sglang 0.1.14__py3-none-any.whl → 0.1.16__py3-none-any.whl

sglang/srt/models/gemma.py

@@ -4,29 +4,25 @@
 from typing import Optional, Tuple
 
 import torch
-from sglang.srt.layers.logits_processor import LogitsProcessor
-from sglang.srt.layers.radix_attention import RadixAttention
 from torch import nn
 from transformers import PretrainedConfig
 from vllm.config import LoRAConfig
-from vllm.model_executor.input_metadata import InputMetadata
+from vllm.distributed import get_tensor_model_parallel_world_size
 from vllm.model_executor.layers.activation import GeluAndMul
 from vllm.model_executor.layers.layernorm import RMSNorm
 from vllm.model_executor.layers.linear import (
-    LinearMethodBase,
     MergedColumnParallelLinear,
     QKVParallelLinear,
     RowParallelLinear,
 )
+from vllm.model_executor.layers.quantization.base_config import QuantizationConfig
 from vllm.model_executor.layers.rotary_embedding import get_rope
 from vllm.model_executor.layers.vocab_parallel_embedding import VocabParallelEmbedding
-from vllm.model_executor.parallel_utils.parallel_state import (
-    get_tensor_model_parallel_world_size,
-)
-from vllm.model_executor.weight_utils import (
-    default_weight_loader,
-    hf_model_weights_iterator,
-)
+
+from sglang.srt.layers.logits_processor import LogitsProcessor
+from sglang.srt.layers.radix_attention import RadixAttention
+from sglang.srt.managers.router.model_runner import InputMetadata
+from sglang.srt.weight_utils import default_weight_loader, hf_model_weights_iterator
 
 
 class GemmaMLP(nn.Module):
@@ -34,17 +30,20 @@ class GemmaMLP(nn.Module):
         self,
         hidden_size: int,
         intermediate_size: int,
-        linear_method: Optional[LinearMethodBase] = None,
+        quant_config: Optional[QuantizationConfig] = None,
     ) -> None:
         super().__init__()
         self.gate_up_proj = MergedColumnParallelLinear(
             hidden_size,
             [intermediate_size] * 2,
             bias=False,
-            linear_method=linear_method,
+            quant_config=quant_config,
         )
         self.down_proj = RowParallelLinear(
-            intermediate_size, hidden_size, bias=False, linear_method=linear_method
+            intermediate_size,
+            hidden_size,
+            bias=False,
+            quant_config=quant_config,
         )
         self.act_fn = GeluAndMul()
 
@@ -65,7 +64,7 @@ class GemmaAttention(nn.Module):
         layer_id: int = 0,
         max_position_embeddings: int = 8192,
         rope_theta: float = 10000,
-        linear_method: Optional[LinearMethodBase] = None,
+        quant_config: Optional[QuantizationConfig] = None,
     ) -> None:
         super().__init__()
         self.hidden_size = hidden_size
@@ -95,13 +94,13 @@ class GemmaAttention(nn.Module):
             self.total_num_heads,
             self.total_num_kv_heads,
             bias=False,
-            linear_method=linear_method,
+            quant_config=quant_config,
         )
         self.o_proj = RowParallelLinear(
             self.total_num_heads * self.head_dim,
             hidden_size,
             bias=False,
-            linear_method=linear_method,
+            quant_config=quant_config,
         )
 
