Merge pull request #50 from alxndrTL/mup-2

muP for Mamba and Mamba-2

Author: alxndrTL

mambapy/lm.py

@@ -41,12 +41,67 @@ def __init__(self, model_config: Union[MambaConfig, Mamba2Config], vocab_size: i
         self.norm_f = RMSNorm(self.config.d_model, self.config.rms_norm_eps, self.config.mup)
 
         self.lm_head = nn.Linear(self.config.d_model, vocab_size, bias=False)
-        self.embedding.weight = self.lm_head.weight
+        # self.embedding.weight = self.lm_head.weight # weight-tying disabled
+
+        # muP custom initialization
+        if self.config.mup and isinstance(self.config, MambaConfig):
+            for pn, p in self.named_parameters():
+                if any(pn.endswith(w) for w in ['mixer.in_proj.weight', 'mixer.x_proj.weight', 'mixer.dt_proj.weight', 'mixer.out_proj.weight']): # # "hidden weights"
+                    std = self.config.base_std
+
+                    if 'mixer.out_proj.weight' in pn:
+                        std = std / math.sqrt(2 * self.config.n_layers) # scale down std of layers which projects onto the residual stream (not muP related)
+
+                    if 'mixer.dt_proj.weight' in pn:
+                        std = self.config.dt_rank**-0.5 * self.config.dt_scale
+                    torch.nn.init.normal_(p, mean=0.0, std=std / math.sqrt(self.config.mup_width_mult))
+                elif 'mixer.conv1d.weight' in pn:
+                    torch.nn.init.zeros_(p)
+                elif pn == "embedding.weight":
+                    torch.nn.init.normal_(p, mean=0.0, std=self.config.base_std)
+                elif pn == "lm_head.weight":
+                    torch.nn.init.zeros_(p)
+                elif any(pn.endswith(w) for w in ['mixer.A_log', 'mixer.D']):
+                    # keep Mamba default init for these params
+                    pass
+                else:
+                    # here, we only have biases
+                    assert p.dim() == 1, f"a 2d param ({pn}) has not been filtered out for init. please check."
+
+                    if ("in_proj.bias" in pn) or ("out_proj.bias" in pn):
+                        torch.nn.init.zeros_(p)
+        
+        elif self.config.mup and isinstance(self.config, Mamba2Config):
+            for pn, p in self.named_parameters():
+                
+                if any(pn.endswith(w) for w in ['mixer.in_proj.weight', 'mixer.out_proj.weight']): # # "hidden weights"
+                    std = self.config.base_std
+
+                    if 'mixer.out_proj.weight' in pn:
+                        std = std / math.sqrt(2 * self.config.n_layers) # scale down std of layers which projects onto the residual stream (not muP related)
+
+                    torch.nn.init.normal_(p, mean=0.0, std=std / math.sqrt(self.config.mup_width_mult))               
+                elif 'mixer.conv1d.weight' in pn:
+                    torch.nn.init.zeros_(p)
+                elif pn == "embedding.weight":
+                    torch.nn.init.normal_(p, mean=0.0, std=self.config.base_std)
+                elif pn == "lm_head.weight":
+                    torch.nn.init.zeros_(p)
+                elif any(pn.endswith(w) for w in ['mixer.A_log', 'mixer.D', 'mixer.dt_bias']):
+                    # keep Mamba default init for these params
+                    pass
+                else:
+                    # here, we only have biases
+                    assert p.dim() == 1, f"a 2d param ({pn}) has not been filtered out for init. please check."
 
-        self.apply(self._init_weights)
-        for pn, p in self.named_parameters():
-            if pn.endswith('fc_3.weight') or pn.endswith('c_proj.weight'):
-                torch.nn.init.normal_(p, mean=0.0, std=self.config.base_std/math.sqrt(2 * self.config.n_layers))
+                    if ("in_proj.bias" in pn) or ("out_proj.bias" in pn):
+                        torch.nn.init.zeros_(p)
+
+        else:
+            self.apply(self._init_weights)
+            for pn, p in self.named_parameters():
+                if pn.endswith('mixer.out_proj.weight'):
+                    torch.nn.init.normal_(p, mean=0.0, std=self.config.base_std/math.sqrt(2 * self.config.n_layers))
 
     def forward(self, tokens):
         # tokens : (B, L)
@@ -142,13 +197,46 @@ def configure_optimizers(self, weight_decay, learning_rate, betas, device_type):
 
