feat: implement transformer variants (#144)

* added DeepNet
* added Swin v2
* added NormFormer
* added RMSNorm
* added GLU variants

.gitignore

@@ -3,3 +3,4 @@ __pycache__
 .streamlit
 wandb/
 *.egg-info/
+jax_cache/

README.md

@@ -94,26 +94,43 @@ Many thanks to the people who helped make it better:
 
 - the [DALLE-Pytorch](https://discord.gg/xBPBXfcFHd) and [EleutherAI](https://www.eleuther.ai/) communities for testing and exchanging cool ideas
 - [Rohan Anil](https://github.com/rohan-anil) for adding Distributed Shampoo optimizer
+- [Phil Wang](https://github.com/lucidrains) has provided a lot of cool implementations of transformer variants and gives interesting insights with [x-transformers](https://github.com/lucidrains/x-transformers)
 - [Katherine Crowson](https://github.com/crowsonkb) for [super conditioning](https://twitter.com/RiversHaveWings/status/1478093658716966912)
 
 ## Citing DALL·E mini
 
 If you find DALL·E mini useful in your research or wish to refer, please use the following BibTeX entry.
 
-```
+```text
 @misc{Dayma_DALL·E_Mini_2021,
-author = {Dayma, Boris and Patil, Suraj and Cuenca, Pedro and Saifullah, Khalid and Abraham, Tanishq and Lê Khắc, Phúc and Melas, Luke and Ghosh, Ritobrata},
-doi = {10.5281/zenodo.5146400},
-month = {7},
-title = {DALL·E Mini},
-url = {https://github.com/borisdayma/dalle-mini},
-year = {2021}
+      author = {Dayma, Boris and Patil, Suraj and Cuenca, Pedro and Saifullah, Khalid and Abraham, Tanishq and Lê Khắc, Phúc and Melas, Luke and Ghosh, Ritobrata},
+      doi = {10.5281/zenodo.5146400},
+      month = {7},
+      title = {DALL·E Mini},
+      url = {https://github.com/borisdayma/dalle-mini},
+      year = {2021}
 }
 ```
 
 ## References
 
-```
+Original DALL·E from "[Zero-Shot Text-to-Image Generation](https://arxiv.org/abs/2102.12092)" with image quantization from "[Learning Transferable Visual Models From Natural Language Supervision](https://arxiv.org/abs/2103.00020)".
+
+Image encoder from "[Taming Transformers for High-Resolution Image Synthesis](https://arxiv.org/abs/2012.09841v2)".
+
+Sequence to sequence model based on "[BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension](https://arxiv.org/abs/1910.13461v1)" with implementation of a few variants:
+
+- "[GLU Variants Improve Transformer](https://arxiv.org/abs/2002.05202)"
+- "[Deepnet: Scaling Transformers to 1,000 Layers](https://arxiv.org/abs/2203.00555)"
+- "[NormFormer: Improved Transformer Pretraining with Extra Normalization](https://arxiv.org/abs/2110.09456)"
+- "[Swin Transformer: Hierarchical Vision Transformer using Shifted Windows](https://arxiv.org/abs/2103.14030)"
+- "[Root Mean Square Layer Normalization](https://arxiv.org/abs/1910.07467)"
+
+Main optimizer (Distributed Shampoo) from "[Scalable Second Order Optimization for Deep Learning](https://arxiv.org/abs/2002.09018)".
+
+### Citations
+
+```text
 @misc{ramesh2021zeroshot,
       title={Zero-Shot Text-to-Image Generation}, 
       author={Aditya Ramesh and Mikhail Pavlov and Gabriel Goh and Scott Gray and Chelsea Voss and Alec Radford and Mark Chen and Ilya Sutskever},
@@ -124,7 +141,18 @@ year = {2021}
 }
 ```
 
+```text
+@misc{radford2021learning,
+      title={Learning Transferable Visual Models From Natural Language Supervision}, 
+      author={Alec Radford and Jong Wook Kim and Chris Hallacy and Aditya Ramesh and Gabriel Goh and Sandhini Agarwal and Girish Sastry and Amanda Askell and Pamela Mishkin and Jack Clark and Gretchen Krueger and Ilya Sutskever},
+      year={2021},
+      eprint={2103.00020},
+      archivePrefix={arXiv},
+      primaryClass={cs.CV}
+}
 ```
+
+```text
 @misc{esser2021taming,
       title={Taming Transformers for High-Resolution Image Synthesis}, 
       author={Patrick Esser and Robin Rombach and Björn Ommer},
@@ -135,7 +163,7 @@ year = {2021}
 }
 ```
 
-```
+```text
 @misc{lewis2019bart,
       title={BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension}, 
       author={Mike Lewis and Yinhan Liu and Naman Goyal and Marjan Ghazvininejad and Abdelrahman Mohamed and Omer Levy and Ves Stoyanov and Luke Zettlemoyer},
@@ -146,24 +174,64 @@ year = {2021}
 }
 ```
 
-```
-@misc{radford2021learning,
-      title={Learning Transferable Visual Models From Natural Language Supervision}, 
-      author={Alec Radford and Jong Wook Kim and Chris Hallacy and Aditya Ramesh and Gabriel Goh and Sandhini Agarwal and Girish Sastry and Amanda Askell and Pamela Mishkin and Jack Clark and Gretchen Krueger and Ilya Sutskever},
+```text
+@misc{anil2021scalable,
+      title={Scalable Second Order Optimization for Deep Learning},
+      author={Rohan Anil and Vineet Gupta and Tomer Koren and Kevin Regan and Yoram Singer},
       year={2021},
-      eprint={2103.00020},
+      eprint={2002.09018},
       archivePrefix={arXiv},
-      primaryClass={cs.CV}
+      primaryClass={cs.LG}
 }
 ```
 
+```text
+@misc{shazeer2020glu,
+      title={GLU Variants Improve Transformer},
+      author={Noam Shazeer},
+      year={2020},
+      url={https://arxiv.org/abs/2002.05202}    
+}
 ```
-@misc{anil2021scalable,
-      title={Scalable Second Order Optimization for Deep Learning},
-      author={Rohan Anil and Vineet Gupta and Tomer Koren and Kevin Regan and Yoram Singer},
+
+```text
+ @misc{wang_ma_dong_huang_zhang_wei_2022,
+       title={DeepNet: Scaling transformers to 1,000 layers},
+       author={Wang, Hongyu and Ma, Shuming and Dong, Li and Huang, Shaohan and Zhang, Dongdong and Wei, Furu},
+       year={2022},
+       eprint={2203.00555}
+       archivePrefix={arXiv},
+       primaryClass={cs.LG}
+} 
+```
+
+```text
+@misc{shleifer2021normformer,
+      title={NormFormer: Improved Transformer Pretraining with Extra Normalization},
+      author={Sam Shleifer and Jason Weston and Myle Ott},
       year={2021},
-      eprint={2002.09018},
+      eprint={2110.09456},
       archivePrefix={arXiv},
-      primaryClass={cs.LG}
+      primaryClass={cs.CL}
+}
+```
+
+```text
+@inproceedings{liu2021swinv2,
+               title={Swin Transformer V2: Scaling Up Capacity and Resolution}, 
+               author={Ze Liu and Han Hu and Yutong Lin and Zhuliang Yao and Zhenda Xie and Yixuan Wei and Jia Ning and Yue Cao and Zheng Zhang and Li Dong and Furu Wei and Baining Guo},
+               booktitle={International Conference on Computer Vision and Pattern Recognition (CVPR)},
+               year={2022}
+}
+```
+
+```text
+@misc{zhang2019root,
+      title = {Root Mean Square Layer Normalization},
+      author = {Biao Zhang and Rico Sennrich},
+      year = {2019},
+      eprint = {1910.07467},
+      archivePrefix = {arXiv},
+      primaryClass = {cs.LG}
 }
 ```

src/dalle_mini/data.py

@@ -1,3 +1,4 @@
+import random
 from dataclasses import dataclass, field
 from functools import partial
 
@@ -39,6 +40,9 @@ class Dataset:
     multi_hosts: bool = field(init=False)
 
     def __post_init__(self):
+        if self.seed_dataset is None:
+            # create a random seed
+            self.seed_dataset = random.randint(0, 2**32 - 1)
         self.multi_hosts = jax.process_count() > 1
         # feed blank captions only in streaming mode for now
         # otherwise dataset could be cached with same blanked captions
@@ -106,11 +110,10 @@ class Dataset:
         if self.streaming:
             # we need to shuffle early in streaming mode
             if hasattr(self, "train_dataset"):
-                self.train_dataset = self.train_dataset.shuffle(5000, self.seed_dataset)
+                self.train_dataset = self.train_dataset.shuffle(
+                    buffer_size=5000, seed=self.seed_dataset
+                )
         else:
-            # prepare rng for later shuffling
-            if self.seed_dataset is None:
-                self.seed_dataset = np.random.get_state()[1][0]
             self.rng_dataset = jax.random.PRNGKey(self.seed_dataset)
 
