fix(server): fix warpers on CPU (#472)

Closes #471

server/text_generation_server/models/__init__.py

@@ -237,20 +237,12 @@ def get_model(
         )
 
     elif model_type == "t5":
-        if sharded:
         return T5Sharded(
             model_id,
             revision,
             quantize=quantize,
             trust_remote_code=trust_remote_code,
         )
-        else:
-            return Seq2SeqLM(
-                model_id,
-                revision,
-                quantize=quantize,
-                trust_remote_code=trust_remote_code,
-            )
 
     if sharded:
         raise ValueError("sharded is not supported for AutoModel")

server/text_generation_server/utils/logits_process.py

@@ -42,6 +42,7 @@ class StaticWarper:
         self.static_next_logprob = None
 
     def __call__(self, scores):
+        if torch.cuda.is_available():
             if self.cuda_graph is None:
                 self.static_scores = scores
                 self.cuda_graph = torch.cuda.CUDAGraph()
@@ -62,6 +63,11 @@ class StaticWarper:
 
             return self.static_warped_scores, self.static_next_logprob
 
+        # CPU branch
+        for warper in self.warpers:
+            scores = warper(None, scores)
+        return scores, torch.log_softmax(scores, -1)
+
 
 @lru_cache(10)
 def static_warper(