         # create optim groups. Any parameters that is 2D will be weight decayed, otherwise no.
         # i.e. all weight tensors in matmuls + embeddings decay, all biases and layernorms don't.
-        decay_params = [p for n, p in param_dict.items() if p.dim() >= 2]
-        nodecay_params = [p for n, p in param_dict.items() if p.dim() < 2]
+        if self.config.mup and isinstance(self.config, MambaConfig):
+            mup_params_keys = set([pn for pn in param_dict.keys() if any(pn.endswith(w) for w in ['mixer.in_proj.weight', 'mixer.x_proj.weight', 'mixer.dt_proj.weight', 'mixer.out_proj.weight'])])
+            
+            dim2_params_keys = set([pn for pn in param_dict.keys() if param_dict[pn].dim() >= 2])
+            dim2_params_keys = dim2_params_keys.difference(mup_params_keys)
+
+            mup_parameters = [p for n, p in param_dict.items() if n in mup_params_keys]
+            decay_params = [p for n, p in param_dict.items() if n in dim2_params_keys]
+            nodecay_params = [p for n, p in param_dict.items() if p.dim() < 2] # biases and D
+
+            optim_groups = [
+                {'params': mup_parameters, 'weight_decay': weight_decay * self.config.mup_width_mult, 'lr': learning_rate / self.config.mup_width_mult},
+                {'params': decay_params, 'weight_decay': weight_decay, 'lr': learning_rate},
+                {'params': nodecay_params, 'weight_decay': 0.0, 'lr': learning_rate}
+            ]
+
+        elif self.config.mup and isinstance(self.config, Mamba2Config):
+            mup_params_keys = set([pn for pn in param_dict.keys() if any(pn.endswith(w) for w in ['mixer.in_proj.weight', 'mixer.out_proj.weight'])])
+            
+            dim2_params_keys = set([pn for pn in param_dict.keys() if param_dict[pn].dim() >= 2])
+            dim2_params_keys = dim2_params_keys.difference(mup_params_keys)
+
+            mup_parameters = [p for n, p in param_dict.items() if n in mup_params_keys]
+            decay_params = [p for n, p in param_dict.items() if n in dim2_params_keys]
+            nodecay_params = [p for n, p in param_dict.items() if p.dim() < 2] # biases and D and A
+
+            optim_groups = [
+                {'params': mup_parameters, 'weight_decay': weight_decay * self.config.mup_width_mult, 'lr': learning_rate / self.config.mup_width_mult},
+                {'params': decay_params, 'weight_decay': weight_decay, 'lr': learning_rate},
+                {'params': nodecay_params, 'weight_decay': 0.0, 'lr': learning_rate}
+            ]
+        
+        else:
+            decay_params = [p for n, p in param_dict.items() if p.dim() >= 2]
+            nodecay_params = [p for n, p in param_dict.items() if p.dim() < 2]
 
-        optim_groups = [
-            {'params': decay_params, 'weight_decay': weight_decay, 'lr': learning_rate},
-            {'params': nodecay_params, 'weight_decay': 0.0, 'lr': learning_rate}
-        ]
+            optim_groups = [
+                {'params': decay_params, 'weight_decay': weight_decay, 'lr': learning_rate},
+                {'params': nodecay_params, 'weight_decay': 0.0, 'lr': learning_rate}
+            ]
 
         # Create AdamW optimizer and use the fused version if it is available
         fused_available = 'fused' in inspect.signature(torch.optim.AdamW).parameters

mambapy/mamba.py

@@ -50,6 +50,9 @@ class MambaConfig:
     conv_bias: bool = True
     inner_layernorms: bool = False # apply layernorms to internal activations
 
+    mup: bool = False
+    mup_base_width: float = 128 # width=d_model
+
     pscan: bool = True # use parallel scan mode or sequential mode when training
     use_cuda: bool = False # use official CUDA implementation when training (not compatible with (b)float16)
 
@@ -59,6 +62,10 @@ def __post_init__(self):
         if self.dt_rank == 'auto':
             self.dt_rank = math.ceil(self.d_model / 16)
 
+        # muP
+        if self.mup:
+            self.mup_width_mult = self.d_model / self.mup_base_width
+
 class Mamba(nn.Module):
     def __init__(self, config: MambaConfig):
         super().__init__()
@@ -94,7 +101,7 @@ def __init__(self, config: MambaConfig):
         super().__init__()
 
         self.mixer = MambaBlock(config)
-        self.norm = RMSNorm(config.d_model, config.rms_norm_eps)
+        self.norm = RMSNorm(config.d_model, config.rms_norm_eps, config.mup)
 
     def forward(self, x):
         # x : (B, L, D)
@@ -170,9 +177,9 @@ def __init__(self, config: MambaConfig):
 