         self.rotary_emb = get_rope(
@@ -138,7 +137,7 @@ class GemmaDecoderLayer(nn.Module):
         self,
         config: PretrainedConfig,
         layer_id: int = 0,
-        linear_method: Optional[LinearMethodBase] = None,
+        quant_config: Optional[QuantizationConfig] = None,
     ) -> None:
         super().__init__()
         self.hidden_size = config.hidden_size
@@ -150,12 +149,12 @@ class GemmaDecoderLayer(nn.Module):
             layer_id=layer_id,
             max_position_embeddings=config.max_position_embeddings,
             rope_theta=config.rope_theta,
-            linear_method=linear_method,
+            quant_config=quant_config,
         )
         self.mlp = GemmaMLP(
             hidden_size=self.hidden_size,
             intermediate_size=config.intermediate_size,
-            linear_method=linear_method,
+            quant_config=quant_config,
         )
         self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
         self.post_attention_layernorm = RMSNorm(
@@ -191,7 +190,7 @@ class GemmaModel(nn.Module):
     def __init__(
         self,
         config: PretrainedConfig,
-        linear_method: Optional[LinearMethodBase] = None,
+        quant_config: Optional[QuantizationConfig] = None,
     ) -> None:
         super().__init__()
         self.config = config
@@ -202,7 +201,7 @@ class GemmaModel(nn.Module):
         )
         self.layers = nn.ModuleList(
             [
-                GemmaDecoderLayer(config, i, linear_method)
+                GemmaDecoderLayer(config, i, quant_config=quant_config)
                 for i in range(config.num_hidden_layers)
             ]
         )
@@ -263,14 +262,14 @@ class GemmaForCausalLM(nn.Module):
     def __init__(
         self,
         config: PretrainedConfig,
-        linear_method: Optional[LinearMethodBase] = None,
+        quant_config: Optional[QuantizationConfig] = None,
         lora_config: Optional[LoRAConfig] = None,
     ) -> None:
         del lora_config  # Unused.
         super().__init__()
         self.config = config
-        self.linear_method = linear_method
-        self.model = GemmaModel(config, linear_method)
+        self.quant_config = quant_config
+        self.model = GemmaModel(config, quant_config=quant_config)
         self.logits_processor = LogitsProcessor(config)
 
     @torch.no_grad()

sglang/srt/models/llama2.py

@@ -1,34 +1,30 @@
 # Adapted from
 # https://github.com/vllm-project/vllm/blob/671af2b1c0b3ed6d856d37c21a561cc429a10701/vllm/model_executor/models/llama.py#L1
 """Inference-only LLaMA model compatible with HuggingFace weights."""
-from typing import Any, Dict, List, Optional, Tuple
+from typing import Any, Dict, Optional, Tuple
 
 import torch
-from sglang.srt.layers.logits_processor import LogitsProcessor
-from sglang.srt.layers.radix_attention import RadixAttention
-from sglang.srt.managers.router.model_runner import InputMetadata
 from torch import nn
 from transformers import LlamaConfig
+from vllm.distributed import get_tensor_model_parallel_world_size
 from vllm.model_executor.layers.activation import SiluAndMul
 from vllm.model_executor.layers.layernorm import RMSNorm
 from vllm.model_executor.layers.linear import (
-    LinearMethodBase,
     MergedColumnParallelLinear,
     QKVParallelLinear,
     RowParallelLinear,
 )
+from vllm.model_executor.layers.quantization.base_config import QuantizationConfig
 from vllm.model_executor.layers.rotary_embedding import get_rope
 from vllm.model_executor.layers.vocab_parallel_embedding import (
     ParallelLMHead,
     VocabParallelEmbedding,
 )
-from vllm.model_executor.parallel_utils.parallel_state import (
-    get_tensor_model_parallel_world_size,
-)
-from vllm.model_executor.weight_utils import (
-    default_weight_loader,
-    hf_model_weights_iterator,
-)
+
+from sglang.srt.layers.logits_processor import LogitsProcessor
+from sglang.srt.layers.radix_attention import RadixAttention
+from sglang.srt.managers.router.model_runner import InputMetadata
+from sglang.srt.weight_utils import default_weight_loader, hf_model_weights_iterator
 
 
 class LlamaMLP(nn.Module):
@@ -37,17 +33,20 @@ class LlamaMLP(nn.Module):
         hidden_size: int,
         intermediate_size: int,
         hidden_act: str,
-        linear_method: Optional[LinearMethodBase] = None,
+        quant_config: Optional[QuantizationConfig] = None,
     ) -> None:
         super().__init__()
         self.gate_up_proj = MergedColumnParallelLinear(
             hidden_size,
             [intermediate_size] * 2,
             bias=False,
-            linear_method=linear_method,
+            quant_config=quant_config,
         )
         self.down_proj = RowParallelLinear(
-            intermediate_size, hidden_size, bias=False, linear_method=linear_method
+            intermediate_size,
+            hidden_size,
+            bias=False,
+            quant_config=quant_config,
         )
         if hidden_act != "silu":
             raise ValueError(
@@ -73,7 +72,7 @@ class LlamaAttention(nn.Module):
         rope_theta: float = 10000,
         rope_scaling: Optional[Dict[str, Any]] = None,
         max_position_embeddings: int = 8192,
-        linear_method: Optional[LinearMethodBase] = None,
+        quant_config: Optional[QuantizationConfig] = None,
     ) -> None:
         super().__init__()
         self.hidden_size = hidden_size
@@ -104,13 +103,13 @@ class LlamaAttention(nn.Module):
             self.total_num_heads,
             self.total_num_kv_heads,
             bias=False,
-            linear_method=linear_method,
+            quant_config=quant_config,
         )
         self.o_proj = RowParallelLinear(
             self.total_num_heads * self.head_dim,
             hidden_size,
             bias=False,
-            linear_method=linear_method,
+            quant_config=quant_config,
         )
 
         self.rotary_emb = get_rope(
@@ -147,7 +146,7 @@ class LlamaDecoderLayer(nn.Module):
         self,
         config: LlamaConfig,
         layer_id: int = 0,
-        linear_method: Optional[LinearMethodBase] = None,
+        quant_config: Optional[QuantizationConfig] = None,
     ) -> None:
         super().__init__()
         self.hidden_size = config.hidden_size
@@ -162,13 +161,13 @@ class LlamaDecoderLayer(nn.Module):
             rope_theta=rope_theta,
             rope_scaling=rope_scaling,
             max_position_embeddings=max_position_embeddings,
-            linear_method=linear_method,
+            quant_config=quant_config,
         )
         self.mlp = LlamaMLP(
             hidden_size=self.hidden_size,
             intermediate_size=config.intermediate_size,
             hidden_act=config.hidden_act,
-            linear_method=linear_method,
+            quant_config=quant_config,
         )
         self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
         self.post_attention_layernorm = RMSNorm(
@@ -204,7 +203,7 @@ class LlamaModel(nn.Module):
     def __init__(
         self,
         config: LlamaConfig,
-        linear_method: Optional[LinearMethodBase] = None,
+        quant_config: Optional[QuantizationConfig] = None,
     ) -> None:
         super().__init__()
         self.config = config
@@ -216,7 +215,7 @@ class LlamaModel(nn.Module):
         )
         self.layers = nn.ModuleList(
             [
-                LlamaDecoderLayer(config, i, linear_method)
+                LlamaDecoderLayer(config, i, quant_config=quant_config)
                 for i in range(config.num_hidden_layers)
             ]
         )
@@ -250,12 +249,12 @@ class LlamaForCausalLM(nn.Module):
     def __init__(
         self,
         config: LlamaConfig,
-        linear_method: Optional[LinearMethodBase] = None,
+        quant_config: Optional[QuantizationConfig] = None,
     ) -> None:
         super().__init__()
         self.config = config
-        self.linear_method = linear_method
-        self.model = LlamaModel(config, linear_method)
+        self.quant_config = quant_config
+        self.model = LlamaModel(config, quant_config=quant_config)
         self.lm_head = ParallelLMHead(config.vocab_size, config.hidden_size)
         self.logits_processor = LogitsProcessor(config)
 

sglang/srt/models/llava.py

@@ -4,6 +4,11 @@ from typing import List, Optional
 
 import numpy as np
 import torch
+from torch import nn
+from transformers import CLIPVisionModel, LlavaConfig
+from transformers.models.llava.modeling_llava import LlavaMultiModalProjector
+from vllm.model_executor.layers.quantization.base_config import QuantizationConfig
+
 from sglang.srt.managers.router.infer_batch import ForwardMode
 from sglang.srt.managers.router.model_runner import InputMetadata
 from sglang.srt.mm_utils import (
@@ -12,21 +17,14 @@ from sglang.srt.mm_utils import (
     unpad_image_shape,
 )
 from sglang.srt.models.llama2 import LlamaForCausalLM
-from torch import nn
-from transformers import CLIPVisionModel, LlamaConfig, LlavaConfig
-from transformers.models.llava.modeling_llava import LlavaMultiModalProjector
-from vllm.model_executor.layers.linear import LinearMethodBase
-from vllm.model_executor.weight_utils import (
-    default_weight_loader,
-    hf_model_weights_iterator,
-)
+from sglang.srt.weight_utils import default_weight_loader, hf_model_weights_iterator
 
 
 class LlavaLlamaForCausalLM(nn.Module):
     def __init__(
         self,
         config: LlavaConfig,
-        linear_method: Optional[LinearMethodBase] = None,
+        quant_config: Optional[QuantizationConfig] = None,
     ) -> None:
         super().__init__()
         self.config = config
@@ -34,7 +32,7 @@ class LlavaLlamaForCausalLM(nn.Module):
         self.config.vision_config.hidden_size = config.mm_hidden_size
         self.config.text_config.hidden_size = config.hidden_size
         self.multi_modal_projector = LlavaMultiModalProjector(config)
-        self.language_model = LlamaForCausalLM(config, linear_method)
+        self.language_model = LlamaForCausalLM(config, quant_config=quant_config)
         if "unpad" in getattr(config, "mm_patch_merge_type", ""):
             self.language_model.model.image_newline = nn.Parameter(
                 torch.empty(config.text_config.hidden_size, dtype=torch.float16)

