Support MuParametrization and MuTransfer

protein_lm.yml

@@ -20,3 +20,4 @@ dependencies:
       - evaluate
       - pytest
       - fair-esm
+      - mup

protein_lm/modeling/__init__.py

@@ -0,0 +1 @@
+from models import *

protein_lm/modeling/models/__init__.py

@@ -0,0 +1 @@
+from apt import *

protein_lm/modeling/models/apt/__init__.py

@@ -0,0 +1,2 @@
+from config import APTConfig
+from model_pytorch import *

protein_lm/modeling/models/apt/config.py

@@ -11,11 +11,36 @@ def __init__(
         position_embedding="learned",
         tokenizer=None,
         max_sequence_length = 1024,
+        query_zero_init = True,
+        n_layer = None,
+        contact_prediction_head = False,
+        initializer_range = 0.02,
+        # whether to use MuParametrization
+        use_mup = False, 
+        # whether to initialize the output (readout) layer with zero-initialization
+        readout_zero_init = True, 
+        # the output layer multiplier if mup is used, see https://github.com/microsoft/mup/blob/19814971934ef91dd546f88e913fc963e096d11c/mup/layer.py#L56
+        mup_output_mult = 1.0,
+        width_mult_for_weights = 2.0,
+        # rope
+        rope_theta = 0.0,
+        rope_scaling_factor=1,
         **kwargs
     ):
         super().__init__(**kwargs)
         self.nn_model_type = "APT"
         self.position_embedding = position_embedding
         self.tokenizer = tokenizer
         self.max_sequence_length = max_sequence_length
+
+        self.use_mup = use_mup
+        self.query_zero_init = query_zero_init,
+        self.n_layer = n_layer
+        self.contact_prediction_head = contact_prediction_head
+        self.initializer_range = initializer_range
+        self.readout_zero_init = readout_zero_init
+        self.mup_output_mult = mup_output_mult
+        self.width_mult_for_weights = width_mult_for_weights
+        self.rope_theta = rope_theta
+        self.rope_scaling_factor = rope_scaling_factor
 

protein_lm/modeling/models/apt/model_pytorch.py

@@ -1,4 +1,5 @@
 from typing import Optional, Tuple, Union
+import math
 import torch
 from torch import nn
 from torch.nn import CrossEntropyLoss
@@ -8,6 +9,7 @@
 from transformers.pytorch_utils import Conv1D
 from transformers.activations import ACT2FN
 from transformers.utils import logging
+from mup import MuReadout, MuSharedReadout, normal_
 from protein_lm.modeling.utils.rotary_embedding import RotaryEmbedding
 from protein_lm.modeling.utils.rerope_embedding import RectifiedRotaryEmbedding
 from protein_lm.modeling.utils.alibi_embedding import create_alibi_tensor
@@ -16,6 +18,7 @@
 
 logger = logging.get_logger(__name__)
 
+
 class APTAttention(GPT2Attention):
     def __init__(self, config, is_cross_attention=False, layer_idx=None):
         super().__init__(config, is_cross_attention=is_cross_attention, layer_idx=layer_idx)
@@ -42,6 +45,13 @@ def __init__(self, config, is_cross_attention=False, layer_idx=None):
                 f" {self.num_heads})."
             )
 
+        # muP
+        self.use_mup = config.use_mup
+        self.attn_score = nn.Identity() # just for coordcheck
+        self.query = nn.Identity() # just for coordcheck
+        self.key = nn.Identity() # just for coordcheck
+        self.value = nn.Identity() # just for coordcheck
+
         self.scale_attn_weights = config.scale_attn_weights
         self.is_cross_attention = is_cross_attention
 
@@ -55,13 +65,20 @@ def __init__(self, config, is_cross_attention=False, layer_idx=None):
             self.q_attn = Conv1D(self.embed_dim, self.embed_dim)
         else:
             self.c_attn = Conv1D(3 * self.embed_dim, self.embed_dim)
+
         self.c_proj = Conv1D(self.embed_dim, self.embed_dim)
 
-        self.attn_dropout = nn.Dropout(config.attn_pdrop)
-        self.resid_dropout = nn.Dropout(config.resid_pdrop)
+        if self.use_mup:
+            self.attn_dropout = nn.Identity()
+            self.resid_dropout = nn.Identity()
+        else:
+            self.attn_dropout = nn.Dropout(config.attn_pdrop)
+            self.resid_dropout = nn.Dropout(config.resid_pdrop)
 
         self.pruned_heads = set()
 
+
+
         self.rot_emb=None
         if self.position_embedding == "rope":
             self.rot_emb=RotaryEmbedding(dim=self.head_dim)
@@ -72,15 +89,23 @@ def __init__(self, config, is_cross_attention=False, layer_idx=None):
         elif self.position_embedding=="dynamic_rope_scaling":
             self.rot_emb=LlamaDynamicNTKScalingRotaryEmbedding(dim=self.head_dim,max_position_embeddings=self.max_positions,scaling_factor=self.rope_scaling_factor,base=self.rope_theta)
 
-    
+
 
     def _attn(self, query, key, value, attention_mask=None, head_mask=None,alibi_bias=None):
         attn_weights = torch.matmul(query, key.transpose(-1, -2))
 
+        #muP
         if self.scale_attn_weights:
-            attn_weights = attn_weights / torch.full(
-                [], value.size(-1) ** 0.5, dtype=attn_weights.dtype, device=attn_weights.device
-            )
+            if self.use_mup:
+                attn_weights = attn_weights / torch.full(
+                    [], value.size(-1), dtype=attn_weights.dtype, device=attn_weights.device
+                )
+            else:
+                attn_weights = attn_weights / torch.full(
+                    [], value.size(-1) ** 0.5, dtype=attn_weights.dtype, device=attn_weights.device
+                )
+
+        attn_weights = self.attn_score(attn_weights)
 
         # Layer-wise attention scaling
         if self.scale_attn_by_inverse_layer_idx:
@@ -97,7 +122,7 @@ def _attn(self, query, key, value, attention_mask=None, head_mask=None,alibi_bia
             attn_weights = torch.where(causal_mask, attn_weights.to(attn_weights.dtype), mask_value)
         if alibi_bias is not None:
             attn_weights = attn_weights + alibi_bias[:,:,:attn_weights.size(-1)]
-        
+
         if attention_mask is not None:
             # Apply the attention mask
             attn_weights = attn_weights + attention_mask
@@ -150,7 +175,7 @@ def _upcast_and_reordered_attn(self, query, key, value, attention_mask=None, hea
 
         if alibi_bias is not None:
             attn_weights = attn_weights + alibi_bias[:,:,:attn_weights.size(-1)]
-        
+
         if attention_mask is not None:
             # Apply the attention mask
             attn_weights = attn_weights + attention_mask
@@ -171,7 +196,7 @@ def _upcast_and_reordered_attn(self, query, key, value, attention_mask=None, hea
 
         return attn_output, attn_weights
 
-    
+
     def forward(
         self,
         hidden_states: Optional[Tuple[torch.FloatTensor]],
@@ -202,11 +227,15 @@ def forward(
         query = self._split_heads(query, self.num_heads, self.head_dim)
         key = self._split_heads(key, self.num_heads, self.head_dim)
         value = self._split_heads(value, self.num_heads, self.head_dim)
-
+
+        query = self.query(query)
+        key = self.key(key)
+        value = self.value(value)
+
         kv_seq_len=key.shape[-2]
         if layer_past is not None:
             kv_seq_len+=layer_past[0].shape[-2]
-      
+
         # Apply rope embedding to query and key
         if self.rot_emb:
             bsz, q_len, _ = hidden_states.size()
@@ -225,7 +254,6 @@ def forward(
             key = torch.cat((past_key, key), dim=-2)
             value = torch.cat((past_value, value), dim=-2)
 
-
         if use_cache is True:
             present = (key, value)
         else:
@@ -251,10 +279,20 @@ class APTMLP(nn.Module):
     def __init__(self, intermediate_size, config):
         super().__init__()
         embed_dim = config.hidden_size
+
+        #muP
+        use_mup = config.use_mup
+
         self.c_fc = Conv1D(intermediate_size, embed_dim)
+
         self.c_proj = Conv1D(embed_dim, intermediate_size)
+
         self.act = ACT2FN[config.activation_function]
-        self.dropout = nn.Dropout(config.resid_pdrop)
+
+        if use_mup:
+            self.dropout = nn.Identity()
+        else:
+            self.dropout = nn.Dropout(config.resid_pdrop)
 
     def forward(self, hidden_states: Optional[Tuple[torch.FloatTensor]]) -> torch.FloatTensor:
         hidden_states = self.c_fc(hidden_states)
@@ -270,6 +308,9 @@ def __init__(self, config, layer_idx=None):
         hidden_size = config.hidden_size
         inner_dim = config.n_inner if config.n_inner is not None else 4 * hidden_size
 
