multimodal: precompute hash for MultimodalDataItem (#14354)

Signed-off-by: Feng Su <sufeng@linux.alibaba.com>
Signed-off-by: Junjie Mao <junjie.mao@linux.alibaba.com>

docs/references/environment_variables.md

@@ -36,6 +36,7 @@ SGLang supports various environment variables that can be used to configure its
 | `SGLANG_SCHEDULER_RECV_SKIPPER_WEIGHT_DECODE` | Weight increment for decode forward mode in scheduler recv skipper. Works with `--scheduler-recv-interval` to control polling frequency during decode phase. | `1` |
 | `SGLANG_SCHEDULER_RECV_SKIPPER_WEIGHT_VERIFY` | Weight increment for target verify forward mode in scheduler recv skipper. Works with `--scheduler-recv-interval` to control polling frequency during verification phase. | `1` |
 | `SGLANG_SCHEDULER_RECV_SKIPPER_WEIGHT_NONE` | Weight increment when forward mode is None in scheduler recv skipper. Works with `--scheduler-recv-interval` to control polling frequency when no specific forward mode is active. | `1` |
+| `SGLANG_MM_PRECOMPUTE_HASH` | Enable precomputing of hash values for MultimodalDataItem | `false` |
 
 
 ## DeepGEMM Configuration (Advanced Optimization)

python/sglang/srt/environ.py

@@ -330,6 +330,7 @@ class Envs:
     SGLANG_IMAGE_MAX_PIXELS = EnvInt(16384 * 28 * 28)
     SGLANG_RESIZE_RESAMPLE = EnvStr("")
     SGLANG_MM_BUFFER_SIZE_MB = EnvInt(0)
+    SGLANG_MM_PRECOMPUTE_HASH = EnvBool(False)
 
     # VLM Item CUDA IPC Transport
     SGLANG_USE_CUDA_IPC_TRANSPORT=EnvBool(False)

python/sglang/srt/managers/schedule_batch.py

@@ -243,6 +243,9 @@ class MultimodalDataItem:
         """
         Set the pad value after first hashing the data
         """
+        if self.pad_value is not None:
+            return
+
         from sglang.srt.managers.mm_utils import hash_feature
 
         if self.hash is None:

python/sglang/srt/managers/tokenizer_manager.py

@@ -73,7 +73,7 @@ from sglang.srt.managers.io_struct import (
 from sglang.srt.managers.mm_utils import TensorTransportMode
 from sglang.srt.managers.multimodal_processor import get_mm_processor, import_processors
 from sglang.srt.managers.request_metrics_exporter import RequestMetricsExporterManager
-from sglang.srt.managers.schedule_batch import RequestStage
+from sglang.srt.managers.schedule_batch import MultimodalDataItem, RequestStage
 from sglang.srt.managers.scheduler import is_health_check_generate_req
 from sglang.srt.managers.scheduler_input_blocker import input_blocker_guard_region
 from sglang.srt.managers.tokenizer_communicator_mixin import TokenizerCommunicatorMixin
@@ -653,6 +653,14 @@ class TokenizerManager(TokenizerCommunicatorMixin, TokenizerManagerMultiItemMixi
 
             if mm_inputs and "input_ids" in mm_inputs:
                 input_ids = mm_inputs["input_ids"]
+            if (
+                envs.SGLANG_MM_PRECOMPUTE_HASH.get()
+                and mm_inputs
+                and "mm_items" in mm_inputs
+            ):
+                for item in mm_inputs["mm_items"]:
+                    if isinstance(item, MultimodalDataItem):
+                        item.set_pad_value()
         else:
             mm_inputs = None