sglang 0.2.14.post1__py3-none-any.whl → 0.2.14.post2__py3-none-any.whl

sglang/srt/models/llama2.py

@@ -357,6 +357,9 @@ class LlamaForCausalLM(nn.Module):
                 # Models trained using ColossalAI may include these tensors in
                 # the checkpoint. Skip them.
                 return
+            if name.startswith("model.vision_tower") and name not in params_dict:
+                return
+
             for param_name, weight_name, shard_id in stacked_params_mapping:
                 if weight_name not in name:
                     continue
@@ -364,8 +367,6 @@ class LlamaForCausalLM(nn.Module):
                 # Skip loading extra bias for GPTQ models.
                 if name.endswith(".bias") and name not in params_dict:
                     continue
-                if name.startswith("model.vision_tower") and name not in params_dict:
-                    continue
                 param = params_dict[name]
                 weight_loader = param.weight_loader
                 weight_loader(param, loaded_weight, shard_id)
@@ -374,8 +375,6 @@ class LlamaForCausalLM(nn.Module):
                 # Skip loading extra bias for GPTQ models.
                 if name.endswith(".bias") and name not in params_dict:
                     return
-                if name.startswith("model.vision_tower") and name not in params_dict:
-                    return
                 param = params_dict[name]
                 weight_loader = getattr(param, "weight_loader", default_weight_loader)
                 weight_loader(param, loaded_weight)

sglang/srt/models/llama_classification.py

@@ -103,8 +103,6 @@ class LlamaForClassification(nn.Module):
                 # Skip loading extra bias for GPTQ models.
                 if name.endswith(".bias") and name not in params_dict:
                     continue
-                if name.startswith("model.vision_tower") and name not in params_dict:
-                    continue
                 param = params_dict[name]
                 weight_loader = param.weight_loader
                 weight_loader(param, loaded_weight, shard_id)
@@ -113,8 +111,6 @@ class LlamaForClassification(nn.Module):
                 # Skip loading extra bias for GPTQ models.
                 if name.endswith(".bias") and name not in params_dict:
                     continue
-                if name.startswith("model.vision_tower") and name not in params_dict:
-                    continue
                 param = params_dict[name]
                 weight_loader = getattr(param, "weight_loader", default_weight_loader)
                 weight_loader(param, loaded_weight)

sglang/srt/models/llama_embedding.py

@@ -57,6 +57,9 @@ class LlamaEmbeddingModel(nn.Module):
                 # Models trained using ColossalAI may include these tensors in
                 # the checkpoint. Skip them.
                 return
+            if name.startswith("model.vision_tower") and name not in params_dict:
+                return
+
             for param_name, weight_name, shard_id in stacked_params_mapping:
                 if weight_name not in name:
                     continue
@@ -64,8 +67,6 @@ class LlamaEmbeddingModel(nn.Module):
                 # Skip loading extra bias for GPTQ models.
                 if name.endswith(".bias") and name not in params_dict:
                     continue
-                if name.startswith("model.vision_tower") and name not in params_dict:
-                    continue
                 param = params_dict[name]
                 weight_loader = param.weight_loader
                 weight_loader(param, loaded_weight, shard_id)
@@ -74,8 +75,6 @@ class LlamaEmbeddingModel(nn.Module):
                 # Skip loading extra bias for GPTQ models.
                 if name.endswith(".bias") and name not in params_dict:
                     return
-                if name.startswith("model.vision_tower") and name not in params_dict:
-                    return
                 param = params_dict[name]
                 weight_loader = getattr(param, "weight_loader", default_weight_loader)
                 weight_loader(param, loaded_weight)

sglang/srt/models/llava.py

@@ -28,7 +28,6 @@ from transformers import (
     LlavaConfig,
     MistralConfig,
     Qwen2Config,
-    SiglipVisionConfig,
     SiglipVisionModel,
 )
 from transformers.models.llava.modeling_llava import LlavaMultiModalProjector
@@ -47,32 +46,19 @@ from sglang.srt.models.mistral import MistralForCausalLM
 from sglang.srt.models.qwen2 import Qwen2ForCausalLM
 
 
-class LlavaLlamaForCausalLM(nn.Module):
-    def __init__(
+class LlavaBaseForCausalLM(nn.Module):
+    def pad_input_ids(
         self,
-        config: LlavaConfig,
-        quant_config: Optional[QuantizationConfig] = None,
-        cache_config: Optional[CacheConfig] = 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.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)
-            )
-
-    def pad_input_ids(self, input_ids, pad_value, pt_shape=None, image_size=None):
-
+        input_ids: List[int],
+        pad_value: List[int],
+        pixel_values: List,
+        image_sizes: List[List[int]],
+    ):
         # hardcode for spatial_unpad + anyres
-        image_aspect_ratio = "anyres" if len(image_size) == 1 else "pad"
+        image_aspect_ratio = "anyres" if len(image_sizes) == 1 else "pad"
         offset_list = []
-        for image_s in image_size:
-            if len(image_size) > 16:
+        for image_s in image_sizes:
+            if len(image_sizes) > 16:
                 # 2x2 pooling with stride 2
                 new_image_feature_len = (
                     math.ceil(self.image_size / self.patch_size / 2) ** 2
@@ -153,17 +139,15 @@ class LlavaLlamaForCausalLM(nn.Module):
         if input_metadata.forward_mode == ForwardMode.EXTEND:
             bs = input_metadata.batch_size
 
-            # Embed text input
+            # Embed text inputs
             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()
+
+            # Whether the requests need vision inputs
+            max_image_offset = np.array(
+                [max(image_offsets[i]) if image_offsets[i] else -1 for i in range(bs)]
             )
-            # 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
+            start_positions = positions[input_metadata.extend_start_loc].cpu().numpy()
+            need_vision = start_positions <= max_image_offset
 
             if need_vision.any():
                 pixel_values = [pixel_values[i] for i in range(bs) if need_vision[i]]
@@ -332,31 +316,35 @@ class LlavaLlamaForCausalLM(nn.Module):
                         new_image_features.append(image_feature)
                     image_features = new_image_features
 
+                # Fill in the placeholder for the image
                 extend_start_loc_cpu = input_metadata.extend_start_loc.cpu().numpy()
+                prefix_lens_cpu = input_metadata.extend_prefix_lens.cpu().numpy()
                 pt = 0
                 for i in range(bs):
                     if not need_vision[i]:
                         continue
 
                     start_idx = extend_start_loc_cpu[i]
-                    pad_dim = image_features[pt].shape[-1]  # 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:
-                        for j, image_off in enumerate(image_offsets[i]):
-                            # print("actual image_features length: ", image_features[pt][j].shape[0])
-                            pad_len = image_features[pt][j].shape[0]
-                            input_embeds[
-                                start_idx + image_off : start_idx + image_off + pad_len
-                            ] = image_features[pt][j]
-                    except RuntimeError as e:
-                        print(f"RuntimeError in llava image encoding: {e}")
-                        print(image_features[pt].shape)
-                        print(input_embeds.shape)
-                        print(start_idx, image_offsets[i])
+                    prefix_len = prefix_lens_cpu[i]
+
+                    # Multiple images
+                    for j, image_offset in enumerate(image_offsets[i]):
+                        if image_offset < prefix_len:
+                            continue
+
+                        tmp_image_feature = image_features[pt][j]
+                        pad_len = tmp_image_feature.shape[0]
+
+                        left_idx = start_idx + (image_offset - prefix_len)
+                        right_idx = start_idx + (image_offset - prefix_len) + pad_len
+                        try:
+                            input_embeds[left_idx:right_idx] = tmp_image_feature
+                        except RuntimeError as e:
+                            print(f"RuntimeError in image encoding: {e}")
+                            print(f"{input_embeds.shape=}, {tmp_image_feature.shape=}")
+                            print(
+                                f"{start_idx=}, {image_offset=}, {prefix_len=}, {pad_len=}"
+                            )
                     pt += 1
 
             return self.language_model(
@@ -366,8 +354,9 @@ class LlavaLlamaForCausalLM(nn.Module):
             return self.language_model(input_ids, positions, input_metadata)
 
     def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
-        # load clip vision model by cfg['mm_vision_tower']:
-        #   huggingface_name or path_of_clip_relative_to_llava_model_dir
+        # Load clip vision model by cfg['mm_vision_tower']:
+        # huggingface_name or path_of_clip_relative_to_llava_model_dir
+        # We put the initialization here instead of __init__ to allow it being reused by other subclasses.
         vision_path = self.config.mm_vision_tower
         if "clip" in vision_path:
             self.vision_tower = CLIPVisionModel.from_pretrained(
@@ -422,21 +411,41 @@ class LlavaLlamaForCausalLM(nn.Module):
         # load language model
         self.language_model.load_weights(weights)
 
-        monkey_path_clip_vision_embed_forward()
-
     @property
     def num_patches_per_side(self):
         return self.image_size // self.patch_size
 
 
-class LlavaQwenForCausalLM(LlavaLlamaForCausalLM):
+class LlavaLlamaForCausalLM(LlavaBaseForCausalLM):
     def __init__(
         self,
         config: LlavaConfig,
         quant_config: Optional[QuantizationConfig] = None,
         cache_config: Optional[CacheConfig] = None,
     ) -> None:
-        super().__init__(config, quant_config=quant_config, cache_config=cache_config)
+        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.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)
+            )
+
+
+class LlavaQwenForCausalLM(LlavaBaseForCausalLM):
+    def __init__(
+        self,
+        config: LlavaConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+        cache_config: Optional[CacheConfig] = None,
+    ) -> None:
+        super().__init__()
+
         self.config = config
         self.vision_tower = None
         if getattr(self.config, "vision_config", None) is None:
@@ -462,14 +471,15 @@ class LlavaQwenForCausalLM(LlavaLlamaForCausalLM):
             )
 
 
-class LlavaMistralForCausalLM(LlavaLlamaForCausalLM):
+class LlavaMistralForCausalLM(LlavaBaseForCausalLM):
     def __init__(
         self,
         config: LlavaConfig,
         quant_config: Optional[QuantizationConfig] = None,
         cache_config: Optional[CacheConfig] = None,
     ) -> None:
-        super().__init__(config, quant_config=quant_config, cache_config=cache_config)
+        super().__init__()
+
         self.config = config
         self.vision_tower = None
         if getattr(self.config, "vision_config", None) is None:
@@ -495,36 +505,4 @@ class LlavaMistralForCausalLM(LlavaLlamaForCausalLM):
             )
 
 
-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 = [LlavaLlamaForCausalLM, LlavaQwenForCausalLM, LlavaMistralForCausalLM]

sglang/srt/models/llavavid.py

@@ -26,11 +26,6 @@ from vllm.config import CacheConfig
 from vllm.model_executor.layers.quantization.base_config import QuantizationConfig
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 
-from sglang.srt.mm_utils import (
-    get_anyres_image_grid_shape,
-    unpad_image,
-    unpad_image_shape,
-)
 from sglang.srt.model_executor.forward_batch_info import ForwardMode, InputMetadata
 from sglang.srt.models.llama2 import LlamaForCausalLM
 
@@ -59,23 +54,14 @@ class LlavaVidForCausalLM(nn.Module):
                 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):
+    def pad_input_ids(
+        self,
+        input_ids: List[int],
+        pad_value: List[int],
+        pixel_values: List,
+        image_sizes: List[List[int]],
+    ):
         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)
@@ -87,7 +73,7 @@ class LlavaVidForCausalLM(nn.Module):
             + pad_ids[:new_image_feature_len]
             + input_ids[offset + 1 :]
         )
-        return new_input_ids, offset
+        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)
@@ -133,22 +119,18 @@ class LlavaVidForCausalLM(nn.Module):
         if input_metadata.forward_mode == ForwardMode.EXTEND:
             bs = input_metadata.batch_size
 
-            # Embed text input
+            # Embed text inputs
             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()
+            # Whether the requests need vision inputs
+            max_image_offset = np.array(
+                [max(image_offsets[i]) if image_offsets[i] else -1 for i in range(bs)]
             )
-            # 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
+            start_positions = positions[input_metadata.extend_start_loc].cpu().numpy()
+            need_vision = start_positions <= max_image_offset
 
             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 ########
 
@@ -183,31 +165,36 @@ class LlavaVidForCausalLM(nn.Module):
                     new_image_features.append(image_feature.flatten(0, 1))
                 image_features = new_image_features
 
+                # Fill in the placeholder for the image
                 extend_start_loc_cpu = input_metadata.extend_start_loc.cpu().numpy()
+                prefix_lens_cpu = input_metadata.extend_prefix_lens.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
+                    prefix_len = prefix_lens_cpu[i]
+
+                    # Multiple images
+                    for image_offset in image_offsets[i]:
+                        if image_offset < prefix_len:
+                            continue
+
+                        tmp_image_feature = image_features[pt]
+                        pad_len = tmp_image_feature.shape[0]
+
+                        left_idx = start_idx + (image_offset - prefix_len)
+                        right_idx = start_idx + (image_offset - prefix_len) + pad_len
+                        try:
+                            input_embeds[left_idx:right_idx] = tmp_image_feature
+                        except RuntimeError as e:
+                            print(f"RuntimeError in image encoding: {e}")
+                            print(f"{input_embeds.shape=}, {tmp_image_feature.shape=}")
+                            print(
+                                f"{start_idx=}, {image_offset=}, {prefix_len=}, {pad_len=}"
+                            )
+                        pt += 1
 
             return self.language_model(
                 input_ids, positions, input_metadata, input_embeds=input_embeds
@@ -216,8 +203,9 @@ class LlavaVidForCausalLM(nn.Module):
             return self.language_model(input_ids, positions, input_metadata)
 
     def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
-        # load clip vision model by cfg['mm_vision_tower']:
-        #   huggingface_name or path_of_clip_relative_to_llava_model_dir
+        # Load clip vision model by cfg['mm_vision_tower']:
+        # huggingface_name or path_of_clip_relative_to_llava_model_dir
+        # We put the initialization here instead of __init__ to allow it being reused by other subclasses.
         vision_path = self.config.mm_vision_tower
         self.vision_tower = CLIPVisionModel.from_pretrained(
             vision_path, torch_dtype=torch.float16
@@ -271,43 +259,9 @@ class LlavaVidForCausalLM(nn.Module):
         # load language model
         self.language_model.load_weights(weights)
 
-        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

sglang/srt/models/qwen2.py

@@ -312,6 +312,9 @@ class Qwen2ForCausalLM(nn.Module):
                 # Models trained using ColossalAI may include these tensors in
                 # the checkpoint. Skip them.
                 continue
+            if name.startswith("model.vision_tower") and name not in params_dict:
+                continue
+
             for param_name, weight_name, shard_id in stacked_params_mapping:
                 if weight_name not in name:
                     continue
@@ -319,8 +322,6 @@ class Qwen2ForCausalLM(nn.Module):
                 # Skip loading extra bias for GPTQ models.
                 if name.endswith(".bias") and name not in params_dict:
                     continue
-                if name.startswith("model.vision_tower") and name not in params_dict:
-                    continue
                 param = params_dict[name]
                 weight_loader = param.weight_loader
                 weight_loader(param, loaded_weight, shard_id)
@@ -329,8 +330,6 @@ class Qwen2ForCausalLM(nn.Module):
                 # Skip loading extra bias for GPTQ models.
                 if name.endswith(".bias") and name not in params_dict:
                     continue
-                if name.startswith("model.vision_tower") and name not in params_dict:
-                    continue
                 param = params_dict[name]
                 weight_loader = getattr(param, "weight_loader", default_weight_loader)
                 weight_loader(param, loaded_weight)

sglang/srt/models/qwen2_moe.py

@@ -401,24 +401,12 @@ class Qwen2MoeForCausalLM(nn.Module):
             ("gate_up_proj", "up_proj", 1),
         ]
 
-        expert_params_mapping = [
-            # These are the weights for the experts
-            # (param_name, weight_name, expert_id, shard_id)
-            (
-                (
-                    "experts.w13_weight"
-                    if weight_name in ["gate_proj", "up_proj"]
-                    else "experts.w2_weight"
-                ),
-                f"experts.{expert_id}.{weight_name}.weight",
-                expert_id,
-                shard_id,
-            )
-            for expert_id in range(self.config.num_experts)
-            for shard_id, weight_name in enumerate(
-                ["gate_proj", "down_proj", "up_proj"]
-            )
-        ]
+        expert_params_mapping = FusedMoE.make_expert_params_mapping(
+            ckpt_gate_proj_name="gate_proj",
+            ckpt_down_proj_name="down_proj",
+            ckpt_up_proj_name="up_proj",
+            num_experts=self.config.num_experts,
+        )
 
         params_dict = dict(self.named_parameters())
         for name, loaded_weight in weights:
@@ -458,7 +446,7 @@ class Qwen2MoeForCausalLM(nn.Module):
                     weight_loader(
                         param,
                         loaded_weight,
-                        weight_name,
+                        name,
                         shard_id=shard_id,
                         expert_id=expert_id,
                     )

sglang/srt/models/yivl.py

@@ -24,10 +24,7 @@ from vllm.config import CacheConfig
 from vllm.model_executor.layers.quantization.base_config import QuantizationConfig
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 
-from sglang.srt.models.llava import (
-    LlavaLlamaForCausalLM,
-    monkey_path_clip_vision_embed_forward,
-)
+from sglang.srt.models.llava import LlavaLlamaForCausalLM
 
 
 class YiVLForCausalLM(LlavaLlamaForCausalLM):
@@ -50,7 +47,7 @@ class YiVLForCausalLM(LlavaLlamaForCausalLM):
             self.config._name_or_path,
             torch_dtype=torch.float16,
             subfolder=self.vision_tower_subfolder,
-        ).cuda()
+        ).to("cuda")
 
         self.vision_tower.eval()
 
@@ -94,8 +91,6 @@ class YiVLForCausalLM(LlavaLlamaForCausalLM):
         # load language model
         self.language_model.load_weights(weights)
 
-        monkey_path_clip_vision_embed_forward()
-
 
 class YiVLMultiModalProjector(nn.Module):
     def __init__(self, config: LlavaConfig):

sglang/srt/server.py

@@ -335,12 +335,12 @@ def launch_server(
     pipe_detoken_reader, pipe_detoken_writer = mp.Pipe(duplex=False)
 
     if server_args.dp_size == 1:
-        start_process = start_controller_process_single
+        start_controller_process = start_controller_process_single
     else:
-        start_process = start_controller_process_multi
+        start_controller_process = start_controller_process_multi
 
     proc_controller = mp.Process(
-        target=start_process,
+        target=start_controller_process,
         args=(server_args, port_args, pipe_controller_writer, model_overide_args),
     )
     proc_controller.start()