sglang 0.4.6.post3__py3-none-any.whl → 0.4.6.post5__py3-none-any.whl

sglang/srt/models/exaone.py

@@ -307,9 +307,14 @@ class ExaoneForCausalLM(nn.Module):
         self.transformer = ExaoneModel(
             config, quant_config=quant_config, prefix=add_prefix("transformer", prefix)
         )
-        self.lm_head = ParallelLMHead(
-            config.vocab_size, config.hidden_size, prefix=add_prefix("lm_head", prefix)
-        )
+        if self.config.tie_word_embeddings:
+            self.lm_head = self.transformer.wte
+        else:
+            self.lm_head = ParallelLMHead(
+                config.vocab_size,
+                config.hidden_size,
+                prefix=add_prefix("lm_head", prefix),
+            )
         self.logits_processor = LogitsProcessor(config)
 
     @torch.no_grad()

sglang/srt/models/gemma3_mm.py

@@ -21,7 +21,7 @@ from typing import Dict, Iterable, List, Optional, Set, Tuple, TypedDict
 
 import torch
 from torch import nn
-from transformers import AutoModel, Gemma3Config, PreTrainedModel
+from transformers import Gemma3Config, PreTrainedModel
 
 from sglang.srt.hf_transformers_utils import get_processor
 from sglang.srt.layers.layernorm import Gemma3RMSNorm
@@ -42,6 +42,7 @@ from sglang.srt.model_loader.weight_utils import (
     maybe_remap_kv_scale_name,
 )
 from sglang.srt.models.gemma3_causal import Gemma3ForCausalLM
+from sglang.srt.models.siglip import SiglipVisionModel
 from sglang.srt.utils import add_prefix
 
 logger = logging.getLogger(__name__)
@@ -118,6 +119,7 @@ class Gemma3ForConditionalGeneration(PreTrainedModel):
         ".k_proj.",
         ".v_proj.",
         ".o_proj.",
+        ".out_proj.",
     ]
     bitsandbytes_stacked_params_mapping = {
         # shard_name, weight_name, index
@@ -126,6 +128,7 @@ class Gemma3ForConditionalGeneration(PreTrainedModel):
         "v_proj": ("qkv_proj", 2),
         "gate_proj": ("gate_up_proj", 0),
         "up_proj": ("gate_up_proj", 1),
+        "out_proj": ("proj", 0),
     }
 
     packed_modules_mapping = {
@@ -161,20 +164,21 @@ class Gemma3ForConditionalGeneration(PreTrainedModel):
         super().__init__(config=config)
         self.config = config
         self.quant_config = quant_config
-        # Vision components
-        # TODO: replace with vision attention
-        # self.vision_tower = SiglipVisionModel(
-        #     config.vision_config,
-        #     quant_config,
-        #     prefix=add_prefix("vision_tower", prefix),
-        # )
-        self.vision_tower = AutoModel.from_config(config=config.vision_config)
+
+        self.vision_tower = SiglipVisionModel(
+            config=config.vision_config,
+            quant_config=quant_config,
+            prefix=add_prefix("vision_tower", prefix),
+        )
+
         self.multi_modal_projector = Gemma3MultiModalProjector(config)
         self.vocab_size = config.text_config.vocab_size
 
         # Text model
         self.language_model = Gemma3ForCausalLM(
-            config.text_config, quant_config, prefix=add_prefix("model", prefix)
+            config.text_config,
+            quant_config,
+            prefix=add_prefix("language_model", prefix),
         )
         if self.language_model.logits_processor.logit_scale:
             logit_scale = getattr(config, "logit_scale", 1.0)
@@ -278,13 +282,28 @@ class Gemma3ForConditionalGeneration(PreTrainedModel):
         Returns:
             image_features (`torch.Tensor`): Image feature tensor of shape `(num_images, image_length, embed_dim)`).
         """
-        pixel_values = torch.stack(
-            flatten_nested_list([item.pixel_values for item in items]), dim=0
-        )
-        pixel_values = pixel_values.to(device=self.vision_tower.device)
-        pixel_values = pixel_values.to(dtype=self.language_model.dtype())
+        if any(item.precomputed_features is not None for item in items):
+            if not all(item.precomputed_features is not None for item in items):
+                raise NotImplementedError(
+                    "MM inputs where only some items are precomputed."
+                )
+            return torch.concat([item.precomputed_features for item in items])
 
