Fix cuda memory error for Qwen3 non-quantized (#2987)

* Update KvCache initialization in Qwen3 model to use a fixed max position embedding value of 512

* add doc

candle-transformers/src/models/qwen3.rs

@@ -157,7 +157,9 @@ impl Qwen3Attention {
         // Necessary because the hidden_size in the config isn't always accurate
         let hidden_size = head_dim * cfg.num_attention_heads;
 
-        let kv_cache = KvCache::new(2, cfg.max_position_embeddings);
+        // Initialize KV cache with 512 tokens capacity to reduce initial memory allocation.
+        // The cache will grow in chunks of 512 tokens when needed.
+        let kv_cache = KvCache::new(2, 512);
 
         Ok(Self {
             q_proj,