-|Open In Colab|   |Replicate|   |Join us on Discord|
+|Colab|   |Replicate|   |Discord|
 
-This is a fast, minimal implementation of Boris Dayma’s `DALL·E
+This is a fast, minimal port of Boris Dayma’s `DALL·E
 Mega <https://github.com/borisdayma/dalle-mini>`__. It has been stripped
 down for inference and converted to PyTorch. The only third party
 dependencies are numpy, requests, pillow and torch.
 
 To generate a 4x4 grid of DALL·E Mega images it takes: - 89 sec with a
 T4 in Colab - 48 sec with a P100 in Colab - 14 sec with an A100 on
-Replicate - TBD with an H100 (@NVIDIA?)
+Replicate
 
 The flax model and code for converting it to torch can be found
 `here <https://github.com/kuprel/min-dalle-flax>`__.
@@ -32,47 +32,58 @@ images.
 
    from min_dalle import MinDalle
 
-   model = MinDalle(is_mega=True, models_root='./pretrained')
+   model = MinDalle(
+       is_mega=True, 
+       is_reusable=True,
+       models_root='./pretrained'
+   )
 
 The required models will be downloaded to ``models_root`` if they are
 not already there. Once everything has finished initializing, call
-``generate_image`` with some text and a seed as many times as you want.
+``generate_image`` with some text as many times as you want. Use a
+positive ``seed`` for reproducible results. Higher values for
+``log2_supercondition_factor`` result in better agreement with the text
+but a narrower variety of generated images. Every image token is sampled
+from the top-:math:`k` most probable tokens.
 
 .. code:: python
 
-   text = 'Dali painting of WALL·E'
-   image = model.generate_image(text, seed=0, grid_size=4)
-   display(image)
-
-.. code:: python
-
-   text = 'Rusty Iron Man suit found abandoned in the woods being reclaimed by nature'
-   image = model.generate_image(text, seed=0, grid_size=3)
-   display(image)
-
-.. code:: python
+   image = model.generate_image(
+       text='Nuclear explosion broccoli',
+       seed=-1,
+       grid_size=4,
+       log2_k=6,
+       log2_supercondition_factor=5,
+       is_verbose=False
+   )
 
-   text = 'court sketch of godzilla on trial'
-   image = model.generate_image(text, seed=6, grid_size=3)
    display(image)
 
-.. code:: python
+Interactive
+~~~~~~~~~~~
 
-   text = 'a funeral at Whole Foods'
-   image = model.generate_image(text, seed=10, grid_size=3)
-   display(image)
+If the model is being used interactively (e.g. in a notebook)
+``generate_image_stream`` can be used to generate a stream of images as
+the model is decoding. The detokenizer adds a slight delay for each
+image. Setting ``log2_mid_count`` to 3 results in a total of
+``2 ** 3 = 8`` generated images. The only valid values for
+``log2_mid_count`` are 0, 1, 2, 3, and 4. This is implemented in the
+colab.
 
 .. code:: python
 
-   text = 'Jesus turning water into wine on Americas Got Talent'
-   image = model.generate_image(text, seed=2, grid_size=3)
-   display(image)
-
-.. code:: python
+   image_stream = model.generate_image_stream(
+       text='Dali painting of WALL·E',
+       seed=-1,
+       grid_size=3,
+       log2_mid_count=3,
+       log2_k=6,
+       log2_supercondition_factor=3,
+       is_verbose=False
+   )
 
-   text = 'cctv footage of Yoda robbing a liquor store'
-   image = model.generate_image(text, seed=0, grid_size=3)
-   display(image)
+   for image in image_stream:
+       display(image)
 
 Command Line
 ~~~~~~~~~~~~
@@ -81,15 +92,11 @@ Use ``image_from_text.py`` to generate images from the command line.
 
 .. code:: bash
 
-   $ python image_from_text.py --text='artificial intelligence' --no-mega --seed=7
-
-.. code:: bash
-
-   $ python image_from_text.py --text='trail cam footage of gollum eating watermelon' --mega --seed=1 --grid-size=3
+   $ python image_from_text.py --text='artificial intelligence' --no-mega
 
-.. |Open In Colab| image:: https://colab.research.google.com/assets/colab-badge.svg
+.. |Colab| image:: https://colab.research.google.com/assets/colab-badge.svg
    :target: https://colab.research.google.com/github/kuprel/min-dalle/blob/main/min_dalle.ipynb
 .. |Replicate| image:: https://replicate.com/kuprel/min-dalle/badge
    :target: https://replicate.com/kuprel/min-dalle
-.. |Join us on Discord| image:: https://img.shields.io/discord/823813159592001537?color=5865F2&logo=discord&logoColor=white
-   :target: https://discord.gg/xBPBXfcFHd
+.. |Discord| image:: https://img.shields.io/discord/823813159592001537?color=5865F2&logo=discord&logoColor=white
+   :target: https://discord.com/channels/823813159592001537/912729332311556136

cog.yaml

@@ -6,8 +6,8 @@ build:
     - "libgl1-mesa-glx"
     - "libglib2.0-0"
   python_packages:
-    - "min-dalle==0.2.29"
+    - "min-dalle==0.2.35"
   run:
-    - pip install torch==1.10.0+cu113 -f https://download.pytorch.org/whl/torch_stable.html
+    - pip install torch==1.11.0+cu113 -f https://download.pytorch.org/whl/torch_stable.html
 
 predict: "replicate_predictor.py:ReplicatePredictor"

min_dalle/min_dalle.py

@@ -18,13 +18,15 @@ MIN_DALLE_REPO = 'https://huggingface.co/kuprel/min-dalle/resolve/main/'
 class MinDalle:
     def __init__(
         self,
-        is_mega: bool, 
-        is_reusable: bool = True,
         models_root: str = 'pretrained',
+        dtype: torch.dtype = torch.float32,
+        is_mega: bool = True, 
+        is_reusable: bool = True,
         is_verbose = True
     ):
         self.is_mega = is_mega
         self.is_reusable = is_reusable
+        self.dtype = dtype
         self.is_verbose = is_verbose
         self.text_token_count = 64
         self.layer_count = 24 if is_mega else 12
@@ -34,7 +36,6 @@ class MinDalle:
         self.text_vocab_count = 50272 if is_mega else 50264
         self.image_vocab_count = 16415 if is_mega else 16384
 
-        if self.is_verbose: print("initializing MinDalle")
         model_name = 'dalle_bart_{}'.format('mega' if is_mega else 'mini')
         dalle_path = os.path.join(models_root, model_name)
         vqgan_path = os.path.join(models_root, 'vqgan')
@@ -105,7 +106,7 @@ class MinDalle:
             text_token_count = self.text_token_count,
             text_vocab_count = self.text_vocab_count,
             layer_count = self.layer_count
-        )
+        ).to(self.dtype).eval()
         params = torch.load(self.encoder_params_path)
         self.encoder.load_state_dict(params, strict=False)
         del params
@@ -123,7 +124,7 @@ class MinDalle:
             glu_embed_count = self.glu_embed_count,
             layer_count = self.layer_count,
             start_token = self.image_vocab_count
-        )
+        ).to(self.dtype).eval()
         params = torch.load(self.decoder_params_path)
         self.decoder.load_state_dict(params, strict=False)
         del params
@@ -134,7 +135,7 @@ class MinDalle:
         is_downloaded = os.path.exists(self.detoker_params_path)
         if not is_downloaded: self.download_detokenizer()
         if self.is_verbose: print("initializing VQGanDetokenizer")
-        self.detokenizer = VQGanDetokenizer()
+        self.detokenizer = VQGanDetokenizer().to(self.dtype).eval()
         params = torch.load(self.detoker_params_path)
         self.detokenizer.load_state_dict(params)
         del params
@@ -184,38 +185,41 @@ class MinDalle:
 
         if not self.is_reusable: self.init_encoder()
         if is_verbose: print("encoding text tokens")
-        encoder_state = self.encoder.forward(text_tokens)
+        with torch.cuda.amp.autocast(dtype=self.dtype):
+            encoder_state = self.encoder.forward(text_tokens)
         if not self.is_reusable: del self.encoder
         if torch.cuda.is_available(): torch.cuda.empty_cache()
 
         if not self.is_reusable: self.init_decoder()
 
-        encoder_state, attention_mask, attention_state, image_tokens = ( 
-            self.decoder.decode_initial(
-                seed, 
-                grid_size ** 2, 
-                text_tokens, 
-                encoder_state
+        with torch.cuda.amp.autocast(dtype=self.dtype):
+            encoder_state, attention_mask, attention_state, image_tokens = ( 
+                self.decoder.decode_initial(
+                    seed, 
+                    grid_size ** 2, 
+                    text_tokens, 
+                    encoder_state
+                )
             )
-        )
 
         row_count = 16
         for row_index in range(row_count):
             if is_verbose: 
                 print('sampling row {} of {}'.format(row_index + 1, row_count))
-            attention_state, image_tokens = self.decoder.decode_row(
-                row_index,
-                log2_k,
-                log2_supercondition_factor,
-                encoder_state,
-                attention_mask,
-                attention_state,
-                image_tokens
-            )
-            if ((row_index + 1) * (2 ** log2_mid_count)) % row_count == 0:
-                tokens = image_tokens[:, 1:]
-                image = self.image_from_tokens(grid_size, tokens, is_verbose)
-                yield image
+            with torch.cuda.amp.autocast(dtype=self.dtype):
+                attention_state, image_tokens = self.decoder.decode_row(
+                    row_index,
+                    log2_k,
+                    log2_supercondition_factor,
+                    encoder_state,
+                    attention_mask,
+                    attention_state,
+                    image_tokens
+                )
+                if ((row_index + 1) * (2 ** log2_mid_count)) % row_count == 0:
+                    tokens = image_tokens[:, 1:]
+                    image = self.image_from_tokens(grid_size, tokens, is_verbose)
+                    yield image
 
 
     def generate_image(

min_dalle/models/dalle_bart_decoder.py

@@ -40,7 +40,8 @@ class DecoderSelfAttention(AttentionBase):
         queries = self.q_proj.forward(decoder_state)
         attn_mask = self.token_indices < token_index + 1
         attn_mask = attn_mask[None][[0] * decoder_state.shape[0]]
-        attention_state[:, token_index] = torch.cat([keys, values])
+        attn_state_new = torch.cat([keys, values]).to(attention_state.dtype)
+        attention_state[:, token_index] = attn_state_new
         batch_count = decoder_state.shape[0]
         keys = attention_state[:batch_count]
         values = attention_state[batch_count:]

min_dalle/models/vqgan_detokenizer.py

@@ -82,7 +82,7 @@ class Upsample(Module):
         self.conv = Conv2d(n, n, 3, padding=1)
 
     def forward(self, x: Tensor) -> Tensor:
-        x = self.upsample.forward(x)
+        x = self.upsample.forward(x.to(torch.float32))
         x = self.conv.forward(x)
         return x
 

replicate_predictor.py

@@ -19,9 +19,9 @@ class ReplicatePredictor(BasePredictor):
             default=True
         ),
         grid_size: int = Input(
-            description='Size of the image grid',
+            description='Size of the image grid.  4x4 takes about 15 seconds, 8x8 takes about 45 seconds',
             ge=1,
-            le=4,
+            le=8,
             default=4
         ),
         log2_supercondition_factor: int = Input(

setup.py

@@ -4,8 +4,8 @@ from pathlib import Path
 setuptools.setup(
     name='min-dalle',
     description = 'min(DALL·E)',
-    long_description=(Path(__file__).parent / "README.rst").read_text(),
-    version='0.2.29',
+    # long_description=(Path(__file__).parent / "README.rst").read_text(),
+    version='0.2.35',
     author='Brett Kuprel',
     author_email='brkuprel@gmail.com',
     url='https://github.com/kuprel/min-dalle',
@@ -15,8 +15,8 @@ setuptools.setup(
     ],
     license='MIT',
     install_requires=[
-        'torch>=1.10.0',
-        'typing_extensions>=4.1.0',
+        'torch>=1.11',
+        'typing_extensions>=4.1',
         'numpy>=1.21',
         'pillow>=7.1',
         'requests>=2.23'