Fix stop in /v1/completions , fix ts_first_token

refact_scratchpads/scratchpad_hf.py

@@ -37,6 +37,7 @@ def __init__(
 
         self._stop_lf = False
         self._stop_lf_lf = False
+        self._prev_token: Optional[int] = None
         self._stop_tokens: Set[int] = set()
         if isinstance(stop_tokens, str):
             stop_tokens = [stop_tokens]
@@ -80,10 +81,11 @@ def after_token_selection(self, m, chosen_token: th.Tensor, **unused) -> Dict[st
         if t in self._stop_tokens:
             self.finish_reason = "stoptoken"
 
-        t_str = self._tokenizer.decode([t])
-        if self._stop_lf and t_str.startswith("\n"):
+        couple_of_tokens_decoded = self._tokenizer.decode(([self._prev_token] if self._prev_token is not None else []) + [t])
+        self._prev_token = t
+        if self._stop_lf and ("\n" in couple_of_tokens_decoded):
             self.finish_reason = "stop-lf"
-        if self._stop_lf_lf and t_str.startswith("\n\n"):
+        if self._stop_lf_lf and ("\n\n" in couple_of_tokens_decoded):
             self.finish_reason = "stop-lflf"
 
         self._tokens_produced += 1

self_hosting_machinery/inference/inference_hf.py

@@ -78,6 +78,8 @@ def __init__(self, tokenizer, request_id: str, upload_proxy: UploadProxy,
         self.upload_proxy_args = upload_proxy_args
 
     def put(self, value):
+        if self.upload_proxy_args.get("ts_first_token", 0) == 0:
+            self.upload_proxy_args["ts_first_token"] = time.time()
         super().put(value)
 
     def on_finalized_text(self, text: str, stream_end: bool = False):

self_hosting_machinery/webgui/selfhost_fastapi_completions.py

@@ -62,7 +62,7 @@ def clamp(self):
 
 
 class NlpCompletion(NlpSamplingParams):
-    model: str = Query(default="", regex="^[a-z/A-Z0-9_\.]+$")
+    model: str = Query(default=Required, regex="^[a-z/A-Z0-9_\.\-]+$")
     prompt: str
     n: int = 1
     echo: bool = False