         # filter data

src/dalle_mini/model/configuration.py

@@ -44,15 +44,12 @@ class DalleBartConfig(PretrainedFromWandbMixin, PretrainedConfig):
         decoder_layers=12,
         decoder_ffn_dim=4096,
         decoder_attention_heads=16,
-        encoder_layerdrop=0.0,
-        decoder_layerdrop=0.0,
         activation_function="gelu",
         d_model=1024,
         dropout=0.1,
         attention_dropout=0.0,
         activation_dropout=0.0,
         init_std=0.02,
-        classifier_dropout=0.0,
         scale_embedding=False,
         gradient_checkpointing=False,
         use_cache=True,
@@ -60,9 +57,38 @@ class DalleBartConfig(PretrainedFromWandbMixin, PretrainedConfig):
         forced_eos_token_id=None,
         tie_word_embeddings=False,  # different modalities and sizes
         do_sample=True,
+        # transformer variants
+        head_scale=False,  # used in NormFormer
+        ln_type="layernorm",  # layer normalization type, "rmsnorm", "layernorm"
+        ln_positions="deepnet",  # layer normalization positions, "normformer", "swinv2", "deepnet" (same as post-ln)
+        use_cosine_attention=False,  # used in Swin v2
+        tau_init=0.05,  # used only in cosine attention (Swin v2)
+        use_deepnet_scaling=False,  # used in Deepnet
+        use_glu=False,  # "GLU Variants Improve Transformer"
         **kwargs,
     ):
+        # text normalizer
         self.normalize_text = normalize_text
+
+        # transformer variants
+        self.head_scale = head_scale  # per Normformer
+        assert ln_type in [
+            "rmsnorm",
+            "layernorm",
+        ], "ln_type must be 'rmsnorm' or 'layernorm'"
+        self.ln_type = ln_type
+        assert ln_positions in [
+            "normformer",
+            "swinv2",
+            "deepnet",
+        ], "ln_positions must be 'normformer', 'swinv2' or 'deepnet'"
+        self.ln_positions = ln_positions
+        self.use_cosine_attention = use_cosine_attention
+        self.tau_init = tau_init
+        self.use_deepnet_scaling = use_deepnet_scaling
+        self.use_glu = use_glu
+
+        # common parameters
         self.encoder_vocab_size = encoder_vocab_size
         self.image_vocab_size = image_vocab_size
         self.image_length = image_length
@@ -79,9 +105,6 @@ class DalleBartConfig(PretrainedFromWandbMixin, PretrainedConfig):
         self.activation_dropout = activation_dropout
         self.activation_function = activation_function
         self.init_std = init_std
-        self.encoder_layerdrop = encoder_layerdrop
-        self.decoder_layerdrop = decoder_layerdrop
-        self.classifier_dropout = classifier_dropout
         self.use_cache = use_cache
         self.gradient_checkpointing = gradient_checkpointing
         self.scale_embedding = (

src/dalle_mini/model/modeling.py

@@ -18,7 +18,7 @@ import math
 import os
 from functools import partial
 from pickle import UnpicklingError
-from typing import Dict, Optional, Tuple, Union
+from typing import Any, Dict, Optional, Tuple, Union
 
 import flax
 import flax.linen as nn
@@ -26,7 +26,9 @@ import jax
 import jax.numpy as jnp
 import msgpack.exceptions
 from flax.core.frozen_dict import unfreeze
-from flax.linen import make_causal_mask
+from flax.linen import combine_masks, make_causal_mask
+from flax.linen import partitioning as nn_partitioning
+from flax.linen.attention import dot_product_attention_weights
 from flax.serialization import from_bytes
 from flax.traverse_util import flatten_dict, unflatten_dict
 from jax import lax
@@ -42,6 +44,8 @@ from transformers.file_utils import (
 )
 from transformers.generation_flax_utils import FlaxSampleOutput
 from transformers.modeling_flax_outputs import (
+    FlaxBaseModelOutput,
+    FlaxBaseModelOutputWithPastAndCrossAttentions,
     FlaxCausalLMOutputWithCrossAttentions,
     FlaxSeq2SeqLMOutput,
 )
@@ -49,11 +53,7 @@ from transformers.modeling_flax_utils import ACT2FN
 from transformers.models.bart.modeling_flax_bart import (
     FlaxBartAttention,
     FlaxBartDecoder,
-    FlaxBartDecoderLayer,
-    FlaxBartDecoderLayerCollection,
     FlaxBartEncoder,
-    FlaxBartEncoderLayer,
-    FlaxBartEncoderLayerCollection,
     FlaxBartForConditionalGeneration,
     FlaxBartForConditionalGenerationModule,
     FlaxBartModule,
@@ -66,13 +66,124 @@ from .utils import PretrainedFromWandbMixin
 
 logger = logging.get_logger(__name__)
 
+remat = nn_partitioning.remat
+
+
+# deepnet initialization
+def deepnet_init(gain=1):
+    init = jax.nn.initializers.glorot_normal()
+
+    def _init(*args, **kwargs):
+        return gain * init(*args, **kwargs)
+
+    return _init
+
+
+# deepnet gain
+deepnet_gain = {
+    "encoder": {
+        "alpha": lambda config: 0.81
+        * (config.encoder_layers**4 * config.decoder_layers) ** 0.0625,
+        "beta": lambda config: 0.87
+        * (config.encoder_layers**4 * config.decoder_layers) ** -0.0625,
+    },
+    "decoder": {
+        "alpha": lambda config: (3 * config.decoder_layers) ** 0.25,
+        "beta": lambda config: (12 * config.decoder_layers) ** -0.25,
+    },
+}
+
+
+class RMSNorm(nn.Module):
+    """
+    From "Root Mean Square Layer Normalization" by https://arxiv.org/abs/1910.07467
+
+    Adapted from flax.linen.LayerNorm
+    """
+
+    epsilon: float = 1e-6
+    dtype: Any = jnp.float32
+    param_dtype: Any = jnp.float32
+    use_scale: bool = True
+    scale_init: Any = jax.nn.initializers.ones
+
+    @nn.compact
+    def __call__(self, x):
+        reduction_axes = (-1,)
+        feature_axes = (-1,)
+
+        rms_sq = self._compute_rms_sq(x, reduction_axes)
+
+        return self._normalize(
+            self,
+            x,
+            rms_sq,
+            reduction_axes,
+            feature_axes,
+            self.dtype,
+            self.param_dtype,
+            self.epsilon,
+            self.use_scale,
+            self.scale_init,
+        )
+
+    def _compute_rms_sq(self, x, axes):
+        x = jnp.asarray(x, jnp.promote_types(jnp.float32, jnp.result_type(x)))
+        rms_sq = jnp.mean(jax.lax.square(x), axes)
+        return rms_sq
+
+    def _normalize(
+        self,
+        mdl,
+        x,
+        rms_sq,
+        reduction_axes,
+        feature_axes,
+        dtype,
+        param_dtype,
+        epsilon,
+        use_scale,
+        scale_init,
+    ):
+        reduction_axes = nn.normalization._canonicalize_axes(x.ndim, reduction_axes)
+        feature_axes = nn.normalization._canonicalize_axes(x.ndim, feature_axes)
+        stats_shape = list(x.shape)
+        for axis in reduction_axes:
+            stats_shape[axis] = 1
+        rms_sq = rms_sq.reshape(stats_shape)
+        feature_shape = [1] * x.ndim
+        reduced_feature_shape = []
+        for ax in feature_axes:
+            feature_shape[ax] = x.shape[ax]
+            reduced_feature_shape.append(x.shape[ax])
+        mul = lax.rsqrt(rms_sq + epsilon)
+        if use_scale:
+            scale = mdl.param(
+                "scale", scale_init, reduced_feature_shape, param_dtype
+            ).reshape(feature_shape)
+            mul *= scale
+        y = mul * x
+        return jnp.asarray(y, dtype)
+
+
+def norm(type, *args, **kwargs):
+    if type == "rmsnorm":
+        return RMSNorm(*args, **kwargs)
+    elif type == "layernorm":
+        return nn.LayerNorm(*args, **kwargs)
+    else:
+        raise ValueError(f"Unknown norm type {type}")
+
 
 class FlaxBartAttention(FlaxBartAttention):
     """
     Edits:
     - causal mask is used only in decoder and considers image_length
+    - scale attention heads per NormFormer paper
     """
 
+    is_encoder: bool = False
+
     def setup(self) -> None:
         self.head_dim = self.embed_dim // self.num_heads
         if self.head_dim * self.num_heads != self.embed_dim:
@@ -86,142 +197,667 @@ class FlaxBartAttention(FlaxBartAttention):
             self.embed_dim,
             use_bias=self.bias,
             dtype=self.dtype,
-            kernel_init=jax.nn.initializers.normal(self.config.init_std),
         )
 
