Fix incorrect KV indices creation when page_size=32 in TRTLLM MLA backend (#11985)

python/sglang/srt/layers/attention/flashinfer_mla_backend.py

@@ -9,19 +9,12 @@ and uses BatchMLAPaged wrapper for decoding.
 More details can be found in https://docs.flashinfer.ai/api/mla.html
 """
 
-import os
 from dataclasses import dataclass
 from functools import partial
 from typing import TYPE_CHECKING, Callable, Optional, Union
 
 import torch
 
-if os.environ["SGLANG_ENABLE_TORCH_COMPILE"] == "1":
-    import logging
-
-    torch._logging.set_logs(dynamo=logging.ERROR)
-    torch._dynamo.config.suppress_errors = True
-
 from sglang.srt.environ import envs
 from sglang.srt.layers.attention.base_attn_backend import AttentionBackend
 from sglang.srt.layers.attention.flashinfer_backend import (
@@ -45,6 +38,12 @@ if TYPE_CHECKING:
     from sglang.srt.model_executor.model_runner import ModelRunner
     from sglang.srt.speculative.spec_info import SpecInput
 
+if envs.SGLANG_ENABLE_TORCH_COMPILE.get():
+    import logging
+
+    torch._logging.set_logs(dynamo=logging.ERROR)
+    torch._dynamo.config.suppress_errors = True
+
 if is_flashinfer_available():
     from flashinfer import (
         BatchMLAPagedAttentionWrapper,