sglang 0.4.6.post4__py3-none-any.whl → 0.4.7__py3-none-any.whl

sglang/srt/models/qwen3_moe.py

@@ -17,21 +17,32 @@
 
 """Inference-only Qwen3MoE model compatible with HuggingFace weights."""
 
-from functools import partial
+import logging
 from typing import Any, Dict, Iterable, Optional, Tuple
 
 import torch
-import torch.nn.functional as F
 from torch import nn
 
 from sglang.srt.distributed import (
+    get_pp_group,
     get_tensor_model_parallel_rank,
     get_tensor_model_parallel_world_size,
+    parallel_state,
     split_tensor_along_last_dim,
     tensor_model_parallel_all_gather,
     tensor_model_parallel_all_reduce,
 )
 from sglang.srt.layers.activation import SiluAndMul
+from sglang.srt.layers.communicator import LayerCommunicator, LayerScatterModes
+from sglang.srt.layers.dp_attention import (
+    attn_tp_all_gather,
+    attn_tp_reduce_scatter,
+    dp_gather_partial,
+    dp_scatter,
+    get_attention_tp_rank,
+    get_attention_tp_size,
+    get_local_attention_dp_size,
+)
 from sglang.srt.layers.layernorm import RMSNorm
 from sglang.srt.layers.linear import (
     MergedColumnParallelLinear,
@@ -39,52 +50,73 @@ from sglang.srt.layers.linear import (
     ReplicatedLinear,
     RowParallelLinear,
 )
-from sglang.srt.layers.logits_processor import LogitsProcessor
-from sglang.srt.layers.moe.ep_moe.layer import EPMoE
+from sglang.srt.layers.logits_processor import LogitsProcessor, LogitsProcessorOutput
+from sglang.srt.layers.moe.ep_moe.layer import get_moe_impl_class
+from sglang.srt.layers.moe.ep_moe.token_dispatcher import DeepEPDispatcher
 from sglang.srt.layers.moe.fused_moe_triton import FusedMoE
+from sglang.srt.layers.moe.topk import select_experts
 from sglang.srt.layers.quantization.base_config import QuantizationConfig
 from sglang.srt.layers.radix_attention import RadixAttention
 from sglang.srt.layers.rotary_embedding import get_rope
+from sglang.srt.layers.utils import get_layer_id
 from sglang.srt.layers.vocab_parallel_embedding import (
     ParallelLMHead,
     VocabParallelEmbedding,
 )
+from sglang.srt.managers.expert_distribution import (
+    get_global_expert_distribution_recorder,
+)
+from sglang.srt.managers.expert_location import ModelConfigForExpertLocation
+from sglang.srt.managers.expert_location_dispatch import ExpertLocationDispatchInfo
 from sglang.srt.managers.schedule_batch import global_server_args_dict
-from sglang.srt.model_executor.forward_batch_info import ForwardBatch
+from sglang.srt.model_executor.forward_batch_info import (
+    ForwardBatch,
+    ForwardMode,
+    PPProxyTensors,
+)
 from sglang.srt.model_loader.weight_utils import default_weight_loader
 from sglang.srt.models.qwen2_moe import Qwen2MoeMLP as Qwen3MoeMLP
 from sglang.srt.models.qwen2_moe import Qwen2MoeModel
-from sglang.srt.utils import add_prefix
+from sglang.srt.two_batch_overlap import MaybeTboDeepEPDispatcher
+from sglang.srt.utils import DeepEPMode, add_prefix, is_non_idle_and_non_empty
 
 Qwen3MoeConfig = None
 
+logger = logging.getLogger(__name__)
+
 
 class Qwen3MoeSparseMoeBlock(nn.Module):
     def __init__(
         self,
+        layer_id: int,
         config: Qwen3MoeConfig,
         quant_config: Optional[QuantizationConfig] = None,
         prefix: str = "",
     ):
         super().__init__()
         self.tp_size = get_tensor_model_parallel_world_size()
-
+        self.layer_id = layer_id
         if self.tp_size > config.num_experts:
             raise ValueError(
                 f"Tensor parallel size {self.tp_size} is greater than "
                 f"the number of experts {config.num_experts}."
             )
 
