sglang 0.4.6.post4__py3-none-any.whl → 0.4.7__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_causal.py

@@ -277,6 +277,13 @@ class Gemma3Attention(nn.Module):
         k = k.permute(0, 2, 1, 3)
 
         attn_output = self.attn(q, k, v, forward_batch=forward_batch)
+
+        # Compatible with triton backend which returns [1, s, h, head_dim]
+        if attn_output.dim() == 4 and attn_output.shape[0] == 1:
+            attn_output = attn_output.squeeze(0)
+            attn_output = attn_output.flatten(-2, -1)
+        # [s, h * head_dim]
+
         output, _ = self.o_proj(attn_output)
         return output
 

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 +282,33 @@ 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())
+        # 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_values_batch in all_pixel_values:
+            # Normalize input shape to [batch_size, channels, height, width]
+            if pixel_values_batch.dim() == 5:
+                pixel_values_batch = pixel_values_batch.squeeze(0)
+            elif pixel_values_batch.dim() == 3:
+                pixel_values_batch = pixel_values_batch.unsqueeze(0)
+            elif pixel_values_batch.dim() != 4:
+                raise ValueError(
+                    f"Unexpected pixel_values shape: {pixel_values_batch.shape}"
+                )
+
+            # Process each image in the batch
+            batch_size = pixel_values_batch.shape[0]
+            for i in range(batch_size):
+                pixel_value = pixel_values_batch[i : i + 1]  # Keep batch dimension as 1
+                pixel_value = pixel_value.to(
+                    device=self.vision_tower.device, dtype=self.language_model.dtype()
+                )
+                vision_output = self.vision_tower(pixel_values=pixel_value)
+                vision_outputs_list.append(vision_output)
 
-        vision_outputs = self.vision_tower(pixel_values=pixel_values).last_hidden_state
+        # 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 +384,14 @@ 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 +405,33 @@ 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 +442,3 @@ class Gemma3ForConditionalGeneration(PreTrainedModel):
 
 
 EntryClass = Gemma3ForConditionalGeneration
-
-AutoModel.register(Gemma3Config, Gemma3ForConditionalGeneration, exist_ok=True)

sglang/srt/models/idefics2.py

@@ -0,0 +1,342 @@
+# Copyright 2023 The SGLang team.
+# Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved.
+#
+# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
+# and OPT implementations in this library. It has been modified from its
+# original forms to accommodate minor architectural differences compared
+# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Optional
+
+import torch
+from torch import nn
+from transformers import PretrainedConfig
+
+from sglang.srt.layers.activation import get_act_fn
+from sglang.srt.layers.attention.vision import VisionAttention
+from sglang.srt.layers.linear import ColumnParallelLinear, RowParallelLinear
+from sglang.srt.layers.quantization.base_config import QuantizationConfig
+from sglang.srt.utils import add_prefix
+
+
+class Idefics2VisionMLP(nn.Module):
+
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        self.config = config
+        self.activation_fn = get_act_fn(config.hidden_act)
+        self.fc1 = ColumnParallelLinear(
+            config.hidden_size,
+            config.intermediate_size,
+            bias=True,
+            quant_config=quant_config,
+            prefix=add_prefix("fc1", prefix),
+        )
+        self.fc2 = RowParallelLinear(
+            config.intermediate_size,
+            config.hidden_size,
+            bias=True,
+            quant_config=quant_config,
+            prefix=add_prefix("fc2", prefix),
+        )
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states, _ = self.fc1(hidden_states)
+        hidden_states = self.activation_fn(hidden_states)
+        hidden_states, _ = self.fc2(hidden_states)
+        return hidden_states
+
+
+class Idefics2EncoderLayer(nn.Module):
+
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        self.embed_dim = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.self_attn = VisionAttention(
+            embed_dim=config.hidden_size,
+            num_heads=self.num_heads,
+            projection_size=config.intermediate_size,
+            use_qkv_parallel=True,
+            quant_config=quant_config,
+            dropout=config.attention_dropout,
+            qkv_backend="sdpa",
+            softmax_in_single_precision=True,
+            flatten_batch=False,
+            prefix=add_prefix("self_attn", prefix),
+        )
+        self.layer_norm1 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
+        self.mlp = Idefics2VisionMLP(
+            config,
+            quant_config=quant_config,
+            prefix=add_prefix("mlp", prefix),
+        )
+        self.layer_norm2 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        cu_seqlens: torch.Tensor,
+    ) -> torch.Tensor:
+        """
+        Args:
+            hidden_states (`torch.FloatTensor`):
+                Input to the layer of shape `(batch, seq_len, embed_dim)`.
+
+        """
+        residual = hidden_states
+        hidden_states = self.layer_norm1(hidden_states)
+        hidden_states = self.self_attn(hidden_states, cu_seqlens=cu_seqlens)
+
+        hidden_states = residual + hidden_states
+        residual = hidden_states
+        hidden_states = self.layer_norm2(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+        return hidden_states
+
+
+class Idefics2Encoder(nn.Module):
+    """
+    Transformer encoder consisting of `config.num_hidden_layers` self attention
+    layers. Each layer is a
+    [`Idefics2EncoderLayer`].
+
+    Args:
+        config: Idefics2Config
+    """
+
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+
+        self.config = config
+        self.layers = nn.ModuleList(
+            [
+                Idefics2EncoderLayer(
+                    config,
+                    quant_config=quant_config,
+                    prefix=add_prefix(f"layers.{i}", prefix),
+                )
+                for i in range(config.num_hidden_layers)
+            ]
+        )
+
+    def forward(
+        self,
+        inputs_embeds: torch.Tensor,
+        cu_seqlens: torch.Tensor,
+    ) -> torch.Tensor:
+        r"""
+        Args:
+            inputs_embeds (torch.Tensor):
+                Optionally, instead of passing `input_ids` you can choose to
+                directly pass an embedded representation.
+                This is useful if you want more control over how to convert
+                `input_ids` indices into associated vectorsthan the model's
+                internal embedding lookup matrix.
+        """
+        hidden_states = inputs_embeds
+        for encoder_layer in self.layers:
+            layer_outputs = encoder_layer(
+                hidden_states,
+                cu_seqlens=cu_seqlens,
+            )
+            hidden_states = layer_outputs
+        return hidden_states
+
+
+class Idefics2VisionEmbeddings(nn.Module):
+    """
+    This is a modified version of `siglip.modelign_siglip.SiglipVisionEmbeddings
+    ` to enable images of variable
+    resolution.
+
+    The modifications are adapted from [Patch n' Pack: NaViT, a Vision
+    Transformer for any Aspect Ratio and Resolution](https://arxiv.org/abs/2307.06304)
+    which allows treating images in their native aspect ratio and without the
+    need to resize them to the same fixed size. In particular, we start from the
+    original pre-trained SigLIP model(which uses images of fixed-size square
+    images) and adapt it by training on images of variable resolutions.
+    """
+
+    def __init__(self, config: PretrainedConfig):
+        super().__init__()
+        self.embed_dim = config.hidden_size
+        self.image_size = config.image_size
+        self.patch_size = config.patch_size
+        self.patch_embedding = nn.Conv2d(
+            in_channels=config.num_channels,
+            out_channels=self.embed_dim,
+            kernel_size=self.patch_size,
+            stride=self.patch_size,
+            padding="valid",
+        )
+        self.num_patches_per_side = self.image_size // self.patch_size
+        self.num_patches = self.num_patches_per_side**2
+        self.num_positions = self.num_patches
+        self.position_embedding = nn.Embedding(self.num_positions, self.embed_dim)
