sglang 0.2.15__py3-none-any.whl → 0.3.1__py3-none-any.whl

sglang/srt/models/gemma.py

@@ -37,7 +37,6 @@ from sglang.srt.layers.activation import GeluAndMul
 from sglang.srt.layers.layernorm import RMSNorm
 from sglang.srt.layers.logits_processor import LogitsProcessor
 from sglang.srt.layers.radix_attention import RadixAttention
-from sglang.srt.layers.sampler import Sampler
 from sglang.srt.model_executor.forward_batch_info import InputMetadata
 
 
@@ -288,7 +287,6 @@ class GemmaForCausalLM(nn.Module):
         self.quant_config = quant_config
         self.model = GemmaModel(config, quant_config=quant_config)
         self.logits_processor = LogitsProcessor(config)
-        self.sampler = Sampler()
 
     @torch.no_grad()
     def forward(
@@ -299,11 +297,9 @@ class GemmaForCausalLM(nn.Module):
         input_embeds: torch.Tensor = None,
     ) -> torch.Tensor:
         hidden_states = self.model(input_ids, positions, input_metadata, input_embeds)
-        logits_output = self.logits_processor(
+        return self.logits_processor(
             input_ids, hidden_states, self.model.embed_tokens.weight, input_metadata
         )
-        sample_output = self.sampler(logits_output, input_metadata.sampling_info)
-        return (sample_output, logits_output)
 
     def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
         stacked_params_mapping = [

sglang/srt/models/gemma2.py

@@ -37,7 +37,6 @@ from sglang.srt.layers.activation import GeluAndMul
 from sglang.srt.layers.layernorm import GemmaRMSNorm
 from sglang.srt.layers.logits_processor import LogitsProcessor
 from sglang.srt.layers.radix_attention import RadixAttention
-from sglang.srt.layers.sampler import Sampler
 from sglang.srt.model_executor.forward_batch_info import InputMetadata
 
 
@@ -347,7 +346,6 @@ class Gemma2ForCausalLM(nn.Module):
         self.quant_config = quant_config
         self.model = Gemma2Model(config, cache_config, quant_config)
         self.logits_processor = LogitsProcessor(config)
-        self.sampler = Sampler()
 
     @torch.no_grad()
     def forward(
@@ -358,11 +356,9 @@ class Gemma2ForCausalLM(nn.Module):
         input_embeds: torch.Tensor = None,
     ) -> torch.Tensor:
         hidden_states = self.model(input_ids, positions, input_metadata, input_embeds)
-        logits_output = self.logits_processor(
+        return self.logits_processor(
             input_ids, hidden_states, self.model.embed_tokens.weight, input_metadata
         )
-        sample_output = self.sampler(logits_output, input_metadata.sampling_info)
-        return sample_output, logits_output
 
     def get_attention_sliding_window_size(self):
         return get_attention_sliding_window_size(self.config)

sglang/srt/models/gpt_bigcode.py

@@ -35,7 +35,6 @@ from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 from sglang.srt.layers.activation import get_act_fn
 from sglang.srt.layers.logits_processor import LogitsProcessor
 from sglang.srt.layers.radix_attention import RadixAttention
-from sglang.srt.layers.sampler import Sampler
 from sglang.srt.model_executor.forward_batch_info import InputMetadata
 
 
@@ -262,7 +261,6 @@ class GPTBigCodeForCausalLM(nn.Module):
         if lora_config:
             self.unpadded_vocab_size += lora_config.lora_extra_vocab_size
         self.logits_processor = LogitsProcessor(config)
-        self.sampler = Sampler()
 
     @torch.no_grad()
     def forward(
@@ -272,11 +270,9 @@ class GPTBigCodeForCausalLM(nn.Module):
         input_metadata: InputMetadata,
     ) -> torch.Tensor:
         hidden_states = self.transformer(input_ids, positions, input_metadata)
-        logits_output = self.logits_processor(
+        return self.logits_processor(
             input_ids, hidden_states, self.lm_head.weight, input_metadata
         )
-        sample_output = self.sampler(logits_output, input_metadata.sampling_info)
-        return sample_output, logits_output
 
     def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
         params_dict = dict(self.named_parameters(remove_duplicate=False))

sglang/srt/models/grok.py

@@ -46,7 +46,6 @@ from sglang.srt.layers.fused_moe import FusedMoE
 from sglang.srt.layers.layernorm import RMSNorm
 from sglang.srt.layers.logits_processor import LogitsProcessor
 from sglang.srt.layers.radix_attention import RadixAttention
-from sglang.srt.layers.sampler import Sampler
 from sglang.srt.model_executor.forward_batch_info import InputMetadata
 
 
@@ -298,7 +297,6 @@ class Grok1ForCausalLM(nn.Module):
         self.model = Grok1Model(config, quant_config=quant_config)
         self.lm_head = ParallelLMHead(config.vocab_size, config.hidden_size)
         self.logits_processor = LogitsProcessor(config)
-        self.sampler = Sampler()
 
         # Monkey patch _prepare_weights to load pre-sharded weights
         setattr(DefaultModelLoader, "_prepare_weights", _prepare_presharded_weights)
@@ -315,11 +313,9 @@ class Grok1ForCausalLM(nn.Module):
         input_embeds: torch.Tensor = None,
     ) -> torch.Tensor:
         hidden_states = self.model(input_ids, positions, input_metadata, input_embeds)
-        logits_output = self.logits_processor(
+        return self.logits_processor(
             input_ids, hidden_states, self.lm_head.weight, input_metadata
         )
-        sample_output = self.sampler(logits_output, input_metadata.sampling_info)
-        return sample_output, logits_output
 
     def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
         stacked_params_mapping = [

sglang/srt/models/internlm2.py

@@ -40,7 +40,6 @@ from sglang.srt.layers.activation import SiluAndMul
 from sglang.srt.layers.layernorm import RMSNorm
 from sglang.srt.layers.logits_processor import LogitsProcessor
 from sglang.srt.layers.radix_attention import RadixAttention
-from sglang.srt.layers.sampler import Sampler
 from sglang.srt.model_executor.forward_batch_info import InputMetadata
 
 
@@ -263,7 +262,6 @@ class InternLM2ForCausalLM(nn.Module):
         self.model = InternLM2Model(config, quant_config)
         self.output = ParallelLMHead(config.vocab_size, config.hidden_size)
         self.logits_processor = LogitsProcessor(config)
-        self.sampler = Sampler()
 
     @torch.no_grad()
     def forward(
@@ -274,11 +272,9 @@ class InternLM2ForCausalLM(nn.Module):
         input_embeds: torch.Tensor = None,
     ) -> torch.Tensor:
         hidden_states = self.model(input_ids, positions, input_metadata, input_embeds)
-        logits_output = self.logits_processor(
+        return self.logits_processor(
             input_ids, hidden_states, self.output.weight, input_metadata
         )
-        sample_output = self.sampler(logits_output, input_metadata.sampling_info)
-        return sample_output, logits_output
 
     def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
         stacked_params_mapping = [

sglang/srt/models/{llama2.py → llama.py}

@@ -41,7 +41,8 @@ from sglang.srt.layers.activation import SiluAndMul
 from sglang.srt.layers.layernorm import RMSNorm
 from sglang.srt.layers.logits_processor import LogitsProcessor, LogitsProcessorOutput
 from sglang.srt.layers.radix_attention import RadixAttention
-from sglang.srt.layers.sampler import Sampler
+from sglang.srt.layers.torchao_utils import apply_torchao_config_
+from sglang.srt.managers.schedule_batch import global_server_args_dict
 from sglang.srt.model_executor.forward_batch_info import InputMetadata
 
 
@@ -295,15 +296,16 @@ class LlamaForCausalLM(nn.Module):
         config: LlamaConfig,
         quant_config: Optional[QuantizationConfig] = None,
         cache_config: Optional[CacheConfig] = None,
-        efficient_weight_load=False,
     ) -> None:
         super().__init__()
         self.config = config
         self.quant_config = quant_config
+        self.torchao_config = global_server_args_dict["torchao_config"]
         self.model = LlamaModel(config, quant_config=quant_config)
         self.lm_head = ParallelLMHead(config.vocab_size, config.hidden_size)
         self.logits_processor = LogitsProcessor(config)
-        self.sampler = Sampler()
+
+        self.param_dict = dict(self.named_parameters())
 
     @torch.no_grad()
     def forward(
@@ -314,13 +316,35 @@ class LlamaForCausalLM(nn.Module):
         input_embeds: torch.Tensor = None,
     ) -> LogitsProcessorOutput:
         hidden_states = self.model(input_ids, positions, input_metadata, input_embeds)
-        logits_output = self.logits_processor(
+        return self.logits_processor(
             input_ids, hidden_states, self.lm_head.weight, input_metadata
         )
-        sample_output = self.sampler(logits_output, input_metadata.sampling_info)
-        return sample_output, logits_output
+
+    def get_hidden_dim(self, module_name):
+        if module_name in ["q_proj", "o_proj", "qkv_proj"]:
+            return self.config.hidden_size, self.config.hidden_size
+        elif module_name in ["kv_proj"]:
+            return self.config.hidden_size, self.config.hidden_size // (
+                self.config.num_attention_heads // self.config.num_key_value_heads
+            )
+        elif module_name == "gate_up_proj":
+            return self.config.hidden_size, self.config.intermediate_size
+        elif module_name == "down_proj":
+            return self.config.intermediate_size, self.config.hidden_size
+        else:
+            raise NotImplementedError()
 
     def get_module_name(self, name):
+        params_mapping = {
+            "q_proj": "qkv_proj",
+            "k_proj": "qkv_proj",
+            "v_proj": "qkv_proj",
+            "gate_proj": "gate_up_proj",
+            "up_proj": "gate_up_proj",
+        }
+        return params_mapping.get(name, name)
+
+    def get_module_name_from_weight_name(self, name):
         stacked_params_mapping = [
             # (param_name, shard_name, shard_id, num_shard)
             ("qkv_proj", "q_proj", "q", 3),
@@ -341,28 +365,26 @@ class LlamaForCausalLM(nn.Module):
         params_dict = dict(self.named_parameters())
         return len(params_dict)
 
-    def load_weights(
-        self, weights: Iterable[Tuple[str, torch.Tensor]], name=None, loaded_weight=None
-    ):
+    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
         stacked_params_mapping = [
             # (param_name, shard_name, shard_id)
-            ("qkv_proj", "q_proj", "q"),
-            ("qkv_proj", "k_proj", "k"),
-            ("qkv_proj", "v_proj", "v"),
-            ("gate_up_proj", "gate_proj", 0),
-            ("gate_up_proj", "up_proj", 1),
+            (".qkv_proj", ".q_proj", "q"),
+            (".qkv_proj", ".k_proj", "k"),
+            (".qkv_proj", ".v_proj", "v"),
+            (".gate_up_proj", ".gate_proj", 0),
+            (".gate_up_proj", ".up_proj", 1),
         ]
-        params_dict = dict(self.named_parameters())
+        params_dict = self.param_dict
 
-        def load_weights_per_param(name, loaded_weight):
+        for name, loaded_weight in weights:
             if "rotary_emb.inv_freq" in name or "projector" in name:
-                return
+                continue
             if "rotary_emb.cos_cached" in name or "rotary_emb.sin_cached" in name:
                 # Models trained using ColossalAI may include these tensors in
                 # the checkpoint. Skip them.
-                return
+                continue
             if name.startswith("model.vision_tower") and name not in params_dict:
-                return
+                continue
 
             for param_name, weight_name, shard_id in stacked_params_mapping:
                 if weight_name not in name:
@@ -378,16 +400,16 @@ class LlamaForCausalLM(nn.Module):
             else:
                 # Skip loading extra bias for GPTQ models.
                 if name.endswith(".bias") and name not in params_dict:
-                    return
+                    continue
                 param = params_dict[name]
                 weight_loader = getattr(param, "weight_loader", default_weight_loader)
                 weight_loader(param, loaded_weight)
 
-        if name is None or loaded_weight is None:
-            for name, loaded_weight in weights:
-                load_weights_per_param(name, loaded_weight)
-        else:
-            load_weights_per_param(name, loaded_weight)
+        apply_torchao_config_(self, params_dict, set(["proj.weight"]))
+
+
+class Phi3ForCausalLM(LlamaForCausalLM):
+    pass
 
 
-EntryClass = LlamaForCausalLM
+EntryClass = [LlamaForCausalLM, Phi3ForCausalLM]

sglang/srt/models/llama_classification.py

@@ -16,17 +16,15 @@ limitations under the License.
 from typing import Iterable, Optional, Tuple
 
 import torch
-import tqdm
 from torch import nn
 from transformers import LlamaConfig
 from vllm.config import CacheConfig
-from vllm.distributed import get_tensor_model_parallel_rank
 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.layers.logits_processor import LogitsProcessorOutput
 from sglang.srt.model_executor.forward_batch_info import InputMetadata
-from sglang.srt.models.llama2 import LlamaModel
+from sglang.srt.models.llama import LlamaForCausalLM, LlamaModel
 
 
 class LlamaForClassification(nn.Module):
@@ -42,10 +40,12 @@ class LlamaForClassification(nn.Module):
         self.model = LlamaModel(config, quant_config=quant_config)
 
         self.classification_head = nn.Linear(
-            config.hidden_size, config.classification_out_size
+            config.hidden_size, config.classification_out_size, bias=False
         )
         self.eos_token_id = config.eos_token_id
 
+        self.param_dict = dict(self.named_parameters())
+
     @torch.no_grad()
     def forward(
         self,
@@ -65,7 +65,7 @@ class LlamaForClassification(nn.Module):
                 (input_metadata.batch_size, self.config.classification_out_size)
             ).to(input_ids.device)
 
-        return LogitsProcessorOutput(
+        logits_output = LogitsProcessorOutput(
             next_token_logits=scores,
             next_token_logprobs=scores,
             normalized_prompt_logprobs=scores,
@@ -74,46 +74,20 @@ class LlamaForClassification(nn.Module):
             output_top_logprobs=None,
         )
 
+        return logits_output
+
     def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
-        stacked_params_mapping = [
-            # (param_name, shard_name, shard_id)
-            ("qkv_proj", "q_proj", "q"),
-            ("qkv_proj", "k_proj", "k"),
-            ("qkv_proj", "v_proj", "v"),
-            ("gate_up_proj", "gate_proj", 0),
-            ("gate_up_proj", "up_proj", 1),
-        ]
-        params_dict = dict(self.named_parameters())
-        if get_tensor_model_parallel_rank() == 0:
-            weights = tqdm.tqdm(weights, total=int(len(params_dict) * 1.5))
-        for name, loaded_weight in weights:
-            if "rotary_emb.inv_freq" in name or "projector" in name:
-                continue
-            if "rotary_emb.cos_cached" in name or "rotary_emb.sin_cached" in name:
-                # Models trained using ColossalAI may include these tensors in
-                # the checkpoint. Skip them.
-                continue
-            if "lm_head" in name:
-                continue
+        params_dict = self.param_dict
 
-            for param_name, weight_name, shard_id in stacked_params_mapping:
-                if weight_name not in name:
-                    continue
-                name = name.replace(weight_name, param_name)
-                # Skip loading extra bias for GPTQ models.
-                if name.endswith(".bias") and name not in params_dict:
-                    continue
-                param = params_dict[name]
-                weight_loader = param.weight_loader
-                weight_loader(param, loaded_weight, shard_id)
-                break
-            else:
-                # Skip loading extra bias for GPTQ models.
-                if name.endswith(".bias") and name not in params_dict:
-                    continue
+        for name, loaded_weight in weights:
+            if "classification_head" in name:
                 param = params_dict[name]
                 weight_loader = getattr(param, "weight_loader", default_weight_loader)
                 weight_loader(param, loaded_weight)
+            elif "lm_head" in name:
+                continue
+            else:
+                LlamaForCausalLM.load_weights(self, [(name, loaded_weight)])
 
 
 EntryClass = LlamaForClassification

sglang/srt/models/llama_embedding.py

@@ -1,4 +1,4 @@
-from typing import Iterable, Optional, Tuple
+from typing import Iterable, Tuple
 
 import torch
 from torch import nn
@@ -7,7 +7,7 @@ from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 
 from sglang.srt.layers.pooler import EmbeddingPoolerOutput, Pooler, PoolingType
 from sglang.srt.model_executor.model_runner import InputMetadata
-from sglang.srt.models.llama2 import LlamaForCausalLM, LlamaModel
+from sglang.srt.models.llama import LlamaModel
 
 
 class LlamaEmbeddingModel(nn.Module):
@@ -16,7 +16,6 @@ class LlamaEmbeddingModel(nn.Module):
         config: LlamaConfig,
         quant_config=None,
         cache_config=None,
-        efficient_weight_load=False,
     ) -> None:
         super().__init__()
         self.model = LlamaModel(config, quant_config=quant_config)
@@ -86,6 +85,8 @@ class LlamaEmbeddingModel(nn.Module):
             load_weights_per_param(name, loaded_weight)
 
 
-EntryClass = LlamaEmbeddingModel
-# compat: e5-mistral model.config class == MistralModel
-EntryClassRemapping = [("MistralModel", LlamaEmbeddingModel)]
+class MistralModel(LlamaEmbeddingModel):
+    pass
+
+
+EntryClass = [LlamaEmbeddingModel, MistralModel]

sglang/srt/models/llava.py

@@ -41,7 +41,7 @@ from sglang.srt.mm_utils import (
     unpad_image_shape,
 )
 from sglang.srt.model_executor.forward_batch_info import ForwardMode, InputMetadata
-from sglang.srt.models.llama2 import LlamaForCausalLM
+from sglang.srt.models.llama import LlamaForCausalLM
 from sglang.srt.models.mistral import MistralForCausalLM
 from sglang.srt.models.qwen2 import Qwen2ForCausalLM
 
@@ -136,8 +136,14 @@ class LlavaBaseForCausalLM(nn.Module):
         image_sizes: Optional[List[List[int]]] = None,
         image_offsets: Optional[List[int]] = None,
     ) -> torch.Tensor:
-        if input_metadata.forward_mode == ForwardMode.EXTEND:
+        if input_metadata.forward_mode.is_extend():
             bs = input_metadata.batch_size
+            # Got List[List[str]] extend it to List[str]
+            # The length of the List should be equal to batch size
+            modalities_list = []
+            for modalities in input_metadata.modalities:
+                if modalities is not None:
+                    modalities_list.extend(modalities)
 
             # Embed text inputs
             input_embeds = self.language_model.model.embed_tokens(input_ids)
@@ -179,11 +185,14 @@ class LlavaBaseForCausalLM(nn.Module):
                     new_image_features = []
                     height = width = self.num_patches_per_side
                     for image_idx, image_feature in enumerate(image_features):
-                        if len(image_sizes[image_idx]) == 1:
+                        if modalities_list[image_idx] == "image":
                             image_aspect_ratio = (
                                 self.config.image_aspect_ratio
                             )  # single image
-                        else:
+                        elif (
+                            modalities_list[image_idx] == "multi-images"
+                            or modalities_list[image_idx] == "video"
+                        ):
                             image_aspect_ratio = "pad"  # multi image
                         # image_aspect_ratio = (
                         #     "anyres" if len(image_sizes[image_idx]) == 1 else "pad"
@@ -191,6 +200,7 @@ class LlavaBaseForCausalLM(nn.Module):
                         if (
                             image_feature.shape[0] > 1
                             and "anyres" in image_aspect_ratio
+                            and modalities_list[image_idx] == "image"
                         ):
                             base_image_feature = image_feature[0]
                             image_feature = image_feature[1:]
@@ -290,7 +300,7 @@ class LlavaBaseForCausalLM(nn.Module):
                             )
                             image_feature = image_feature.unsqueeze(0)
                         else:
-                            if image_feature.shape[0] > 16:  # video
+                            if modalities_list[image_idx] == "video":  # video
                                 # 2x2 pooling
                                 num_of_frames = image_feature.shape[0]
                                 image_feature = image_feature.view(
@@ -312,6 +322,21 @@ class LlavaBaseForCausalLM(nn.Module):
                                     .transpose(1, 2)
                                     .contiguous()
                                 )  # N, C, H*W
+                            if "unpad" in self.mm_patch_merge_type:
+                                image_feature = torch.cat(
+                                    (
+                                        image_feature,
+                                        # Expand to (bs, 1, hidden_dim) and concat at the end of the image tokens
+                                        self.language_model.model.image_newline[
+                                            None, None
+                                        ].expand(
+                                            image_feature.shape[0],
+                                            1,
+                                            image_feature.shape[-1],
+                                        ),
+                                    ),
+                                    dim=1,
+                                )
 
                         new_image_features.append(image_feature)
                     image_features = new_image_features
@@ -350,7 +375,7 @@ class LlavaBaseForCausalLM(nn.Module):
             return self.language_model(
                 input_ids, positions, input_metadata, input_embeds=input_embeds
             )
-        elif input_metadata.forward_mode == ForwardMode.DECODE:
+        elif input_metadata.forward_mode.is_decode():
             return self.language_model(input_ids, positions, input_metadata)
 
     def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
@@ -395,21 +420,19 @@ class LlavaBaseForCausalLM(nn.Module):
             "model.mm_projector.0": "multi_modal_projector.linear_1",
             "model.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).
+            "model.image_newline": "language_model.model.image_newline",
         }
         params_dict = dict(self.named_parameters())
-        weights = list(weights)
         for name, loaded_weight in weights:
-            # FIXME: why projector weights read two times?
-            if "projector" in name or "vision_tower" in name:
+            if "projector" in name or "vision_tower" in name or "image_newline" in name:
                 for weight_name, param_name in projector_weights.items():
                     if weight_name in name:
                         name = name.replace(weight_name, param_name)
                 param = params_dict[name]
                 weight_loader = getattr(param, "weight_loader", default_weight_loader)
                 weight_loader(param, loaded_weight)
-
-        # load language model
-        self.language_model.load_weights(weights)
+            else:
+                self.language_model.load_weights([(name, loaded_weight)])
 
     @property
     def num_patches_per_side(self):
@@ -429,6 +452,7 @@ class LlavaLlamaForCausalLM(LlavaBaseForCausalLM):
         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", ""):
@@ -448,9 +472,9 @@ class LlavaQwenForCausalLM(LlavaBaseForCausalLM):
 
         self.config = config
         self.vision_tower = None
+
         if getattr(self.config, "vision_config", None) is None:
             self.config.vision_config = CLIPVisionConfig(self.config.mm_vision_tower)
-
         if getattr(self.config, "text_config", None) is None:
             self.config.text_config = Qwen2Config(self.config._name_or_path)
 
@@ -459,7 +483,6 @@ class LlavaQwenForCausalLM(LlavaBaseForCausalLM):
 
         if getattr(self.config, "projector_hidden_act", None) is None:
             self.config.projector_hidden_act = "gelu"
-
         if getattr(self.config, "image_token_index", None) is None:
             self.config.image_token_index = 151646
 
@@ -482,9 +505,9 @@ class LlavaMistralForCausalLM(LlavaBaseForCausalLM):
 
         self.config = config
         self.vision_tower = None
+
         if getattr(self.config, "vision_config", None) is None:
             self.config.vision_config = CLIPVisionConfig(self.config.mm_vision_tower)
-
         if getattr(self.config, "text_config", None) is None:
             self.config.text_config = MistralConfig(self.config._name_or_path)
 
@@ -493,7 +516,6 @@ class LlavaMistralForCausalLM(LlavaBaseForCausalLM):
 
         if getattr(self.config, "projector_hidden_act", None) is None:
             self.config.projector_hidden_act = "gelu"
-
         if getattr(self.config, "image_token_index", None) is None:
             self.config.image_token_index = 32000
 

sglang/srt/models/llavavid.py

@@ -27,7 +27,7 @@ from vllm.model_executor.layers.quantization.base_config import QuantizationConf
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 
 from sglang.srt.model_executor.forward_batch_info import ForwardMode, InputMetadata
-from sglang.srt.models.llama2 import LlamaForCausalLM
+from sglang.srt.models.llama import LlamaForCausalLM
 
 
 class LlavaVidForCausalLM(nn.Module):
@@ -116,7 +116,7 @@ class LlavaVidForCausalLM(nn.Module):
         image_sizes: Optional[List[List[int]]] = None,
         image_offsets: Optional[List[int]] = None,
     ) -> torch.Tensor:
-        if input_metadata.forward_mode == ForwardMode.EXTEND:
+        if input_metadata.forward_mode.is_extend():
             bs = input_metadata.batch_size
 
             # Embed text inputs
@@ -199,7 +199,7 @@ class LlavaVidForCausalLM(nn.Module):
             return self.language_model(
                 input_ids, positions, input_metadata, input_embeds=input_embeds
             )
-        elif input_metadata.forward_mode == ForwardMode.DECODE:
+        elif input_metadata.forward_mode.is_decode():
             return self.language_model(input_ids, positions, input_metadata)
 
     def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
@@ -239,12 +239,12 @@ class LlavaVidForCausalLM(nn.Module):
             "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).
+            "model.image_newline": "language_model.model.image_newline",
         }
         params_dict = dict(self.named_parameters())
-        weights = list(weights)
         for name, loaded_weight in weights:
             # FIXME: why projector weights read two times?
-            if "projector" in name or "vision_tower" in name:
+            if "projector" in name or "vision_tower" in name or "image_newline" in name:
                 for weight_name, param_name in projector_weights.items():
                     if weight_name in name:
                         name = name.replace(weight_name, param_name)
@@ -255,9 +255,8 @@ class LlavaVidForCausalLM(nn.Module):
                     continue
                 weight_loader = getattr(param, "weight_loader", default_weight_loader)
                 weight_loader(param, loaded_weight)
-
-        # load language model
-        self.language_model.load_weights(weights)
+            else:
+                self.language_model.load_weights([(name, loaded_weight)])
 
     @property
     def num_patches_per_side(self):

sglang/srt/models/minicpm.py

@@ -39,7 +39,6 @@ from sglang.srt.layers.activation import SiluAndMul
 from sglang.srt.layers.layernorm import RMSNorm
 from sglang.srt.layers.logits_processor import LogitsProcessor
 from sglang.srt.layers.radix_attention import RadixAttention
-from sglang.srt.layers.sampler import Sampler
 from sglang.srt.model_executor.forward_batch_info import InputMetadata
 
 
@@ -298,7 +297,6 @@ class MiniCPMForCausalLM(nn.Module):
         self.scale_width = self.config.hidden_size / self.config.dim_model_base
 
         self.logits_processor = LogitsProcessor(config)
-        self.sampler = Sampler()
 
     @torch.no_grad()
     def forward(
@@ -316,11 +314,9 @@ class MiniCPMForCausalLM(nn.Module):
             lm_head_weight = self.model.embed_tokens.weight
         else:
             lm_head_weight = self.lm_head.weight
-        logits_output = self.logits_processor(
+        return self.logits_processor(
             input_ids, hidden_states, lm_head_weight, input_metadata
         )
-        sample_output = self.sampler(logits_output, input_metadata.sampling_info)
-        return sample_output, logits_output
 
     def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
         stacked_params_mapping = [