-        MoEImpl = EPMoE if global_server_args_dict["enable_ep_moe"] else FusedMoE
-
-        self.experts = MoEImpl(
-            num_experts=config.num_experts,
+        self.experts = get_moe_impl_class()(
+            num_experts=config.num_experts
+            + global_server_args_dict["ep_num_redundant_experts"],
             top_k=config.num_experts_per_tok,
+            layer_id=layer_id,
             hidden_size=config.hidden_size,
             intermediate_size=config.moe_intermediate_size,
             renormalize=config.norm_topk_prob,
             quant_config=quant_config,
             prefix=add_prefix("experts", prefix),
+            **(
+                dict(deepep_mode=DeepEPMode[global_server_args_dict["deepep_mode"]])
+                if global_server_args_dict["enable_deepep_moe"]
+                else {}
+            ),
         )
 
         self.gate = ReplicatedLinear(
@@ -95,7 +127,45 @@ class Qwen3MoeSparseMoeBlock(nn.Module):
             prefix=add_prefix("gate", prefix),
         )
 
-    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        if global_server_args_dict["enable_deepep_moe"]:
+            # TODO: we will support tp < ep in the future
+            self.ep_size = get_tensor_model_parallel_world_size()
+            self.num_experts = (
+                config.num_experts + global_server_args_dict["ep_num_redundant_experts"]
+            )
+            self.top_k = config.num_experts_per_tok
+            self.renormalize = config.norm_topk_prob
+
+            self.deepep_dispatcher = MaybeTboDeepEPDispatcher(
+                group=parallel_state.get_tp_group().device_group,
+                router_topk=self.top_k,
+                permute_fusion=True,
+                num_experts=self.num_experts,
+                num_local_experts=config.num_experts // self.tp_size,
+                hidden_size=config.hidden_size,
+                params_dtype=config.torch_dtype,
+                deepep_mode=DeepEPMode[global_server_args_dict["deepep_mode"]],
+                async_finish=True,  # TODO
+                return_recv_hook=True,
+            )
+
+    def forward(
+        self, hidden_states: torch.Tensor, forward_batch: Optional[ForwardBatch] = None
+    ) -> torch.Tensor:
+
+        if not global_server_args_dict["enable_deepep_moe"]:
+            return self.forward_normal(hidden_states)
+        else:
+            return self.forward_deepep(hidden_states, forward_batch)
+
+    def get_moe_weights(self):
+        return [
+            x.data
+            for name, x in self.experts.named_parameters()
+            if name not in ["correction_bias"]
+        ]
+
+    def forward_normal(self, hidden_states: torch.Tensor) -> torch.Tensor:
         num_tokens, hidden_dim = hidden_states.shape
         hidden_states = hidden_states.view(-1, hidden_dim)
 
@@ -109,6 +179,165 @@ class Qwen3MoeSparseMoeBlock(nn.Module):
 
         return final_hidden_states.view(num_tokens, hidden_dim)
 
