Fixing codellama loads by using purely AutoTokenizer.

- The need for the slow tokenizer default stems from back
  when llama 1 was introduced and all the flags where not
  supported in `tokenizers`.

- Fixes #1891

server/text_generation_server/models/flash_llama.py

@@ -3,7 +3,6 @@ import torch.distributed
 
 from opentelemetry import trace
 from transformers import AutoConfig, AutoTokenizer, GenerationConfig
-from transformers.models.llama import LlamaTokenizer
 from typing import Optional
 
 from text_generation_server.models import FlashCausalLM
@@ -41,22 +40,13 @@ class FlashLlama(FlashCausalLM):
         else:
             raise NotImplementedError("FlashLlama is only available on GPU")
 
-        try:
+        tokenizer = AutoTokenizer.from_pretrained(
-            tokenizer = LlamaTokenizer.from_pretrained(
+            model_id,
-                model_id,
+            revision=revision,
-                revision=revision,
+            padding_side="left",
-                padding_side="left",
+            truncation_side="left",
-                truncation_side="left",
+            trust_remote_code=trust_remote_code,
-                trust_remote_code=trust_remote_code,
+        )
-            )
-        except Exception:
-            tokenizer = AutoTokenizer.from_pretrained(
-                model_id,
-                revision=revision,
-                padding_side="left",
-                truncation_side="left",
-                trust_remote_code=trust_remote_code,
-            )
         try:
             generation_config = GenerationConfig.from_pretrained(
                 model_id, revision=revision, trust_remote_code=trust_remote_code