feat(server): cleanup flash neox loading (#139)

server/text_generation_server/models/flash_neox.py

@@ -450,8 +450,6 @@ class FlashNeoX(Model):
             next_batch_input_ids = next_batch_input_ids[0].view(1)
             next_batch_past_key_values = next_batch_past_key_values[0]
 
-        print(next_batch_input_ids.shape)
-
         next_batch = FlashNeoXBatch(
             batch_id=batch.batch_id,
             requests=next_batch_requests,
@@ -507,6 +505,7 @@ class FlashNeoXSharded(FlashNeoX):
             rank=self.rank,
             world_size=self.world_size,
         )
+        model.post_load_weights()
         self.model = model.eval().to(dtype)
         torch.distributed.barrier(group=self.process_group)
         super(FlashNeoX, self).__init__(

server/text_generation_server/models/flash_neox_modeling.py

@@ -1,6 +1,8 @@
 import torch
 import torch.distributed
 
+from torch.nn import functional as F
+
 from torch import nn
 from transformers.activations import ACT2FN
 from transformers.modeling_utils import PreTrainedModel
@@ -24,13 +26,11 @@ class FastLinear(nn.Linear):
         dtype=None,
     ) -> None:
         super(FastLinear, self).__init__(in_features, out_features, bias, device, dtype)
-        self.swap_dims = True
+
+    def transpose_weight(self):
+        self.weight = nn.Parameter(self.weight.T)
 
     def forward(self, input: torch.Tensor) -> torch.Tensor:
-        if self.swap_dims:
-            self.weight = nn.Parameter(self.weight.T)
-            self.swap_dims = False
-
         if self.bias is not None:
             return torch.addmm(self.bias, input, self.weight)
         return torch.matmul(input, self.weight)
@@ -120,6 +120,10 @@ class TensorParallelEmbedding(nn.Embedding):
         self.min_id = self.tp_rank * block_size
         self.max_id = (self.tp_rank + 1) * block_size
 
+        # Additional entry that will map to zero
+        # Used for masking
+        self.null_idx = block_size
+
         super().__init__(
             block_size,
             embedding_dim,
@@ -133,15 +137,19 @@ class TensorParallelEmbedding(nn.Embedding):
             dtype=dtype,
         )
 
+    def add_null_idx(self):
+        """Additional 0 entry used for masking"""
+        self.weight = nn.Parameter(F.pad(self.weight, (0, 0, 0, 1)))
+
     def forward(self, input: torch.Tensor) -> torch.Tensor:
-        # `0` if input is in the correct interval, else `1`
+        # default all out of bounds values to `self.null_idx` that will then be mapped to 0
-        input_mask = torch.logical_or(self.min_id > input, input >= self.max_id)
         # translate for [0, self.max_id - self.min_id[
-        input = input - self.min_id
+        input = torch.where(
-        # default all out of bounds values to `0`
+            (self.min_id > input) | (input >= self.max_id),
-        input[input_mask] = 0
+            self.null_idx,
+            input - self.min_id,
+        )
         out = super().forward(input)
-        out[input_mask] = 0.0
         torch.distributed.all_reduce(out, group=self.process_group)
         return out
 
@@ -214,11 +222,9 @@ class FlashNeoxAttention(torch.nn.Module):
                 hidden_size,
                 process_group=process_group,
             )
-        self.swap_dims = True
 
-    # TODO: remove and swap dims when loading weights
+    def shuffle_qkv_dims(self):
-    def _swap_dims(self):
+        """Swap dims to avoid an additional permute"""
-        """Swap dims for the first inference to avoid an additional permute"""
         self.query_key_value.weight = torch.nn.Parameter(
             self.query_key_value.weight.view(
                 self.num_heads, 3, self.head_size, self.hidden_size
@@ -231,7 +237,6 @@ class FlashNeoxAttention(torch.nn.Module):
             .permute(1, 0, 2)
             .reshape(-1)
         )
-        self.swap_dims = False
 
     def forward(
         self,
@@ -244,9 +249,6 @@ class FlashNeoxAttention(torch.nn.Module):
         layer_past_present_indices,
         cu_seqlens_q,
     ):
-        if self.swap_dims:
-            self._swap_dims()
-
         qkv = self.query_key_value(hidden_states)
         qkv = qkv.view(-1, 3, self.num_heads, self.head_size)
         qkv_rot = self.rotary_emb(qkv, cos, sin)
@@ -329,7 +331,6 @@ class FlashMLP(nn.Module):
                 hidden_size,
                 process_group=process_group,
             )
-        self.heuristic = "auto"
         self.process_group = process_group
 
     def forward(self, hidden_states):
@@ -531,6 +532,25 @@ class FlashGPTNeoXModel(FlashGPTNeoXPreTrainedModel):
         self.head_size = self.layers[0].attention.head_size
         self.num_heads = self.layers[0].attention.num_heads
 
+    def post_load_weights(self):
+        if isinstance(self.embed_in, TensorParallelEmbedding):
+            self.embed_in.add_null_idx()
+        for layer in self.layers:
+            layer: FlashNeoXLayer
+            layer.attention.shuffle_qkv_dims()
+            layer.attention.query_key_value.transpose_weight()
+            layer.attention.dense.transpose_weight()
+            layer.mlp.dense_h_to_4h.transpose_weight()
+            layer.mlp.dense_4h_to_h.transpose_weight()
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
+        model = super(FlashGPTNeoXModel, cls).from_pretrained(
+            pretrained_model_name_or_path, *model_args, **kwargs
+        )
+        model.post_load_weights()
+        return model
+
     def forward(
         self,
         input_ids,
@@ -627,6 +647,18 @@ class FlashGPTNeoXForCausalLM(FlashGPTNeoXPreTrainedModel):
                 config.hidden_size, config.vocab_size, bias=False
             )
 
+    def post_load_weights(self):
+        self.gpt_neox.post_load_weights()
+        self.embed_out.transpose_weight()
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
+        model = super(FlashGPTNeoXForCausalLM, cls).from_pretrained(
+            pretrained_model_name_or_path, *model_args, **kwargs
+        )
+        model.post_load_weights()
+        return model
+
     def forward(
         self,
         input_ids,