sglang 0.5.2rc2__py3-none-any.whl → 0.5.3__py3-none-any.whl

sglang/srt/models/dots_ocr.py

@@ -0,0 +1,173 @@
+# coding=utf-8
+# Adapted from Qwen2.5-VL SGLang implementation
+
+import logging
+from typing import Iterable, List, Optional, Tuple
+
+import torch
+import torch.nn as nn
+from transformers.activations import ACT2FN
+
+from sglang.srt.configs import DotsOCRConfig
+from sglang.srt.layers.logits_processor import LogitsProcessor
+from sglang.srt.layers.quantization.base_config import QuantizationConfig
+from sglang.srt.layers.vocab_parallel_embedding import ParallelLMHead
+from sglang.srt.managers.mm_utils import (
+    MultiModalityDataPaddingPatternMultimodalTokens,
+    general_mm_embed_routine,
+)
+from sglang.srt.managers.schedule_batch import MultimodalDataItem, MultimodalInputs
+from sglang.srt.model_executor.forward_batch_info import ForwardBatch
+from sglang.srt.model_loader.weight_utils import default_weight_loader
+from sglang.srt.models.dots_vlm_vit import DotsVisionTransformer
+from sglang.srt.models.qwen2 import Qwen2ForCausalLM
+from sglang.srt.utils import add_prefix
+from sglang.srt.utils.hf_transformers_utils import get_processor
+
+logger = logging.getLogger(__name__)
+
+
+class DotsOCRForCausalLM(nn.Module):
+    def __init__(
+        self,
+        config: DotsOCRConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        self.config = config
+
+        # Initialize vision transformer
+        self.visual = DotsVisionTransformer(
+            config.vision_config,
+        )
+
+        # Initialize language model
+        self.model = Qwen2ForCausalLM(config, quant_config)
+
+        # Initialize LM head
+        if config.tie_word_embeddings:
+            self.lm_head = self.model.embed_tokens
+        else:
+            self.lm_head = ParallelLMHead(
+                config.vocab_size,
+                config.hidden_size,
+                quant_config=quant_config,
+                prefix=add_prefix("lm_head", prefix),
+            )
+
+        self.logits_processor = LogitsProcessor(config)
+
+    def pad_input_ids(self, input_ids: List[int], mm_inputs: MultimodalInputs):
+        pattern = MultiModalityDataPaddingPatternMultimodalTokens()
+        return pattern.pad_input_tokens(input_ids, mm_inputs)
+
+    def get_image_feature(self, items: List[MultimodalDataItem]) -> torch.Tensor:
+        # Extract pixel values and grid information (following reference pattern)
+        pixel_values = torch.cat([item.feature for item in items], dim=0).type(
+            self.visual.dtype
+        )
+        image_grid_thw = torch.concat(
+            [item.image_grid_thw for item in items], dim=0
+        ).to(self.visual.device)
+
+        # Add dimension checks like in reference code
+        assert pixel_values.dim() == 2, f"{pixel_values.dim()=}"
+        assert image_grid_thw.dim() == 2, f"{image_grid_thw.dim()=}"
+
+        # Process through vision tower
+        image_embeds = self.visual(pixel_values, image_grid_thw)
+
+        # Ensure consistent dtype for FlashInfer compatibility
+        # Force bfloat16 to match model's expected dtype
+        if hasattr(self.model, "embed_tokens"):
+            target_dtype = self.model.embed_tokens.weight.dtype
+            if image_embeds.dtype != target_dtype:
+                image_embeds = image_embeds.to(target_dtype)
+
+        return image_embeds
+
+    def _pad_vit_attn_dummy_heads(self, name: str, loaded_weight: torch.Tensor):
+        """pad attn qkv weights for dummy heads"""
+        num_dummy_heads = self.config.vision_config.num_dummy_heads
+        if num_dummy_heads == 0:
+            return loaded_weight
+        head_dim = self.config.vision_config.head_dim
+
+        if "attn.qkv_proj" in name:
+            wq, wk, wv = loaded_weight.chunk(3, dim=0)
+            if name.endswith(".weight"):
+                dummy_shape = [num_dummy_heads, head_dim, wq.shape[-1]]
+            elif name.endswith(".bias"):
+                dummy_shape = [num_dummy_heads, head_dim]
+            else:
+                raise RuntimeError(f"Unsupported weight with name={name}")
+            pad_func = lambda x: torch.cat(
+                [x.unflatten(0, (-1, head_dim)), x.new_zeros(dummy_shape)], dim=0
+            ).flatten(0, 1)
+            wq, wk, wv = pad_func(wq), pad_func(wk), pad_func(wv)
+            loaded_weight = torch.cat([wq, wk, wv], dim=0)
+        if "attn.proj.weight" in name:
+            padded_weight = loaded_weight.new_zeros(
+                loaded_weight.shape[0], head_dim * num_dummy_heads
+            )
+            loaded_weight = torch.cat([loaded_weight, padded_weight], dim=-1)
+        if "attn.q_norm.weight" in name or "attn.k_norm.weight" in name:
+            padded_weight = loaded_weight.new_zeros(head_dim * num_dummy_heads)
+            loaded_weight = torch.cat([loaded_weight, padded_weight], dim=0)
+        return loaded_weight
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        forward_batch: ForwardBatch,
+        **kwargs: object,
+    ) -> torch.Tensor:
+        hidden_states = general_mm_embed_routine(
+            input_ids=input_ids,
+            positions=positions,
+            forward_batch=forward_batch,
+            multimodal_model=self,
+            language_model=self.model,
+        )
+        return hidden_states
+
+    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
+        """Load weights for the model, separating vision and language weights"""
+        weights = list(weights)
+
+        # Separate vision tower weights and language model weights
+        vision_weights = []
+        language_weights = []
+
+        for name, loaded_weight in weights:
+            if name.startswith("vision_tower."):
+                vision_name = name.replace(r"attn.qkv.", r"attn.qkv_proj.")
+
+                vision_weights.append((vision_name, loaded_weight))
+            else:
+                # All other weights go to language model
+                language_weights.append((name, loaded_weight))
+
+        # Load vision tower weights
+        vision_state_dict = dict(vision_weights)
+        params_dict = dict(self.named_parameters(remove_duplicate=False))
+
+        for name, loaded_weight in vision_state_dict.items():
+            name = name.replace("vision_tower", "visual")
+            if name not in params_dict:
+                raise ValueError(f"Weight {name} not found in params_dict")
+            param = params_dict[name]
+            weight_loader = getattr(param, "weight_loader", default_weight_loader)
+            loaded_weight = self._pad_vit_attn_dummy_heads(name, loaded_weight)
+            weight_loader(param, loaded_weight)
+
+        if language_weights:
+            self.model.load_weights(language_weights)
+
+    def get_embed_and_head(self):
+        return self.model.embed_tokens.weight, self.lm_head.weight
+
+
+EntryClass = [DotsOCRForCausalLM]

