Fixup residual, initial block attention config

server/text_generation_server/models/custom_modeling/flash_phi3small_modeling.py

@@ -19,6 +19,7 @@
 # limitations under the License.
 
 from typing import List, Optional, Tuple
+from dataclasses import dataclass
 
 import torch
 import torch.distributed
@@ -50,6 +51,14 @@ if SYSTEM == "rocm":
         raise ImportError(f"Could not load `vllm._custom_C`. Full error: {e}")
 
 
+@dataclass
+class BlockSparseAttentionConfig:
+    block_size: int
+    homo_head_pattern: bool
+    num_local_blocks: int
+    vert_stride: int
+
+
 def load_attention(config, prefix, weights):
     # Only defined in granite.
     bias = getattr(config, "attention_bias", False)
@@ -69,6 +78,7 @@ class FlashPhi3SmallAttention(torch.nn.Module):
         self,
         prefix: str,
         config,
+        layer_id: int,
         weights,
     ):
         super().__init__()
@@ -83,7 +93,10 @@ class FlashPhi3SmallAttention(torch.nn.Module):
             device=weights.device,
         )
 
-        self.softmax_scale = self.head_size**-0.5
+        if hasattr(config, "mup_use_scaling") and config.mup_use_scaling:
+            self.softmax_scale = self.head_size / config.mup_attn_multiplier
+        else:
+            self.softmax_scale = self.head_size**-0.5
 
         if self.num_heads % weights.process_group.size() != 0:
             raise ValueError(
@@ -102,11 +115,25 @@ class FlashPhi3SmallAttention(torch.nn.Module):
 
         self.query_key_value = load_attention(config, prefix, weights)
 
+        is_dense = getattr(config, "dense_attention_every_n_layers", False) and (
+            (layer_id + 1) % config.dense_attention_every_n_layers == 0
+        )
+
+        if is_dense:
+            self.blocksparse_config = None
+        else:
+            self.blocksparse_config = BlockSparseAttentionConfig(
+                block_size=config.blocksparse_block_size,
+                homo_head_pattern=config.blocksparse_homo_head_pattern,
+                num_local_blocks=config.blocksparse_num_local_blocks,
+                vert_stride=config.blocksparse_vert_stride,
+            )
+
         self.o_proj = TensorParallelRowLinear.load(
             config,
             prefix=f"{prefix}.dense",
             weights=weights,
-            bias=False,
+            bias=True,
         )
 
         self.num_groups = self.num_heads // self.num_key_value_heads
@@ -246,10 +273,13 @@ class Phi3SmallMLP(nn.Module):
 
 
 class FlashPhi3SmallLayer(nn.Module):
-    def __init__(self, prefix, config, weights):
+    def __init__(self, prefix, config, layer_id: int, weights):
         super().__init__()
         self.self_attn = FlashPhi3SmallAttention(
-            prefix=f"{prefix}.self_attn", config=config, weights=weights
+            prefix=f"{prefix}.self_attn",
+            config=config,
+            layer_id=layer_id,
+            weights=weights,
         )
         self.mlp = Phi3SmallMLP(prefix=f"{prefix}.mlp", config=config, weights=weights)
 
@@ -267,7 +297,6 @@ class FlashPhi3SmallLayer(nn.Module):
     def forward(
         self,
         hidden_states,
-        residual,
         cos,
         sin,
         cu_seqlen_prefill,
@@ -277,7 +306,8 @@ class FlashPhi3SmallLayer(nn.Module):
         input_lengths,
         max_s,
     ):
-        normed_hidden_states, res = self.input_layernorm(hidden_states, residual)
+        residual = hidden_states
+        normed_hidden_states, res = self.input_layernorm(hidden_states, None)
 
         # Self Attention
         attn_output = self.self_attn(
@@ -294,12 +324,13 @@ class FlashPhi3SmallLayer(nn.Module):
 
         # faster post attention rms norm
         normed_attn_res_output, attn_res = self.post_attention_layernorm(
-            attn_output, res
+            attn_output, residual
         )
 
         mlp_output = self.mlp(normed_attn_res_output)
+        mlp_output = attn_res + mlp_output
 
-        return mlp_output, attn_res
+        return mlp_output
 
 
 class FlashPhi3SmallModel(torch.nn.Module):
@@ -318,12 +349,13 @@ class FlashPhi3SmallModel(torch.nn.Module):
                         else f"{prefix}.model.layers.{layer_id}"
                     ),
                     config=config,
+                    layer_id=layer_id,
                     weights=weights,
                 )
                 for layer_id in range(config.num_hidden_layers)
             ]
         )
-        self.norm = FastLayerNorm.load(
+        self.norm = nn.LayerNorm.load(
             prefix=(
                 "model.final_layernorm"
                 if not prefix
@@ -360,11 +392,9 @@ class FlashPhi3SmallModel(torch.nn.Module):
             position_ids, max_s, hidden_states.dtype
         )
 
-        residual = None
         for i, layer in enumerate(self.layers):
-            hidden_states, residual = layer(
+            hidden_states = layer(
                 hidden_states,
-                residual,
                 cos,
                 sin,
                 cu_seqlen_prefill,
@@ -375,7 +405,7 @@ class FlashPhi3SmallModel(torch.nn.Module):
                 max_s,
             )
 
-        hidden_states, _ = self.norm(hidden_states, residual)
+        hidden_states = self.norm(hidden_states)
 
         return hidden_states