sglang 0.5.2rc1__py3-none-any.whl → 0.5.3rc0__py3-none-any.whl

sglang/srt/layers/quantization/unquant.py

@@ -9,6 +9,8 @@ from torch.nn.parameter import Parameter
 
 from sglang.srt.custom_op import CustomOp
 from sglang.srt.layers.amx_utils import _amx_process_weight_after_loading
+from sglang.srt.layers.moe import MoeRunner, MoeRunnerBackend, MoeRunnerConfig
+from sglang.srt.layers.moe.moe_runner.triton import TritonMoeQuantInfo
 from sglang.srt.layers.quantization.base_config import (
     FusedMoEMethodBase,
     LinearMethodBase,
@@ -24,8 +26,10 @@ from sglang.srt.utils import (
 )
 
 if TYPE_CHECKING:
-    from sglang.srt.layers.moe.moe_runner import MoeRunnerConfig
-    from sglang.srt.layers.moe.topk import TopKOutput
+    from sglang.srt.layers.moe.token_dispatcher import (
+        CombineInput,
+        StandardDispatchOutput,
+    )
 
 has_triton_kernels = importlib.util.find_spec("triton_kernels") is not None
 
@@ -155,7 +159,7 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
         layer: torch.nn.Module,
         num_experts: int,
         hidden_size: int,
-        intermediate_size: int,
+        intermediate_size_per_partition: int,
         params_dtype: torch.dtype,
         with_bias: bool = False,
         **extra_weight_attrs,
@@ -163,7 +167,7 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
         self.with_bias = with_bias
 
         # Fused gate_up_proj (column parallel)
-        w13_weight_n, w13_weight_k = 2 * intermediate_size, hidden_size
+        w13_weight_n, w13_weight_k = 2 * intermediate_size_per_partition, hidden_size
         if self.use_triton_kernels:
             w13_weight_n, w13_weight_k = w13_weight_k, w13_weight_n
         w13_weight = torch.nn.Parameter(
@@ -175,7 +179,11 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
 
         if self.with_bias:
             w13_weight_bias = torch.nn.Parameter(
-                torch.empty(num_experts, 2 * intermediate_size, dtype=torch.float32),
+                torch.empty(
+                    num_experts,
+                    2 * intermediate_size_per_partition,
+                    dtype=torch.float32,
+                ),
                 requires_grad=False,
             )
             layer.register_parameter("w13_weight_bias", w13_weight_bias)
@@ -184,7 +192,7 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
         # down_proj (row parallel)
         w2_weight_n, w2_weight_k = (
             hidden_size,
-            intermediate_size,
+            intermediate_size_per_partition,
         )
         if self.use_triton_kernels:
             w2_weight_n, w2_weight_k = w2_weight_k, w2_weight_n
@@ -222,33 +230,40 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
 
         return
 
+    def create_moe_runner(
+        self, layer: torch.nn.Module, moe_runner_config: MoeRunnerConfig
+    ):
+        self.moe_runner_config = moe_runner_config
+        self.runner = MoeRunner(MoeRunnerBackend.TRITON, moe_runner_config)
+
     def apply(
         self,
         layer: torch.nn.Module,
-        x: torch.Tensor,
-        topk_output: TopKOutput,
-        moe_runner_config: MoeRunnerConfig,
-    ) -> torch.Tensor:
+        dispatch_output: StandardDispatchOutput,
+    ) -> CombineInput:
 
         return self.forward(
-            x=x,
             layer=layer,
-            topk_output=topk_output,
-            moe_runner_config=moe_runner_config,
+            dispatch_output=dispatch_output,
         )
 
     def forward_cuda(
         self,
         layer: torch.nn.Module,
-        x: torch.Tensor,
-        topk_output: TopKOutput,
-        moe_runner_config: MoeRunnerConfig,
-    ) -> torch.Tensor:
+        dispatch_output: StandardDispatchOutput,
+    ) -> CombineInput:
+
+        from sglang.srt.layers.moe.token_dispatcher import StandardCombineInput
+
+        x = dispatch_output.hidden_states
+        topk_output = dispatch_output.topk_output
+
+        moe_runner_config = self.moe_runner_config
 
         if self.use_triton_kernels:
             if self.with_bias:
                 assert self.triton_kernel_moe_with_bias_forward is not None
-                return self.triton_kernel_moe_with_bias_forward(
+                output = self.triton_kernel_moe_with_bias_forward(
                     hidden_states=x,
                     w1=layer.w13_weight,
                     w2=layer.w2_weight,
@@ -261,13 +276,14 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
                 )
             else:
                 assert self.triton_kernel_moe_forward is not None
-                return self.triton_kernel_moe_forward(
+                output = self.triton_kernel_moe_forward(
                     hidden_states=x,
                     w1=layer.w13_weight,
                     w2=layer.w2_weight,
                     topk_output=topk_output,
                     moe_runner_config=moe_runner_config,
                 )
+            return StandardCombineInput(hidden_states=output)
         else:
             if _use_aiter:
                 assert not moe_runner_config.no_combine, "unsupported"
@@ -284,7 +300,7 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
                     topk_weights = torch.ones_like(
                         topk_weights, dtype=torch.float32
                     )  # topk_weights must be FP32 (float32)
-                return fused_moe(
+                output = fused_moe(
                     x,
                     layer.w13_weight,
                     layer.w2_weight,
@@ -296,28 +312,30 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
                         else ActivationType.Gelu
                     ),
                 )
+                return StandardCombineInput(hidden_states=output)
             else:
-                from sglang.srt.layers.moe.fused_moe_triton.fused_moe import (
-                    fused_experts,
-                )
 
-                return fused_experts(
-                    hidden_states=x,
-                    w1=layer.w13_weight,
-                    w2=layer.w2_weight,
-                    b1=getattr(layer, "w13_weight_bias", None),
+                quant_info = TritonMoeQuantInfo(
+                    w13_weight=layer.w13_weight,
+                    w2_weight=layer.w2_weight,
+                    b13=getattr(layer, "w13_weight_bias", None),
                     b2=getattr(layer, "w2_weight_bias", None),
-                    topk_output=topk_output,
-                    moe_runner_config=moe_runner_config,
                 )
+                return self.runner.run(dispatch_output, quant_info)
 
     def forward_cpu(
         self,
         layer: torch.nn.Module,
-        x: torch.Tensor,
-        topk_output: TopKOutput,
-        moe_runner_config: MoeRunnerConfig,
-    ) -> torch.Tensor:
+        dispatch_output: StandardDispatchOutput,
+    ) -> CombineInput:
+
+        from sglang.srt.layers.moe.token_dispatcher import StandardCombineInput
+
+        x = dispatch_output.hidden_states
+        topk_output = dispatch_output.topk_output
+
+        moe_runner_config = self.moe_runner_config
+
         assert (
             moe_runner_config.activation == "silu"
         ), f"activation = {moe_runner_config.activation} is not supported."
@@ -332,7 +350,7 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
             x, topk_weights = apply_topk_weights_cpu(
                 moe_runner_config.apply_router_weight_on_input, topk_weights, x
             )
-            return torch.ops.sgl_kernel.fused_experts_cpu(
+            output = torch.ops.sgl_kernel.fused_experts_cpu(
                 x,
                 layer.w13_weight,
                 layer.w2_weight,
@@ -348,33 +366,103 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
                 None,  # a2_scale
                 True,  # is_vnni
             )
+            return StandardCombineInput(hidden_states=output)
         else:
             from sglang.srt.layers.moe.fused_moe_native import moe_forward_native
 
-            return moe_forward_native(
+            output = moe_forward_native(
                 layer,
                 x,
                 topk_output,
                 moe_runner_config,
             )
+            return StandardCombineInput(hidden_states=output)
 
     def forward_npu(
         self,
         layer: torch.nn.Module,
-        x: torch.Tensor,
-        topk_output: TopKOutput,
-        moe_runner_config: MoeRunnerConfig,
-    ) -> torch.Tensor:
-        from sglang.srt.layers.moe.fused_moe_native import moe_forward_native
+        dispatch_output: StandardDispatchOutput,
+    ) -> CombineInput:
+
+        import torch_npu
+
+        from sglang.srt.layers.moe.token_dispatcher import StandardCombineInput
+
+        x = dispatch_output.hidden_states
+        topk_weights, topk_ids, _ = dispatch_output.topk_output
+
+        original_dtype = x.dtype
+        num_tokens = x.shape[0]
+        topk_weights = topk_weights.to(x.dtype)
+        topk_ids = topk_ids.to(torch.int32)
+        num_experts = layer.num_experts
+        top_k = layer.top_k
+        row_idx_len = num_tokens * top_k
+        row_idx = (
+            torch.arange(0, row_idx_len, dtype=torch.int32, device=topk_weights.device)
+            .view(top_k, -1)
+            .permute(1, 0)
+            .contiguous()
+        )
 
