Upload MDLM

README.md

@@ -0,0 +1,199 @@
+---
+library_name: transformers
+tags: []
+---
+
+# Model Card for Model ID
+
+<!-- Provide a quick summary of what the model is/does. -->
+
+
+
+## Model Details
+
+### Model Description
+
+<!-- Provide a longer summary of what this model is. -->
+
+This is the model card of a 🤗 transformers model that has been pushed on the Hub. This model card has been automatically generated.
+
+- **Developed by:** [More Information Needed]
+- **Funded by [optional]:** [More Information Needed]
+- **Shared by [optional]:** [More Information Needed]
+- **Model type:** [More Information Needed]
+- **Language(s) (NLP):** [More Information Needed]
+- **License:** [More Information Needed]
+- **Finetuned from model [optional]:** [More Information Needed]
+
+### Model Sources [optional]
+
+<!-- Provide the basic links for the model. -->
+
+- **Repository:** [More Information Needed]
+- **Paper [optional]:** [More Information Needed]
+- **Demo [optional]:** [More Information Needed]
+
+## Uses
+
+<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
+
+### Direct Use
+
+<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
+
+[More Information Needed]
+
+### Downstream Use [optional]
+
+<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
+
+[More Information Needed]
+
+### Out-of-Scope Use
+
+<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
+
+[More Information Needed]
+
+## Bias, Risks, and Limitations
+
+<!-- This section is meant to convey both technical and sociotechnical limitations. -->
+
+[More Information Needed]
+
+### Recommendations
+
+<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
+
+Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
+
+## How to Get Started with the Model
+
+Use the code below to get started with the model.
+
+[More Information Needed]
+
+## Training Details
+
+### Training Data
+
+<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
+
+[More Information Needed]
+
+### Training Procedure
+
+<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
+
+#### Preprocessing [optional]
+
+[More Information Needed]
+
+
+#### Training Hyperparameters
+
+- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
+
+#### Speeds, Sizes, Times [optional]
+
+<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
+
+[More Information Needed]
+
+## Evaluation
+
+<!-- This section describes the evaluation protocols and provides the results. -->
+
+### Testing Data, Factors & Metrics
+
+#### Testing Data
+
+<!-- This should link to a Dataset Card if possible. -->
+
+[More Information Needed]
+
+#### Factors
+
+<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
+
+[More Information Needed]
+
+#### Metrics
+
+<!-- These are the evaluation metrics being used, ideally with a description of why. -->
+
+[More Information Needed]
+
+### Results
+
+[More Information Needed]
+
+#### Summary
+
+
+
+## Model Examination [optional]
+
+<!-- Relevant interpretability work for the model goes here -->
+
+[More Information Needed]
+
+## Environmental Impact
+
+<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
+
+Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
+
+- **Hardware Type:** [More Information Needed]
+- **Hours used:** [More Information Needed]
+- **Cloud Provider:** [More Information Needed]
+- **Compute Region:** [More Information Needed]
+- **Carbon Emitted:** [More Information Needed]
+
+## Technical Specifications [optional]
+
+### Model Architecture and Objective
+
+[More Information Needed]
+
+### Compute Infrastructure
+
+[More Information Needed]
+
+#### Hardware
+
+[More Information Needed]
+
+#### Software
+
+[More Information Needed]
+
+## Citation [optional]
+
+<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
+
+**BibTeX:**
+
+[More Information Needed]
+
+**APA:**
+
+[More Information Needed]
+
+## Glossary [optional]
+
+<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
+
+[More Information Needed]
+
+## More Information [optional]
+
+[More Information Needed]
+
+## Model Card Authors [optional]
+
+[More Information Needed]
+
+## Model Card Contact
+
+[More Information Needed]

config.json

@@ -0,0 +1,22 @@
+{
+  "_name_or_path": "kuleshov-group/mdlm-owt",
+  "architectures": [
+    "MDLM"
+  ],
+  "auto_map": {
+    "AutoConfig": "configuration_mdlm.MDLMConfig",
+    "AutoModelForMaskedLM": "modeling_mdlm.MDLM"
+  },
+  "cond_dim": 128,
+  "dropout": 0.1,
+  "hidden_dim": 768,
+  "model_length": 1024,
+  "model_type": "mdlm",
+  "n_blocks": 12,
+  "n_heads": 12,
+  "return_dict": false,
+  "time_conditioning": false,
+  "torch_dtype": "float32",
+  "transformers_version": "4.38.2",
+  "vocab_size": 50258
+}

configuration_mdlm.py

@@ -0,0 +1,31 @@
+"""MDLM config for Hugging Face.
+
+"""
+
+import transformers
+
+
+class MDLMConfig(transformers.PretrainedConfig):
+  """Hugging Face configuration class for MDLM."""
+  model_type = "mdlm"
+
+  def __init__(
+    self,
+    vocab_size: int = 50258,
+    model_length: int = 1024,
+    hidden_dim: int = 768,
+    cond_dim: int = 129,
+    n_blocks: int = 12,
+    n_heads: int = 12,
+    dropout: float = 0.1,
+    time_conditioning: bool = False,
+    ** kwargs):
+    super().__init__(**kwargs)
+    self.vocab_size = vocab_size
+    self.model_length = model_length
+    self.hidden_dim = hidden_dim
+    self.cond_dim = cond_dim
+    self.n_blocks = n_blocks
+    self.n_heads = n_heads
+    self.dropout = dropout
+    self.time_conditioning = time_conditioning

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:47149e73f7552f39ea9776dbe74d925d25237bcf2ed2e2ec03cdff9d51c82aa4
+size 678522728

modeling_mdlm.py

@@ -0,0 +1,450 @@
+"""MDLM model for Hugging Face.
+
+"""
+import itertools
+import math
+import typing
+
+import flash_attn
+import flash_attn.layers.rotary
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import transformers
+from einops import rearrange
+from transformers import modeling_outputs
+
+from .configuration_mdlm import MDLMConfig
+
+# Flags required to enable jit fusion kernels
+torch._C._jit_set_profiling_mode(False)
+torch._C._jit_set_profiling_executor(False)
+torch._C._jit_override_can_fuse_on_cpu(True)
+torch._C._jit_override_can_fuse_on_gpu(True)
+
+
+def bias_dropout_add_scale(
+    x: torch.Tensor,
+    bias: typing.Optional[torch.Tensor],
+    scale: torch.Tensor,
+    residual: typing.Optional[torch.Tensor],
+    prob: float,
+    training: bool) -> torch.Tensor:
+  if bias is not None:
+    out = scale * F.dropout(x + bias, p=prob,
+                            training=training)
+  else:
+    out = scale * F.dropout(x, p=prob, training=training)
+
+  if residual is not None:
+    out = residual + out
+  return out
+
+
+def get_bias_dropout_add_scale(training):
+  def _bias_dropout_add(x, bias, scale, residual, prob):
+    return bias_dropout_add_scale(
+      x, bias, scale, residual, prob, training)
+
+  return _bias_dropout_add
+
+
+# function overload
+def modulate(x: torch.Tensor,
+             shift: torch.Tensor,
+             scale: torch.Tensor) -> torch.Tensor:
+  return x * (1 + scale) + shift
+
+
+@torch.jit.script
+def bias_dropout_add_scale_fused_train(
+    x: torch.Tensor,
+    bias: typing.Optional[torch.Tensor],
+    scale: torch.Tensor,
+    residual: typing.Optional[torch.Tensor],
+    prob: float) -> torch.Tensor:
+  return bias_dropout_add_scale(
+    x, bias, scale, residual, prob, True)
+
+
+@torch.jit.script
+def bias_dropout_add_scale_fused_inference(
+    x: torch.Tensor,
+    bias: typing.Optional[torch.Tensor],
+    scale: torch.Tensor,
+    residual: typing.Optional[torch.Tensor],
+    prob: float) -> torch.Tensor:
+  return bias_dropout_add_scale(
+    x, bias, scale, residual, prob, False)
+
+
+@torch.jit.script
+def modulate_fused(x: torch.Tensor,
+                   shift: torch.Tensor,
+                   scale: torch.Tensor) -> torch.Tensor:
+  return modulate(x, shift, scale)
+
+
+class Rotary(torch.nn.Module):
+  def __init__(self, dim, base=10_000):
+    super().__init__()
+    inv_freq = 1.0 / (
+          base ** (torch.arange(0, dim, 2).float() / dim))
+    self.register_buffer('inv_freq', inv_freq)
+    self.seq_len_cached = None
+    self.cos_cached = None
+    self.sin_cached = None
+
+  def forward(self, x, seq_dim=1):
+    seq_len = x.shape[seq_dim]
+    if seq_len != self.seq_len_cached:
+      self.seq_len_cached = seq_len
+      t = torch.arange(x.shape[seq_dim],
+                       device=x.device).type_as(
