refactor(disagg): register SGLANG_ENCODER_MM_LOAD_WORKERS in Envs (#36006)

python/sglang/srt/disaggregation/encoder/preprocessor.py

@@ -10,7 +10,6 @@ import asyncio
 import concurrent.futures
 import functools
 import logging
-import os
 from dataclasses import dataclass
 from typing import Callable, List, Optional, Tuple, Union
 
@@ -149,7 +148,7 @@ class EncoderPreprocessor:
             max_workers=envs.SGLANG_ENCODER_PREPROC_WORKERS.get()
         )
         self.io_executor = concurrent.futures.ThreadPoolExecutor(
-            max_workers=int(os.environ.get("SGLANG_ENCODER_MM_LOAD_WORKERS", 4))
+            max_workers=envs.SGLANG_ENCODER_MM_LOAD_WORKERS.get()
         )
 
     # ------------------------------------------------------------------

python/sglang/srt/environ.py

@@ -1235,6 +1235,7 @@ class Envs:
     SGLANG_ENCODER_IMAGE_PROCESSOR_USE_GPU = EnvBool(False)
     SGLANG_ENCODER_MAX_BATCH_SIZE = EnvInt(8)
     SGLANG_ENCODER_PREPROC_WORKERS = EnvInt(8)
+    SGLANG_ENCODER_MM_LOAD_WORKERS = EnvInt(4)
     # EncoderBootstrapServer health-check tuning.  Interval == 0 disables it.
     SGLANG_ENCODER_BOOTSTRAP_HEALTH_CHECK_INTERVAL = EnvFloat(10.0)
     SGLANG_ENCODER_BOOTSTRAP_HEALTH_CHECK_TIMEOUT = EnvFloat(2.0)