+    def forward_deepep(
+        self, hidden_states: torch.Tensor, forward_batch: ForwardBatch
+    ) -> torch.Tensor:
+        forward_mode = forward_batch.forward_mode
+        if is_non_idle_and_non_empty(forward_mode, hidden_states):
+            # router_logits: (num_tokens, n_experts)
+            router_logits, _ = self.gate(hidden_states)
+
+            topk_weights, topk_idx = select_experts(
+                hidden_states=hidden_states,
+                router_logits=router_logits,
+                top_k=self.top_k,
+                use_grouped_topk=False,
+                renormalize=self.renormalize,
+                num_token_non_padded=forward_batch.num_token_non_padded,
+                expert_location_dispatch_info=ExpertLocationDispatchInfo.init_new(
+                    layer_id=self.layer_id,
+                ),
+            )
+        else:
+            topk_idx = torch.full(
+                (0, self.top_k), -1, dtype=torch.int, device=hidden_states.device
+            )
+            topk_weights = torch.empty(
+                (0, self.top_k), dtype=torch.float32, device=hidden_states.device
+            )
+        if self.ep_size > 1:
+            # TODO(ch-wan): allow users to set num_max_dispatch_tokens_per_rank value
+            (
+                hidden_states,
+                topk_idx,
+                topk_weights,
+                reorder_topk_ids,
+                num_recv_tokens_per_expert,
+                seg_indptr,
+                masked_m,
+                expected_m,
+            ) = self.deepep_dispatcher.dispatch(
+                hidden_states=hidden_states,
+                topk_idx=topk_idx,
+                topk_weights=topk_weights,
+                forward_mode=forward_mode,
+            )
+        final_hidden_states = self.experts(
+            hidden_states=hidden_states,
+            topk_idx=topk_idx,
+            topk_weights=topk_weights,
+            reorder_topk_ids=reorder_topk_ids,
+            seg_indptr=seg_indptr,
+            masked_m=masked_m,
+            expected_m=expected_m,
+            num_recv_tokens_per_expert=num_recv_tokens_per_expert,
+            forward_mode=forward_mode,
+        )
+        if self.ep_size > 1:
+            final_hidden_states = self.deepep_dispatcher.combine(
+                hidden_states=final_hidden_states,
+                topk_idx=topk_idx,
+                topk_weights=topk_weights,
+                forward_mode=forward_mode,
+            )
+        return final_hidden_states
+
+    def op_gate(self, state):
+        if is_non_idle_and_non_empty(
+            state.forward_batch.forward_mode, state.hidden_states_mlp_input
+        ):
+            # router_logits: (num_tokens, n_experts)
+            state.router_logits, _ = self.gate(state.hidden_states_mlp_input)
+        else:
+            state.router_logits = None
+
+    def op_select_experts(self, state):
+        router_logits = state.pop("router_logits")
+        hidden_states = state.hidden_states_mlp_input
+        if router_logits is not None:
+            with get_global_expert_distribution_recorder().with_current_layer(
+                self.layer_id
+            ):
+                state.topk_weights_local, state.topk_idx_local = select_experts(
+                    hidden_states=hidden_states,
+                    router_logits=router_logits,
+                    top_k=self.top_k,
+                    use_grouped_topk=False,
+                    renormalize=self.renormalize,
+                    num_token_non_padded=state.forward_batch.num_token_non_padded,
+                    expert_location_dispatch_info=ExpertLocationDispatchInfo.init_new(
+                        layer_id=self.layer_id,
+                    ),
+                )
+        else:
+            state.topk_idx_local = torch.full(
+                (0, self.top_k), -1, dtype=torch.int, device=hidden_states.device
+            )
+            state.topk_weights_local = torch.empty(
+                (0, self.top_k), dtype=torch.float32, device=hidden_states.device
+            )
+
+    def op_dispatch_a(self, state):
+        if self.ep_size > 1:
+            # TODO(ch-wan): allow users to set num_max_dispatch_tokens_per_rank value
+            self.deepep_dispatcher.dispatch_a(
+                hidden_states=state.pop("hidden_states_mlp_input"),
+                topk_idx=state.pop("topk_idx_local"),
+                topk_weights=state.pop("topk_weights_local"),
+                forward_mode=state.forward_batch.forward_mode,
+                tbo_subbatch_index=state.get("tbo_subbatch_index"),
+            )
+
+    def op_dispatch_b(self, state):
+        if self.ep_size > 1:
+            with get_global_expert_distribution_recorder().with_current_layer(
+                self.layer_id
+            ):
+                (
+                    state.hidden_states_experts_input,
+                    state.topk_idx_dispatched,
+                    state.topk_weights_dispatched,
+                    state.reorder_topk_ids,
+                    state.num_recv_tokens_per_expert,
+                    state.seg_indptr,
+                    state.masked_m,
+                    state.expected_m,
+                ) = self.deepep_dispatcher.dispatch_b(
+                    tbo_subbatch_index=state.get("tbo_subbatch_index"),
+                )
+
+    def op_experts(self, state):
+        state.hidden_states_experts_output = self.experts(
+            hidden_states=state.pop("hidden_states_experts_input"),
+            topk_idx=state.topk_idx_dispatched,
+            topk_weights=state.topk_weights_dispatched,
+            reorder_topk_ids=state.pop("reorder_topk_ids"),
+            seg_indptr=state.pop("seg_indptr"),
+            masked_m=state.pop("masked_m"),
+            expected_m=state.pop("expected_m"),
+            num_recv_tokens_per_expert=state.pop("num_recv_tokens_per_expert"),
+            forward_mode=state.forward_batch.forward_mode,
+        )
+
+    def op_combine_a(self, state):
+        if self.ep_size > 1:
+            self.deepep_dispatcher.combine_a(
+                hidden_states=state.pop("hidden_states_experts_output"),
+                topk_idx=state.pop("topk_idx_dispatched"),
+                topk_weights=state.pop("topk_weights_dispatched"),
+                forward_mode=state.forward_batch.forward_mode,
+                tbo_subbatch_index=state.get("tbo_subbatch_index"),
+            )
+
+    def op_combine_b(self, state):
+        if self.ep_size > 1:
+            state.hidden_states_after_combine = self.deepep_dispatcher.combine_b(
+                tbo_subbatch_index=state.get("tbo_subbatch_index"),
+            )
+
+    def op_output(self, state):
+        state.hidden_states_mlp_output = state.pop("hidden_states_after_combine")
+
 
