[Bug Fix] Validate tokenizer-dependent features with skip_tokenizer_init (#27882)

Co-authored-by: Randall <randall@iterationlab.com>
Co-authored-by: Cursor <cursoragent@cursor.com>

python/sglang/srt/sampling/sampling_params.py

@@ -29,6 +29,36 @@ TOP_K_ALL = 1 << 30
 logger = logging.getLogger(__name__)
 
 
+def raise_if_tokenizer_required(
+    tokenizer, stop_strs, stop_regex_strs, min_new_tokens=0
+):
+    """Raise ValueError if tokenizer-dependent features are used without a tokenizer.
+
+    String-based stop conditions (stop_strs, stop_regex_strs) require tokenizer.decode()
+    to convert output token IDs to text for matching. min_new_tokens requires the
+    tokenizer's eos_token_id to penalize. When skip_tokenizer_init=True, these cannot
+    be used.
+    """
+    if tokenizer is not None:
+        return
+
+    if stop_strs:
+        raise ValueError(
+            f"stop={stop_strs!r} is unavailable when skip_tokenizer_init=True "
+            "(requires tokenizer to decode tokens to text for matching)."
+        )
+    if stop_regex_strs:
+        raise ValueError(
+            f"stop_regex={stop_regex_strs!r} is unavailable when skip_tokenizer_init=True "
+            "(requires tokenizer to decode tokens to text for matching)."
+        )
+    if min_new_tokens > 0:
+        raise ValueError(
+            f"min_new_tokens={min_new_tokens} is unavailable when skip_tokenizer_init=True "
+            "(requires tokenizer for eos_token_id)."
+        )
+
+
 class SamplingParams:
     """
     The sampling parameters.
@@ -210,6 +240,11 @@ class SamplingParams:
 
             self.stop_regex_max_len = stop_regex_max_len
 
+        # Validate tokenizer is available for tokenizer-dependent features
+        raise_if_tokenizer_required(
+            tokenizer, self.stop_strs, self.stop_regex_strs, self.min_new_tokens
+        )
+
 
 # This function gets a strict upperbound on the maximum number of tokens that would need
 # to be buffered to match the input regex string

python/sglang/srt/server_args.py

@@ -4354,6 +4354,11 @@ class ServerArgs:
                 )
                 self.enable_dynamic_batch_tokenizer = False
 
+            logger.info(
+                "skip_tokenizer_init=True: string-based stop conditions (stop, stop_regex) "
+                "and min_new_tokens are unavailable."
+            )
+
     def _handle_environment_variables(self):
         envs.SGLANG_ENABLE_TORCH_COMPILE.set("1" if self.enable_torch_compile else "0")
         if self.mamba_ssm_dtype is not None:
@@ -4681,6 +4686,14 @@ class ServerArgs:
                     self.preferred_sampling_params
                 )
 
+            # Validate preferred_sampling_params doesn't use tokenizer-dependent features
+            if self.skip_tokenizer_init:
+                from sglang.srt.sampling.sampling_params import SamplingParams
+
+                test_params = SamplingParams(**self.preferred_sampling_params)
+                # raises if tokenizer-dependent features used
+                test_params.normalize(None)
+
     def _handle_crash_dump_env(self):
         if not self.crash_dump_folder:
             return