fix: LlamaTokenizerFast to AutoTokenizer at flash_mistral.py (#1637)

# What does this PR do?

A few cases where you're using a mistral structure or mixtral structure
but not a llama tokenizer, why not make it to call the AutoTokenizer in
exception handling.

Similar PR #619

@Narsil

server/text_generation_server/models/flash_mistral.py

@@ -6,7 +6,7 @@ import numpy as np
 
 from dataclasses import dataclass
 from opentelemetry import trace
-from transformers import PreTrainedTokenizerBase
+from transformers import PreTrainedTokenizerBase, AutoTokenizer
 from transformers.models.llama import LlamaTokenizerFast
 from typing import Optional, Tuple, Type
 
@@ -317,6 +317,7 @@ class BaseFlashMistral(FlashCausalLM):
         else:
             raise NotImplementedError("FlashMistral is only available on GPU")
 
+        try:
             tokenizer = LlamaTokenizerFast.from_pretrained(
                 model_id,
                 revision=revision,
@@ -324,6 +325,14 @@ class BaseFlashMistral(FlashCausalLM):
                 truncation_side="left",
                 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,
+            )
 
         config = config_cls.from_pretrained(
             model_id, revision=revision, trust_remote_code=trust_remote_code