 class Qwen3MoeAttention(nn.Module):
     def __init__(
@@ -128,20 +357,23 @@ class Qwen3MoeAttention(nn.Module):
     ) -> None:
         super().__init__()
         self.hidden_size = hidden_size
-        self.tp_size = get_tensor_model_parallel_world_size()
+
+        attn_tp_rank = get_attention_tp_rank()
+        attn_tp_size = get_attention_tp_size()
+
         self.total_num_heads = num_heads
-        assert self.total_num_heads % self.tp_size == 0
-        self.num_heads = self.total_num_heads // self.tp_size
+        assert self.total_num_heads % attn_tp_size == 0
+        self.num_heads = self.total_num_heads // attn_tp_size
         self.total_num_kv_heads = num_kv_heads
-        if self.total_num_kv_heads >= self.tp_size:
+        if self.total_num_kv_heads >= attn_tp_size:
             # Number of KV heads is greater than TP size, so we partition
             # the KV heads across multiple tensor parallel GPUs.
-            assert self.total_num_kv_heads % self.tp_size == 0
+            assert self.total_num_kv_heads % attn_tp_size == 0
         else:
             # Number of KV heads is less than TP size, so we replicate
             # the KV heads across multiple tensor parallel GPUs.
-            assert self.tp_size % self.total_num_kv_heads == 0
-        self.num_kv_heads = max(1, self.total_num_kv_heads // self.tp_size)
+            assert attn_tp_size % self.total_num_kv_heads == 0
+        self.num_kv_heads = max(1, self.total_num_kv_heads // attn_tp_size)
         self.head_dim = head_dim or hidden_size // self.total_num_heads
         self.q_size = self.num_heads * self.head_dim
         self.kv_size = self.num_kv_heads * self.head_dim
@@ -157,6 +389,8 @@ class Qwen3MoeAttention(nn.Module):
             self.total_num_kv_heads,
             bias=attention_bias,
             quant_config=quant_config,
+            tp_rank=attn_tp_rank,
+            tp_size=attn_tp_size,
             prefix=add_prefix("qkv_proj", prefix),
         )
 
@@ -165,6 +399,9 @@ class Qwen3MoeAttention(nn.Module):
             hidden_size,
             bias=attention_bias,
             quant_config=quant_config,
+            tp_rank=attn_tp_rank,
+            tp_size=attn_tp_size,
+            reduce_results=False,
             prefix=add_prefix("o_proj", prefix),
         )
 
@@ -198,20 +435,54 @@ class Qwen3MoeAttention(nn.Module):
         k = k_by_head.view(k.shape)
         return q, k
 
-    def forward(
+    def op_prepare(self, state):
+        state.attn_intermediate_state = self.forward_prepare(
+            positions=state.positions,
+            hidden_states=state.pop("hidden_states_after_comm_pre_attn"),
+            forward_batch=state.forward_batch,
+        )
+
+    def op_core(self, state):
+        state.hidden_states_after_attn = self.forward_core(
+            state.pop("attn_intermediate_state")
+        )
+
+    def forward_prepare(
         self,
         positions: torch.Tensor,
         hidden_states: torch.Tensor,
         forward_batch: ForwardBatch,
-    ) -> torch.Tensor:
+    ):
+        if hidden_states.shape[0] == 0:
+            return hidden_states, forward_batch, None
         qkv, _ = self.qkv_proj(hidden_states)
         q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
         q, k = self._apply_qk_norm(q, k)
         q, k = self.rotary_emb(positions, q, k)
-        attn_output = self.attn(q, k, v, forward_batch)
+        inner_state = q, k, v, forward_batch
+        return None, forward_batch, inner_state
+
+    def forward_core(self, intermediate_state):
+        hidden_states, forward_batch, inner_state = intermediate_state
+        if inner_state is None:
+            return hidden_states
+        attn_output = self.attn(*inner_state)
         output, _ = self.o_proj(attn_output)
         return output
 
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        forward_batch: ForwardBatch,
+    ) -> torch.Tensor:
+        s = self.forward_prepare(
+            positions=positions,
+            hidden_states=hidden_states,
+            forward_batch=forward_batch,
+        )
+        return self.forward_core(s)
+
 
 class Qwen3MoeDecoderLayer(nn.Module):
     def __init__(
@@ -222,6 +493,7 @@ class Qwen3MoeDecoderLayer(nn.Module):
         prefix: str = "",
     ) -> None:
         super().__init__()
+        self.config = config
         self.hidden_size = config.hidden_size
         rope_theta = getattr(config, "rope_theta", 10000)
         rope_scaling = getattr(config, "rope_scaling", None)
@@ -246,15 +518,26 @@ class Qwen3MoeDecoderLayer(nn.Module):
             prefix=add_prefix("self_attn", prefix),
         )
 
-        # Note: Qwen/Qwen2-57B-A14B-Instruct does not have
-        # `mlp_only_layers` in the config.
-        mlp_only_layers = (
-            [] if not hasattr(config, "mlp_only_layers") else config.mlp_only_layers
+        self.layer_id = layer_id
+
+        self.attn_tp_size = get_attention_tp_size()
+        self.attn_tp_rank = get_attention_tp_rank()
+        self.local_dp_size = get_local_attention_dp_size()
+
+        # Qwen3MoE all layers are sparse and have no nextn now
+        self.is_layer_sparse = True
+        is_previous_layer_sparse = True
+
+        self.layer_scatter_modes = LayerScatterModes.init_new(
+            layer_id=layer_id,
+            num_layers=config.num_hidden_layers,
+            is_layer_sparse=self.is_layer_sparse,
+            is_previous_layer_sparse=is_previous_layer_sparse,
         )
-        if (layer_id not in mlp_only_layers) and (
-            config.num_experts > 0 and (layer_id + 1) % config.decoder_sparse_step == 0
-        ):
+
+        if self.is_layer_sparse:
             self.mlp = Qwen3MoeSparseMoeBlock(
+                layer_id=self.layer_id,
                 config=config,
                 quant_config=quant_config,
                 prefix=add_prefix("mlp", prefix),
@@ -272,30 +555,102 @@ class Qwen3MoeDecoderLayer(nn.Module):
             config.hidden_size, eps=config.rms_norm_eps
         )
 
