Fix grammar backend (#2018)

python/sglang/srt/managers/schedule_batch.py

@@ -37,7 +37,6 @@ import torch
 
 from sglang.global_config import global_config
 from sglang.srt.configs.model_config import ModelConfig
-from sglang.srt.constrained.grammar import Grammar
 from sglang.srt.mem_cache.base_prefix_cache import BasePrefixCache
 from sglang.srt.mem_cache.chunk_cache import ChunkCache
 from sglang.srt.mem_cache.memory_pool import BaseTokenToKVPool, ReqToTokenPool
@@ -249,7 +248,7 @@ class Req:
         self.embedding = None
 
         # Constrained decoding
-        self.grammar: Optional[Grammar] = None
+        self.grammar = None
 
         # The number of cached tokens, that were already cached in the KV cache
         self.cached_tokens = 0
@@ -359,8 +358,6 @@ class Req:
                     return
 
     def jump_forward_and_retokenize(self, jump_forward_str, next_state):
-        assert self.grammar is not None and self.tokenizer is not None
-
         if self.origin_input_text is None:
             # Recovering text can only use unpadded ids
             self.origin_input_text = self.tokenizer.decode(
@@ -809,9 +806,10 @@ class ScheduleBatch:
 
         for i, req in enumerate(self.reqs):
             if req.grammar is not None:
-                jump_helper = req.grammar.try_jump(req.tokenizer)
-                if jump_helper.can_jump():
-                    suffix_ids = jump_helper.suffix_ids
+                jump_helper = req.grammar.try_jump_forward(req.tokenizer)
+                if jump_helper:
+                    suffix_ids, _ = jump_helper
+
                     # Current ids, for cache and revert
                     cur_all_ids = tuple(req.origin_input_ids + req.output_ids)[:-1]
                     cur_output_ids = req.output_ids
@@ -827,6 +825,8 @@ class ScheduleBatch:
                         next_state,
                     ) = req.grammar.jump_forward_str_state(jump_helper)
 
+                    # Make the incrementally decoded text part of jump_forward_str
+                    # so that the UTF-8 will not corrupt
                     jump_forward_str = new_text + jump_forward_str
                     if not req.jump_forward_and_retokenize(
                         jump_forward_str, next_state