updt to include low precision groupnorm;

Author: root

llmfoundry/models/layers/attention.py

@@ -12,10 +12,10 @@
 import transformers
 from einops import rearrange
 from packaging import version
-from torch import nn
 
 from llmfoundry.models.layers.fc import FC_CLASS_REGISTRY
-from llmfoundry.models.layers.norm import NORM_CLASS_REGISTRY
+from llmfoundry.models.layers.norm import (NORM_CLASS_REGISTRY, LPLayerNorm,
+                                           low_precision_groupnorm)
 
 
 def is_flash_v2_installed(v2_version: str = '2.0.0'):
@@ -498,6 +498,47 @@ def triton_flash_attn_fn(
     return output, None, past_key_value
 
 
+def _expand_params(heads: int, param: Optional[torch.Tensor] = None):
+    if param is None:
+        return None
+    return param.repeat(heads)
+
+
+def _apply_qk_gn(
+    query: torch.Tensor,
+    key: torch.Tensor,
+    n_heads: int,
+    kv_n_heads: int,
+    q_ln: nn.Module,
+    k_ln: nn.Module,
+):
+    dtype = query.dtype
+
+    w = _expand_params(n_heads, q_ln.weight)
+    b = _expand_params(n_heads, q_ln.bias)
+    if isinstance(q_ln, LPLayerNorm):
+        query = low_precision_groupnorm(query, n_heads, w, b,
+                                        eps=q_ln.eps).to(dtype)
+    elif isinstance(q_ln, nn.LayerNorm):
+        query = nn.functional.group_norm(query, n_heads, w, b, eps=q_ln.eps)
+    else:
+        raise ValueError(
+            f'qk_gn not applicable for given q_ln type ({type(q_ln)=}).')
+
+    w = _expand_params(kv_n_heads, k_ln.weight)
+    b = _expand_params(kv_n_heads, k_ln.bias)
+    if isinstance(k_ln, LPLayerNorm):
+        key = low_precision_groupnorm(key, kv_n_heads, w, b,
+                                      eps=k_ln.eps).to(dtype)
+    elif isinstance(k_ln, nn.LayerNorm):
+        key = nn.functional.group_norm(key, kv_n_heads, w, b, eps=k_ln.eps)
+    else:
+        raise ValueError(
+            f'qk_gn not applicable for given k_ln type ({type(k_ln)=}).')
+
+    return query, key
+
+
 class GroupedQueryAttention(nn.Module):
     """Grouped Query Attention (GQA) is a generalization of Multi-head (MHA).
 
@@ -629,16 +670,22 @@ def forward(
 
         key_padding_mask = attention_mask
 
-        if self.qk_ln or self.qk_gn:
+        if self.qk_gn:
+            # Applying groupnorm to qk
+            query, key = _apply_qk_gn(
+                query,
+                key,
+                self.n_heads,
+                self.kv_n_heads,
+                self.q_ln,
+                self.k_ln,
+            )
+
+        if self.qk_ln:
             # Applying layernorm to qk
-            q_shape, k_shape = query.shape, key.shape
-            if self.qk_gn:
-                b, s = query.shape[:2]
-                query = query.view(b, s, self.n_heads, -1)
-                key = key.view(b, s, self.kv_n_heads, -1)
             dtype = query.dtype
-            query = self.q_ln(query).to(dtype).view(q_shape)
-            key = self.k_ln(key).to(dtype).view(k_shape)
+            query = self.q_ln(query).to(dtype)
+            key = self.k_ln(key).to(dtype)
 
         if rotary_emb_w_meta_info is not None:
             rotary_emb = rotary_emb_w_meta_info['rotary_emb']

llmfoundry/models/layers/norm.py

@@ -53,6 +53,29 @@ def forward(self, x: torch.Tensor) -> torch.Tensor:
             )
 
 
+def low_precision_groupnorm(
+    x: torch.Tensor,
+    groups: int,
+    weight: Optional[torch.Tensor] = None,
+    bias: Optional[torch.Tensor] = None,
+    eps: float = 1e-05,
+):
+    device = x.device
+    downcast_x = _cast_if_autocast_enabled(x)
+    downcast_weight = _cast_if_autocast_enabled(
+        weight) if weight is not None else weight
+    downcast_bias = _cast_if_autocast_enabled(
+        bias) if bias is not None else bias
+    with torch.autocast(enabled=False, device_type=device.type):
+        return torch.nn.functional.group_norm(
+            downcast_x,
+            groups,
+            downcast_weight,
+            downcast_bias,
+            eps,
+        )
+
+
 def rms_norm(x: torch.Tensor,
              weight: Optional[torch.Tensor] = None,
              eps: float = 1e-5) -> torch.Tensor: