Merge branch 'v0.5.4_dev_maxiao' into 'v0.5.4_dev'

修复w8a8_marlin tp pp

See merge request OpenDAS/sglang!30

python/sglang/srt/layers/quantization/compressed_tensors/compressed_tensors_moe_marlin.py

@@ -15,6 +15,7 @@ from sglang.srt.layers.quantization.base_config import FusedMoEMethodBase
 
 from sglang.srt.utils import set_weight_attrs
 from sglang.srt.layers.moe import MoeRunner, MoeRunnerBackend, MoeRunnerConfig
+from sglang.srt.layers.moe.utils import get_moe_a2a_backend
 try:
     from lmslim.layers.fused_moe.fuse_moe_int8_marlin import fused_experts_impl_int8_marlin
 except Exception:
@@ -77,7 +78,7 @@ class CompressedTensorsW8A8Int8MarlinMoEMethod(CompressedTensorsMarlinMoEMethod)
             "weights")
         self.input_quant = self.quant_config.target_scheme_map["Linear"].get(
             "input_activations")
-        self.use_deepep = True
+        self.use_deepep = get_moe_a2a_backend().is_deepep()
         per_channel = (
             self.weight_quant.strategy == QuantizationStrategy.CHANNEL
             and self.input_quant.strategy == QuantizationStrategy.TOKEN)