-        vision_outputs = self.vision_tower(pixel_values=pixel_values).last_hidden_state
+        # Process images one by one to handle flatten_batch=True constraint in vision_tower
+        all_pixel_values = flatten_nested_list([item.pixel_values for item in items])
+        vision_outputs_list = []
+
+        for pixel_value in all_pixel_values:
+            # Add batch dimension for single image processing
+            pixel_value_batch = pixel_value.unsqueeze(0)
+            pixel_value_batch = pixel_value_batch.to(device=self.vision_tower.device)
+            pixel_value_batch = pixel_value_batch.to(dtype=self.language_model.dtype())
+
+            vision_output = self.vision_tower(pixel_values=pixel_value_batch)
+            vision_outputs_list.append(vision_output)
+
+        # Concatenate all vision outputs
+        vision_outputs = torch.cat(vision_outputs_list, dim=0)
         image_features = self.multi_modal_projector(vision_outputs)
         return image_features
 
@@ -360,6 +379,14 @@ class Gemma3ForConditionalGeneration(PreTrainedModel):
         return self.language_model.tie_weights()
 
     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", "up_proj", 1),
+            ("gate_up_proj", "gate_proj", 0),
+        ]
         """Load weights for the model."""
         params_dict = dict(self.named_parameters())
         loaded_params: Set[str] = set()
@@ -373,21 +400,33 @@ class Gemma3ForConditionalGeneration(PreTrainedModel):
                 loaded_params.update(causal_loaded_params)
                 continue
             else:
-                # Skip lm_head.weight as it's tied with embed_tokens
-                if "lm_head.weight" in name:
-                    continue
-
-                # Skip loading extra bias for GPTQ models
-                if name.endswith(".bias") and name not in params_dict:
-                    continue
-
-                # Remapping the name of FP8 kv-scale
-                name = maybe_remap_kv_scale_name(name, params_dict)
-                if name is None:
-                    continue
-                param = params_dict[name]
-                weight_loader = getattr(param, "weight_loader", default_weight_loader)
-                weight_loader(param, loaded_weight)
+                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:
+                    if "vision_model" in name:
+                        # adapt to VisionAttention
+                        name = name.replace(".self_attn.out_proj", ".self_attn.proj")
+                    # Skip loading extra bias for GPTQ models
+                    if name.endswith(".bias") and name not in params_dict:
+                        continue
+                    # Remapping the name of FP8 kv-scale
+                    name = maybe_remap_kv_scale_name(name, params_dict)
+                    if name is None:
+                        continue
+                    param = params_dict[name]
+                    weight_loader = getattr(
+                        param, "weight_loader", default_weight_loader
+                    )
+                    weight_loader(param, loaded_weight)
                 loaded_params.add(name)
         unloaded_params = params_dict.keys() - loaded_params
         if unloaded_params:
@@ -398,5 +437,3 @@ class Gemma3ForConditionalGeneration(PreTrainedModel):
 
 
 EntryClass = Gemma3ForConditionalGeneration
-
-AutoModel.register(Gemma3Config, Gemma3ForConditionalGeneration, exist_ok=True)

sglang/srt/models/llama.py

@@ -45,6 +45,7 @@ from sglang.srt.layers.vocab_parallel_embedding import (
     ParallelLMHead,
     VocabParallelEmbedding,
 )
+from sglang.srt.managers.schedule_batch import global_server_args_dict
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch, PPProxyTensors
 from sglang.srt.model_loader.weight_utils import (
     default_weight_loader,
@@ -420,6 +421,7 @@ class LlamaForCausalLM(nn.Module):
                 config.hidden_size,
                 quant_config=quant_config,
                 prefix=add_prefix("lm_head", prefix),
+                use_attn_tp_group=global_server_args_dict["enable_dp_lm_head"],
             )
         self.logits_processor = LogitsProcessor(config)
         self.pooler = Pooler(pooling_type=PoolingType.LAST, normalize=True)

sglang/srt/models/llama4.py

@@ -30,9 +30,9 @@ from sglang.srt.distributed import (
 from sglang.srt.layers.dp_attention import (
     dp_gather_partial,
     dp_scatter,
-    get_attention_dp_size,
     get_attention_tp_rank,
     get_attention_tp_size,
+    get_local_attention_dp_size,
 )
 from sglang.srt.layers.layernorm import RMSNorm
 from sglang.srt.layers.linear import (
@@ -52,7 +52,15 @@ from sglang.srt.model_executor.forward_batch_info import (
     PPProxyTensors,
 )
 from sglang.srt.models.llama import LlamaForCausalLM, LlamaMLP
-from sglang.srt.utils import add_prefix, fast_topk, get_compiler_backend, make_layers
+from sglang.srt.utils import (
+    add_prefix,
+    fast_topk,
+    get_compiler_backend,
+    is_cuda,
+    make_layers,
+)
+
+_is_cuda = is_cuda()
 
 logger = logging.getLogger(__name__)
 
@@ -131,7 +139,7 @@ class Llama4MoE(nn.Module):
         return out_aD
 
     def _forward_core(self, hidden_states, forward_mode: ForwardMode):
-        if hidden_states.shape[0] < 4:
+        if hidden_states.shape[0] < 4 and _is_cuda:
             return self._forward_core_shared_routed_overlap(hidden_states)
         else:
             return self._forward_core_normal(hidden_states)
@@ -198,7 +206,6 @@ class Llama4Attention(nn.Module):
         self.use_rope = int((layer_id + 1) % 4 != 0)
         self.use_qk_norm = config.use_qk_norm and self.use_rope
 
-        self.dp_size = get_attention_dp_size()
         attn_tp_rank = get_attention_tp_rank()
         attn_tp_size = get_attention_tp_size()
 
@@ -342,7 +349,7 @@ class Llama4DecoderLayer(nn.Module):
         rope_theta = config.rope_theta
         rope_scaling = config.rope_scaling
         max_position_embeddings = config.max_position_embeddings
-        self.dp_size = get_attention_dp_size()
+        self.local_dp_size = get_local_attention_dp_size()
         self.attn_tp_size = get_attention_tp_size()
         self.attn_tp_rank = get_attention_tp_rank()
 
@@ -405,7 +412,7 @@ class Llama4DecoderLayer(nn.Module):
         # Gather
         if get_tensor_model_parallel_world_size() > 1:
             # all gather and all reduce
-            if self.dp_size != 1:
+            if self.local_dp_size != 1:
                 if self.attn_tp_rank == 0:
                     hidden_states += residual
                 hidden_states, local_hidden_states = (
@@ -428,9 +435,9 @@ class Llama4DecoderLayer(nn.Module):
         # Fully Connected
         hidden_states = self.feed_forward(hidden_states, forward_batch)
 
-        # TODO(ch-wan): ues reduce-scatter in MLP to avoid this scatter
+        # TODO(ch-wan): use reduce-scatter in MLP to avoid this scatter
         # Scatter
-        if self.dp_size != 1:
+        if self.local_dp_size != 1:
             # important: forward batch.gathered_buffer is used both after scatter and after gather.
             # be careful about this!
             hidden_states, global_hidden_states = (

sglang/srt/models/llava.py

@@ -15,7 +15,8 @@
 
 import math
 import re
-from typing import Iterable, List, Optional, Tuple
+from functools import lru_cache
+from typing import Dict, Iterable, List, Optional, Tuple, Type, Union
 
 import numpy as np
 import torch
@@ -28,10 +29,18 @@ from transformers import (
     Qwen2Config,
     SiglipVisionModel,
 )
+from transformers.models.auto.modeling_auto import AutoModel, AutoModelForCausalLM
 from transformers.models.llava.modeling_llava import LlavaMultiModalProjector
 
+# leave till last and symbol only in case circular import
+import sglang.srt.models as sgl_models
 from sglang.srt.layers.quantization.base_config import QuantizationConfig
-from sglang.srt.managers.schedule_batch import Modality, MultimodalInputs
+from sglang.srt.managers.mm_utils import general_mm_embed_routine
+from sglang.srt.managers.schedule_batch import (
+    Modality,
+    MultimodalDataItem,
+    MultimodalInputs,
+)
 from sglang.srt.mm_utils import (
     get_anyres_image_grid_shape,
     unpad_image,
@@ -42,7 +51,7 @@ from sglang.srt.model_loader.weight_utils import default_weight_loader
 from sglang.srt.models.llama import LlamaForCausalLM
 from sglang.srt.models.mistral import MistralForCausalLM
 from sglang.srt.models.qwen2 import Qwen2ForCausalLM
-from sglang.srt.utils import add_prefix, flatten_nested_list
+from sglang.srt.utils import add_prefix, flatten_nested_list, logger
 
 
 class LlavaBaseForCausalLM(nn.Module):
@@ -114,10 +123,18 @@ class LlavaBaseForCausalLM(nn.Module):
         image_inputs.image_offsets = offset_list
         return input_ids
 
-    def encode_images(self, pixel_values: torch.Tensor) -> torch.Tensor:
+    def encode_images(
+        self, pixel_values: Union[torch.Tensor, List[torch.Tensor]]
+    ) -> torch.Tensor:
+        """
+        encode images by vision tower and multimodal projector
+        Args:
+            pixel_values: torch.Tensor or List[torch.Tensor]: each tensor for an input image
+        Returns:
+            torch.Tensor: encoded image features from the input image; if multiple, flattened by seq_len axis
+        """
         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:]
@@ -128,7 +145,6 @@ class LlavaBaseForCausalLM(nn.Module):
                 f"Unexpected select feature strategy: {self.config.vision_feature_select_strategy}"
             )
         image_features = self.multi_modal_projector(selected_image_feature)
-
         return image_features
 
