[10/n] decouple quantization impl from vllm dependency - fix import (#13524)

python/sglang/srt/layers/quantization/fp8.py

@@ -10,35 +10,12 @@ import torch.nn.functional as F
 from torch.nn import Module
 from torch.nn.parameter import Parameter
 
-from sglang.srt.distributed import get_tp_group
-from sglang.srt.distributed.device_communicators.pynccl_allocator import (
-    use_symmetric_memory,
-)
-from sglang.srt.layers.dp_attention import is_allocation_symmetric
-
-try:
-    from vllm.model_executor.layers.quantization.utils.marlin_utils_fp8 import (
-        apply_fp8_marlin_linear,
-        prepare_fp8_layer_for_marlin,
-    )
-
-    MARLIN_FP8_AVAILABLE = True
-except ImportError:
-    MARLIN_FP8_AVAILABLE = False
-
-    def dummy_func(*args, **kwargs):
-        raise ImportError(
-            "marlin FP8 requires some operators from vllm. Please install vllm."
-        )
-
-    apply_fp8_marlin_linear = prepare_fp8_layer_for_marlin = dummy_func
-
-
 from sglang.srt.distributed import get_tensor_model_parallel_world_size, get_tp_group
 from sglang.srt.distributed.device_communicators.pynccl_allocator import (
     use_symmetric_memory,
 )
 from sglang.srt.layers.amx_utils import _amx_process_weight_after_loading
+from sglang.srt.layers.dp_attention import is_allocation_symmetric
 from sglang.srt.layers.moe import MoeRunner, MoeRunnerBackend, MoeRunnerConfig
 from sglang.srt.layers.moe.moe_runner.deep_gemm import DeepGemmMoeQuantInfo
 from sglang.srt.layers.moe.moe_runner.triton import TritonMoeQuantInfo
@@ -69,6 +46,10 @@ from sglang.srt.layers.quantization.fp8_utils import (
     normalize_e4m3fn_to_e4m3fnuz,
 )
 from sglang.srt.layers.quantization.kv_cache import BaseKVCacheMethod
+from sglang.srt.layers.quantization.marlin_utils_fp8 import (
+    apply_fp8_marlin_linear,
+    prepare_fp8_layer_for_marlin,
+)
 from sglang.srt.layers.quantization.unquant import UnquantizedLinearMethod
 from sglang.srt.layers.quantization.utils import (
     all_close_1d,
@@ -228,7 +209,7 @@ class Fp8LinearMethod(LinearMethodBase):
         # For GPUs that lack FP8 hardware support, we can leverage the Marlin
         # kernel for fast weight-only FP8 quantization
         self.use_marlin = False
-        if _is_cuda and MARLIN_FP8_AVAILABLE:
+        if _is_cuda:
             force_marlin = get_bool_env_var("SGLANG_FORCE_FP8_MARLIN")
             auto_enable = can_auto_enable_marlin_fp8()
             self.use_marlin = force_marlin or auto_enable