+        self.layer_communicator = LayerCommunicator(
+            layer_scatter_modes=self.layer_scatter_modes,
+            input_layernorm=self.input_layernorm,
+            post_attention_layernorm=self.post_attention_layernorm,
+        )
+
     def forward(
         self,
         positions: torch.Tensor,
         hidden_states: torch.Tensor,
         forward_batch: ForwardBatch,
         residual: Optional[torch.Tensor],
-    ) -> torch.Tensor:
-        # Self Attention
-        if residual is None:
-            residual = hidden_states
-            hidden_states = self.input_layernorm(hidden_states)
-        else:
-            hidden_states, residual = self.input_layernorm(hidden_states, residual)
-        hidden_states = self.self_attn(
-            positions=positions,
-            hidden_states=hidden_states,
-            forward_batch=forward_batch,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+
+        hidden_states, residual = self.layer_communicator.prepare_attn(
+            hidden_states, residual, forward_batch
+        )
+
+        if hidden_states.shape[0] != 0:
+            hidden_states = self.self_attn(
+                positions=positions,
+                hidden_states=hidden_states,
+                forward_batch=forward_batch,
+            )
+
+        hidden_states, residual = self.layer_communicator.prepare_mlp(
+            hidden_states, residual, forward_batch
+        )
+
+        hidden_states = self.mlp(hidden_states, forward_batch)
+
+        hidden_states, residual = self.layer_communicator.postprocess_layer(
+            hidden_states, residual, forward_batch
         )
 
-        # Fully Connected
-        hidden_states, residual = self.post_attention_layernorm(hidden_states, residual)
-        hidden_states = self.mlp(hidden_states)
         return hidden_states, residual
 
+    def op_comm_prepare_attn(
+        self,
+        state,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        forward_batch: ForwardBatch,
+        residual: Optional[torch.Tensor],
+        tbo_subbatch_index: Optional[int] = None,
+    ):
+        state.hidden_states_after_comm_pre_attn, state.residual_after_input_ln = (
+            self.layer_communicator.prepare_attn(hidden_states, residual, forward_batch)
+        )
+        state.update(
+            dict(
+                forward_batch=forward_batch,
+                positions=positions,
+                tbo_subbatch_index=tbo_subbatch_index,
+            )
+        )
+
+    def op_comm_prepare_mlp(self, state):
+        state.hidden_states_mlp_input, state.residual_after_comm_pre_mlp = (
+            self.layer_communicator.prepare_mlp(
+                state.pop("hidden_states_after_attn"),
+                state.pop("residual_after_input_ln"),
+                state.forward_batch,
+            )
+        )
+
+    def op_mlp(self, state):
+        hidden_states = state.pop("hidden_states_mlp_input")
+        state.hidden_states_mlp_output = self.mlp(
+            hidden_states, state.forward_batch.forward_mode
+        )
+
+    def op_comm_postprocess_layer(self, state):
+        hidden_states, residual = self.layer_communicator.postprocess_layer(
+            state.pop("hidden_states_mlp_output"),
+            state.pop("residual_after_comm_pre_mlp"),
+            state.forward_batch,
+        )
+
+        output = dict(
+            positions=state.positions,
+            hidden_states=hidden_states,
+            residual=residual,
+            forward_batch=state.forward_batch,
+            tbo_subbatch_index=state.tbo_subbatch_index,
+        )
+
+        state.clear(
+            expect_keys={
+                "positions",
+                "forward_batch",
+                "tbo_subbatch_index",
+            }
+        )
+        return output
+
 
 class Qwen3MoeModel(Qwen2MoeModel):
     def __init__(
@@ -313,7 +668,6 @@ class Qwen3MoeModel(Qwen2MoeModel):
 
 
 class Qwen3MoeForCausalLM(nn.Module):
-
     fall_back_to_pt_during_load = False
 
     def __init__(
@@ -323,6 +677,7 @@ class Qwen3MoeForCausalLM(nn.Module):
         prefix: str = "",
     ) -> None:
         super().__init__()
+        self.pp_group = get_pp_group()
         self.config = config
         self.quant_config = quant_config
         self.model = Qwen3MoeModel(
@@ -333,6 +688,7 @@ class Qwen3MoeForCausalLM(nn.Module):
             config.hidden_size,
             quant_config=quant_config,
             prefix=add_prefix("lm_head", prefix),
+            use_attn_tp_group=global_server_args_dict["enable_dp_lm_head"],
         )
         self.logits_processor = LogitsProcessor(config)
 
@@ -343,12 +699,31 @@ class Qwen3MoeForCausalLM(nn.Module):
         positions: torch.Tensor,
         forward_batch: ForwardBatch,
         input_embeds: torch.Tensor = None,
+        pp_proxy_tensors: Optional[PPProxyTensors] = None,
     ) -> torch.Tensor:
-        hidden_states = self.model(input_ids, positions, forward_batch, input_embeds)
-        return self.logits_processor(
-            input_ids, hidden_states, self.lm_head, forward_batch
+        hidden_states = self.model(
+            input_ids,
+            positions,
+            forward_batch,
+            input_embeds,
+            pp_proxy_tensors=pp_proxy_tensors,
         )
 
+        if self.pp_group.is_last_rank:
+            return self.logits_processor(
+                input_ids, hidden_states, self.lm_head, forward_batch
+            )
+        else:
+            return hidden_states
+
+    @property
+    def start_layer(self):
+        return self.model.start_layer
+
+    @property
+    def end_layer(self):
+        return self.model.end_layer
+
     def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
         stacked_params_mapping = [
             # (param_name, shard_name, shard_id)
@@ -359,9 +734,7 @@ class Qwen3MoeForCausalLM(nn.Module):
             ("gate_up_proj", "up_proj", 1),
         ]
 
-        MoEImpl = EPMoE if global_server_args_dict["enable_ep_moe"] else FusedMoE
-
-        expert_params_mapping = MoEImpl.make_expert_params_mapping(
+        expert_params_mapping = get_moe_impl_class().make_expert_params_mapping(
             ckpt_gate_proj_name="gate_proj",
             ckpt_down_proj_name="down_proj",
             ckpt_up_proj_name="up_proj",
@@ -370,6 +743,17 @@ class Qwen3MoeForCausalLM(nn.Module):
 
         params_dict = dict(self.named_parameters())
         for name, loaded_weight in weights:
+            layer_id = get_layer_id(name)
+            if (
+                layer_id is not None
+                and hasattr(self.model, "start_layer")
+                and (
+                    layer_id < self.model.start_layer
+                    or layer_id >= self.model.end_layer
+                )
+            ):
+                continue
+
             if "rotary_emb.inv_freq" in name:
                 continue
             for param_name, weight_name, shard_id in stacked_params_mapping:
@@ -418,11 +802,29 @@ class Qwen3MoeForCausalLM(nn.Module):
                     if name not in params_dict:
                         continue
 
-                    param = params_dict[name]
-                    weight_loader = getattr(
-                        param, "weight_loader", default_weight_loader
-                    )
-                    weight_loader(param, loaded_weight)
+                    if name in params_dict.keys():
+                        param = params_dict[name]
+                        weight_loader = getattr(
+                            param, "weight_loader", default_weight_loader
+                        )
+                        weight_loader(param, loaded_weight)
+                    else:
+                        logger.warning(f"Parameter {name} not found in params_dict")
+
+        # TODO mimic deepseek
+        self.routed_experts_weights_of_layer = {
+            layer_id: self.model.layers[layer_id].mlp.get_moe_weights()
+            for layer_id in range(self.start_layer, self.end_layer)
+            if isinstance(self.model.layers[layer_id].mlp, Qwen3MoeSparseMoeBlock)
+        }
+
+    @classmethod
+    def get_model_config_for_expert_location(cls, config):
+        return ModelConfigForExpertLocation(
+            num_layers=config.num_hidden_layers,
+            num_logical_experts=config.num_experts,
+            num_groups=None,
+        )
 
 
 EntryClass = Qwen3MoeForCausalLM