Fix InternVL2 model sharding (#481)

Author: pufanyi

lmms_eval/models/internvl2.py

@@ -119,12 +119,55 @@ def load_video(video_path, bound=None, input_size=448, max_num=1, num_segments=3
     return pixel_values, num_patches_list
 
 
+import math
 from datetime import timedelta
 
 from accelerate.state import AcceleratorState
 from accelerate.utils import InitProcessGroupKwargs
 
 
+# The reason for writing the code this way is to avoid errors that occur during multi-GPU inference due to tensors not being on the same device. By ensuring that the first and last layers of the large language model (LLM) are on the same device, we prevent such errors.
+def split_model(model_name, num_layers=None):
+    device_map = {}
+    world_size = torch.cuda.device_count()
+    if num_layers is None:
+        num_layers = {
+            "InternVL2_5-1B": 24,
+            "InternVL2_5-2B": 24,
+            "InternVL2_5-4B": 36,
+            "InternVL2_5-8B": 32,
+            "InternVL2_5-26B": 48,
+            "InternVL2_5-38B": 64,
+            "InternVL2_5-78B": 80,
+            "InternVL2-1B": 24,
+            "InternVL2-2B": 24,
+            "InternVL2-4B": 32,
+            "InternVL2-8B": 32,
+            "InternVL2-26B": 48,
+            "InternVL2-40B": 60,
+            "InternVL2-Llama3-76B": 80,
+        }[model_name]
+    # Since the first GPU will be used for ViT, treat it as half a GPU.
+    num_layers_per_gpu = math.ceil(num_layers / (world_size - 0.5))
+    num_layers_per_gpu = [num_layers_per_gpu] * world_size
+    num_layers_per_gpu[0] = math.ceil(num_layers_per_gpu[0] * 0.5)
+    layer_cnt = 0
+    for i, num_layer in enumerate(num_layers_per_gpu):
+        for j in range(num_layer):
+            device_map[f"language_model.model.layers.{layer_cnt}"] = i
+            layer_cnt += 1
+    device_map["vision_model"] = 0
+    device_map["mlp1"] = 0
+    device_map["language_model.model.tok_embeddings"] = 0
+    device_map["language_model.model.embed_tokens"] = 0
+    device_map["language_model.output"] = 0
+    device_map["language_model.model.norm"] = 0
+    device_map["language_model.lm_head"] = 0
+    device_map[f"language_model.model.layers.{num_layers - 1}"] = 0
+
+    return device_map
+
+
 @register_model("internvl2")
 class InternVL2(lmms):
     def __init__(
@@ -134,13 +177,14 @@ def __init__(
         device: str = "cuda:0",
         device_map: str = "cuda:0",
         batch_size: str = "1",
+        num_frame: int = 32,
+        num_layers=None,
         **kwargs,
     ):
         super().__init__()
 
         self.path = pretrained
-        self._model = AutoModel.from_pretrained(self.path, torch_dtype=torch.bfloat16, low_cpu_mem_usage=True, trust_remote_code=True, device_map=device_map).eval()
-        self._tokenizer = AutoTokenizer.from_pretrained(self.path, trust_remote_code=True, device_map=device_map)
+        self.num_frame = num_frame
 
         batch_size = int(batch_size)
         assert batch_size == 1, f"Batch size should be 1 for InternVL2, but got {batch_size}."
@@ -154,11 +198,15 @@ def __init__(
             self.device_map = f"cuda:{accelerator.local_process_index}"
         elif accelerator.num_processes == 1 and device_map == "auto":
             self._device = torch.device(device)
+            device_map = split_model(pretrained.split("/")[-1], num_layers=num_layers)
             self.device_map = device_map
         else:
             self._device = torch.device(f"cuda:{accelerator.local_process_index}")
             self.device_map = f"cuda:{accelerator.local_process_index}"
 
+        self._model = AutoModel.from_pretrained(self.path, torch_dtype=torch.bfloat16, low_cpu_mem_usage=True, trust_remote_code=True, device_map=device_map).eval()
+        self._tokenizer = AutoTokenizer.from_pretrained(self.path, trust_remote_code=True, device_map=device_map)
+
         if accelerator.num_processes > 1:
             assert accelerator.distributed_type in [DistributedType.FSDP, DistributedType.MULTI_GPU, DistributedType.DEEPSPEED], "Unsupported distributed type provided. Only DDP and FSDP are supported."
             # If you want to use DistributedType.DEEPSPEED, you have to run accelerate config before using the model
@@ -269,7 +317,7 @@ def generate_until(self, requests) -> List[str]:
             elif self.modality == "video":
                 assert len(visuals) == 1, f"Only one video is supported, but got {len(visuals)} videos."
                 video_path = visuals[0]
-                pixel_values, num_patches_list = load_video(video_path, num_segments=8, max_num=1)
+                pixel_values, num_patches_list = load_video(video_path, num_segments=self.num_frame)
                 pixel_values = pixel_values.to(torch.bfloat16).cuda()
                 video_prefix = "".join([f"Frame{i+1}: <image>\n" for i in range(len(num_patches_list))])
                 question = video_prefix + contexts