fix: add @ModuleInfo to BertModel.pooler and NomicBertModel.pooler

These Linear? properties are replaced by QuantizedLinear during
quantization via Module.update(modules:). Without @ModuleInfo,
the non-throwing update(modules:) wrapper hits try! on the thrown
needModuleInfo error, causing SIGABRT when loading quantized
embedding models.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: jowharshamshiri

Libraries/MLXEmbedders/Models/Bert.swift

@@ -259,7 +259,7 @@ public class BertModel: Module, EmbeddingModel {
     @ModuleInfo(key: "embeddings") fileprivate var embedder: BertEmbedding
 
     /// A linear layer used to "pool" the [CLS] token into a single sentence vector.
-    let pooler: Linear?
+    @ModuleInfo var pooler: Linear?
 
     /// The stack of Transformer layers.
     fileprivate let encoder: Encoder
@@ -280,10 +280,10 @@ public class BertModel: Module, EmbeddingModel {
 
         if lmHead {
             _lmHead.wrappedValue = LMHead(config)
-            self.pooler = nil
+            _pooler.wrappedValue = nil
         } else {
             // Pooler projects the [CLS] token to a hidden state of the same size
-            pooler = Linear(config.embedDim, config.embedDim)
+            _pooler.wrappedValue = Linear(config.embedDim, config.embedDim)
             _lmHead.wrappedValue = nil
         }
     }

Libraries/MLXEmbedders/Models/NomicBert.swift

@@ -671,7 +671,7 @@ public class NomicBertModel: Module, EmbeddingModel {
 
     /// The optional pooler layer.
     /// Used to extract a single vector representation for the whole sequence (usually from the [CLS] token).
-    let pooler: Linear?
+    @ModuleInfo var pooler: Linear?
 
     /// The stack of Transformer blocks.
     fileprivate let encoder: Encoder
@@ -696,9 +696,9 @@ public class NomicBertModel: Module, EmbeddingModel {
 
         // Initialize Pooler (for sentence embeddings)
         if pooler {
-            self.pooler = Linear(config.embedDim, config.embedDim, bias: false)
+            _pooler.wrappedValue = Linear(config.embedDim, config.embedDim, bias: false)
         } else {
-            self.pooler = nil
+            _pooler.wrappedValue = nil
         }
 
         // Initialize LM Head (for training/masked prediction)