Fix cuda graph with flashinfer (#675)

benchmark/gsm8k/bench_sglang.py

@@ -64,7 +64,7 @@ def main(args):
     @sgl.function
     def few_shot_gsm8k(s, question):
         s += few_shot_examples + question
-        s += sgl.gen("answer", max_tokens=256, stop="Question")
+        s += sgl.gen("answer", max_tokens=512, stop="Question")
 
     #####################################
     ########## SGL Program End ##########

python/sglang/srt/managers/controller/cuda_graph_runner.py

@@ -150,8 +150,8 @@ class CudaGraphRunner:
         index = bisect.bisect_left(self.batch_size_list, raw_bs)
         bs = self.batch_size_list[index]
         if bs != raw_bs:
-            self.seq_lens.zero_()
-            self.position_ids_offsets.fill_(1)
+            self.seq_lens.fill_(1)
+            self.position_ids_offsets.zero_()
             self.out_cache_loc.zero_()
 
         # Common inputs