tokenized inputs from pb

server/text_generation_server/models/causal_lm.py

@@ -83,7 +83,7 @@ class CausalLMBatch(Batch):
         for i, r in enumerate(pb.requests):
             requests_idx_mapping[r.id] = i
             inputs.append(r.inputs)
-            next_token_choosers.append(NextTokenChooser.from_pb(r.parameters, device))
+            next_token_choosers.append(NextTokenChooser.from_pb(r.parameters, device, tokenizer))
             stopping_criteria = StoppingCriteria.from_pb(
                 r.stopping_parameters, tokenizer
             )

server/text_generation_server/models/flash_causal_lm.py

@@ -303,7 +303,7 @@ class FlashCausalLMBatch(Batch):
             max_length = max(max_length, input_length + max_new_tokens)
 
         next_token_chooser = HeterogeneousNextTokenChooser.from_pb(
-            next_token_chooser_parameters, dtype, device
+            next_token_chooser_parameters, dtype, device, tokenizer
         )
         start_slots = torch.tensor(start_slots, dtype=torch.int64)
 
@@ -634,6 +634,8 @@ class FlashCausalLMBatch(Batch):
             next_token_chooser_parameters,
             dtype=batches[0].next_token_chooser.dtype,
             device=batches[0].next_token_chooser.device,
+            # todo - determine how to obtain access to a tokenizer here
+            tokenizer=...
         )
 
         # Needed to avoid dropping blocks when the batches will go out of scope

server/text_generation_server/models/galactica.py

@@ -91,7 +91,7 @@ class GalacticaCausalLMBatch(CausalLMBatch):
             requests_idx_mapping[r.id] = i
             # Add escape_custom_split_sequence to the CausalLMBatch logic
             inputs.append(escape_custom_split_sequence(r.inputs))
-            next_token_choosers.append(NextTokenChooser.from_pb(r.parameters, device))
+            next_token_choosers.append(NextTokenChooser.from_pb(r.parameters, device, tokenizer))
             stopping_criteria = StoppingCriteria.from_pb(
                 r.stopping_parameters, tokenizer
             )

server/text_generation_server/models/seq2seq_lm.py

@@ -92,7 +92,7 @@ class Seq2SeqLMBatch(Batch):
             inputs.append(r.inputs)
             requests_idx_mapping[r.id] = i
             decoder_input_lengths.append(1)
-            next_token_choosers.append(NextTokenChooser.from_pb(r.parameters, device))
+            next_token_choosers.append(NextTokenChooser.from_pb(r.parameters, device, tokenizer))
             stopping_criteria = StoppingCriteria.from_pb(
                 r.stopping_parameters, tokenizer
             )

server/text_generation_server/utils/tokens.py

@@ -33,7 +33,7 @@ class NextTokenChooser:
             do_sample=False,
             seed=0,
             device="cpu",
-            logit_bias=None,
+            logit_bias={},
     ):
         self.watermark_processor = (
             WatermarkLogitsProcessor(device=device) if watermark else None
@@ -85,6 +85,7 @@ class NextTokenChooser:
             cls,
             pb: generate_pb2.NextTokenChooserParameters,
             device: torch.device,
+            tokenizer: PreTrainedTokenizerBase,
     ) -> "NextTokenChooser":
         return NextTokenChooser(
             watermark=pb.watermark,
@@ -96,7 +97,7 @@ class NextTokenChooser:
             do_sample=pb.do_sample,
             seed=pb.seed,
             device=device,
-            logit_bias=pb.logit_bias,
+            logit_bias=dict([(tuple(tokenizer.encode(bias.string, add_special_tokens=False).input_ids[0]), bias.bias) for  bias in pb.logit_bias]),
         )
 
 
@@ -285,6 +286,7 @@ bias
             pb: List[generate_pb2.NextTokenChooserParameters],
             dtype: torch.dtype,
             device: torch.device,
+            tokenizer: PreTrainedTokenizerBase,
     ) -> "HeterogeneousNextTokenChooser":
         return HeterogeneousNextTokenChooser(
             watermark=[pb_.watermark for pb_ in pb],
@@ -297,7 +299,7 @@ bias
             seeds=[pb_.seed for pb_ in pb],
             device=device,
             dtype=dtype,
-            logit_bias={},
+            logit_bias=[dict([(tuple(tokenizer.encode(bias.string, add_special_tokens=False).input_ids[0]), bias.bias) for bias in pb_.logit_bias])  for pb_ in pb],
         )