初始化项目，由ModelHub XC社区提供模型

Model: ayh015/myLightningOPD
Source: Original Platform

slime/backends/fsdp_utils/models/__init__.py

@@ -0,0 +1,3 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+

slime/backends/fsdp_utils/models/qwen3_moe.py

@@ -0,0 +1,128 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from transformers.models.qwen3_moe.modeling_qwen3_moe import Qwen3MoeMLP
+
+from slime.backends.fsdp_utils.kernels.fused_experts import (
+    DownProjFunction,
+    GateUpProjFunction,
+    MoeSumReduceFunction,
+    SiluAndMulFunction,
+)
+
+
+def fused_experts_impl(
+    hidden_states: torch.Tensor,
+    w1: torch.Tensor,
+    w2: torch.Tensor,
+    topk_weights: torch.Tensor,
+    topk_ids: torch.Tensor,
+):
+    assert hidden_states.shape[1] == w1.shape[2], "Hidden size mismatch"
+    assert topk_weights.shape == topk_ids.shape, "topk shape mismatch"
+    assert hidden_states.is_contiguous(), "Hidden_states must be contiguous"
+    assert w1.is_contiguous(), "Expert weights1 must be contiguous"
+    assert w2.is_contiguous(), "Expert weights2 must be contiguous"
+    assert hidden_states.dtype in [torch.bfloat16]
+
+    intermediate_cache1 = GateUpProjFunction.apply(
+        hidden_states,
+        w1,
+        topk_weights,
+        topk_ids,
+    )
+    intermediate_cache2 = SiluAndMulFunction.apply(intermediate_cache1)
+    intermediate_cache3 = DownProjFunction.apply(
+        intermediate_cache2,
+        w2,
+        topk_weights,
+        topk_ids,
+    )
+    output_hidden_states = MoeSumReduceFunction.apply(
+        intermediate_cache3,
+        hidden_states.shape,
+    )
+    return output_hidden_states
+
+
+class StandardDispatcher:
+    def __init__(self, num_experts: int, num_local_experts: int):
+        self.moe_ep_size = 1
+        self.num_experts = num_experts
+        self.num_local_experts = num_local_experts
+        self.moe_ep_rank = 0
+        self.local_expert_mapping = None
+
+        if self.moe_ep_size > 1:
+            self.local_expert_mapping = torch.full((self.num_experts,), -1, dtype=torch.int32, device="cuda")
+            self.local_expert_mapping[
+                self.moe_ep_rank * self.num_local_experts : (self.moe_ep_rank + 1) * self.num_local_experts
+            ] = torch.arange(0, self.num_local_experts, dtype=torch.int32, device="cuda")
+
+    def dispatch(self, topk_ids) -> torch.Tensor:
+        if self.local_expert_mapping is not None:
+            return self.local_expert_mapping[topk_ids]
+        return topk_ids
+
+
+class Qwen3MoeSparseMoeBlock(nn.Module):
+    dispatcher = None
+    runner = None
+
+    def __init__(self, config):
+        super().__init__()
+        self.num_experts = config.num_experts
+        self.top_k = config.num_experts_per_tok
+        self.norm_topk_prob = config.norm_topk_prob
+
+        # gating
+        self.gate = nn.Linear(config.hidden_size, config.num_experts, bias=False)
+
+        self.experts = nn.ModuleList(
+            [Qwen3MoeMLP(config, intermediate_size=config.moe_intermediate_size) for _ in range(self.num_experts)]
+        )
+
+        if Qwen3MoeSparseMoeBlock.dispatcher is None:
+            Qwen3MoeSparseMoeBlock.dispatcher = StandardDispatcher(
+                num_experts=config.num_experts, num_local_experts=config.num_experts
+            )
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        batch_size, sequence_length, hidden_dim = hidden_states.shape
+        hidden_states = hidden_states.view(-1, hidden_dim)
+        # router_logits: (batch * sequence_length, n_experts)
+        router_logits = self.gate(hidden_states)
+
+        routing_weights = F.softmax(router_logits, dim=1, dtype=torch.float)
+        routing_weights, selected_experts = torch.topk(routing_weights, self.top_k, dim=-1)
+
+        if self.norm_topk_prob:  # only diff with mixtral sparse moe block!
+            routing_weights /= routing_weights.sum(dim=-1, keepdim=True)
+        # we cast back to the input dtype
+        routing_weights = routing_weights.to(hidden_states.dtype)
+
+        selected_experts = Qwen3MoeSparseMoeBlock.dispatcher.dispatch(selected_experts)
+
+        w13_weight = torch.stack(
+            [torch.cat([layer.gate_proj.weight, layer.up_proj.weight], dim=0) for layer in self.experts]
+        )
+        w2_weight = torch.stack([layer.down_proj.weight for layer in self.experts], dim=0)
+
+        final_hidden_states = fused_experts_impl(
+            hidden_states.to(torch.bfloat16),
+            w13_weight,
+            w2_weight,
+            routing_weights,
+            selected_experts,
+        )
+
+        return final_hidden_states, router_logits
+
+
+def apply_true_on_policy_patch_for_qwen3_moe():
+    from transformers.models.qwen3_moe import modeling_qwen3_moe
+
+    modeling_qwen3_moe.Qwen3MoeSparseMoeBlock = Qwen3MoeSparseMoeBlock

slime/backends/fsdp_utils/models/qwen3_moe_hf.py

@@ -0,0 +1,46 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+import torch
+import torch.nn.functional as F
+
+
+def apply_fsdp_moe_patch():
+
+    from transformers.models.qwen3_moe import modeling_qwen3_moe
+
+    def _forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        batch_size, sequence_length, hidden_dim = hidden_states.shape
+        hidden_states = hidden_states.view(-1, hidden_dim)
+        router_logits = self.gate(hidden_states)
+
+        routing_weights = F.softmax(router_logits, dim=1, dtype=torch.float)
+        routing_weights, selected_experts = torch.topk(routing_weights, self.top_k, dim=-1)
+        if self.norm_topk_prob:
+            routing_weights /= routing_weights.sum(dim=-1, keepdim=True)
+        routing_weights = routing_weights.to(hidden_states.dtype)
+
+        final_hidden_states = torch.zeros(
+            (batch_size * sequence_length, hidden_dim), dtype=hidden_states.dtype, device=hidden_states.device
+        )
+
+        expert_mask = torch.nn.functional.one_hot(selected_experts, num_classes=self.num_experts).permute(2, 1, 0)
+
+        # Loop over all experts
+        for expert_idx in range(self.num_experts):
+            expert_layer = self.experts[expert_idx]
+            idx, top_x = torch.where(expert_mask[expert_idx])
+
+            if top_x.numel() > 0:
+                current_state = hidden_states[None, top_x].reshape(-1, hidden_dim)
+                current_hidden_states = expert_layer(current_state) * routing_weights[top_x, idx, None]
+                final_hidden_states.index_add_(0, top_x, current_hidden_states.to(hidden_states.dtype))
+            else:
+                # force experts to participate in computation graph
+                dummy_output = expert_layer(hidden_states[:1]) * 0.0
+                final_hidden_states[:1] = final_hidden_states[:1] + dummy_output
+
+        final_hidden_states = final_hidden_states.reshape(batch_size, sequence_length, hidden_dim)
+        return final_hidden_states, router_logits
+
+    modeling_qwen3_moe.Qwen3MoeSparseMoeBlock.forward = _forward