-        self.q_proj, self.k_proj, self.v_proj = dense(), dense(), dense()
-        self.out_proj = dense()
+        gain = deepnet_gain["encoder" if self.is_encoder else "decoder"]["beta"](
+            self.config
+        )
 
+        self.q_proj = dense(
+            kernel_init=deepnet_init()
+            if self.config.use_deepnet_scaling
+            else jax.nn.initializers.normal(self.config.init_std)
+        )
+        self.k_proj = dense(
+            kernel_init=deepnet_init()
+            if self.config.use_deepnet_scaling
+            else jax.nn.initializers.normal(self.config.init_std)
+        )
+        self.v_proj = dense(
+            kernel_init=deepnet_init(gain)
+            if self.config.use_deepnet_scaling
+            else jax.nn.initializers.normal(self.config.init_std)
+        )
+        self.out_proj = dense(
+            kernel_init=deepnet_init(gain)
+            if self.config.use_deepnet_scaling
+            else jax.nn.initializers.normal(self.config.init_std)
+        )
         self.dropout_layer = nn.Dropout(rate=self.dropout)
 
+        if self.config.head_scale:
+            self.head_scale = self.param(
+                "head_scale", jax.nn.initializers.ones, (1, 1, self.num_heads, 1)
+            )
+
+        if self.config.use_cosine_attention:
+            self.tau = self.param(
+                "tau",
+                jax.nn.initializers.constant(self.config.tau_init),
+                (1, self.num_heads, 1, 1),
+            )
+
         if self.causal:
             # used only in decoder
             self.causal_mask = make_causal_mask(
                 jnp.ones((1, self.config.image_length), dtype="bool"), dtype="bool"
             )
 
+    def __call__(
+        self,
+        hidden_states: jnp.ndarray,
+        key_value_states: Optional[jnp.ndarray] = None,
+        attention_mask: Optional[jnp.ndarray] = None,
+        init_cache: bool = False,
+        deterministic: bool = True,
+    ) -> Tuple[jnp.ndarray]:
+        """Input shape: Batch x Time x Channel"""
+
+        # if key_value_states are provided this layer is used as a cross-attention layer
+        # for the decoder
+        is_cross_attention = key_value_states is not None
+        batch_size = hidden_states.shape[0]
+
+        # get query proj
+        query_states = self.q_proj(hidden_states)
+        # get key, value proj
+        if is_cross_attention:
+            # cross_attentions
+            key_states = self.k_proj(key_value_states)
+            value_states = self.v_proj(key_value_states)
+        else:
+            # self_attention
+            key_states = self.k_proj(hidden_states)
+            value_states = self.v_proj(hidden_states)
 
-class FlaxBartEncoderLayer(FlaxBartEncoderLayer):
-    """
-    Edits:
-    - no bias
-    - use custom FlaxBartAttention
-    """
+        query_states = self._split_heads(query_states)
+        key_states = self._split_heads(key_states)
+        value_states = self._split_heads(value_states)
 
-    def setup(self) -> None:
-        self.embed_dim = self.config.d_model
-        self.self_attn = FlaxBartAttention(
-            config=self.config,
-            embed_dim=self.embed_dim,
-            num_heads=self.config.encoder_attention_heads,
-            dropout=self.config.attention_dropout,
-            bias=False,
+        # handle cache prepare causal attention mask
+        if self.causal:
+            query_length, key_length = query_states.shape[1], key_states.shape[1]
+            if self.has_variable("cache", "cached_key"):
+                mask_shift = self.variables["cache"]["cache_index"]
+                max_decoder_length = self.variables["cache"]["cached_key"].shape[1]
+                causal_mask = lax.dynamic_slice(
+                    self.causal_mask,
+                    (0, 0, mask_shift, 0),
+                    (1, 1, query_length, max_decoder_length),
+                )
+            else:
+                causal_mask = self.causal_mask[:, :, :query_length, :key_length]
+            causal_mask = jnp.broadcast_to(
+                causal_mask, (batch_size,) + causal_mask.shape[1:]
+            )
+
+        # combine masks if needed
+        if attention_mask is not None and self.causal:
+            attention_mask = jnp.broadcast_to(
+                jnp.expand_dims(attention_mask, axis=(-3, -2)), causal_mask.shape
+            )
+            attention_mask = combine_masks(attention_mask, causal_mask)
+        elif self.causal:
+            attention_mask = causal_mask
+        elif attention_mask is not None:
+            attention_mask = jnp.expand_dims(attention_mask, axis=(-3, -2))
+
+        # During fast autoregressive decoding, we feed one position at a time,
+        # and cache the keys and values step by step.
+        if self.causal and (self.has_variable("cache", "cached_key") or init_cache):
+            key_states, value_states, attention_mask = self._concatenate_to_cache(
+                key_states, value_states, query_states, attention_mask
+            )
+
+        # Convert the boolean attention mask to an attention bias.
+        if attention_mask is not None:
+            # attention mask in the form of attention bias
+            attention_bias = lax.select(
+                attention_mask > 0,
+                jnp.full(attention_mask.shape, 0.0).astype(self.dtype),
+                jnp.full(attention_mask.shape, float("-inf")).astype(self.dtype),
+            )
+        else:
+            attention_bias = None
+
+        dropout_rng = None
+        if not deterministic and self.dropout > 0.0:
+            dropout_rng = self.make_rng("dropout")
+
+        if self.config.use_cosine_attention:
+            # normalize q and k
+            query_states = query_states / (
+                jnp.linalg.norm(query_states, axis=-1, keepdims=True) + 1e-8
+            )
+            key_states = key_states / (
+                jnp.linalg.norm(key_states, axis=-1, keepdims=True) + 1e-8
+            )
+        attn_weights = dot_product_attention_weights(
+            query_states,
+            key_states,
+            bias=attention_bias,
+            dropout_rng=dropout_rng,
+            dropout_rate=self.dropout,
+            broadcast_dropout=True,
+            deterministic=deterministic,
             dtype=self.dtype,
+            precision=None,
         )
-        self.self_attn_layer_norm = nn.LayerNorm(dtype=self.dtype, epsilon=1e-05)
-        self.dropout_layer = nn.Dropout(rate=self.config.dropout)
-        self.activation_fn = ACT2FN[self.config.activation_function]
-        self.activation_dropout_layer = nn.Dropout(rate=self.config.activation_dropout)
-        self.fc1 = nn.Dense(
-            self.config.encoder_ffn_dim,
+        if self.config.use_cosine_attention:
+            # divide by tau
+            attn_weights = attn_weights / jnp.maximum(self.tau, 0.01)
+
+        attn_output = jnp.einsum("...hqk,...khd->...qhd", attn_weights, value_states)
+        if self.config.head_scale:
+            # per Normformer
+            attn_output = attn_output * self.head_scale
+        attn_output = self._merge_heads(attn_output)
+        attn_output = self.out_proj(attn_output)
+
+        return attn_output, attn_weights
+
+
+class GLU(nn.Module):
+    """From "GLU Variants Improve Transformer" by https://arxiv.org/abs/2002.05202"""
+
+    config: DalleBartConfig
+    ffn_dim: int
+    embed_dim: int
+    dtype: jnp.dtype = jnp.float32
+    is_encoder: bool = False
+
+    @nn.compact
+    def __call__(self, x: jnp.ndarray, deterministic: bool = True) -> jnp.ndarray:
+
+        gain = deepnet_gain["encoder" if self.is_encoder else "decoder"]["beta"](
+            self.config
+        )
+
+        if self.config.ln_positions in ["normformer"]:
+            x = norm(
+                self.config.ln_type, dtype=self.dtype, epsilon=1e-05, use_scale=False
+            )(x)
+        w = nn.Dense(
+            self.ffn_dim,
+            dtype=self.dtype,
+            use_bias=False,
+            kernel_init=deepnet_init(gain)
+            if self.config.use_deepnet_scaling
+            else jax.nn.initializers.normal(self.config.init_std),
+        )(x)
+        w = ACT2FN[self.config.activation_function](w)
+        v = nn.Dense(
+            self.ffn_dim,
             dtype=self.dtype,
             use_bias=False,
-            kernel_init=jax.nn.initializers.normal(self.config.init_std),
+            kernel_init=deepnet_init(gain)
+            if self.config.use_deepnet_scaling
+            else jax.nn.initializers.normal(self.config.init_std),
+        )(x)
+        x = w * v
+        if self.config.ln_positions in ["normformer"]:
+            x = norm(
+                self.config.ln_type, dtype=self.dtype, epsilon=1e-05, use_scale=False
+            )(x)
+        x = nn.Dropout(rate=self.config.activation_dropout)(
+            x, deterministic=deterministic
         )
-        self.fc2 = nn.Dense(
+
+        x = nn.Dense(
             self.embed_dim,
             dtype=self.dtype,
             use_bias=False,
-            kernel_init=jax.nn.initializers.normal(self.config.init_std),
+            kernel_init=deepnet_init(gain)
+            if self.config.use_deepnet_scaling
+            else jax.nn.initializers.normal(self.config.init_std),
+        )(x)
+        if self.config.ln_positions in ["swinv2"]:
+            x = norm(self.config.ln_type, dtype=self.dtype, epsilon=1e-05)(x)
+        x = nn.Dropout(rate=self.config.dropout)(x, deterministic=deterministic)
+        return x
+
+
+class FFN(nn.Module):
+    """Simple FFN layer"""
+
+    config: DalleBartConfig
+    ffn_dim: int
+    embed_dim: int
+    dtype: jnp.dtype = jnp.float32
+    is_encoder: bool = False
+
+    @nn.compact
+    def __call__(self, x: jnp.ndarray, deterministic: bool = True) -> jnp.ndarray:
+
+        gain = deepnet_gain["encoder" if self.is_encoder else "decoder"]["beta"](
+            self.config
         )
-        self.final_layer_norm = nn.LayerNorm(dtype=self.dtype, epsilon=1e-05)
+        if self.config.ln_positions in ["normformer"]:
+            x = norm(
+                self.config.ln_type, dtype=self.dtype, epsilon=1e-05, use_scale=False
+            )(x)
+        x = nn.Dense(
+            self.ffn_dim,
+            dtype=self.dtype,
+            use_bias=False,
+            kernel_init=deepnet_init(gain)
+            if self.config.use_deepnet_scaling
+            else jax.nn.initializers.normal(self.config.init_std),
+        )(x)
+        x = ACT2FN[self.config.activation_function](x)
+        if self.config.ln_positions in ["normformer"]:
+            x = norm(
+                self.config.ln_type, dtype=self.dtype, epsilon=1e-05, use_scale=False
+            )(x)
+        x = nn.Dropout(rate=self.config.activation_dropout)(
+            x, deterministic=deterministic
+        )
+        x = nn.Dense(
+            self.embed_dim,
+            dtype=self.dtype,
+            use_bias=False,
+            kernel_init=deepnet_init(gain)
+            if self.config.use_deepnet_scaling
+            else jax.nn.initializers.normal(self.config.init_std),
+        )(x)
+        if self.config.ln_positions in ["swinv2"]:
+            x = norm(self.config.ln_type, dtype=self.dtype, epsilon=1e-05)(x)
+        x = nn.Dropout(rate=self.config.dropout)(x, deterministic=deterministic)
+        return x
 