-        return moe_forward_native(
-            layer,
-            x,
-            topk_output,
-            moe_runner_config,
+        hidden_states, expanded_row_idx, expanded_expert_idx = (
+            torch_npu.npu_moe_init_routing(
+                x, row_idx=row_idx, expert_idx=topk_ids, active_num=num_tokens
+            )
+        )
+
+        expert_tokens = torch_npu.npu_moe_compute_expert_tokens(
+            expanded_expert_idx, num_experts
         )
 
-    def forward_tpu(self, *args, **kwargs) -> torch.Tensor:
+        expert_tokens = expert_tokens.to(torch.int64)
+        if layer.w13_weight.shape[-1] == layer.hidden_size:
+            w13 = layer.w13_weight.transpose(1, 2)
+            w2 = layer.w2_weight.transpose(1, 2)
+
+        # gmm1: gate_up_proj
+        hidden_states = torch_npu.npu_grouped_matmul(
+            x=[hidden_states],
+            weight=[w13],
+            split_item=2,
+            group_list_type=0,
+            group_type=0,
+            group_list=expert_tokens,
+            output_dtype=original_dtype,
+        )[0]
+
+        # act_fn:
+        if self.moe_runner_config.activation == "silu":
+            hidden_states = torch_npu.npu_swiglu(hidden_states)
+        else:
+            from sglang.srt.layers.activation import GeluAndMul
+
+            hidden_states = GeluAndMul()(hidden_states)
+
+        # gmm2: down_proj
+        hidden_states = torch_npu.npu_grouped_matmul(
+            x=[hidden_states],
+            weight=[w2],
+            split_item=2,
+            group_list_type=0,
+            group_type=0,
+            group_list=expert_tokens,
+            output_dtype=original_dtype,
+        )[0]
+
+        final_hidden_states = torch_npu.npu_moe_finalize_routing(
+            hidden_states,
+            skip1=None,
+            skip2=None,
+            bias=None,
+            scales=topk_weights,
+            expanded_src_to_dst_row=expanded_row_idx,
+            export_for_source_row=topk_ids,
+        )
+
+        return StandardCombineInput(hidden_states=final_hidden_states)
+
+    def forward_tpu(self, *args, **kwargs) -> CombineInput:
         raise NotImplementedError("The TPU backend currently does not support MoE.")
 
     forward_native = forward_cpu

sglang/srt/layers/quantization/w4afp8.py

@@ -17,12 +17,19 @@ from sglang.srt.layers.quantization.base_config import (
 from sglang.srt.layers.quantization.fp8 import Fp8LinearMethod
 from sglang.srt.layers.quantization.unquant import UnquantizedLinearMethod
 from sglang.srt.layers.quantization.utils import is_layer_skipped
-from sglang.srt.utils import set_weight_attrs
+from sglang.srt.utils import is_npu, set_weight_attrs
+
+_is_npu = is_npu()
+if not _is_npu:
+    from sglang.srt.layers.moe.cutlass_w4a8_moe import cutlass_w4a8_moe
 
 if TYPE_CHECKING:
     from sglang.srt.layers.moe import MoeRunnerConfig
     from sglang.srt.layers.moe.ep_moe.layer import EPMoE
-    from sglang.srt.layers.moe.topk import StandardTopKOutput
+    from sglang.srt.layers.moe.token_dispatcher import (
+        CombineInput,
+        StandardDispatchOutput,
+    )
 
 ACTIVATION_SCHEMES = ["static", "dynamic"]
 
@@ -133,7 +140,7 @@ class W4AFp8MoEMethod(FusedMoEMethodBase):
         layer: EPMoE,
         num_experts: int,
         hidden_size: int,
-        intermediate_size: int,
+        intermediate_size_per_partition: int,
         params_dtype: torch.dtype,
         **extra_weight_attrs,
     ):
@@ -145,7 +152,7 @@ class W4AFp8MoEMethod(FusedMoEMethodBase):
         w13_weight = torch.nn.Parameter(
             torch.empty(
                 num_experts,
-                intermediate_size * 2,
+                intermediate_size_per_partition * 2,
                 hidden_size // 2,
                 dtype=torch.int8,
             ),
@@ -159,7 +166,7 @@ class W4AFp8MoEMethod(FusedMoEMethodBase):
             torch.empty(
                 num_experts,
                 hidden_size,
-                intermediate_size // 2,
+                intermediate_size_per_partition // 2,
                 dtype=torch.int8,
             ),
             requires_grad=False,
@@ -173,7 +180,7 @@ class W4AFp8MoEMethod(FusedMoEMethodBase):
         w13_weight_scale = torch.nn.Parameter(
             torch.zeros(
                 num_experts,
-                2 * intermediate_size,
+                2 * intermediate_size_per_partition,
                 hidden_size // self.quant_config.group_size,
                 dtype=torch.float32,
             ),
@@ -186,7 +193,7 @@ class W4AFp8MoEMethod(FusedMoEMethodBase):
             torch.zeros(
                 num_experts,
                 hidden_size,
-                intermediate_size // self.quant_config.group_size,
+                intermediate_size_per_partition // self.quant_config.group_size,
                 dtype=torch.float32,
             ),
             requires_grad=False,
@@ -220,13 +227,13 @@ class W4AFp8MoEMethod(FusedMoEMethodBase):
         )
         self.c_strides1 = torch.full(
             (num_experts, 3),
-            2 * intermediate_size,
+            2 * intermediate_size_per_partition,
             device=device,
             dtype=torch.int64,
         )
         self.a_strides2 = torch.full(
             (num_experts, 3),
-            intermediate_size,
+            intermediate_size_per_partition,
             device=device,
             dtype=torch.int64,
         )
@@ -282,16 +289,22 @@ class W4AFp8MoEMethod(FusedMoEMethodBase):
         )
         layer.w2_input_scale = Parameter(new_w2_input_scale, requires_grad=False)
 
+    def create_moe_runner(
+        self, layer: torch.nn.Module, moe_runner_config: MoeRunnerConfig
+    ):
+        self.moe_runner_config = moe_runner_config
+
     def apply(
         self,
         layer: EPMoE,
-        x: torch.Tensor,
-        topk_output: StandardTopKOutput,
-        moe_runner_config: MoeRunnerConfig,
-    ) -> torch.Tensor:
+        dispatch_output: StandardDispatchOutput,
+    ) -> CombineInput:
 
-        # TODO(ch-wan): move it out of this class
         from sglang.srt.layers.moe.cutlass_w4a8_moe import cutlass_w4a8_moe
+        from sglang.srt.layers.moe.token_dispatcher import StandardCombineInput
+
+        x = dispatch_output.hidden_states
+        topk_output = dispatch_output.topk_output
 
         topk_weights, topk_ids, _ = topk_output
         local_topk_ids = topk_ids
@@ -328,6 +341,6 @@ class W4AFp8MoEMethod(FusedMoEMethodBase):
             layer.w13_input_scale,
             layer.w2_input_scale,
         )
-        if moe_runner_config.routed_scaling_factor is not None:
-            output *= moe_runner_config.routed_scaling_factor
-        return output
+        if self.moe_runner_config.routed_scaling_factor is not None:
+            output *= self.moe_runner_config.routed_scaling_factor
+        return StandardCombineInput(hidden_states=output)

sglang/srt/layers/quantization/w8a8_fp8.py

@@ -5,6 +5,8 @@ from typing import TYPE_CHECKING, Any, Dict, List, Optional
 import torch
 from torch.nn.parameter import Parameter
 