     @torch.no_grad()
@@ -583,4 +599,239 @@ class LlavaMistralForCausalLM(LlavaBaseForCausalLM):
             )
 
 
-EntryClass = [LlavaLlamaForCausalLM, LlavaQwenForCausalLM, LlavaMistralForCausalLM]
+class LlavaForConditionalGeneration(LlavaBaseForCausalLM):
+    """
+    An adaptor class to enable support for multiple mmlm such as mistral-community/pixtral-12b
+    It follows the structure of (vision_tower, multi_modal_projector, language_model)
+
+    Once a model config is loaded, text_config and vision_config will be extracted, and
+    LlavaForConditionalGeneration will load the language_model and vision_tower models
+    according to config.
+    """
+
+    MULTIMODAL_PROJECTOR_TYPE = LlavaMultiModalProjector
+
+    @property
+    def dtype(self):
+        return self.torch_dtype
+
+    def pad_input_ids(self, input_ids: List[int], image_inputs: MultimodalInputs):
+        if hasattr(self.vision_tower, "pad_input_ids"):
+            return self.vision_tower.pad_input_ids(input_ids, image_inputs)
+        else:
+            return super().pad_input_ids(input_ids, image_inputs)
+
+    def _get_sgl_model_cls(self, config, auto_model_type: Type[AutoModel] = AutoModel):
+        """
+        Get the SGLang model implementation class according to config.
+
+        Args:
+            config: The config object of the model.
+            auto_model_type: The type of the auto model.
+
+        Returns:
+            The SGLang model implementation class.
+        """
+        config_cls_name = config.__class__.__name__
+        arch_name_mapping = self._config_cls_name_to_arch_name_mapping(auto_model_type)
+        if arch := arch_name_mapping.get(config_cls_name):
+            if isinstance(arch, tuple):
+                arch = arch[0]
+                logger.warning(
+                    f"Multiple {auto_model_type.__name__} models found for submodule config `{config_cls_name}`, defaulting to [0]: {arch.__name__}"
+                )
+            try:
+                return sgl_models.registry.ModelRegistry.resolve_model_cls(arch)[0]
+            except Exception as e:
+                raise ValueError(
+                    f"{auto_model_type.__name__} found a corresponding model `{arch}` for config class `{config_cls_name}`, but failed to load it from SGLang ModelRegistry. \n{e}"
+                )
+        else:
+            raise ValueError(
+                f"{auto_model_type.__name__} cannot find a corresponding model for config class `{config_cls_name}`"
+            )
+
+    @lru_cache
+    def _config_cls_name_to_arch_name_mapping(
+        self, auto_model_type: Type[AutoModel]
+    ) -> Dict[str, str]:
+        mapping = {}
+        for config_cls, archs in auto_model_type._model_mapping.items():
+            if isinstance(archs, tuple):
+                mapping[config_cls.__name__] = tuple(arch.__name__ for arch in archs)
+            else:
+                mapping[config_cls.__name__] = archs.__name__
+        return mapping
+
+    def __init__(
+        self,
+        config: LlavaConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+
+        assert hasattr(config, "text_config")
+        assert hasattr(config, "vision_config")
+        self.config = config
+        self.text_config = self.config.text_config
+        self.vision_config = self.config.vision_config
+        self.torch_dtype = getattr(self.config, "torch_dtype")
+
+        if not getattr(self.text_config, "torch_dtype"):
+            self.text_config.torch_dtype = self.torch_dtype
+        if not getattr(self.vision_config, "torch_dtype"):
+            self.vision_config.torch_dtype = self.torch_dtype
+
+        if not hasattr(self.config, "vocab_size"):
+            self.config.vocab_size = self.text_config.vocab_size
+        if not hasattr(self.config, "image_aspect_ratio"):
+            self.config.image_aspect_ratio = "anyres"
+        if not hasattr(self.config, "image_grid_pinpoints"):
+            # from transformers.models.llava_onevision.configuration_llava_onevision import LlavaOnevisionConfig
+            # self.config.image_grid_pinpoints = LlavaOnevisionConfig().image_grid_pinpoints
+            self.config.image_grid_pinpoints = [
+                [96, 96],
+                [224, 224],
+                [384, 384],
+                [512, 512],
+                [768, 768],
+                [1024, 1024],
+            ]
+        if not hasattr(self.config, "mm_patch_merge_type"):
+            self.config.mm_patch_merge_type = "flat"
+        if not hasattr(self.config, "image_token_index"):
+            self.config.image_token_index = 10
+        if not hasattr(self.config, "projector_hidden_act"):
+            self.config.projector_hidden_act = "gelu"
+