         # used in jamba
         if self.config.inner_layernorms:
-            self.dt_layernorm = RMSNorm(self.config.dt_rank, config.rms_norm_eps)
-            self.B_layernorm = RMSNorm(self.config.d_state, config.rms_norm_eps)
-            self.C_layernorm = RMSNorm(self.config.d_state, config.rms_norm_eps)
+            self.dt_layernorm = RMSNorm(self.config.dt_rank, config.rms_norm_eps, config.mup)
+            self.B_layernorm = RMSNorm(self.config.d_state, config.rms_norm_eps, config.mup)
+            self.C_layernorm = RMSNorm(self.config.d_state, config.rms_norm_eps, config.mup)
         else:
             self.dt_layernorm = None
             self.B_layernorm = None
@@ -407,13 +414,21 @@ def ssm_step(self, x, h):
         return y, h
 
 class RMSNorm(nn.Module):
-    def __init__(self, d_model: int, eps: float = 1e-5):
+    def __init__(self, d_model: int, eps: float = 1e-5, use_mup: bool = False):
         super().__init__()
 
+        self.use_mup = use_mup
         self.eps = eps
-        self.weight = nn.Parameter(torch.ones(d_model))
+
+        # https://arxiv.org/abs/2404.05728, RMSNorm gains prevents muTransfer (section 4.2.3)
+        if not use_mup:
+            self.weight = nn.Parameter(torch.ones(d_model))
 
     def forward(self, x):
         output = x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
 
-        return output * self.weight
+        if not self.use_mup:
+            return output * self.weight
+        else:
+            return output
+    

mambapy/mamba2.py

@@ -3,7 +3,7 @@
 """
 
 adapted from https://github.com/state-spaces/mamba/blob/main/mamba_ssm/modules/mamba2_simple.py
-It justs implements a config similar to what's being done in mamba.py.
+It justs implements a config similar to what's being done in mamba.py, as well as supports muP.
 
 """
 
@@ -56,6 +56,9 @@ class Mamba2Config:
     bias: bool = False
     conv_bias: bool = True
 
+    mup: bool = False
+    mup_base_width: float = 128 # width=d_model
+
     chunk_size: int = 256
     use_mem_eff_path: bool = True
     dtype=None
@@ -68,6 +71,10 @@ def __post_init__(self):
 
         assert (self.d_inner / self.d_head) % 8 == 0, "requierement of causal_conv1d"
 
+        # muP
+        if self.mup:
+            self.mup_width_mult = self.d_model / self.mup_base_width
+
 class Mamba2(nn.Module):
     def __init__(self, config: Mamba2Config):
         super().__init__()
@@ -103,7 +110,7 @@ def __init__(self, config: Mamba2Config):
         super().__init__()
 
         self.mixer = Mamba2Block(config)
-        self.norm = RMSNorm(config.d_model, config.rms_norm_eps)
+        self.norm = RMSNorm(config.d_model, config.rms_norm_eps, config.mup)
 
     def forward(self, x):
         # x : (B, L, D)
@@ -152,6 +159,7 @@ def __init__(self, config: Mamba2Config):
             nn.init.uniform_(self.conv1d.weight, -self.config.conv_init, self.config.conv_init)
         # self.conv1d.weight._no_weight_decay = True
 
+        # todo : mup init + lr
         if self.config.learnable_init_states:
             self.init_states = nn.Parameter(torch.zeros(self.config.n_heads, self.config.d_head, self.config.d_state, **factory_kwargs))
             self.init_states._no_weight_decay = True
@@ -267,12 +275,20 @@ def forward(self, u, seq_idx=None):
 
 # taken straight from https://github.com/johnma2006/mamba-minimal/blob/master/model.py
 class RMSNorm(nn.Module):
-    def __init__(self, d_model: int, eps: float = 1e-5):
+    def __init__(self, d_model: int, eps: float = 1e-5, use_mup: bool = False):
         super().__init__()
 
+        self.use_mup = use_mup
         self.eps = eps
-        self.weight = nn.Parameter(torch.ones(d_model))
+
+        # https://arxiv.org/abs/2404.05728, RMSNorm gains prevents muTransfer (section 4.2.3)
+        if not use_mup:
+            self.weight = nn.Parameter(torch.ones(d_model))
 
     def forward(self, x):
         output = x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
-        return output * self.weight
+
+        if not self.use_mup:
+            return output * self.weight
+        else:
+            return output