+        self.inv_freq)
+      freqs = torch.einsum("i,j->ij", t,
+                           self.inv_freq.clone())
+      emb = torch.cat((freqs, freqs), dim=-1).to(x.device)
+      # dims are: batch, seq_len, qkv, head, dim
+      self.cos_cached = emb.cos()[None, :, None, None,
+                        :].repeat(1, 1, 3, 1, 1)
+      self.sin_cached = emb.sin()[None, :, None, None,
+                        :].repeat(1, 1, 3, 1, 1)
+      # This makes the transformation on v an identity.
+      self.cos_cached[:, :, 2, :, :].fill_(1.)
+      self.sin_cached[:, :, 2, :, :].fill_(0.)
+
+    return self.cos_cached, self.sin_cached
+
+
+def rotate_half(x):
+  x1, x2 = x[..., : x.shape[-1] // 2], x[...,
+                                       x.shape[-1] // 2:]
+  return torch.cat((-x2, x1), dim=-1)
+
+
+def apply_rotary_pos_emb(qkv, cos, sin):
+  cos = cos[0, :, 0, 0, :cos.shape[-1] // 2]
+  sin = sin[0, :, 0, 0, :sin.shape[-1] // 2]
+  return flash_attn.layers.rotary.apply_rotary_emb_qkv_(qkv,
+                                                        cos,
+                                                        sin)
+# function overload
+def modulate(x, shift, scale):
+  return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1)
+
+
+############################################################
+#                          Layers                          #
+############################################################
+class LayerNorm(nn.Module):
+  def __init__(self, dim):
+    super().__init__()
+    self.weight = nn.Parameter(torch.ones([dim]))
+    self.dim = dim
+
+  def forward(self, x):
+    with torch.cuda.amp.autocast(enabled=False):
+      x = F.layer_norm(x.float(), [self.dim])
+    return x * self.weight[None, None, :]
+
+
+def residual_linear(x, W, x_skip, residual_scale):
+  """x_skip + residual_scale * W @ x"""
+  dim_out, dim_in = W.shape[0], W.shape[1]
+  return torch.addmm(
+    x_skip.view(-1, dim_out),
+    x.view(-1, dim_in),
+    W.T,
+    alpha=residual_scale).view(*x.shape[:-1], dim_out)
+
+
+############################################################
+#     Embedding Layers for Timesteps and Class Labels      #
+############################################################
+class TimestepEmbedder(nn.Module):
+  """
+  Embeds scalar timesteps into vector representations.
+  """
+
+  def __init__(self, hidden_size,
+               frequency_embedding_size=256):
+    super().__init__()
+    self.mlp = nn.Sequential(
+      nn.Linear(frequency_embedding_size, hidden_size,
+                bias=True),
+      nn.SiLU(),
+      nn.Linear(hidden_size, hidden_size, bias=True))
+    self.frequency_embedding_size = frequency_embedding_size
+
+  @staticmethod
+  def timestep_embedding(t, dim, max_period=10000):
+    """
+    Create sinusoidal timestep embeddings.
+    :param t: a 1-D Tensor of N indices, one per batch
+              element. These may be fractional.
+    :param dim: the dimension of the output.
+    :param max_period: controls the minimum frequency of the
+    embeddings.
+    :return: an (N, D) Tensor of positional embeddings.
+    """
+    # https://github.com/openai/glide-text2im/blob/main/glide_text2im/nn.py
+    half = dim // 2
+    freqs = torch.exp(
+      - math.log(max_period)
+      * torch.arange(start=0, end=half, dtype=torch.float32)
+      / half).to(device=t.device)
+    args = t[:, None].float() * freqs[None]
+    embedding = torch.cat(
+      [torch.cos(args), torch.sin(args)], dim=-1)
+    if dim % 2:
+      embedding = torch.cat(
+        [embedding,
+         torch.zeros_like(embedding[:, :1])], dim=-1)
+    return embedding
+
+  def forward(self, t):
+    t_freq = self.timestep_embedding(t,
+                                     self.frequency_embedding_size)
