fix(server): Check for device type correctly when determining initial padding (#16)

AFAIK there is no torch device type called "gpu".

server/text_generation/models/causal_lm.py

@@ -65,7 +65,7 @@ class CausalLMBatch:
             )
             all_logprobs.append(None)
 
-        pad_to_multiple_of = 8 if "gpu" in str(device) else None
+        pad_to_multiple_of = 8 if device.type == "cuda" else None
         tokenized_inputs = tokenizer(
             inputs,
             return_tensors="pt",

server/text_generation/models/seq2seq_lm.py

@@ -77,7 +77,7 @@ class Seq2SeqLMBatch:
             decoder_logprobs.append(None)
 
         # Tokenize batch
-        pad_to_multiple_of = 8 if "gpu" in str(device) else None
+        pad_to_multiple_of = 8 if device.type == "cuda" else None
         tokenized_inputs = tokenizer(
             inputs,
             return_tensors="pt",