 
-class FlaxBartEncoderLayerCollection(FlaxBartEncoderLayerCollection):
+class FlaxBartEncoderLayer(nn.Module):
     """
     Edits:
-    - use custom FlaxBartEncoderLayer
-    - allow Gradient Checkpointing (nn.remat)
+    - no bias
+    - use custom FlaxBartAttention
     """
 
-    def setup(self):
-        layer_module = (
-            nn.remat(FlaxBartEncoderLayer, concrete=True)
-            if self.config.gradient_checkpointing
-            else FlaxBartEncoderLayer
+    config: DalleBartConfig
+    dtype: jnp.dtype = jnp.float32
+    add_norm: bool = False
+    use_scale: bool = True
+
+    @nn.compact
+    def __call__(
+        self,
+        hidden_states: jnp.ndarray,
+        attention_mask: jnp.ndarray,
+        output_attentions: bool = True,
+        deterministic: bool = True,
+    ) -> Tuple[jnp.ndarray]:
+
+        res_gain = (
+            deepnet_gain["encoder"]["alpha"](self.config)
+            if self.config.use_deepnet_scaling
+            else 1
         )
-        self.layers = [
-            layer_module(self.config, name=str(i), dtype=self.dtype)
-            for i in range(self.config.encoder_layers)
-        ]
-        self.layerdrop = self.config.encoder_layerdrop
+
+        embed_dim = self.config.d_model
+        residual = hidden_states
+        if self.config.ln_positions in ["normformer"]:
+            hidden_states = norm(self.config.ln_type, dtype=self.dtype, epsilon=1e-05)(
+                hidden_states
+            )
+        hidden_states, attn_weights = FlaxBartAttention(
+            config=self.config,
+            embed_dim=embed_dim,
+            num_heads=self.config.encoder_attention_heads,
+            dropout=self.config.attention_dropout,
+            bias=False,
+            dtype=self.dtype,
+            is_encoder=True,
+        )(hidden_states=hidden_states, attention_mask=attention_mask)
+
+        if self.config.ln_positions in ["normformer", "swinv2"]:
+            hidden_states = norm(self.config.ln_type, dtype=self.dtype, epsilon=1e-05)(
+                hidden_states
+            )
+        hidden_states = nn.Dropout(rate=self.config.dropout)(
+            hidden_states, deterministic=deterministic
+        )
+        hidden_states = residual * res_gain + hidden_states
+        if self.config.ln_positions in ["deepnet"]:
+            hidden_states = norm(self.config.ln_type, dtype=self.dtype, epsilon=1e-05)(
+                hidden_states
+            )
+
+        residual = hidden_states
+        ff_block = (
+            GLU(
+                config=self.config,
+                ffn_dim=self.config.encoder_ffn_dim,
+                embed_dim=embed_dim,
+                dtype=self.dtype,
+                is_encoder=True,
+            )
+            if self.config.use_glu
+            else FFN(
+                config=self.config,
+                ffn_dim=self.config.encoder_ffn_dim,
+                embed_dim=embed_dim,
+                dtype=self.dtype,
+                is_encoder=True,
+            )
+        )
+        hidden_states = ff_block(hidden_states, deterministic=deterministic)
+        hidden_states = residual * res_gain + hidden_states
+        if self.add_norm or self.config.ln_positions in ["deepnet"]:
+            use_scale = self.use_scale or self.config.ln_positions == "deepnet"
+            hidden_states = norm(
+                self.config.ln_type,
+                dtype=self.dtype,
+                epsilon=1e-05,
+                use_scale=use_scale,
+            )(hidden_states)
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (attn_weights,)
+
+        return outputs
 
 
-class FlaxBartDecoderLayer(FlaxBartDecoderLayer):
+class FlaxBartDecoderLayer(nn.Module):
     """
     Edits:
     - no bias
-    - uses custom FlaxBartAttention
+    - use custom FlaxBartAttention
     """
 
-    def setup(self) -> None:
-        self.embed_dim = self.config.d_model
-        self.self_attn = FlaxBartAttention(
+    config: DalleBartConfig
+    dtype: jnp.dtype = jnp.float32
+    add_norm: bool = False
+    use_scale: bool = False
+
+    @nn.compact
+    def __call__(
+        self,
+        hidden_states: jnp.ndarray,
+        attention_mask: jnp.ndarray,
+        encoder_hidden_states: Optional[jnp.ndarray] = None,
+        encoder_attention_mask: Optional[jnp.ndarray] = None,
+        init_cache: bool = False,
+        output_attentions: bool = True,
+        deterministic: bool = True,
+    ) -> Tuple[jnp.ndarray]:
+
+        res_gain = (
+            deepnet_gain["decoder"]["alpha"](self.config)
+            if self.config.use_deepnet_scaling
+            else 1
+        )
+
+        embed_dim = self.config.d_model
+        residual = hidden_states
+
+        # Self Attention
+        if self.config.ln_positions in ["normformer"]:
+            hidden_states = norm(
+                self.config.ln_type,
+                dtype=self.dtype,
+                epsilon=1e-05,
+                use_scale=False,
+            )(hidden_states)
+        hidden_states, attn_weights = FlaxBartAttention(
             config=self.config,
-            embed_dim=self.embed_dim,
+            embed_dim=embed_dim,
             num_heads=self.config.decoder_attention_heads,
             dropout=self.config.attention_dropout,
             causal=True,
             bias=False,
             dtype=self.dtype,
+            is_encoder=False,
+        )(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            init_cache=init_cache,
         )
-        self.dropout_layer = nn.Dropout(rate=self.config.dropout)
-        self.activation_fn = ACT2FN[self.config.activation_function]
-        self.activation_dropout_layer = nn.Dropout(rate=self.config.activation_dropout)
 
