feat(server): avoid manipulating position_ids for non-applicable models

Currently position_ids are maintained/updated in the CausalLM case but this is unnecessary for models like BLOOM which don't use them.

server/text_generation_server/models/causal_lm.py

@@ -1,3 +1,5 @@
+import inspect
+
 import torch
 
 from dataclasses import dataclass
@@ -26,7 +28,7 @@ class CausalLMBatch(Batch):
     # Decoder values
     input_ids: torch.Tensor
     attention_mask: torch.Tensor
-    position_ids: torch.Tensor
+    position_ids: Optional[torch.Tensor]
     past_key_values: Optional[List[Tuple]]
 
     # All tokens
@@ -62,6 +64,7 @@ class CausalLMBatch(Batch):
         pb: generate_pb2.Batch,
         tokenizer: PreTrainedTokenizerBase,
         device: torch.device,
+        use_position_ids: bool = False,
     ) -> "CausalLMBatch":
         inputs = []
         next_token_choosers = []
@@ -105,10 +108,13 @@ class CausalLMBatch(Batch):
         )
         # Copy tokenizer attention_mask into fully allocated attention_mask
         attention_mask[:, :max_input_length] = tokenized_inputs["attention_mask"]
+        all_input_ids = tokenized_inputs["input_ids"].unsqueeze(-1)
 
+        if use_position_ids:
             position_ids = tokenized_inputs["attention_mask"].long().cumsum(-1) - 1
             position_ids.masked_fill_(tokenized_inputs["attention_mask"] == 0, 1)
-        all_input_ids = tokenized_inputs["input_ids"].unsqueeze(-1)
+        else:
+            position_ids = None
 
         return cls(
             batch_id=pb.id,
@@ -204,6 +210,7 @@ class CausalLMBatch(Batch):
                 batch_left_offset : -batch.padding_right_offset,
             ]
 
+            if batch.position_ids is not None:
                 # Create empty tensor
                 # position_ids is always of shape [batch_size, 1]
                 if position_ids is None:
@@ -328,6 +335,10 @@ class CausalLM(Model):
             tokenizer=tokenizer, device=device, decode_buffer=decode_buffer
         )
 
+        self.use_position_ids = "position_ids" in set(
+            inspect.signature(self.model.forward).parameters.keys()
+        )
+
     @property
     def batch_type(self) -> Type[CausalLMBatch]:
         return CausalLMBatch
@@ -494,7 +505,8 @@ class CausalLM(Model):
         if len(next_batch_keep_indices) != len(batch):
             # Apply indices to attention mask, past key values and other items that need to be cached
             next_batch_attention_mask = batch.attention_mask[next_batch_keep_indices]
-            next_batch_position_ids = batch.position_ids[next_batch_keep_indices]
+            next_batch_position_ids = batch.position_ids[next_batch_keep_indices] \
+                if batch.position_ids is not None else None
             # Force past to be of dim [batch_size, num_heads, ...] for easy indexing
             next_batch_past_key_values = [
                 [
@@ -522,6 +534,7 @@ class CausalLM(Model):
         next_batch_attention_mask[:, -batch.padding_right_offset] = 1
 
         # Update position_ids
+        if next_batch_position_ids is not None:
             next_batch_position_ids = next_batch_position_ids[:, -1:] + 1
 
         next_batch = CausalLMBatch(

server/text_generation_server/models/model.py

@@ -23,6 +23,7 @@ class Model(ABC):
         self.all_special_ids = set(tokenizer.all_special_ids)
         self.device = device
         self.decode_buffer = decode_buffer
+        self.use_position_ids = False
 
     @property
     @abstractmethod

server/text_generation_server/server.py

@@ -39,8 +39,9 @@ class TextGenerationService(generate_pb2_grpc.TextGenerationServiceServicer):
         return generate_pb2.ClearCacheResponse()
 
     async def Prefill(self, request, context):
+        kwargs = {"use_position_ids": True} if self.model.use_position_ids else {}
         batch = self.model.batch_type.from_pb(
-            request.batch, self.model.tokenizer, self.model.device
+            request.batch, self.model.tokenizer, self.model.device, **kwargs,
         )
 
         generations, next_batch = self.model.generate_token(batch)