+        self.vision_feature_layer = getattr(self.config, "vision_feature_layer", -1)
+        self.vision_feature_select_strategy = getattr(
+            self.config, "vision_feature_select_strategy", "full"
+        )
+        self.image_size = self.vision_config.image_size
+        self.patch_size = self.vision_config.patch_size
+
+        self.mm_patch_merge_type = self.config.mm_patch_merge_type
+        self.image_aspect_ratio = self.config.image_aspect_ratio
+        self.image_grid_pinpoints = self.config.image_grid_pinpoints
+
+        self.image_feature_len = int((self.image_size // self.patch_size) ** 2)
+
+        self.multi_modal_projector = self.MULTIMODAL_PROJECTOR_TYPE(config)
+
+        language_model_cls = self._get_sgl_model_cls(
+            self.text_config, AutoModelForCausalLM
+        )
+        vision_model_cls = self._get_sgl_model_cls(self.vision_config, AutoModel)
+        self.language_model = language_model_cls(
+            self.text_config,
+            quant_config=quant_config,
+            prefix=add_prefix("language_model", prefix),
+        )
+        self.vision_tower = vision_model_cls(
+            self.vision_config,
+            quant_config=quant_config,
+            prefix=add_prefix("vision_tower", prefix),
+        )
+
+        if "unpad" in getattr(self.config, "mm_patch_merge_type", ""):
+            self.language_model.model.image_newline = nn.Parameter(
+                torch.empty(self.text_config.hidden_size, dtype=self.torch_dtype)
+            )
+
+    def get_image_feature(self, items: List[MultimodalDataItem]) -> torch.Tensor:
+        """Extract features from image inputs.
+
+        Args:
+            items: List of MultimodalDataItem objects containing image data
+                Note that an item can be either "image" or "multi-images"
+
+        Returns:
+            torch.Tensor: features from image inputs, concatenated
+        """
+        features = []
+        for item in items:
+            # in each item, we assume pixel_values is always batched
+            pixel_values, image_sizes = item.pixel_values, item.image_sizes
+            image_outputs = self.vision_tower(
+                pixel_values, image_sizes, output_hidden_states=True
+            )
+            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: {self.vision_feature_select_strategy}"
+                )
+            features.append(
+                self.multi_modal_projector(selected_image_feature.squeeze(0))
+            )
+        ret = torch.cat(features, dim=0)
+        return ret
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        forward_batch: ForwardBatch,
+        get_embedding: bool = False,
+    ):
+        hidden_states = general_mm_embed_routine(
+            input_ids=input_ids,
+            forward_batch=forward_batch,
+            get_embedding=get_embedding,
+            language_model=self.language_model,
+            image_data_embedding_func=self.get_image_feature,
+            placeholder_tokens=None,  # using mm_item.pad_value
+            positions=positions,
+        )
+
+        return hidden_states
+
+    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
+        """Load weights for LlavaForConditionalGeneration.
+
+        Unlike the base class implementation, this one doesn't need to handle
+        weight name remapping as the weights are already properly structured with
+        'language_model' and 'vision_tower' prefixes in the safetensors files.
+        """
+        if (
+            self.vision_feature_select_strategy == "patch"
+            or self.vision_feature_select_strategy == "full"
+        ):
+            pass
+        elif self.vision_feature_select_strategy == "cls_patch":
+            self.image_feature_len += 1
+        else:
+            raise ValueError(
+                f"Unexpected select feature: {self.vision_feature_select_strategy}"
+            )
+
+        # Create dictionaries for direct parameter loading
+        params_dict = dict(self.named_parameters())
+
+        # Load weights directly without remapping
+        for name, loaded_weight in weights:
+            for part in ("language_model", "vision_tower"):
+                if name.startswith(part):
+                    name = name[len(part + ".") :]
+                    getattr(self, part).load_weights([(name, loaded_weight)])
+                    break
+            else:
+                param = params_dict[name]
+                weight_loader = getattr(param, "weight_loader", default_weight_loader)
+                weight_loader(param, loaded_weight)
+
+
+EntryClass = [
+    LlavaLlamaForCausalLM,
+    LlavaQwenForCausalLM,
+    LlavaMistralForCausalLM,
+    LlavaForConditionalGeneration,
+]