+    t_emb = self.mlp(t_freq)
+    return t_emb
+
+
+class LabelEmbedder(nn.Module):
+  """Embeds class labels into vector representations.
+
+  Also handles label dropout for classifier-free guidance.
+  """
+
+  def __init__(self, num_classes, cond_size):
+    super().__init__()
+    self.embedding_table = nn.Embedding(num_classes + 1,
+                                        cond_size)
+    self.num_classes = num_classes
+
+    # TODO think of initializing with 0.02 std deviation like in original DiT paper
+
+  def forward(self, labels):
+    embeddings = self.embedding_table(labels)
+    return embeddings
+
+
+############################################################
+#                        Core Model                        #
+############################################################
+
+
+class DDiTBlock(nn.Module):
+  def __init__(self, dim, n_heads, cond_dim, mlp_ratio=4,
+               dropout=0.1):
+    super().__init__()
+    self.n_heads = n_heads
+
+    self.norm1 = LayerNorm(dim)
+    self.attn_qkv = nn.Linear(dim, 3 * dim, bias=False)
+    self.attn_out = nn.Linear(dim, dim, bias=False)
+    self.dropout1 = nn.Dropout(dropout)
+
+    self.norm2 = LayerNorm(dim)
+    self.mlp = nn.Sequential(
+      nn.Linear(dim, mlp_ratio * dim, bias=True),
+      nn.GELU(approximate='tanh'),
+      nn.Linear(mlp_ratio * dim, dim, bias=True))
+    self.dropout2 = nn.Dropout(dropout)
+    self.dropout = dropout
+
+    self.adaLN_modulation = nn.Linear(cond_dim, 6 * dim,
+                                      bias=True)
+    self.adaLN_modulation.weight.data.zero_()
+    self.adaLN_modulation.bias.data.zero_()
+
+  def _get_bias_dropout_scale(self):
+    if self.training:
+      return bias_dropout_add_scale_fused_train
+    else:
+      return bias_dropout_add_scale_fused_inference
+
+  def forward(self, x, rotary_cos_sin, c, seqlens=None):
+    batch_size, seq_len = x.shape[0], x.shape[1]
+
+    bias_dropout_scale_fn = self._get_bias_dropout_scale()
+
+    (shift_msa, scale_msa, gate_msa, shift_mlp,
+     scale_mlp, gate_mlp) = self.adaLN_modulation(c)[:,
+                            None].chunk(6, dim=2)
+
+    # attention operation
+    x_skip = x
+    x = modulate_fused(self.norm1(x), shift_msa, scale_msa)
+
+    qkv = self.attn_qkv(x)
+    qkv = rearrange(qkv,
+                    'b s (three h d) -> b s three h d',
+                    three=3,
+                    h=self.n_heads)
+    with torch.cuda.amp.autocast(enabled=False):
+      cos, sin = rotary_cos_sin
+      qkv = apply_rotary_pos_emb(
+        qkv, cos.to(qkv.dtype), sin.to(qkv.dtype))
+    qkv = rearrange(qkv, 'b s ... -> (b s) ...')
+    if seqlens is None:
+      cu_seqlens = torch.arange(
+        0, (batch_size + 1) * seq_len, step=seq_len,
+        dtype=torch.int32, device=qkv.device)
+    else:
+      cu_seqlens = seqlens.cumsum(-1)
+    x = flash_attn.flash_attn_interface.flash_attn_varlen_qkvpacked_func(
+      qkv, cu_seqlens, seq_len, 0., causal=False)
+
+    x = rearrange(x, '(b s) h d -> b s (h d)', b=batch_size)
+
+    x = bias_dropout_scale_fn(self.attn_out(x),
+                              None,
+                              gate_msa,
+                              x_skip,
+                              self.dropout)
+
+    # mlp operation
+    x = bias_dropout_scale_fn(
+      self.mlp(modulate_fused(
+        self.norm2(x), shift_mlp, scale_mlp)),
+      None, gate_mlp, x, self.dropout)
+    return x
+
+
+class EmbeddingLayer(nn.Module):
+  def __init__(self, dim, vocab_dim):
+    super().__init__()
+    self.embedding = nn.Parameter(
+      torch.empty((vocab_dim, dim)))
+    torch.nn.init.kaiming_uniform_(self.embedding,
+                                   a=math.sqrt(5))
+
+  def forward(self, x):
+    return self.embedding[x]
+
+
