Incorrectly judging the request as a bad request (#4121)

* Incorrectly judging the request as a Bad Request

* fix

* final

Author: lvhan028

lmdeploy/pytorch/engine/engine.py

@@ -875,7 +875,7 @@ def _make_infer_outputs(
             # logprobs
             num_logprobs = msg.sampling_param.num_logprobs
             cur_logprobs = None
-            if num_logprobs >= 0:
+            if logprobs is not None:
                 cur_logprobs = (logprobs.vals[idx][:num_logprobs + 1], logprobs.indices[idx][:num_logprobs + 1])
 
             req_metrics = RequestMetrics(new_token_timestamp, msg.engine_events)

lmdeploy/serve/openai/serving_chat_completion.py

@@ -15,9 +15,13 @@ def check_request(request: ChatCompletionRequest, engine_config: 'TurbomindEngin
     try:
         logprobs_mode = engine_config.logprobs_mode
         logprobs = request.logprobs
-        top_logprobs = request.top_logprobs
-        if logprobs_mode is None and (logprobs is not None or top_logprobs is not None):
-            return 'Logprobs or top_logprobs requested but not enabled logprobs_mode in engine configuration.'
+        top_logprobs = request.top_logprobs or 0
+        if logprobs_mode is None and (logprobs or top_logprobs > 0):
+            return (f'Logprobs({logprobs})/top_logprobs({top_logprobs}) requested '
+                    'but not enabled logprobs_mode in engine configuration')
+        if logprobs_mode is not None and (top_logprobs < 0 or (not logprobs and top_logprobs > 0)):
+            return (f'Invalid logprobs({logprobs})/top_logprobs({top_logprobs}) requested '
+                    'when logprobs_mode is enabled in engine configuration.')
     except AttributeError:
         pass
 

lmdeploy/serve/openai/serving_completion.py

@@ -14,9 +14,11 @@ def check_request(request: CompletionRequest, engine_config: 'TurbomindEngineCon
     # Check logprobs settings
     try:
         logprobs_mode = engine_config.logprobs_mode
-        logprobs = request.logprobs
-        if logprobs_mode is None and logprobs:
-            return 'logprobs requested but not enabled logprobs_mode in engine configuration.'
+        logprobs = request.logprobs or 0
+        if logprobs > 0 and logprobs_mode is None:
+            return f'logprobs({logprobs}) requested but not enabled logprobs_mode in engine configuration.'
+        if logprobs_mode is not None and logprobs < 0:
+            return 'logprobs must be non-negative when logprobs_mode is enabled in engine configuration.'
     except AttributeError:
         pass
 

lmdeploy/serve/openai/serving_generate.py

@@ -16,7 +16,7 @@ def check_request(request: GenerateReqInput, engine_config: 'TurbomindEngineConf
         logprobs_mode = engine_config.logprobs_mode
         return_logprob = request.return_logprob
         if logprobs_mode is None and return_logprob:
-            return 'return_logprob requested but not enabled logprobs_mode in engine configuration.'
+            return f'return_logprob({return_logprob}) requested but not enabled logprobs_mode in engine configuration.'
     except AttributeError:
         pass