sglang/srt/models/llavavid.py

@@ -0,0 +1,307 @@
+"""Inference-only LLaVa video model compatible with HuggingFace weights."""
+
+import os
+from typing import List, Optional
+
+import numpy as np
+import torch
+from torch import nn
+from transformers import CLIPVisionModel, LlamaConfig, LlavaConfig
+from transformers.models.llava.modeling_llava import LlavaMultiModalProjector
+from vllm.model_executor.layers.quantization.base_config import QuantizationConfig
+
+from sglang.srt.managers.router.infer_batch import ForwardMode
+from sglang.srt.managers.router.model_runner import InputMetadata
+from sglang.srt.mm_utils import (
+    get_anyres_image_grid_shape,
+    unpad_image,
+    unpad_image_shape,
+)
+from sglang.srt.models.llama2 import LlamaForCausalLM
+from sglang.srt.weight_utils import default_weight_loader, hf_model_weights_iterator
+
+
+class LlavaVidForCausalLM(nn.Module):
+    def __init__(
+        self,
+        config: LlavaConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+    ) -> None:
+        super().__init__()
+        self.config = config
+        self.vision_tower = None
+        self.config.vision_config.hidden_size = config.mm_hidden_size
+        self.config.text_config.hidden_size = config.hidden_size
+        self.multi_modal_projector = LlavaMultiModalProjector(config)
+        self.mm_spatial_pool_stride = getattr(self.config, "mm_spatial_pool_stride", 2)
+        self.resampler = nn.AvgPool2d(
+            kernel_size=self.mm_spatial_pool_stride, stride=self.mm_spatial_pool_stride
+        )
+        self.language_model = LlamaForCausalLM(config, quant_config=quant_config)
+        self.num_frames = getattr(self.config, "num_frames", 16)
+        if "unpad" in getattr(config, "mm_patch_merge_type", ""):
+            self.language_model.model.image_newline = nn.Parameter(
+                torch.empty(config.text_config.hidden_size, dtype=torch.float16)
+            )
+
+    def pad_input_ids(self, input_ids, pad_value, pt_shape=None, image_size=None):
+        new_image_feature_len = self.image_feature_len
+        # now only support spatial_unpad + anyres
+        # if self.mm_patch_merge_type.startswith("spatial"):
+        #     height = width = self.num_patches_per_side
+        #     if pt_shape[0] > 1:
+        #         if self.image_aspect_ratio == "anyres":
+        #             num_patch_width, num_patch_height = get_anyres_image_grid_shape(
+        #                 image_size,
+        #                 self.image_grid_pinpoints,
+        #                 self.vision_tower.config.image_size,
+        #             )
+        #         if "unpad" in self.mm_patch_merge_type:
+        #             h = num_patch_height * height
+        #             w = num_patch_width * width
+        #             new_h, new_w = unpad_image_shape(h, w, image_size)
+        #             new_image_feature_len += new_h * (new_w + 1)
+
+        pad_ids = pad_value * (
+            (new_image_feature_len + len(pad_value)) // len(pad_value)
+        )
+        # print(input_ids)
+        offset = input_ids.index(self.config.image_token_index)
+        # old_len + pad_len - 1, because we need to remove image_token_id
+        new_input_ids = (
+            input_ids[:offset]
+            + pad_ids[:new_image_feature_len]
+            + input_ids[offset + 1 :]
+        )
+        return new_input_ids, offset
+
+    def encode_images(self, pixel_values: torch.Tensor) -> torch.Tensor:
+        image_outputs = self.vision_tower(pixel_values, output_hidden_states=True)
+        # NOTE: This is not memory efficient. (output_hidden_states=True) will save all the hidden stated.
+
+        selected_image_feature = image_outputs.hidden_states[self.vision_feature_layer]
+        if self.vision_feature_select_strategy in ["default", "patch"]:
+            selected_image_feature = selected_image_feature[:, 1:]
+        elif self.vision_feature_select_strategy == "full":
+            selected_image_feature = selected_image_feature
+        else:
+            raise ValueError(
+                f"Unexpected select feature strategy: {self.config.vision_feature_select_strategy}"
+            )
+
+        height = width = self.num_patches_per_side
+        num_of_frames = selected_image_feature.shape[0]
+        selected_image_feature = selected_image_feature.view(
+            num_of_frames, height, width, -1
+        )
+        selected_image_feature = selected_image_feature.permute(0, 3, 1, 2).contiguous()
+        selected_image_feature = (
+            self.resampler(selected_image_feature)
+            .flatten(2)
+            .transpose(1, 2)
+            .contiguous()
+        )
+
+        image_features = self.multi_modal_projector(selected_image_feature)
+
+        return image_features
+
+    def forward(
+        self,
+        input_ids: torch.LongTensor,
+        positions: torch.Tensor,
+        input_metadata: InputMetadata,
+        pixel_values: Optional[List[Optional[np.array]]] = None,
+        image_sizes: Optional[List[List[int]]] = None,
+        image_offsets: Optional[List[int]] = None,
+    ) -> torch.Tensor:
+        if input_metadata.forward_mode == ForwardMode.EXTEND:
+            bs = input_metadata.batch_size
+
+            # Embed text input
+            input_embeds = self.language_model.model.embed_tokens(input_ids)
+
+            # Embed vision input
+            need_vision = (
+                (positions[input_metadata.extend_start_loc] < self.image_feature_len)
+                .cpu()
+                .numpy()
+            )
+            # FIXME: We need to substract the length of the system prompt
+            has_pixel = np.array([pixel_values[i] is not None for i in range(bs)])
+            need_vision = need_vision & has_pixel
+
+            if need_vision.any():
+                pixel_values = [pixel_values[i] for i in range(bs) if need_vision[i]]
+                image_sizes = [image_sizes[i] for i in range(bs) if need_vision[i]]
+
+                ########## Encode Image ########
+
+                if pixel_values[0].ndim == 4:
+                    # llava-hd: BS, num_patch, C=3, H=336, W=336, num_patch obtained from process_images
+                    np.concatenate(pixel_values, axis=0)
+                    # ndim=4
+                    concat_images = torch.tensor(
+                        np.concatenate(pixel_values, axis=0),
+                        device=self.vision_tower.device,
+                    )
+                    # image_features = self.encode_images(concat_images)
+                    # split_sizes = [image.shape[0] for image in pixel_values]
+                    # image_features = torch.split(image_features, split_sizes, dim=0)
+                    image_features = self.encode_images(
+                        concat_images
+                    )  # , prompts)#, image_counts, long_video=long_video)
+                    split_sizes = [image.shape[0] for image in pixel_values]
+                    image_features = torch.split(image_features, split_sizes, dim=0)
+
+                    # hd image_features: BS, num_patch, 576, 4096
+                else:
+                    # normal pixel: BS, C=3, H=336, W=336
+                    pixel_values = torch.tensor(
+                        np.array(pixel_values), device=self.vision_tower.device
+                    )
+                    image_features = self.encode_images(pixel_values)
+                    # image_features: BS, 576, 4096
+
+                new_image_features = []
+                for image_idx, image_feature in enumerate(image_features):
+                    new_image_features.append(image_feature.flatten(0, 1))
+                image_features = new_image_features
+
+                extend_start_loc_cpu = input_metadata.extend_start_loc.cpu().numpy()
+                pt = 0
+                for i in range(bs):
+                    if not need_vision[i]:
+                        continue
+
+                    start_idx = extend_start_loc_cpu[i]
+                    pad_len, pad_dim = image_features[pt].shape  # 576, 4096
+                    dim = input_embeds.shape[1]
+                    assert (
+                        pad_dim == dim
+                    ), "invalid pad_dim={}, input_embed_dim={}!".format(pad_dim, dim)
+                    # Fill in the placeholder for the image
+                    try:
+                        input_embeds[
+                            start_idx
+                            + image_offsets[i] : start_idx
+                            + image_offsets[i]
+                            + pad_len
+                        ] = image_features[pt]
+                    except RuntimeError as e:
+                        print(f"RuntimeError in llava image encoding: {e}")
+                        print(input_embeds.shape)
+                        print(start_idx, image_offsets[i])
+                    pt += 1
+
+            return self.language_model(
+                input_ids, positions, input_metadata, input_embeds=input_embeds