+        #muP
+        self.use_mup = config.use_mup
+
         self.ln_1 = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
         self.attn = APTAttention(config, layer_idx=layer_idx)
         self.ln_2 = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
@@ -354,23 +395,32 @@ def __init__(self, config):
         super().__init__(config)
 
         self.embed_dim = config.hidden_size
+        use_mup = config.use_mup
 
         self.wte = nn.Embedding(config.vocab_size, self.embed_dim)
+
         self.position_embedding = config.position_embedding if hasattr(config, "position_embedding") else "learned"
-        
+
         if self.position_embedding=="learned" or self.position_embedding == 'rope' or self.position_embedding == 'rerope' or self.position_embedding=="linear_rope_scaling" or self.position_embedding =="dynamic_rope_scaling":
             self.wpe = nn.Embedding(config.max_position_embeddings, self.embed_dim)
             self.alibi = None
         elif self.position_embedding=="alibi":
+            #muP TO DO: check proper behavior in alibi case
             maxpos = config.n_positions
             attn_heads = config.n_head
             alibi = create_alibi_tensor(attn_heads,maxpos)
             self.register_buffer('alibi',alibi)
         else:
             raise Exception(f'position_embedding {self.position_embedding} not supported. Please select one of learned, rope, rerope, linear rope, dynamic rope or alibi')
-
-        self.drop = nn.Dropout(config.embd_pdrop)
+
+        #muP
+        if use_mup:
+            self.drop = nn.Identity()
+        else:
+            self.drop = nn.Dropout(config.embd_pdrop)
+
         self.h = nn.ModuleList([APTBlock(config, layer_idx=i) for i in range(config.num_hidden_layers)])
+
         self.ln_f = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_epsilon)
 
         # Model parallel
@@ -477,7 +527,7 @@ def forward(
             hidden_states = inputs_embeds + position_embeds
         else:
             hidden_states = inputs_embeds
-        
+
 
         if token_type_ids is not None:
             token_type_embeds = self.wte(token_type_ids)
@@ -593,19 +643,78 @@ class APTLMHeadModel(GPT2PreTrainedModel):
     def __init__(self, config):
         super().__init__(config)
         self.transformer = APTModel(config)
-        self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
+
+        # muP
+        # TO DO: look into weight tying
+        # TO DO: if weight tying is used, APTMuSharedReadout with the proper tied weight should be used instead
+        self.lm_head = MuReadout(config.n_embd,
+                                 config.vocab_size,
+                                 bias=False,
+                                 readout_zero_init=config.readout_zero_init,
+                                 output_mult=config.output_mult)
+
+        # mup
+        # note that this has to be run after mup.set_base_shape for it to work
+        # see https://github.com/microsoft/mup#basic-usage
+        # not sure if this is required here
+        self.apply(self._init_weights)
 
         # Model parallel
         self.model_parallel = False
         self.device_map = None
 
-        self.contact_head=ContactPredictionHead(config.num_hidden_layers * config.num_attention_heads,
-                                            prepend_bos=True,
-                                            append_eos=True,
-                                            eos_idx=2)        
+        # mup implementation does not currently support this
+        if config.contact_prediction_head:
+            self.contact_head=ContactPredictionHead(config.num_hidden_layers * config.num_attention_heads,
+                                                prepend_bos=True,
+                                                append_eos=True,
+                                                eos_idx=2)        
 
         # Initialize weights and apply final processing
         self.post_init()
+
+    # mup
+    # general function for mup-specific weight initialization
+    def _init_weights(self, module):
+        if isinstance(module, (MuReadout, MuSharedReadout)) and self.config.readout_zero_init:
+            module.weight.data.zero_()
+        elif isinstance(module, (nn.Linear, Conv1D)):
+            if hasattr(module.weight, 'infshape'):
+                normal_(module.weight, mean=0.0, std=self.config.initializer_range)
+            else:
+                module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+        elif isinstance(module, nn.LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+
+        if isinstance(module, APTAttention):
+            if hasattr(module, "q_attn"):
+                # cross attention case
+                if self.config.query_zero_init:
+                    # q_attn same as last third of c_attn in no cross attention case -- zero initialization
+                    self.q_attn.weight.data = 0
+            else:
+                if self.config.query_zero_init:
+                    _, fanout = module.c_attn.weight.shape
+                    assert fanout % 3 == 0
+                    module.c_attn.weight.data[:, :fanout//3] = 0
+
+        depth_std = self.config.initializer_range / math.sqrt(2 * self.config.n_layer)
+        for name, p in module.named_parameters():
+            if "c_proj" in name and "weight" in name:
+                # Special Scaled Initialization --> There are 2 Layer Norms per Transformer Block
+                if hasattr(p, 'infshape'):
+                    normal_(p, mean=0.0, std=depth_std)
+                else:
+                    p.data.normal_(mean=0.0, std=depth_std)
+
+
     def forward(
         self,
         input_ids: Optional[torch.LongTensor] = None,

protein_lm/modeling/scripts/train.py

@@ -53,6 +53,8 @@ def train(
             config_dict["wandb"],
         )
 
+    # TO DO: add support for mup's optimizers in case use_mup is used, see e.g. https://github.com/microsoft/mup/blob/19814971934ef91dd546f88e913fc963e096d11c/mup/optim.py
+    # available via mup.optim
     trainer = Trainer(
         model=model,
         args=training_args,