[PD]decode: add CLIP_MAX_NEW_TOKEN for pop_preallocated (#8866)

python/sglang/srt/disaggregation/decode.py

@@ -51,7 +51,7 @@ from sglang.srt.mem_cache.base_prefix_cache import BasePrefixCache
 from sglang.srt.mem_cache.memory_pool import KVCache, ReqToTokenPool
 from sglang.srt.model_executor.forward_batch_info import ForwardMode
 from sglang.srt.torch_memory_saver_adapter import TorchMemorySaverAdapter
-from sglang.srt.utils import require_mlp_sync
+from sglang.srt.utils import get_int_env_var, require_mlp_sync
 
 logger = logging.getLogger(__name__)
 
@@ -59,6 +59,10 @@ if TYPE_CHECKING:
     from sglang.srt.managers.schedule_batch import Req
     from sglang.srt.managers.scheduler import Scheduler
 
+DECODE_CLIP_MAX_NEW_TOKEN = get_int_env_var(
+    "SGLANG_CLIP_MAX_NEW_TOKENS_ESTIMATION", 4096
+)
+
 
 class DecodeReqToTokenPool:
     """
@@ -384,7 +388,10 @@ class DecodePreallocQueue:
                 max(
                     required_tokens_for_request,
                     origin_input_len
-                    + decode_req.req.sampling_params.max_new_tokens
+                    + min(
+                        decode_req.req.sampling_params.max_new_tokens,
+                        DECODE_CLIP_MAX_NEW_TOKEN,
+                    )
                     - retractable_tokens,
                 )
                 > allocatable_tokens
@@ -433,7 +440,7 @@ class DecodePreallocQueue:
         need_space_for_single_req = (
             max(
                 [
-                    x.sampling_params.max_new_tokens
+                    min(x.sampling_params.max_new_tokens, DECODE_CLIP_MAX_NEW_TOKEN)
                     + len(x.origin_input_ids)
                     - retractable_tokens
                     for x in self.scheduler.running_batch.reqs