20260801-moe model impl

need to add aux loss for load balancing

astrai/model/components/__init__.py

@@ -3,7 +3,7 @@ from astrai.model.components.attention import GQA, MLA
 from astrai.model.components.decoder_block import DecoderBlock
 from astrai.model.components.embedding import Embedding
 from astrai.model.components.linear import Linear
-from astrai.model.components.mlp import MLP
+from astrai.model.components.mlp import MLP, DeepSeekMoE
 from astrai.model.components.norm import RMSNorm
 from astrai.model.components.rope import (
     RotaryEmbedding,
@@ -14,6 +14,7 @@ __all__ = [
     "Linear",
     "RMSNorm",
     "MLP",
+    "DeepSeekMoE",
     "Embedding",
     "GQA",
     "MLA",

astrai/model/components/decoder_block.py

@@ -26,7 +26,20 @@ class DecoderBlock(nn.Module):
         self.attention = AttnFactory.create(config.attn_type, **cfg, layer_id=layer_id)
         self.input_norm = RMSNorm(config.hidden_size, config.rms_norm_eps)
         self.post_attention_norm = RMSNorm(config.hidden_size, config.rms_norm_eps)
-        self.mlp = FFNFactory.create(config.ffn_type, **cfg)
+        ffn_type = self._resolve_ffn_type(config, layer_id)
+        self.mlp = FFNFactory.create(ffn_type, **cfg)
+
+    @staticmethod
+    def _resolve_ffn_type(config, layer_id: int) -> str:
+        if config.ffn_type != "moe":
+            return config.ffn_type
+        mlp_only = config.mlp_only_layers or []
+        if layer_id in mlp_only:
+            return "mlp"
+        if config.decoder_sparse_step > 1:
+            if (layer_id + 1) % config.decoder_sparse_step != 0:
+                return "mlp"
+        return "moe"
 
     def forward(
         self,

astrai/model/components/mlp.py

@@ -1,3 +1,5 @@
+from typing import Optional
+
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
@@ -36,6 +38,9 @@ class DeepSeekMoE(nn.Module):
         n_activated_experts: int = 2,
         topk_method: str = "greedy",
         n_layers: int = 1,
+        moe_intermediate_size: Optional[int] = None,
+        shared_expert_intermediate_size: Optional[int] = None,
+        norm_topk_prob: bool = False,
     ):
         super().__init__()
         self.dim = dim
@@ -43,6 +48,10 @@ class DeepSeekMoE(nn.Module):
         self.n_shared_experts = n_shared_experts
         self.n_activated_experts = n_activated_experts
         self.topk_method = topk_method
+        self.norm_topk_prob = norm_topk_prob
+
+        expert_dim_ffn = moe_intermediate_size if moe_intermediate_size is not None else dim_ffn
+        shared_dim_ffn = shared_expert_intermediate_size if shared_expert_intermediate_size is not None else dim_ffn
 
         self.router = Linear(dim, n_routed_experts, bias=False)
         moe_scale = 1 / max(n_shared_experts, 1) + 1 / n_activated_experts
@@ -50,13 +59,13 @@ class DeepSeekMoE(nn.Module):
 
         self.shared_experts = nn.ModuleList(
             [
-                MLP(dim, dim_ffn, down_init_std=down_init_std)
+                MLP(dim, shared_dim_ffn, down_init_std=down_init_std)
                 for _ in range(n_shared_experts)
             ]
         )
         self.routed_experts = nn.ModuleList(
             [
-                MLP(dim, dim_ffn, down_init_std=down_init_std)
+                MLP(dim, expert_dim_ffn, down_init_std=down_init_std)
                 for _ in range(n_routed_experts)
             ]
         )
@@ -84,7 +93,8 @@ class DeepSeekMoE(nn.Module):
         router_probs = torch.softmax(router_logits.float(), dim=-1).to(x.dtype)
 
         topk_weights, topk_indices = torch.topk(router_probs, K, dim=-1)
-        topk_weights = topk_weights / topk_weights.sum(dim=-1, keepdim=True)
+        if self.norm_topk_prob:
+            topk_weights = topk_weights / topk_weights.sum(dim=-1, keepdim=True)
 
         output = torch.zeros(N, D, device=x.device, dtype=x.dtype)
         for expert_idx in range(self.n_routed_experts):