+class DDitFinalLayer(nn.Module):
+  def __init__(self, hidden_size, out_channels, cond_dim):
+    super().__init__()
+    self.norm_final = LayerNorm(hidden_size)
+    self.linear = nn.Linear(hidden_size, out_channels)
+    self.linear.weight.data.zero_()
+    self.linear.bias.data.zero_()
+
+    self.adaLN_modulation = nn.Linear(cond_dim,
+                                      2 * hidden_size,
+                                      bias=True)
+    self.adaLN_modulation.weight.data.zero_()
+    self.adaLN_modulation.bias.data.zero_()
+
+  def forward(self, x, c):
+    shift, scale = self.adaLN_modulation(c)[:, None].chunk(
+      2, dim=2)
+    x = modulate_fused(self.norm_final(x), shift, scale)
+    x = self.linear(x)
+    return x
+
+
+class DITBackbone(nn.Module):
+  def __init__(
+      self,
+      config: MDLMConfig):
+    super().__init__()
+
+    self.config = config
+    self.vocab_size = config.vocab_size
+
+    self.vocab_embed = EmbeddingLayer(
+      config.hidden_dim,
+      config.vocab_size)
+    self.sigma_map = TimestepEmbedder(
+      config.cond_dim)
+    self.rotary_emb = Rotary(
+      config.hidden_dim // config.n_heads)
+
+    blocks = []
+    for _ in range(config.n_blocks):
+      blocks.append(DDiTBlock(config.hidden_dim,
+                              config.n_heads,
+                              config.cond_dim,
+                              dropout=config.dropout))
+    self.blocks = nn.ModuleList(blocks)
+
+    self.output_layer = DDitFinalLayer(
+      config.hidden_dim,
+      config.vocab_size,
+      config.cond_dim)
+
+  def _get_bias_dropout_scale(self):
+    if self.training:
+      return bias_dropout_add_scale_fused_train
+    else:
+      return bias_dropout_add_scale_fused_inference
+
+  def forward(self, indices, sigma,
+              output_hidden_states=False):
+    if not self.config.time_conditioning:
+      sigma = torch.zeros_like(sigma)
+    all_hidden_states = []
+    x = self.vocab_embed(indices)
+    if output_hidden_states:
+      all_hidden_states.append(x)
+    c = F.silu(self.sigma_map(sigma))
+
+    rotary_cos_sin = self.rotary_emb(x)
+
+    with torch.cuda.amp.autocast(dtype=torch.bfloat16):
+      for i in range(len(self.blocks)):
+        x = self.blocks[i](x, rotary_cos_sin, c,
+                           seqlens=None)
+        if output_hidden_states:
+          all_hidden_states.append(x)
+      logits = self.output_layer(x, c)
+    return logits, all_hidden_states
+
+class MDLM(transformers.PreTrainedModel):
+  """HF-compatible model."""
+  config_class = MDLMConfig
+  base_model_prefix = "mdlm"
+
+  def __init__(
+    self,
+    config: MDLMConfig):
+    super().__init__(config)
+    self.backbone = DITBackbone(config)
+
+  def forward(
+      self,
+      input_ids: torch.LongTensor = None,
+      timesteps: torch.FloatTensor = None,
+      output_hidden_states: typing.Optional[bool] = None,
+      return_dict: typing.Optional[bool] = None,
+  ) -> typing.Union[
+    torch.Tensor, typing.Tuple,
+    modeling_outputs.MaskedLMOutput]:
+    """HF-compatible forward method."""
+    output_hidden_states = (
+      output_hidden_states
+      if output_hidden_states is not None
+      else self.config.output_hidden_states
+    )
+    return_dict = return_dict \
+      if return_dict is not None \
+      else self.config.use_return_dict
+
+    logits, all_hidden_states = self.backbone(
+      indices=input_ids,
+      sigma=timesteps,
+      output_hidden_states=output_hidden_states
+    )
+    if return_dict:
+      return modeling_outputs.MaskedLMOutput(
+        logits=logits,
+        hidden_states=all_hidden_states if output_hidden_states else None,
+        loss=None
+      )
+    elif output_hidden_states:
+      return logits, all_hidden_states
+    else:
+      return logits