Add Qwen3-VL vision-language model support

lib/bumblebee.ex

@@ -192,6 +192,8 @@ defmodule Bumblebee do
     "Qwen3Model" => {Bumblebee.Text.Qwen3, :base},
     "Qwen3ForCausalLM" => {Bumblebee.Text.Qwen3, :for_causal_language_modeling},
     "Qwen3ForSequenceClassification" => {Bumblebee.Text.Qwen3, :for_sequence_classification},
+    "Qwen3VLForConditionalGeneration" =>
+      {Bumblebee.Multimodal.Qwen3VL, :for_conditional_generation},
     "ResNetForImageClassification" => {Bumblebee.Vision.ResNet, :for_image_classification},
     "ResNetModel" => {Bumblebee.Vision.ResNet, :base},
     "RobertaForMaskedLM" => {Bumblebee.Text.Roberta, :for_masked_language_modeling},
@@ -242,12 +244,14 @@ defmodule Bumblebee do
 
   @transformers_image_processor_type_to_featurizer %{
     "BlipImageProcessor" => Bumblebee.Vision.BlipFeaturizer,
-    "BitImageProcessor" => Bumblebee.Vision.BitFeaturizer
+    "BitImageProcessor" => Bumblebee.Vision.BitFeaturizer,
+    "Qwen3VLImageProcessor" => Bumblebee.Vision.Qwen3VLFeaturizer
   }
 
   @model_type_to_featurizer %{
     "convnext" => Bumblebee.Vision.ConvNextFeaturizer,
     "deit" => Bumblebee.Vision.DeitFeaturizer,
+    "qwen3_vl" => Bumblebee.Vision.Qwen3VLFeaturizer,
     "resnet" => Bumblebee.Vision.ConvNextFeaturizer,
     "vit" => Bumblebee.Vision.VitFeaturizer,
     "whisper" => Bumblebee.Audio.WhisperFeaturizer
@@ -274,7 +278,9 @@ defmodule Bumblebee do
     "mpnet" => :mpnet,
     "phi" => :code_gen,
     "phi3" => :llama,
+    "qwen2_vl" => :qwen2,
     "qwen3" => :qwen2,
+    "qwen3_vl" => :qwen2,
     "roberta" => :roberta,
     "smollm3" => :smollm3,
     "t5" => :t5,

lib/bumblebee/layers/transformer.ex

@@ -75,6 +75,7 @@ defmodule Bumblebee.Layers.Transformer do
             :num_blocks,
             :rotary_embedding,
             :attention_window_size,
+            :post_block_hook,
             attention_mask: Layers.none(),
             attention_head_mask: Layers.none(),
             attention_relative_bias: nil,
@@ -97,6 +98,7 @@ defmodule Bumblebee.Layers.Transformer do
     cache = opts[:cache]
     rotary_embedding = opts[:rotary_embedding]
     attention_window_size = opts[:attention_window_size]
+    post_block_hook = opts[:post_block_hook]
 
     block_opts = Keyword.take(opts, block_opts_keys)
 
@@ -160,6 +162,14 @@ defmodule Bumblebee.Layers.Transformer do
               ] ++ block_opts
             )
 
+          # Apply post-block hook if provided (e.g., for DeepStack feature injection)
+          hidden_state =
+            if post_block_hook do
+              post_block_hook.(idx, hidden_state)
+            else
+              hidden_state
+            end
+
           cache = Layers.Decoder.put_block_cache(state.cache, idx, block_cache)
 
           %{