Clean up imports (#5467)

python/sglang/srt/layers/moe/ep_moe/layer.py

@@ -2,6 +2,7 @@ import logging
 from typing import Callable, List, Optional, Tuple
 
 import torch
+from torch.nn import Module
 
 try:
     from deep_gemm import (
@@ -13,8 +14,6 @@ try:
 except ImportError:
     use_deep_gemm = False
 
-from torch.nn import Module
-
 from sglang.srt.custom_op import CustomOp
 from sglang.srt.distributed import (
     get_tensor_model_parallel_rank,
@@ -37,21 +36,16 @@ from sglang.srt.layers.quantization.base_config import (
     QuantizeMethodBase,
 )
 from sglang.srt.layers.quantization.fp8 import Fp8Config, Fp8MoEMethod
+from sglang.srt.layers.quantization.fp8_kernel import scaled_fp8_quant
 from sglang.srt.model_executor.forward_batch_info import ForwardMode
-from sglang.srt.utils import DeepEPMode, is_cuda, is_hip, set_weight_attrs
-
-_is_cuda = is_cuda()
-
-if _is_cuda:
-    from sglang.srt.custom_op import scaled_fp8_quant as sgl_scaled_fp8_quant
-else:
-    from vllm import _custom_ops as vllm_ops
-
-logger = logging.getLogger(__name__)
+from sglang.srt.utils import DeepEPMode, is_hip, set_weight_attrs
 
 _is_hip = is_hip()
 
-_buffer = None
+if _is_hip:
+    from vllm._custom_ops import scaled_fp8_quant
+
+logger = logging.getLogger(__name__)
 
 
 class GroupedGemmRunner(torch.nn.Module):
@@ -740,20 +734,12 @@ class Fp8EPMoEMethod(Fp8MoEMethod):
             )
 
             for expert in range(layer.num_experts_per_partition):
-                if _is_cuda:
-                    w13_weight[expert, :, :], layer.w13_weight_scale[expert] = (
-                        sgl_scaled_fp8_quant(layer.w13_weight.data[expert, :, :])
-                    )
-                    w2_weight[expert, :, :], layer.w2_weight_scale[expert] = (
-                        sgl_scaled_fp8_quant(layer.w2_weight.data[expert, :, :])
-                    )
-                else:
-                    w13_weight[expert, :, :], layer.w13_weight_scale[expert] = (
-                        vllm_ops.scaled_fp8_quant(layer.w13_weight.data[expert, :, :])
-                    )
-                    w2_weight[expert, :, :], layer.w2_weight_scale[expert] = (
-                        vllm_ops.scaled_fp8_quant(layer.w2_weight.data[expert, :, :])
-                    )
+                w13_weight[expert, :, :], layer.w13_weight_scale[expert] = (
+                    scaled_fp8_quant(layer.w13_weight.data[expert, :, :])
+                )
+                w2_weight[expert, :, :], layer.w2_weight_scale[expert] = (
+                    scaled_fp8_quant(layer.w2_weight.data[expert, :, :])
+                )
             layer.w13_weight = torch.nn.Parameter(w13_weight, requires_grad=False)
             layer.w2_weight = torch.nn.Parameter(w2_weight, requires_grad=False)
             return

python/sglang/srt/layers/moe/fused_moe_triton/fused_moe.py

@@ -13,6 +13,7 @@ import triton
 import triton.language as tl
 
 from sglang.srt.layers.moe.topk import select_experts
+from sglang.srt.layers.quantization.fp8_kernel import scaled_fp8_quant
 from sglang.srt.utils import (
     direct_register_custom_op,
     get_bool_env_var,
@@ -22,28 +23,25 @@ from sglang.srt.utils import (
 )
 
 _is_hip = is_hip()
-
-
-logger = logging.getLogger(__name__)
-padding_size = 128 if bool(int(os.getenv("MOE_PADDING", "0"))) else 0
-
-enable_moe_align_block_size_triton = bool(
-    int(os.getenv("ENABLE_MOE_ALIGN_BLOCK_SIZE_TRITON", "0"))
-)
-
 _is_cuda = is_cuda()
 
 if _is_cuda:
     from sgl_kernel import gelu_and_mul, silu_and_mul
-
-    from sglang.srt.custom_op import scaled_fp8_quant as sgl_scaled_fp8_quant
 else:
     from vllm import _custom_ops as vllm_ops
+    from vllm._custom_ops import scaled_fp8_quant
 
 if _is_cuda or _is_hip:
     from sgl_kernel import moe_align_block_size as sgl_moe_align_block_size
 
 
+logger = logging.getLogger(__name__)
+padding_size = 128 if bool(int(os.getenv("MOE_PADDING", "0"))) else 0
+enable_moe_align_block_size_triton = bool(
+    int(os.getenv("ENABLE_MOE_ALIGN_BLOCK_SIZE_TRITON", "0"))
+)
+
+
 @triton.jit
 def write_zeros_to_output(
     c_ptr,
@@ -770,14 +768,9 @@ def invoke_fused_moe_kernel(
             # activation tensor-wise fp8 quantization, dynamic or static
             padded_size = padding_size
             # activations apply per-token quantization when weights apply per-channel quantization by default
-            if _is_cuda:
-                A, A_scale = sgl_scaled_fp8_quant(
-                    A, A_scale, use_per_token_if_dynamic=per_channel_quant
-                )
-            else:
-                A, A_scale = vllm_ops.scaled_fp8_quant(
-                    A, A_scale, use_per_token_if_dynamic=per_channel_quant
-                )
+            A, A_scale = scaled_fp8_quant(
+                A, A_scale, use_per_token_if_dynamic=per_channel_quant
+            )
         else:
             # activation block-wise fp8 quantization
             assert len(block_shape) == 2

python/sglang/srt/layers/moe/topk.py

@@ -13,7 +13,6 @@
 # ==============================================================================
 
 import math
-import os
 from typing import Callable, Optional
 
 import torch
@@ -29,6 +28,10 @@ _is_hip = is_hip()
 if _is_cuda:
     from sgl_kernel import moe_fused_gate
 
+if _is_cuda or _is_hip:
+    from sgl_kernel import topk_softmax
+
+
 expert_distribution_recorder = ExpertDistributionRecorder()
 
 
@@ -59,11 +62,6 @@ def fused_topk(
     topk: int,
     renormalize: bool,
 ):
-    if _is_cuda or _is_hip:
-        from sgl_kernel import topk_softmax
-    else:
-        from vllm import _custom_ops as vllm_ops
-
     assert hidden_states.shape[0] == gating_output.shape[0], "Number of tokens mismatch"
 
     M, _ = hidden_states.shape
@@ -76,20 +74,12 @@ def fused_topk(
         M, topk, dtype=torch.int32, device=hidden_states.device
     )
 
-    if _is_cuda or _is_hip:
-        topk_softmax(
-            topk_weights,
-            topk_ids,
-            token_expert_indicies,
-            gating_output.float(),
-        )
-    else:
-        vllm_ops.topk_softmax(
-            topk_weights,
-            topk_ids,
-            token_expert_indicies,
-            gating_output.float(),
-        )
+    topk_softmax(
+        topk_weights,
+        topk_ids,
+        token_expert_indicies,
+        gating_output.float(),
+    )
     del token_expert_indicies
 
     if renormalize: