PaliGemma 2 support (#11142)

docs/source/models/supported_models.rst

@@ -664,9 +664,9 @@ Text Generation (``--task generate``)
     - ✅︎
     - ✅︎
   * - :code:`PaliGemmaForConditionalGeneration`
-    - PaliGemma
+    - PaliGemma, PaliGemma 2
     - T + I\ :sup:`E`
-    - :code:`google/paligemma-3b-pt-224`, :code:`google/paligemma-3b-mix-224`, etc.
+    - :code:`google/paligemma-3b-pt-224`, :code:`google/paligemma-3b-mix-224`, :code:`google/paligemma2-3b-ft-docci-448`, etc.
     - 
     - ✅︎
     - 

examples/offline_inference_vision_language.py

@@ -137,6 +137,18 @@ def run_paligemma(question: str, modality: str):
     return llm, prompt, stop_token_ids
 
 
+# PaliGemma 2
+def run_paligemma2(question: str, modality: str):
+    assert modality == "image"
+
+    # PaliGemma 2 has special prompt format for VQA
+    prompt = "caption en"
+    llm = LLM(model="google/paligemma2-3b-ft-docci-448",
+              mm_cache_preprocessor=args.mm_cache_preprocessor)
+    stop_token_ids = None
+    return llm, prompt, stop_token_ids
+
+
 # Chameleon
 def run_chameleon(question: str, modality: str):
     assert modality == "image"
@@ -473,6 +485,7 @@ model_example_map = {
     "fuyu": run_fuyu,
     "phi3_v": run_phi3v,
     "paligemma": run_paligemma,
+    "paligemma2": run_paligemma2,
     "chameleon": run_chameleon,
     "minicpmv": run_minicpmv,
     "blip-2": run_blip2,

vllm/model_executor/models/paligemma.py

@@ -105,6 +105,11 @@ def input_processor_for_paligemma(ctx: InputContext,
         orig_prompt_ids.remove(hf_config.image_token_index)
 
     new_prompt = f"{image_token_str_pad}{bos_token}{orig_prompt}\n"
+
+    # The PaliGemma 2 tokenizer does not include a starting BOS token
+    if orig_prompt_ids[0] != hf_config.bos_token_id:
+        orig_prompt_ids = [hf_config.bos_token_id] + orig_prompt_ids
+
     new_token_ids = image_token_ids_pad + orig_prompt_ids + [108]  #newline
 
     # NOTE: Create a defensive copy of the original inputs
@@ -149,7 +154,11 @@ class PaliGemmaForConditionalGeneration(nn.Module, SupportsMultiModal,
             projection_dim=config.vision_config.projection_dim)
 
         self.quant_config = quant_config
-        config.text_config.architectures = ["GemmaForCausalLM"]
+
+        if config.text_config.model_type == "gemma":
+            config.text_config.architectures = ["GemmaForCausalLM"]
+        else:
+            config.text_config.architectures = ["Gemma2ForCausalLM"]
         self.language_model = init_vllm_registered_model(
             vllm_config=vllm_config,
             hf_config=config.text_config,