-        self.self_attn_layer_norm = nn.LayerNorm(dtype=self.dtype, epsilon=1e-05)
-        self.encoder_attn = FlaxBartAttention(
-            config=self.config,
-            embed_dim=self.embed_dim,
-            num_heads=self.config.decoder_attention_heads,
-            dropout=self.config.attention_dropout,
-            bias=False,
-            dtype=self.dtype,
+        if self.config.ln_positions in ["normformer", "swinv2"]:
+            hidden_states = norm(self.config.ln_type, dtype=self.dtype, epsilon=1e-05)(
+                hidden_states
+            )
+        hidden_states = nn.Dropout(rate=self.config.dropout)(
+            hidden_states, deterministic=deterministic
         )
-        self.encoder_attn_layer_norm = nn.LayerNorm(dtype=self.dtype, epsilon=1e-05)
-        self.fc1 = nn.Dense(
-            self.config.encoder_ffn_dim,
-            dtype=self.dtype,
-            use_bias=False,
-            kernel_init=jax.nn.initializers.normal(self.config.init_std),
+        hidden_states = residual * res_gain + hidden_states
+        if self.config.ln_positions in ["deepnet"]:
+            hidden_states = norm(self.config.ln_type, dtype=self.dtype, epsilon=1e-05)(
+                hidden_states
+            )
+
+        # Cross Attention
+        cross_attn_weights = None
+        if encoder_hidden_states is not None:
+            residual = hidden_states
+            if self.config.ln_positions in ["normformer"]:
+                hidden_states = norm(
+                    self.config.ln_type,
+                    dtype=self.dtype,
+                    epsilon=1e-05,
+                    use_scale=False,
+                )(hidden_states)
+            hidden_states, cross_attn_weights = FlaxBartAttention(
+                config=self.config,
+                embed_dim=embed_dim,
+                num_heads=self.config.decoder_attention_heads,
+                dropout=self.config.attention_dropout,
+                bias=False,
+                dtype=self.dtype,
+                is_encoder=False,
+            )(
+                hidden_states=hidden_states,
+                key_value_states=encoder_hidden_states,
+                attention_mask=encoder_attention_mask,
+            )
+            if self.config.ln_positions in ["normformer", "swinv2"]:
+                hidden_states = norm(
+                    self.config.ln_type, dtype=self.dtype, epsilon=1e-05
+                )(hidden_states)
+            hidden_states = nn.Dropout(rate=self.config.dropout)(
+                hidden_states, deterministic=deterministic
+            )
+            hidden_states = residual * res_gain + hidden_states
+            if self.config.ln_positions in ["deepnet"]:
+                hidden_states = norm(
+                    self.config.ln_type, dtype=self.dtype, epsilon=1e-05
+                )(hidden_states)
+
+        # Feed forward
+        residual = hidden_states
+        ff_block = (
+            GLU(
+                config=self.config,
+                ffn_dim=self.config.decoder_ffn_dim,
+                embed_dim=embed_dim,
+                dtype=self.dtype,
+                is_encoder=False,
+            )
+            if self.config.use_glu
+            else FFN(
+                config=self.config,
+                ffn_dim=self.config.decoder_ffn_dim,
+                embed_dim=embed_dim,
+                dtype=self.dtype,
+                is_encoder=False,
+            )
         )
-        self.fc2 = nn.Dense(
-            self.embed_dim,
-            dtype=self.dtype,
-            use_bias=False,
-            kernel_init=jax.nn.initializers.normal(self.config.init_std),
+        hidden_states = ff_block(hidden_states, deterministic=deterministic)
+        hidden_states = residual * res_gain + hidden_states
+        if self.add_norm or self.config.ln_positions in ["deepnet"]:
+            use_scale = self.use_scale or self.config.ln_positions == "deepnet"
+            hidden_states = norm(
+                self.config.ln_type,
+                dtype=self.dtype,
+                epsilon=1e-05,
+                use_scale=use_scale,
+            )(hidden_states)
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (attn_weights, cross_attn_weights)
+
+        return outputs
+
+
+class FlaxBartEncoderLayerCollection(nn.Module):
+    config: DalleBartConfig
+    dtype: jnp.dtype = jnp.float32  # the dtype of the computation
+    """
+    Edits:
+    - use custom FlaxBartEncoderLayer
+    - allow Gradient Checkpointing (nn.remat)
+    """
+
+    @nn.compact
+    def __call__(
+        self,
+        hidden_states,
+        attention_mask,
+        deterministic: bool = True,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+    ):
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+
+        n_layers = self.config.encoder_layers
+        layer = (
+            remat(FlaxBartEncoderLayer, static_argnums=(2, 3))
+            if self.config.gradient_checkpointing
+            else FlaxBartEncoderLayer
+        )
+        for i in range(n_layers):
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+            # final layernorm on the output of the last layer
+            # or every 6 layers for Swin v2
+            # ignored args for deepnet which always add a norm with scale
+            add_norm = (i == n_layers - 1) or (
+                (self.config.ln_positions == "swinv2") and ((i + 1) % 6 == 0)
+            )
+            # we don't need to scale the norm for the last layer
+            use_scale = i != n_layers - 1
+            layer_outputs = layer(
+                self.config, dtype=self.dtype, add_norm=add_norm, use_scale=use_scale
+            )(
+                hidden_states,
+                attention_mask,
+                output_attentions,
+                deterministic,
+            )
+            hidden_states = layer_outputs[0]
+            if output_attentions:
+                all_self_attns += (layer_outputs[1],)
+
+        # add hidden states from the last layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        outputs = [
+            hidden_states,
+            all_hidden_states,
+            all_self_attns,
+        ]
+
+        if not return_dict:
+            return tuple(v for v in outputs if v is not None)
+
+        return FlaxBaseModelOutput(
+            last_hidden_state=hidden_states,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
         )
-        self.final_layer_norm = nn.LayerNorm(dtype=self.dtype, epsilon=1e-05)
 
 
-class FlaxBartDecoderLayerCollection(FlaxBartDecoderLayerCollection):
+class FlaxBartDecoderLayerCollection(nn.Module):
+    config: DalleBartConfig
+    dtype: jnp.dtype = jnp.float32  # the dtype of the computation
     """
     Edits:
     - use custom FlaxBartDecoderLayer
     - allow Gradient Checkpointing (nn.remat)
     """
 
-    def setup(self):
-        layer_module = (
-            nn.remat(FlaxBartDecoderLayer, concrete=True)
+    @nn.compact
+    def __call__(
+        self,
+        hidden_states,
+        attention_mask,
+        encoder_hidden_states: Optional[jnp.ndarray] = None,
+        encoder_attention_mask: Optional[jnp.ndarray] = None,
+        deterministic: bool = True,
+        init_cache: bool = False,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+    ):
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        all_cross_attentions = (
+            () if (output_attentions and encoder_hidden_states is not None) else None
+        )
+
+        n_layers = self.config.decoder_layers
+        layer = (
+            remat(FlaxBartDecoderLayer, static_argnums=(4, 5, 6))
             if self.config.gradient_checkpointing
             else FlaxBartDecoderLayer
         )
-        self.layers = [
-            layer_module(self.config, name=str(i), dtype=self.dtype)
-            for i in range(self.config.decoder_layers)
+        for i in range(n_layers):
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+            # final layernorm on the output of the last layer
+            # or every 6 layers for Swin v2
+            add_norm = (i == n_layers - 1) or (
+                (self.config.ln_positions == "swinv2") and ((i + 1) % 6 == 0)
+            )
+            # we don't need to scale the norm for the last layer
+            use_scale = i != n_layers - 1
+            layer_outputs = layer(
+                self.config, dtype=self.dtype, add_norm=add_norm, use_scale=use_scale
+            )(
+                hidden_states,
+                attention_mask,
+                encoder_hidden_states,
+                encoder_attention_mask,
+                init_cache,
+                output_attentions,
+                deterministic,
+            )
+
+            hidden_states = layer_outputs[0]
+            if output_attentions:
+                all_self_attns += (layer_outputs[1],)
+
+                if encoder_hidden_states is not None:
+                    all_cross_attentions += (layer_outputs[2],)
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        outputs = [
+            hidden_states,
+            all_hidden_states,
+            all_self_attns,
+            all_cross_attentions,
         ]
-        self.layerdrop = self.config.decoder_layerdrop
+
+        if not return_dict:
+            return tuple(v for v in outputs if v is not None)
+
+        return FlaxBaseModelOutputWithPastAndCrossAttentions(
+            last_hidden_state=hidden_states,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+            cross_attentions=all_cross_attentions,
+        )
 
 
 class FlaxBartEncoder(FlaxBartEncoder):
@@ -246,10 +882,14 @@ class FlaxBartEncoder(FlaxBartEncoder):
         self.embed_positions = nn.Embed(
             self.config.max_text_length + self.offset,
             embed_dim,
-            embedding_init=jax.nn.initializers.normal(self.config.init_std),
+            embedding_init=deepnet_init()
+            if self.config.use_deepnet_scaling
+            else jax.nn.initializers.normal(self.config.init_std),
         )
         self.layers = FlaxBartEncoderLayerCollection(self.config, self.dtype)
-        self.layernorm_embedding = nn.LayerNorm(dtype=self.dtype, epsilon=1e-05)
+        self.layernorm_embedding = norm(
+            self.config.ln_type, dtype=self.dtype, epsilon=1e-05
+        )
 
