Move args from global_config to environ (#11332)

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

@@ -22,7 +22,7 @@ if os.environ["SGLANG_ENABLE_TORCH_COMPILE"] == "1":
     torch._logging.set_logs(dynamo=logging.ERROR)
     torch._dynamo.config.suppress_errors = True
 
-from sglang.global_config import global_config
+from sglang.srt.environ import envs
 from sglang.srt.layers.attention.base_attn_backend import AttentionBackend
 from sglang.srt.layers.attention.flashinfer_backend import (
     create_flashinfer_kv_indices_triton,
@@ -204,7 +204,7 @@ class FlashInferMLAAttnBackend(AttentionBackend):
         if global_workspace_buffer is None:
             # different from flashinfer zero_init_global_workspace_buffer
             global_workspace_buffer = torch.empty(
-                global_config.flashinfer_workspace_size,
+                envs.SGLANG_FLASHINFER_WORKSPACE_SIZE.get(),
                 dtype=torch.uint8,
                 device=model_runner.device,
             )