+
+    def get_position_ids(
+        self,
+        pixel_values: torch.FloatTensor,
+        patch_attention_mask: torch.BoolTensor,
+        tgt_sizes: Optional[torch.IntTensor] = None,
+    ):
+        batch_size, _, max_im_h, max_im_w = pixel_values.shape
+
+        max_nb_patches_h, max_nb_patches_w = (
+            max_im_h // self.patch_size,
+            max_im_w // self.patch_size,
+        )
+        boundaries = torch.arange(
+            1 / self.num_patches_per_side, 1.0, 1 / self.num_patches_per_side
+        )
+        position_ids = torch.full(
+            size=(batch_size, max_nb_patches_h * max_nb_patches_w), fill_value=0
+        )
+
+        for batch_idx, p_attn_mask in enumerate(patch_attention_mask):
+
+            if tgt_sizes is not None:
+                nb_patches_h = tgt_sizes[batch_idx][0]
+                nb_patches_w = tgt_sizes[batch_idx][1]
+            else:
+                nb_patches_h = p_attn_mask[:, 0].sum()
+                nb_patches_w = p_attn_mask[0].sum()
+            fractional_coords_h = torch.arange(0, 1 - 1e-6, 1 / nb_patches_h)
+            fractional_coords_w = torch.arange(0, 1 - 1e-6, 1 / nb_patches_w)
+            bucket_coords_h = torch.bucketize(
+                fractional_coords_h, boundaries, right=True
+            )
+            bucket_coords_w = torch.bucketize(
+                fractional_coords_w, boundaries, right=True
+            )
+            pos_ids = (
+                bucket_coords_h[:, None] * self.num_patches_per_side + bucket_coords_w
+            ).flatten()
+            position_ids[batch_idx][p_attn_mask.view(-1).cpu()] = pos_ids
+        position_ids = position_ids.to(self.position_embedding.weight.device)
+        return position_ids
+
+    def forward(
+        self,
+        pixel_values: torch.FloatTensor,
+        patch_attention_mask: torch.BoolTensor,
+        tgt_sizes: Optional[torch.IntTensor] = None,
+    ) -> torch.Tensor:
+        target_dtype = self.patch_embedding.weight.dtype
+        pixel_values = pixel_values.to(
+            device=self.patch_embedding.weight.device, dtype=target_dtype
+        )
+        patch_embeds = self.patch_embedding(pixel_values)
+        embeddings = patch_embeds.flatten(2).transpose(1, 2)
+        position_ids = self.get_position_ids(
+            pixel_values, patch_attention_mask, tgt_sizes
+        )
+
+        embeddings = embeddings + self.position_embedding(position_ids)
+        return embeddings
+
+
+class Idefics2VisionTransformer(nn.Module):
+
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+        require_post_norm: bool = True,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+
+        embed_dim = config.hidden_size
+        self.config = config
+        self.embeddings = Idefics2VisionEmbeddings(config)
+        self.encoder = Idefics2Encoder(
+            config=config,
+            quant_config=quant_config,
+            prefix=add_prefix("encoder", prefix),
+        )
+        self.post_layernorm = (
+            nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)
+            if require_post_norm
+            else nn.Identity()
+        )
+
+    def get_input_embeddings(self) -> nn.Embedding:
+        return self.embeddings
+
+    def compute_cu_seqlens(
+        self,
+        tgt_sizes: Optional[torch.Tensor] = None,
+        input_embeds: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        # shape: (batch_size,)
+        if tgt_sizes is not None:
+            seqlen = tgt_sizes[:, 0] * tgt_sizes[:, 1]
+        elif input_embeds is not None:
+            seqlen = torch.full(
+                size=(input_embeds.shape[0],),
+                fill_value=input_embeds.shape[1],
+                dtype=torch.int32,
+                device=input_embeds.device,
+            )
+        else:
+            raise ValueError(
+                "Either `tgt_sizes` or `input_embeds` must be provided to compute cu_seqlens."
+            )
+
+        cu_seqlens = torch.cat(
+            [
+                torch.tensor([0], device=seqlen.device, dtype=torch.int32),
+                torch.cumsum(seqlen, dim=0, dtype=torch.int32),
+            ],
+            dim=0,
+        ).to(seqlen.device)
+        return cu_seqlens
+
+    def forward(
+        self,
+        pixel_values,
+        patch_attention_mask: Optional[torch.BoolTensor] = None,
+        tgt_sizes: Optional[torch.IntTensor] = None,
+    ) -> torch.Tensor:
+        hidden_states = self.embeddings(
+            pixel_values=pixel_values,
+            patch_attention_mask=patch_attention_mask,
+            tgt_sizes=tgt_sizes,
+        )
+        cu_seqlens = self.compute_cu_seqlens(tgt_sizes, hidden_states)
+        encoder_outputs = self.encoder(
+            hidden_states,
+            cu_seqlens=cu_seqlens,
+        )
+        last_hidden_state = self.post_layernorm(encoder_outputs)
+        return last_hidden_state

sglang/srt/models/kimi_vl.py

@@ -144,10 +144,10 @@ class KimiVLForConditionalGeneration(nn.Module):
             .type(self.vision_tower.dtype)
             .to(self.vision_tower.device)
         )
-        image_grid_thws = torch.concat(
-            [item.image_grid_thws for item in items], dim=0
-        ).to(self.vision_tower.device)
-        image_features = self.vision_tower(pixel_values, image_grid_thws)
+        image_grid_hws = torch.cat([item.image_grid_hws for item in items], dim=0).to(
+            self.vision_tower.device
+        )
+        image_features = self.vision_tower(pixel_values, image_grid_hws)
         assert isinstance(image_features, list)
         # lengths = [x.shape[0] for x in image_features]
         res = self.multi_modal_projector(torch.cat(image_features))  # .split(lengths)

sglang/srt/models/llama.py

@@ -17,7 +17,7 @@
 """Inference-only LLaMA model compatible with HuggingFace weights."""
 
 import logging
-from typing import Any, Dict, Iterable, List, Optional, Set, Tuple, Union
+from typing import Any, Dict, Iterable, List, Optional, Tuple, Union
 
 import torch
 from torch import nn

sglang/srt/models/llama4.py

@@ -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)