 
 class FlaxBartDecoder(FlaxBartDecoder):
@@ -276,11 +916,15 @@ class FlaxBartDecoder(FlaxBartDecoder):
         self.embed_positions = nn.Embed(
             self.config.image_length + self.offset,  # image length for BOS
             embed_dim,
-            embedding_init=jax.nn.initializers.normal(self.config.init_std),
+            embedding_init=deepnet_init()
+            if self.config.use_deepnet_scaling
+            else jax.nn.initializers.normal(self.config.init_std),
         )
 
         self.layers = FlaxBartDecoderLayerCollection(self.config, self.dtype)
-        self.layernorm_embedding = nn.LayerNorm(dtype=self.dtype, epsilon=1e-05)
+        self.layernorm_embedding = norm(
+            self.config.ln_type, dtype=self.dtype, epsilon=1e-05
+        )
 
 
 class FlaxBartModule(FlaxBartModule):
@@ -294,12 +938,16 @@ class FlaxBartModule(FlaxBartModule):
         encoder_embed_tokens = nn.Embed(
             self.config.encoder_vocab_size,
             self.config.d_model,
-            embedding_init=jax.nn.initializers.normal(self.config.init_std),
+            embedding_init=deepnet_init()
+            if self.config.use_deepnet_scaling
+            else jax.nn.initializers.normal(self.config.init_std),
         )
         decoder_embed_tokens = nn.Embed(
             self.config.image_vocab_size + 1,  # image vocab size + 1 for BOS
             self.config.d_model,
-            embedding_init=jax.nn.initializers.normal(self.config.init_std),
+            embedding_init=deepnet_init()
+            if self.config.use_deepnet_scaling
+            else jax.nn.initializers.normal(self.config.init_std),
         )
 
         self.encoder = FlaxBartEncoder(
@@ -639,7 +1287,9 @@ class FlaxBartForConditionalGenerationModule(FlaxBartForConditionalGenerationMod
             + 1,  # image vocab size + 1 for BOS to have same size as decoder inputs (for sharding)
             use_bias=False,
             dtype=self.dtype,
-            kernel_init=jax.nn.initializers.normal(self.config.init_std),
+            kernel_init=deepnet_init()
+            if self.config.use_deepnet_scaling
+            else jax.nn.initializers.normal(self.config.init_std),
         )
 
     def __call__(

src/dalle_mini/model/partitions.py

@@ -36,23 +36,21 @@ def _replacement_rules(rules):
 def _get_partition_rules():
     return [
         # embeddings
-        ((r"embed_positions", "embedding"), P("mp", None)),
-        ((r"embed_tokens", "embedding"), P("mp", None)),
-        # self-attention
-        ((r"self_attn", "(q_proj|k_proj|v_proj)", "kernel"), P(None, "mp")),
-        ((r"self_attn", "out_proj", "kernel"), P("mp", None)),
-        # enc-dec attention
-        ((r"encoder_attn", "(q_proj|k_proj|v_proj)", "kernel"), P(None, "mp")),
-        ((r"encoder_attn", "out_proj", "kernel"), P("mp", None)),
+        (("embed_positions", "embedding"), P("mp", None)),
+        (("embed_tokens", "embedding"), P("mp", None)),
+        # attention
+        (("(q_proj|k_proj|v_proj)", "kernel"), P(None, "mp")),
+        (("out_proj", "kernel"), P("mp", None)),
         # FFN
-        ((r"fc1", "kernel"), P(None, "mp")),
-        ((r"fc2", "kernel"), P("mp", None)),
+        (("Dense_0", "kernel"), P(None, "mp")),
+        (("GLU.*", "Dense_1", "kernel"), P(None, "mp")),
+        (("GLU.*", "Dense_2", "kernel"), P("mp", None)),
+        (("FFN.*", "Dense_1", "kernel"), P("mp", None)),
         # layer norms
-        ((r"layernorm_embedding", "(bias|scale)"), None),
-        ((r"self_attn_layer_norm", "(bias|scale)"), None),
-        ((r"encoder_attn_layer_norm", "(bias|scale)"), None),
-        ((r"final_layer_norm", "(bias|scale)"), None),
-        ((r"lm_head", "kernel"), P(None, "mp")),
+        (("(bias|scale)",), None),
+        (("lm_head", "kernel"), P(None, "mp")),
+        # head scale and tau
+        (("(head_scale|tau)",), None),
     ]
 
 
@@ -63,6 +61,6 @@ def set_partitions(in_dict):
     result = {k: replace(k, v) for k, v in initd.items()}
     for k, v in result.items():
         if v == _unmatched:
-            print(k)
+            print(f"Unmatched -> {k}")
     assert _unmatched not in result.values(), "Incomplete partition spec."
     return freeze(unflatten_dict(result))

src/dalle_mini/model/utils.py

@@ -9,13 +9,11 @@ class PretrainedFromWandbMixin:
     @classmethod
     def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
         """
-        Initializes from a wandb artifact, google bucket path or delegates loading to the superclass.
+        Initializes from a wandb artifact or delegates loading to the superclass.
         """
         with tempfile.TemporaryDirectory() as tmp_dir:  # avoid multiple artifact copies
-            if (
-                ":" in pretrained_model_name_or_path
-                and not os.path.isdir(pretrained_model_name_or_path)
-                and not pretrained_model_name_or_path.startswith("gs")
+            if ":" in pretrained_model_name_or_path and not os.path.isdir(
+                pretrained_model_name_or_path
             ):
                 # wandb artifact
                 if wandb.run is not None:
@@ -27,17 +25,3 @@ class PretrainedFromWandbMixin:
             return super(PretrainedFromWandbMixin, cls).from_pretrained(
                 pretrained_model_name_or_path, *model_args, **kwargs
             )
-
-
-def copy_blobs(source_path, dest_path):
-    assert source_path.startswith("gs://")
-    from google.cloud import storage
-
-    bucket_path = Path(source_path[5:])
-    bucket, dir_path = str(bucket_path).split("/", 1)
-    client = storage.Client()
-    bucket = client.bucket(bucket)
-    blobs = client.list_blobs(bucket, prefix=f"{dir_path}/")
-    for blob in blobs:
-        dest_name = str(Path(dest_path) / Path(blob.name).name)
-        blob.download_to_filename(dest_name)

tools/inference/inference_pipeline.ipynb

@@ -47,7 +47,7 @@
    "outputs": [],
    "source": [
     "# Install required libraries\n",
-    "!pip install -q transformers\n",
+    "!pip install -q git+https://github.com/huggingface/transformers.git\n",
     "!pip install -q git+https://github.com/patil-suraj/vqgan-jax.git\n",
     "!pip install -q git+https://github.com/borisdayma/dalle-mini.git"
    ]
@@ -75,7 +75,7 @@
     "# Model references\n",
     "\n",
     "# dalle-mini\n",
-    "DALLE_MODEL = \"dalle-mini/dalle-mini/model-2vm4itcx:latest\"  # can be wandb artifact or 🤗 Hub or local folder or google bucket\n",
+    "DALLE_MODEL = \"dalle-mini/dalle-mini/model-3e2l7fxk:latest\"  # can be wandb artifact or 🤗 Hub or local folder or google bucket\n",
     "DALLE_COMMIT_ID = None\n",
     "\n",
     "# VQGAN model\n",
@@ -302,21 +302,7 @@
    },
    "outputs": [],
    "source": [
-    "tokenized_prompt = processor([prompt])\n",
-    "tokenized_prompt"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {
-    "id": "_Y5dqFj7prMQ"
-   },
-   "source": [
-    "Notes:\n",
-    "\n",
-    "* `0`: BOS, special token representing the beginning of a sequence\n",
-    "* `2`: EOS, special token representing the end of a sequence\n",
-    "* `1`: special token representing the padding of a sequence when requesting a specific length"
+    "tokenized_prompt = processor([prompt])"
    ]
   },
   {
@@ -459,13 +445,6 @@
     "    display(images[idx])\n",
     "    print(f\"Score: {logits[idx]:.2f}\\n\")"
    ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": []
   }
  ],
  "metadata": {

tools/train/config/medium/config.json

@@ -3,7 +3,6 @@
   "activation_function": "gelu",
   "attention_dropout": 0.0,
   "bos_token_id": 16385,
-  "classifier_dropout": 0.0,
   "d_model": 1408,
   "decoder_attention_heads": 16,
   "decoder_ffn_dim": 4096,

