Fixes for worker prompt truncation in ChatML case (#3673)

Co-authored-by: Andreas Koepf <andreas.koepf@provisio.com>

Author: olliestanley

inference/worker/__main__.py

@@ -34,7 +34,7 @@ def main():
         tokenizer = None
     else:
         tokenizer: transformers.PreTrainedTokenizer = transformers.AutoTokenizer.from_pretrained(model_config.model_id)
-        logger.warning(f"Tokenizer {tokenizer.name_or_path} vocab size: {tokenizer.vocab_size}")
+        logger.warning(f"Tokenizer {tokenizer.name_or_path} vocab size: {len(tokenizer)}")
 
     inference_http = utils.HttpClient(
         base_url=settings.inference_server_url,

inference/worker/basic_hf_server.py

@@ -138,7 +138,7 @@ def load_models():
     hf_config = transformers.AutoConfig.from_pretrained(model_config.model_id)
     logger.warning(f"Loading model {model_config.model_id}...")
     tokenizer = transformers.AutoTokenizer.from_pretrained(model_config.model_id)
-    logger.warning(f"tokenizer {tokenizer.name_or_path} has vocab size {tokenizer.vocab_size}")
+    logger.warning(f"tokenizer {tokenizer.name_or_path} has vocab size {len(tokenizer)}")
 
     # see `decode_token` method, taken from HF text-generation-inference
     tokenizer.add_special_tokens({"additional_special_tokens": ["<decode-token>"]})

inference/worker/utils.py

@@ -95,11 +95,14 @@ def get_max_input_length(worker_config: inference.WorkerConfig, plugin_used: boo
     return max_input_length
 
 
-def get_tokens_until(tokens: list[int], target: int | list[int]) -> list[int]:
-    if isinstance(target, int):
-        return tokens[: tokens.index(target)]
-    else:
-        return next((i for i in range(len(tokens) - len(target) + 1) if tokens[i : i + len(target)] == target))
+def get_tokens_until(tokens: list[int], target: list[int]) -> list[int]:
+    if len(target) == 1:
+        return tokens[: tokens.index(target[0])]
+
+    for i in range(len(tokens) - len(target)):
+        if tokens[i : i + len(target)] == target:
+            break
+    return tokens[:i]
 
 
 def truncate_prompt(
@@ -118,8 +121,8 @@ def truncate_prompt(
     """
     with shared_tokenizer_lock:
         ids = tokenizer.encode(prompt)
-        # prompter_prefix_ids could be int or list of ints
-        prompter_prefix_ids = tokenizer.convert_tokens_to_ids(special_tokens["prompter"])
+        # list of int IDs
+        prompter_prefix_ids = tokenizer.encode(special_tokens["prompter"])
 
     system_prompt: str | None = None
     system_tokens: list[int] | None = None
@@ -134,7 +137,9 @@ def truncate_prompt(
 
         num_system_tokens = len(system_tokens) if system_tokens else 0
         # Maximum token allowed for the conversation, ex system prompt
-        max_conversation_length = max_input_length - num_system_tokens
+        # We incorporate a buffer to allow for final inference tokenization differing from ours
+        # This is a slightly hacky workaround and it would be better to find a cleaner solution
+        max_conversation_length = max_input_length - num_system_tokens - int(0.01 * max_input_length)
         ids = ids[-(max_conversation_length - 1) :]
 
         with shared_tokenizer_lock:

oasst-shared/oasst_shared/model_configs.py

@@ -150,4 +150,9 @@ def compat_hash(self) -> str:
         max_input_length=3072,
         max_total_length=4096,
     ),
+    "OA_SFT_CodeLlama_13B_10": ModelConfig(
+        model_id="OpenAssistant/codellama-13b-oasst-sft-v10",
+        max_input_length=8192,
+        max_total_length=12288,
+    ),
 }