+from sglang.srt.layers.moe import MoeRunner, MoeRunnerBackend, MoeRunnerConfig
+from sglang.srt.layers.moe.moe_runner.triton import TritonMoeQuantInfo
 from sglang.srt.layers.parameter import ChannelQuantScaleParameter, ModelWeightParameter
 from sglang.srt.layers.quantization.base_config import (
     FusedMoEMethodBase,
@@ -26,8 +28,10 @@ from sglang.srt.layers.quantization.fp8_utils import (
 from sglang.srt.utils import set_weight_attrs
 
 if TYPE_CHECKING:
-    from sglang.srt.layers.moe.moe_runner import MoeRunnerConfig
-    from sglang.srt.layers.moe.topk import StandardTopKOutput
+    from sglang.srt.layers.moe.token_dispatcher import (
+        CombineInput,
+        StandardDispatchOutput,
+    )
 
 _is_fp8_fnuz = is_fp8_fnuz()
 
@@ -209,7 +213,7 @@ class W8A8FP8MoEMethod(FusedMoEMethodBase):
         layer: torch.nn.Module,
         num_experts: int,
         hidden_size: int,
-        intermediate_size: int,
+        intermediate_size_per_partition: int,
         params_dtype: torch.dtype,
         **extra_weight_attrs,
     ):
@@ -218,7 +222,10 @@ class W8A8FP8MoEMethod(FusedMoEMethodBase):
         # WEIGHTS
         w13_weight = torch.nn.Parameter(
             torch.empty(
-                num_experts, 2 * intermediate_size, hidden_size, dtype=fp8_dtype
+                num_experts,
+                2 * intermediate_size_per_partition,
+                hidden_size,
+                dtype=fp8_dtype,
             ),
             requires_grad=False,
         )
@@ -226,14 +233,21 @@ class W8A8FP8MoEMethod(FusedMoEMethodBase):
         set_weight_attrs(w13_weight, extra_weight_attrs)
 
         w2_weight = torch.nn.Parameter(
-            torch.empty(num_experts, hidden_size, intermediate_size, dtype=fp8_dtype),
+            torch.empty(
+                num_experts,
+                hidden_size,
+                intermediate_size_per_partition,
+                dtype=fp8_dtype,
+            ),
             requires_grad=False,
         )
         layer.register_parameter("w2_weight", w2_weight)
         set_weight_attrs(w2_weight, extra_weight_attrs)
 
         w13_weight_scale = torch.nn.Parameter(
-            torch.ones(num_experts, 2 * intermediate_size, 1, dtype=torch.float32),
+            torch.ones(
+                num_experts, 2 * intermediate_size_per_partition, 1, dtype=torch.float32
+            ),
             requires_grad=False,
         )
         w2_weight_scale = torch.nn.Parameter(
@@ -266,25 +280,26 @@ class W8A8FP8MoEMethod(FusedMoEMethodBase):
             layer.w2_weight_scale.data, requires_grad=False
         )
 
+    def create_moe_runner(
+        self, layer: torch.nn.Module, moe_runner_config: MoeRunnerConfig
+    ):
+        self.moe_runner_config = moe_runner_config
+        self.runner = MoeRunner(MoeRunnerBackend.TRITON, moe_runner_config)
+
     def apply(
         self,
         layer: torch.nn.Module,
-        x: torch.Tensor,
-        topk_output: StandardTopKOutput,
-        moe_runner_config: MoeRunnerConfig,
-    ) -> torch.Tensor:
-        from sglang.srt.layers.moe.fused_moe_triton.fused_moe import fused_experts
+        dispatch_output: StandardDispatchOutput,
+    ) -> CombineInput:
 
-        return fused_experts(
-            x,
-            layer.w13_weight,
-            layer.w2_weight,
-            topk_output=topk_output,
-            moe_runner_config=moe_runner_config,
+        quant_info = TritonMoeQuantInfo(
+            w13_weight=layer.w13_weight,
+            w2_weight=layer.w2_weight,
             use_fp8_w8a8=True,
             per_channel_quant=True,
-            w1_scale=(layer.w13_weight_scale),
-            w2_scale=(layer.w2_weight_scale),
-            a1_scale=layer.w13_input_scale,
+            w13_scale=layer.w13_weight_scale,
+            w2_scale=layer.w2_weight_scale,
+            a13_scale=layer.w13_input_scale,
             a2_scale=layer.w2_input_scale,
         )
+        return self.runner.run(dispatch_output, quant_info)