[Model] Support NVLM-D and fix QK Norm in InternViT (vllm-project#9045)

Co-authored-by: Roger Wang <ywang@roblox.com>
Co-authored-by: Isotr0py <mozf@mail2.sysu.edu.cn>

docs/source/models/supported_models.rst

@@ -315,6 +315,9 @@ Multimodal Language Models
 
 .. _supported_vlms:
 
+Text Generation
+---------------
+
 .. list-table::
   :widths: 25 25 25 25 5 5
   :header-rows: 1
@@ -384,7 +387,13 @@ Multimodal Language Models
     - Image
     - :code:`meta-llama/Llama-3.2-90B-Vision-Instruct`, :code:`meta-llama/Llama-3.2-11B-Vision`, etc.
     -
+    -
+  * - :code:`NVLM_D_Model`
+    - NVLM-D 1.0
+    - Image\ :sup:`E+`
+    - :code:`nvidia/NVLM-D-72B`, etc.
     - 
+    - ✅︎
   * - :code:`PaliGemmaForConditionalGeneration`
     - PaliGemma
     - Image\ :sup:`E`

examples/offline_inference_vision_language.py

@@ -18,7 +18,7 @@
 
 
 # LLaVA-1.5
-def run_llava(question, modality):
+def run_llava(question: str, modality: str):
     assert modality == "image"
 
     prompt = f"USER: <image>\n{question}\nASSISTANT:"
@@ -29,7 +29,7 @@ def run_llava(question, modality):
 
 
 # LLaVA-1.6/LLaVA-NeXT
-def run_llava_next(question, modality):
+def run_llava_next(question: str, modality: str):
     assert modality == "image"
 
     prompt = f"[INST] <image>\n{question} [/INST]"
@@ -40,7 +40,7 @@ def run_llava_next(question, modality):
 
 # LlaVA-NeXT-Video
 # Currently only support for video input
-def run_llava_next_video(question, modality):
+def run_llava_next_video(question: str, modality: str):
     assert modality == "video"
 
     prompt = f"USER: <video>\n{question} ASSISTANT:"
@@ -50,7 +50,7 @@ def run_llava_next_video(question, modality):
 
 
 # LLaVA-OneVision
-def run_llava_onevision(question, modality):
+def run_llava_onevision(question: str, modality: str):
 
     if modality == "video":
         prompt = f"<|im_start|>user <video>\n{question}<|im_end|> \
@@ -67,7 +67,7 @@ def run_llava_onevision(question, modality):
 
 
 # Fuyu
-def run_fuyu(question, modality):
+def run_fuyu(question: str, modality: str):
     assert modality == "image"
 
     prompt = f"{question}\n"
@@ -77,7 +77,7 @@ def run_fuyu(question, modality):
 
 
 # Phi-3-Vision
-def run_phi3v(question, modality):
+def run_phi3v(question: str, modality: str):
     assert modality == "image"
 
     prompt = f"<|user|>\n<|image_1|>\n{question}<|end|>\n<|assistant|>\n"  # noqa: E501
@@ -112,7 +112,7 @@ def run_phi3v(question, modality):
 
 
 # PaliGemma
-def run_paligemma(question, modality):
+def run_paligemma(question: str, modality: str):
     assert modality == "image"
 
     # PaliGemma has special prompt format for VQA
@@ -123,7 +123,7 @@ def run_paligemma(question, modality):
 
 
 # Chameleon
-def run_chameleon(question, modality):
+def run_chameleon(question: str, modality: str):
     assert modality == "image"
 
     prompt = f"{question}<image>"
@@ -133,7 +133,7 @@ def run_chameleon(question, modality):
 
 
 # MiniCPM-V
-def run_minicpmv(question, modality):
+def run_minicpmv(question: str, modality: str):
     assert modality == "image"
 
     # 2.0
@@ -176,7 +176,7 @@ def run_minicpmv(question, modality):
 
 
 # InternVL
-def run_internvl(question, modality):
+def run_internvl(question: str, modality: str):
     assert modality == "image"
 
     model_name = "OpenGVLab/InternVL2-2B"
@@ -203,8 +203,32 @@ def run_internvl(question, modality):
     return llm, prompt, stop_token_ids
 
 
+# NVLM-D
+def run_nvlm_d(question: str, modality: str):
+    assert modality == "image"
+
+    model_name = "nvidia/NVLM-D-72B"
+
+    # Adjust this as necessary to fit in GPU
+    llm = LLM(
+        model=model_name,
+        trust_remote_code=True,
+        max_model_len=4096,
+        tensor_parallel_size=4,
+    )
+
+    tokenizer = AutoTokenizer.from_pretrained(model_name,
+                                              trust_remote_code=True)
+    messages = [{'role': 'user', 'content': f"<image>\n{question}"}]
+    prompt = tokenizer.apply_chat_template(messages,
+                                           tokenize=False,
+                                           add_generation_prompt=True)
+    stop_token_ids = None
+    return llm, prompt, stop_token_ids
+
+
 # BLIP-2
-def run_blip2(question, modality):
+def run_blip2(question: str, modality: str):
     assert modality == "image"
 
     # BLIP-2 prompt format is inaccurate on HuggingFace model repository.
@@ -216,7 +240,7 @@ def run_blip2(question, modality):
 
 
 # Qwen
-def run_qwen_vl(question, modality):
+def run_qwen_vl(question: str, modality: str):
     assert modality == "image"
 
     llm = LLM(
@@ -232,7 +256,7 @@ def run_qwen_vl(question, modality):
 
 
 # Qwen2-VL
-def run_qwen2_vl(question, modality):
+def run_qwen2_vl(question: str, modality: str):
     assert modality == "image"
 
     model_name = "Qwen/Qwen2-VL-7B-Instruct"
@@ -252,8 +276,8 @@ def run_qwen2_vl(question, modality):
     return llm, prompt, stop_token_ids
 
 
-# LLama
-def run_mllama(question, modality):
+# LLama 3.2
+def run_mllama(question: str, modality: str):
     assert modality == "image"
 
     model_name = "meta-llama/Llama-3.2-11B-Vision-Instruct"
@@ -287,6 +311,7 @@ def run_mllama(question, modality):
     "minicpmv": run_minicpmv,
     "blip-2": run_blip2,
     "internvl_chat": run_internvl,
+    "NVLM_D": run_nvlm_d,
     "qwen_vl": run_qwen_vl,
     "qwen2_vl": run_qwen2_vl,
     "mllama": run_mllama,

examples/offline_inference_vision_language_multi_image.py

@@ -144,6 +144,39 @@ def load_internvl(question: str, image_urls: List[str]) -> ModelRequestData:
     )
 
 
+def load_nvlm_d(question: str, image_urls: List[str]):
+    model_name = "nvidia/NVLM-D-72B"
+
+    # Adjust this as necessary to fit in GPU
+    llm = LLM(
+        model=model_name,
+        trust_remote_code=True,
+        max_model_len=8192,
+        tensor_parallel_size=4,
+        limit_mm_per_prompt={"image": len(image_urls)},
+        mm_processor_kwargs={"max_dynamic_patch": 4},
+    )
+
+    placeholders = "\n".join(f"Image-{i}: <image>\n"
+                             for i, _ in enumerate(image_urls, start=1))
+    messages = [{'role': 'user', 'content': f"{placeholders}\n{question}"}]
+
+    tokenizer = AutoTokenizer.from_pretrained(model_name,
+                                              trust_remote_code=True)
+    prompt = tokenizer.apply_chat_template(messages,
+                                           tokenize=False,
+                                           add_generation_prompt=True)
+    stop_token_ids = None
+
+    return ModelRequestData(
+        llm=llm,
+        prompt=prompt,
+        stop_token_ids=stop_token_ids,
+        image_data=[fetch_image(url) for url in image_urls],
+        chat_template=None,
+    )
+
+
 def load_qwen2_vl(question, image_urls: List[str]) -> ModelRequestData:
     try:
         from qwen_vl_utils import process_vision_info
@@ -204,6 +237,7 @@ def load_qwen2_vl(question, image_urls: List[str]) -> ModelRequestData:
 model_example_map = {
     "phi3_v": load_phi3v,
     "internvl_chat": load_internvl,
+    "NVLM_D": load_nvlm_d,
     "qwen2_vl": load_qwen2_vl,
     "qwen_vl_chat": load_qwenvl_chat,
 }

vllm/entrypoints/chat_utils.py

@@ -157,7 +157,7 @@ def _placeholder_str(self, modality: ModalityStr,
             if model_type.startswith("llava"):
                 return self._cached_token_str(self._tokenizer,
                                               hf_config.image_token_index)
-            if model_type in ("chameleon", "internvl_chat"):
+            if model_type in ("chameleon", "internvl_chat", "NVLM_D"):
                 return "<image>"
             if model_type == "mllama":
                 return "<|image|>"

vllm/model_executor/layers/layernorm.py

@@ -18,10 +18,16 @@ def __init__(
         self,
         hidden_size: int,
         eps: float = 1e-6,
+        var_hidden_size: Optional[int] = None,
     ) -> None:
         super().__init__()
-        self.weight = nn.Parameter(torch.ones(hidden_size))
+
+        self.hidden_size = hidden_size
         self.variance_epsilon = eps
+        self.variance_size_override = (None if var_hidden_size == hidden_size
+                                       else var_hidden_size)
+
+        self.weight = nn.Parameter(torch.ones(hidden_size))
 
     def forward_native(
         self,
@@ -35,7 +41,23 @@ def forward_native(
             x = x + residual.to(torch.float32)
             residual = x.to(orig_dtype)
 
-        variance = x.pow(2).mean(dim=-1, keepdim=True)
+        hidden_size = x.shape[-1]
+        if hidden_size != self.hidden_size:
+            raise ValueError("Expected hidden_size to be "
+                             f"{self.hidden_size}, but found: {hidden_size}")
+
+        if self.variance_size_override is None:
+            x_var = x
+        else:
+            if hidden_size < self.variance_size_override:
+                raise ValueError(
+                    "Expected hidden_size to be at least "
+                    f"{self.variance_size_override}, but found: {hidden_size}")
+
+            x_var = x[:, :, :self.variance_size_override]
+
+        variance = x_var.pow(2).mean(dim=-1, keepdim=True)
+
         x = x * torch.rsqrt(variance + self.variance_epsilon)
         x = x.to(orig_dtype) * self.weight
         if residual is None:
@@ -48,6 +70,9 @@ def forward_cuda(
         x: torch.Tensor,
         residual: Optional[torch.Tensor] = None,
     ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
+        if self.variance_size_override is not None:
+            return self.forward_native(x, residual)
+
         from vllm import _custom_ops as ops
 
         if residual is not None:
@@ -72,6 +97,9 @@ def forward_xpu(
         x: torch.Tensor,
         residual: Optional[torch.Tensor] = None,
     ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
+        if self.variance_size_override is not None:
+            return self.forward_native(x, residual)
+
         from vllm._ipex_ops import ipex_ops as ops
 
         if residual is not None: