[Fix] fix eos trim inconsistency (#1650)

python/sglang/srt/managers/detokenizer_manager.py

@@ -18,7 +18,7 @@ limitations under the License.
 import dataclasses
 import logging
 from collections import OrderedDict
-from typing import List
+from typing import List, Union
 
 import zmq
 
@@ -29,7 +29,7 @@ from sglang.srt.managers.io_struct import (
     BatchTokenIDOut,
     UpdateWeightReqOutput,
 )
-from sglang.srt.managers.schedule_batch import FINISH_MATCHED_STR
+from sglang.srt.managers.schedule_batch import FINISH_MATCHED_STR, FINISH_MATCHED_TOKEN
 from sglang.srt.server_args import PortArgs, ServerArgs
 from sglang.srt.utils import configure_logger, kill_parent_process
 from sglang.utils import find_printable_text, get_exception_traceback
@@ -75,6 +75,21 @@ class DetokenizerManager:
 
         self.decode_status = LimitedCapacityDict()
 
+    def trim_eos(self, output: Union[str, List[int]], finished_reason, no_eos_trim):
+        if no_eos_trim:
+            return output
+
+        # Trim stop str. TODO(lmzheng): handle the case where multiple stop strs are hit
+        if isinstance(finished_reason, FINISH_MATCHED_STR) and isinstance(output, str):
+            pos = output.find(finished_reason.matched)
+            return output[:pos] if pos != -1 else output
+        if isinstance(finished_reason, FINISH_MATCHED_TOKEN) and isinstance(
+            output, list
+        ):
+            assert len(output) > 0
+            return output[:-1]
+        return output
+
     def event_loop(self):
         """The event loop that handles requests"""
 
@@ -122,7 +137,13 @@ class DetokenizerManager:
                     s = self.decode_status[rid]
                     s.decode_ids = recv_obj.decode_ids[i]
 
-                read_ids.append(s.decode_ids[s.surr_offset :])
+                read_ids.append(
+                    self.trim_eos(
+                        s.decode_ids[s.surr_offset :],
+                        recv_obj.finished_reason[i],
+                        recv_obj.no_eos_trim[i],
+                    )
+                )
                 surr_ids.append(s.decode_ids[s.surr_offset : s.read_offset])
 
             # TODO(lmzheng): handle skip_special_tokens/spaces_between_special_tokens per request
@@ -152,13 +173,13 @@ class DetokenizerManager:
                     else:
                         new_text = find_printable_text(new_text)
 
-                output_strs.append(s.decoded_text + new_text)
+                output_strs.append(
-
+                    self.trim_eos(
-                # Trim stop str. TODO(lmzheng): handle the case where multiple stop strs are hit
+                        s.decoded_text + new_text,
-                if isinstance(recv_obj.finished_reason[i], FINISH_MATCHED_STR):
+                        recv_obj.finished_reason[i],
-                    pos = output_strs[i].find(recv_obj.finished_reason[i].matched)
+                        recv_obj.no_eos_trim[i],
-                    if pos != -1:
+                    )
-                        output_strs[i] = output_strs[i][:pos]
+                )
 
             self.send_to_tokenizer.send_pyobj(
                 BatchStrOut(

python/sglang/srt/managers/io_struct.py

@@ -295,6 +295,7 @@ class BatchTokenIDOut:
     spaces_between_special_tokens: List[bool]
     meta_info: List[Dict]
     finished_reason: List[BaseFinishReason]
+    no_eos_trim: List[bool]
 
 
 @dataclass

python/sglang/srt/managers/scheduler.py

@@ -883,6 +883,7 @@ class Scheduler:
             output_read_offsets = []
             output_skip_special_tokens = []
             output_spaces_between_special_tokens = []
+            output_no_eos_trim = []
         else:  # embedding or reward model
             output_embeddings = []
         unfinished_indices = []
@@ -914,6 +915,7 @@ class Scheduler:
                     output_spaces_between_special_tokens.append(
                         req.sampling_params.spaces_between_special_tokens
                     )
+                    output_no_eos_trim.append(req.sampling_params.no_eos_trim)
 
                     meta_info = {
                         "prompt_tokens": len(req.origin_input_ids),
@@ -961,6 +963,7 @@ class Scheduler:
                         output_spaces_between_special_tokens,
                         output_meta_info,
                         output_finished_reason,
+                        output_no_eos_trim,
                     )
                 )
             else:  # embedding or reward model

python/sglang/srt/openai_api/adapter.py

@@ -493,7 +493,13 @@ def v1_generate_request(
         top_logprobs_nums.append(
             request.logprobs if request.logprobs is not None else 0
         )
-        sampling_params_list.append(
+        sampling_params = []
+        if isinstance(request.no_eos_trim, list):
+            num_reqs = len(request.prompt)
+        else:
+            num_reqs = 1
+        for i in range(num_reqs):
+            sampling_params.append(
                 {
                     "temperature": request.temperature,
                     "max_new_tokens": request.max_tokens,
@@ -508,8 +514,17 @@ def v1_generate_request(
                     "json_schema": request.json_schema,
                     "n": request.n,
                     "ignore_eos": request.ignore_eos,
+                    "no_eos_trim": (
+                        request.no_eos_trim
+                        if not isinstance(request.no_eos_trim, list)
+                        else request.no_eos_trim[i]
+                    ),
                 }
             )
+        if num_reqs == 1:
+            sampling_params_list.append(sampling_params[0])
+        else:
+            sampling_params_list.append(sampling_params)
 
     if len(all_requests) == 1:
         prompt = prompts[0]

python/sglang/srt/openai_api/protocol.py

@@ -174,6 +174,7 @@ class CompletionRequest(BaseModel):
     min_tokens: int = 0
     repetition_penalty: Optional[float] = 1.0
     stop_token_ids: Optional[List[int]] = Field(default_factory=list)
+    no_eos_trim: Union[bool, List[bool]] = False
 
 
 class CompletionResponseChoice(BaseModel):

python/sglang/srt/sampling/sampling_params.py

@@ -40,6 +40,7 @@ class SamplingParams:
         regex: Optional[str] = None,
         n: int = 1,
         json_schema: Optional[str] = None,
+        no_eos_trim: bool = False,
     ) -> None:
         self.temperature = temperature
         self.top_p = top_p
@@ -60,6 +61,7 @@ class SamplingParams:
         self.regex = regex
         self.n = n
         self.json_schema = json_schema
+        self.no_eos_trim = no_eos_trim
 
         # Process some special cases
         if self.temperature < _SAMPLING_EPS:

python/sglang/srt/utils.py

@@ -690,3 +690,10 @@ def pytorch_profile(name, func, *args, data_size=-1):
     prof.export_chrome_trace(f"trace/{name}_{step_counter}.json")
     step_counter += 1
     return result
+
+
+def first_rank_print(*args, **kwargs):
+    if torch.cuda.current_device() == 0:
+        print(*args, **kwargs)
+    else:
+        pass