+            )
+        elif input_metadata.forward_mode == ForwardMode.DECODE:
+            return self.language_model(input_ids, positions, input_metadata)
+
+    def load_weights(
+        self,
+        model_name_or_path: str,
+        cache_dir: Optional[str] = None,
+        load_format: str = "auto",
+        revision: Optional[str] = None,
+    ):
+        # load clip vision model by cfg['mm_vision_tower']:
+        #   huggingface_name or path_of_clip_relative_to_llava_model_dir
+        vision_path = self.config.mm_vision_tower
+        self.vision_tower = CLIPVisionModel.from_pretrained(
+            vision_path, torch_dtype=torch.float16
+        ).cuda()
+        self.vision_tower.eval()
+
+        self.vision_feature_layer = self.config.mm_vision_select_layer
+        self.vision_feature_select_strategy = self.config.mm_vision_select_feature
+        self.image_size = self.vision_tower.config.image_size
+        self.patch_size = self.vision_tower.config.patch_size
+
+        self.mm_patch_merge_type = getattr(self.config, "mm_patch_merge_type", "flat")
+        self.image_aspect_ratio = getattr(self.config, "image_aspect_ratio", "square")
+        self.image_grid_pinpoints = getattr(self.config, "image_grid_pinpoints", None)
+
+        print(f"target_frames: {self.num_frames}")
+        self.image_feature_len = self.num_frames * int(
+            (self.image_size / self.patch_size / self.mm_spatial_pool_stride) ** 2
+        )
+        if self.vision_feature_select_strategy == "patch":
+            pass
+        elif self.vision_feature_select_strategy == "cls_patch":
+            self.image_feature_len += 1
+        else:
+            raise ValueError(f"Unexpected select feature: {self.select_feature}")
+
+        # load mm_projector
+        projector_weights = {
+            "model.mm_projector.0": "multi_modal_projector.linear_1",
+            "model.mm_projector.2": "multi_modal_projector.linear_2",
+            "model.vision_resampler.mm_projector.0": "multi_modal_projector.linear_1",
+            "model.vision_resampler.mm_projector.2": "multi_modal_projector.linear_2",
+            "model.vision_tower.vision_tower": "vision_tower",  # Update the vision tower weights if we find them in the checkpoint (it may be finetuned).
+        }
+        params_dict = dict(self.named_parameters())
+        for name, loaded_weight in hf_model_weights_iterator(
+            model_name_or_path, cache_dir, load_format, revision
+        ):
+            # FIXME: why projector weights read two times?
+            if "projector" in name or "vision_tower" in name:
+                for weight_name, param_name in projector_weights.items():
+                    if weight_name in name:
+                        name = name.replace(weight_name, param_name)
+                if name in params_dict:
+                    param = params_dict[name]
+                else:
+                    print(f"Warning: {name} not found in the model")
+                    continue
+                weight_loader = getattr(param, "weight_loader", default_weight_loader)
+                weight_loader(param, loaded_weight)
+
+        # load language model
+        self.language_model.load_weights(
+            model_name_or_path, cache_dir, load_format, revision
+        )
+
+        monkey_path_clip_vision_embed_forward()
+
+    @property
+    def num_patches_per_side(self):
+        return self.image_size // self.patch_size
+
+
+first_call = True
+
+
+def clip_vision_embed_forward(self, pixel_values: torch.FloatTensor) -> torch.Tensor:
+    batch_size = pixel_values.shape[0]
+
+    # Move this conv layer to CPU to avoid a bug in torch >= 2.1 on A10G.
+    global first_call
+    if first_call:
+        self.patch_embedding.cpu().float()
+        first_call = False
+    pixel_values = pixel_values.to(dtype=torch.float32, device="cpu")
+    patch_embeds = self.patch_embedding(pixel_values).cuda().half()
+
+    patch_embeds = patch_embeds.flatten(2).transpose(1, 2)
+
+    class_embeds = self.class_embedding.expand(batch_size, 1, -1)
+    embeddings = torch.cat([class_embeds, patch_embeds], dim=1)
+    embeddings = embeddings + self.position_embedding(self.position_ids)
+    return embeddings
+
+
+def monkey_path_clip_vision_embed_forward():
+    import transformers
+
+    setattr(
+        transformers.models.clip.modeling_clip.CLIPVisionEmbeddings,
+        "forward",
+        clip_vision_embed_forward,
+    )
+
+
+EntryClass = LlavaVidForCausalLM