tools/train/config/mega/config.json

@@ -3,7 +3,6 @@
   "activation_function": "gelu",
   "attention_dropout": 0.0,
   "bos_token_id": 16385,
-  "classifier_dropout": 0.0,
   "d_model": 2048,
   "decoder_attention_heads": 32,
   "decoder_ffn_dim": 8192,

tools/train/config/micro/config.json

@@ -3,7 +3,6 @@
   "activation_function": "gelu",
   "attention_dropout": 0.0,
   "bos_token_id": 16385,
-  "classifier_dropout": 0.0,
   "d_model": 256,
   "decoder_attention_heads": 2,
   "decoder_ffn_dim": 256,

tools/train/config/mini/config.json

@@ -3,17 +3,14 @@
   "activation_function": "gelu",
   "attention_dropout": 0.0,
   "bos_token_id": 16385,
-  "classifier_dropout": 0.0,
   "d_model": 1024,
   "decoder_attention_heads": 16,
-  "decoder_ffn_dim": 4096,
-  "decoder_layerdrop": 0.0,
+  "decoder_ffn_dim": 2560,
   "decoder_layers": 12,
   "decoder_start_token_id": 16384,
   "dropout": 0.0,
   "encoder_attention_heads": 16,
-  "encoder_ffn_dim": 4096,
-  "encoder_layerdrop": 0.0,
+  "encoder_ffn_dim": 2560,
   "encoder_layers": 12,
   "encoder_vocab_size": 50264,
   "eos_token_id": 16385,

tools/train/train.py

@@ -18,7 +18,6 @@ Training DALL·E Mini.
 Script adapted from run_summarization_flax.py
 """
 
-import copy
 import io
 import logging
 import os
@@ -30,8 +29,10 @@ from pathlib import Path
 from typing import Any, Callable, NamedTuple, Optional
 
 import datasets
+import flax
 import jax
 import jax.numpy as jnp
+import jaxlib
 import numpy as np
 import optax
 import transformers
@@ -405,10 +406,14 @@ class TrainingArguments:
         default=False,
         metadata={"help": "Log model to wandb at `save_steps` frequency."},
     )
-    log_histograms: bool = field(
+    log_norm_steps: int = field(
+        default=True,
+        metadata={"help": "Log parameters and gradients norm at this frequency."},
+    )
+    log_histogram_steps: int = field(
         default=False,
         metadata={
-            "help": "Log parameters and gradients histograms. Slows down training."
+            "help": "Log parameters and gradients histograms at this frequency. Slows down training."
         },
     )
 
@@ -471,6 +476,8 @@ class TrainingArguments:
         ], f"Selected learning rate decay not supported: {self.lr_decay}"
         if self.per_device_eval_batch_size is None:
             self.per_device_eval_batch_size = self.per_device_train_batch_size
+        if self.log_norm_steps is True:
+            self.log_norm_steps = self.logging_steps
         if (
             os.path.exists(self.output_dir)
             and os.listdir(self.output_dir)
@@ -497,48 +504,6 @@ class TrainState(train_state.TrainState):
     train_samples: int = 0  # number of samples seen
 
 
-class MetricsLogger:
-    def __init__(self, step):
-        self.step = step
-        self.time = time.perf_counter()
-        self.state_dict = {}
-
-    def update_state_metrics(self, state):
-        """Update internal state metrics (logged at each call to be used as x-axis)"""
-        self.state_dict = {
-            f'train/{k.split("_")[-1]}': getattr(state, k)
-            for k in ["step", "epoch", "train_time", "train_samples"]
-        }
-        # timing metrics
-        new_step = int(state.step)
-        new_time = time.perf_counter()
-        if new_step > self.step:
-            time_per_step = (new_time - self.time) / (new_step - self.step)
-            self.step = new_step
-            self.time = new_time
-            self.state_dict["train/time_per_step"] = time_per_step
-
-    def log(self, metrics, prefix=None):
-        if jax.process_index() == 0:
-            log_metrics = {}
-            for k, v in metrics.items():
-                if prefix is not None:
-                    k = f"{prefix}/{k}"
-                if "_norm" in k:
-                    log_metrics[f"{k}/"] = unfreeze(v)
-                elif "_hist" in k:
-                    v = jax.tree_map(lambda x: jax.device_get(x), unfreeze(v))
-                    v = jax.tree_map(
-                        lambda x: wandb.Histogram(np_histogram=x),
-                        v,
-                        is_leaf=lambda x: isinstance(x, tuple),
-                    )
-                    log_metrics[f"{k}/"] = v
-                else:
-                    log_metrics[k] = v
-            wandb.log({**log_metrics, **self.state_dict})
-
-
 def main():
     # See all possible arguments by passing the --help flag to this script.
     parser = HfArgumentParser(
@@ -593,9 +558,7 @@ def main():
     # Set up our new model config
     if model_args.config_name:
         config = DalleBartConfig.from_pretrained(model_args.config_name)
-        # initializing params with gradient checkpointing creates issues
-        # we correctly set it later per training_args
-        config.gradient_checkpointing = False
+        config.gradient_checkpointing = training_args.gradient_checkpointing
     else:
         config = None
 
@@ -607,9 +570,7 @@ def main():
             seed=training_args.seed_model,
             dtype=getattr(jnp, model_args.dtype),
             abstract_init=True,  # we overwrite them with loaded checkpoint
-            # initializing params with gradient checkpointing creates issues
-            # we correctly set it later per training_args
-            gradient_checkpointing=False,
+            gradient_checkpointing=training_args.gradient_checkpointing,
         )
     else:
         model = DalleBart(
@@ -619,21 +580,6 @@ def main():
             abstract_init=True,
         )
 
-    # define model eval and train functions
-    eval_fn = model.__call__
-    if training_args.gradient_checkpointing:
-        remat_config = copy.deepcopy(model.config)
-        remat_config.gradient_checkpointing = True
-        remat_model = DalleBart(
-            remat_config,
-            seed=training_args.seed_model,
-            dtype=getattr(jnp, model_args.dtype),
-            init_weights=False,
-        )
-        train_fn = remat_model.__call__
-    else:
-        train_fn = model.__call__
-
     # get model metadata
     model_metadata = model_args.get_metadata()
 
@@ -708,8 +654,16 @@ def main():
                 "len_train_dataset": len_train_dataset,
                 "len_eval_dataset": len_eval_dataset,
                 "batch_size_per_step": batch_size_per_step,
-                "num_params": num_params,
+                "model": {"num_params": num_params, "config": model.config.to_dict()},
                 "num_devices": jax.device_count(),
+                "versions": {
+                    "jax": jax.__version__,
+                    "jaxlib": jaxlib.__version__,
+                    "flax": flax.__version__,
+                    "transformers": transformers.__version__,
+                    "datasets": datasets.__version__,
+                    "wandb": wandb.__version__,
+                },
             }
         )
 
@@ -719,7 +673,7 @@ def main():
         warmup_fn = optax.linear_schedule(
             init_value=0.0,
             end_value=training_args.learning_rate,
-            transition_steps=training_args.warmup_steps,
+            transition_steps=training_args.warmup_steps + 1,  # ensure not 0
         )
         # offset step when resuming
         if model_metadata.get("step", 0):
@@ -867,7 +821,7 @@ def main():
         epoch=None,
         train_time=None,
         train_samples=None,
-        apply_fn=train_fn,
+        apply_fn=model.__call__,
         tx=optimizer,
     )
 
@@ -880,13 +834,13 @@ def main():
             # params have not been initialized yet
             return model.init_weights()
 
-    with maps.mesh(mesh.devices, mesh.axis_names):
+    with mesh:
         logger.info("  Creating state")
         if not model_args.restore_state:
 
             def init_state(params):
                 return TrainState.create(
-                    apply_fn=train_fn,
+                    apply_fn=model.__call__,
                     tx=optimizer,
                     params=maybe_init_params(params),
                     dropout_rng=dropout_rng,
@@ -913,7 +867,7 @@ def main():
 
             def restore_state(params, opt_state):
                 return TrainState(
-                    apply_fn=train_fn,
+                    apply_fn=model.__call__,
                     tx=optimizer,
                     params=params,
                     opt_state=opt_state,
@@ -959,7 +913,7 @@ def main():
         )
 
     # Define gradient update step fn
-    def train_step(state, batch, delta_time):
+    def train_step(state, batch, train_time):
 
         # get a minibatch (one gradient accumulation slice)
         def get_minibatch(batch, grad_idx):
@@ -1048,36 +1002,45 @@ def main():
         state = state.apply_gradients(
             grads=grads,
             dropout_rng=dropout_rng,
-            train_time=state.train_time + delta_time,
+            train_time=train_time,
             train_samples=state.train_samples + batch_size_per_step,
         )
 
-        # get norm and histogram of grads and params
-        zeros_norm = jax.tree_map(lambda _: jnp.float32(0), state.params)
+        metrics = {
+            "loss": loss,
+            "learning_rate": learning_rate_fn(state.step),
+        }
 
-        def maybe_fn(fn, val, zeros):
+        def maybe_fn(fn, val, zeros, freq):
             """Call fn only if it is a logging step"""
             return jax.lax.cond(
-                state.step % training_args.logging_steps == 0,
+                state.step % freq == 0,
                 fn,
                 lambda _: zeros,
                 val,
             )
 
-        def norm(val):
-            return jax.tree_map(lambda x: jnp.linalg.norm(x), val)
+        if training_args.log_norm_steps:
+            zeros_norm = jax.tree_map(lambda _: jnp.float32(0), state.params)
 
-        gradients_norm = maybe_fn(norm, grads, zeros_norm)
-        params_norm = maybe_fn(norm, state.params, zeros_norm)
+            def norm(val):
+                return jax.tree_map(lambda x: jnp.linalg.norm(x), val)
 
-        metrics = {
-            "loss": loss,
-            "learning_rate": learning_rate_fn(state.step),
-            "gradients_norm": gradients_norm,
-            "params_norm": params_norm,
-        }
+            gradients_norm = maybe_fn(
+                norm, grads, zeros_norm, training_args.log_norm_steps
+            )
+            params_norm = maybe_fn(
+                norm, state.params, zeros_norm, training_args.log_norm_steps
+            )
 
-        if training_args.log_histograms:
+            metrics.update(
+                {
+                    "gradients_norm": gradients_norm,
+                    "params_norm": params_norm,
+                }
+            )
+
+        if training_args.log_histogram_steps:
             zeros_hist = jax.tree_map(
                 lambda _: jnp.histogram(jnp.zeros(1), density=True), state.params
             )
@@ -1085,8 +1048,12 @@ def main():
             def histogram(val):
                 return jax.tree_map(lambda x: jnp.histogram(x, density=True), val)
 
-            gradients_hist = maybe_fn(histogram, grads, zeros_hist)
-            params_hist = maybe_fn(histogram, state.params, zeros_hist)
+            gradients_hist = maybe_fn(
+                histogram, grads, zeros_hist, training_args.log_histogram_steps
+            )
+            params_hist = maybe_fn(
+                histogram, state.params, zeros_hist, training_args.log_histogram_steps
+            )
 
             metrics.update(
                 {
@@ -1101,7 +1068,7 @@ def main():
     def eval_step(state, batch):
         def compute_eval_loss(batch):
             batch, labels = batch.pop("labels")
-            logits = eval_fn(**batch, params=state.params, train=False)[0]
+            logits = model(**batch, params=state.params, train=False)[0]
             return loss_fn(logits, labels)
 
         if use_vmap_trick:
@@ -1134,13 +1101,73 @@ def main():
         out_axis_resources=None,
     )
 
+    # define metrics logger
+    class MetricsLogger:
+        def __init__(self, step):
+            # keep state
+            self.state_dict = {}
+            # estimate speed
+            self.step = step
+            self.time = time.perf_counter()
+            self.offset_time = 0.0
+
+        def update_state_metrics(self, state):
+            """Update internal state metrics (logged at each call to be used as x-axis)"""
+            self.state_dict = {
+                f'train/{k.split("_")[-1]}': state[k]
+                for k in ["step", "epoch", "train_time", "train_samples"]
+            }
+            # timing metrics
+            new_step = int(state["step"])
+            new_time = time.perf_counter()
+            if new_step > self.step:
+                # remove time for eval & save
+                delta_time = new_time - self.time - self.offset_time
+                self.offset_time = 0
+                time_per_step = delta_time / (new_step - self.step)
+                self.step = new_step
+                self.time = new_time
+                self.log_time("train_per_step", time_per_step, offset=False)
+                self.log_time("train_per_log", delta_time, offset=False)
+
+        def log_time(self, key, duration, offset=True):
+            wandb.log({f"time/{key}": duration, **self.state_dict})
+            if offset:
+                self.offset_time += duration
+
+        def log(self, metrics, prefix=None):
+            if jax.process_index() == 0:
+                log_metrics = {}
+                for k, v in metrics.items():
+                    if "_norm" in k:
+                        if self.step % training_args.log_norm_steps == 0:
+                            log_metrics[f"{k}/"] = unfreeze(v)
+                    elif "_hist" in k:
+                        if self.step % training_args.log_histogram_steps == 0:
+                            v = jax.tree_map(lambda x: jax.device_get(x), unfreeze(v))
+                            v = jax.tree_map(
+                                lambda x: wandb.Histogram(np_histogram=x),
+                                v,
+                                is_leaf=lambda x: isinstance(x, tuple),
+                            )
+                            log_metrics[f"{k}/"] = v
+                    else:
+                        if prefix is not None:
+                            k = f"{prefix}/{k}"
+                        log_metrics[k] = v
+                wandb.log({**log_metrics, **self.state_dict})
+
+    # keep local copy of state
+    local_state = {
+        k: jax.device_get(getattr(state, k)).item()
+        for k in ["step", "epoch", "train_time", "train_samples"]
+    }
     # init variables
-    last_time = time.perf_counter()
+    start_time = time.perf_counter() - local_state["train_time"]
     train_metrics = None
-    step = int(state.step)
-    metrics_logger = MetricsLogger(step)
+    metrics_logger = MetricsLogger(local_state["step"])
     epochs = tqdm(
-        range(state.epoch, num_epochs),
+        range(local_state["epoch"], num_epochs),
         desc=f"Epoch ... (1/{num_epochs})",
         position=0,
         disable=jax.process_index() > 0,
@@ -1149,6 +1176,7 @@ def main():
     def run_evaluation():
         # ======================== Evaluating ==============================
         if training_args.do_eval:
+            start_eval_time = time.perf_counter()
             eval_loader = dataset.dataloader("eval", eval_batch_size_per_step)
             eval_steps = (
                 len_eval_dataset // eval_batch_size_per_step
@@ -1195,6 +1223,7 @@ def main():
 
             # log metrics
             metrics_logger.log(eval_metrics, prefix="eval")
+            metrics_logger.log_time("eval", time.perf_counter() - start_eval_time)
 
             # Print metrics and update progress bar
             desc = f"Epoch... ({epoch + 1}/{num_epochs} | Eval Loss: {eval_metrics['loss']})"
@@ -1206,6 +1235,7 @@ def main():
     def run_save_model(state, eval_metrics=None):
         if jax.process_index() == 0:
 
+            start_save_time = time.perf_counter()
             output_dir = training_args.output_dir
             use_bucket = output_dir.startswith("gs://")
             if use_bucket:
@@ -1298,13 +1328,15 @@ def main():
                         f"{Path(training_args.output_dir) / 'opt_state.msgpack'}"
                     )
                 wandb.run.log_artifact(artifact_state)
+            metrics_logger.log_time("save_model", time.perf_counter() - start_save_time)
 
     logger.info("  Ready to start training")
-    with maps.mesh(mesh.devices, mesh.axis_names):
+    with mesh:
         for epoch in epochs:
             state.replace(epoch=epoch)
+            local_state["epoch"] = epoch
             # ======================== Training ================================
-            metrics_logger.update_state_metrics(state)
+            metrics_logger.update_state_metrics(local_state)
             metrics_logger.log({})
 
             # Generate an epoch by shuffling sampling indices from the train dataset
@@ -1323,9 +1355,7 @@ def main():
                 disable=jax.process_index() > 0,
             ):
                 # calculate delta time (we have a lag of one step but it's ok)
-                new_time = time.perf_counter()
-                delta_time = new_time - last_time
-                last_time = new_time
+                train_time = time.perf_counter() - start_time
 
                 # set correct shape to batch
                 # - add grad_step dim if gradient_accumulation_steps > 1
@@ -1353,18 +1383,23 @@ def main():
                 batch = freeze(batch)
 
                 # train step
-                state, train_metrics = p_train_step(state, batch, delta_time)
-                step += 1
-
-                if step % training_args.logging_steps == 0 and jax.process_index() == 0:
-                    metrics_logger.update_state_metrics(state)
+                state, train_metrics = p_train_step(state, batch, train_time)
+                local_state["step"] += 1
+                local_state["train_time"] = train_time
+                local_state["train_samples"] += batch_size_per_step
+
+                if (
+                    local_state["step"] % training_args.logging_steps == 0
+                    and jax.process_index() == 0
+                ):
+                    metrics_logger.update_state_metrics(local_state)
                     metrics_logger.log(train_metrics, prefix="train")
 
                 eval_metrics = None
-                if step % training_args.eval_steps == 0:
+                if local_state["step"] % training_args.eval_steps == 0:
                     eval_metrics = run_evaluation()
 
-                if step % training_args.save_steps == 0:
+                if local_state["step"] % training_args.save_steps == 0:
                     run_save_model(state, eval_metrics)
 
             # log final train metrics