sglang/srt/models/dots_vlm.py

@@ -0,0 +1,174 @@
+# Copyright 2025 The RedNote HiLab team.
+# Copyright 2025 The SGLang team.
+#
+# This code is based on the DeepseekVL2ForCausalLM and DotsVisionTransformer
+# implementation in this library.
+#
+# 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.
+"""Inference-only Dots-VL model compatible with HuggingFace weights."""
+
+from typing import Iterable, List, Optional, Tuple
+
+import torch
+from torch import nn
+
+from sglang.srt.configs.dots_vlm import DotsVLMConfig
+from sglang.srt.distributed import parallel_state
+from sglang.srt.layers.quantization.base_config import QuantizationConfig
+from sglang.srt.managers.mm_utils import (
+    MultiModalityDataPaddingPatternMultimodalTokens,
+    general_mm_embed_routine,
+)
+from sglang.srt.managers.schedule_batch import MultimodalDataItem, MultimodalInputs
+from sglang.srt.model_executor.forward_batch_info import ForwardBatch
+from sglang.srt.model_loader.weight_utils import default_weight_loader
+from sglang.srt.models.deepseek_v2 import DeepseekV2ForCausalLM
+
+from .dots_vlm_vit import DotsVisionTransformer
+
+
+class DotsVLMForCausalLM(nn.Module):
+    """DotsVLM model for sglang inference"""
+
+    def __init__(
+        self, config: DotsVLMConfig, quant_config: Optional[QuantizationConfig] = None
+    ) -> None:
+        super().__init__()
+
+        self.config = config
+        self.image_token_id = config.im_span_id
+        self.video_token_id = config.video_span_id
+
+        self.language_model = DeepseekV2ForCausalLM(
+            config.language_config, quant_config
+        )
+
+        # Initialize vision tower (matching transformers naming for weight compatibility)
+        self.vision_tower = DotsVisionTransformer(config.vision_config)
+
+    def _pad_vit_attn_dummy_heads(self, name: str, loaded_weight: torch.Tensor):
+        """pad attn qkv weights for dummy heads"""
+        num_dummy_heads = self.config.vision_config.num_dummy_heads
+        if num_dummy_heads == 0:
+            return loaded_weight
+        head_dim = self.config.vision_config.head_dim
+
+        if "attn.qkv_proj" in name:
+            wq, wk, wv = loaded_weight.chunk(3, dim=0)
+            if name.endswith(".weight"):
+                dummy_shape = [num_dummy_heads, head_dim, wq.shape[-1]]
+            elif name.endswith(".bias"):
+                dummy_shape = [num_dummy_heads, head_dim]
+            else:
+                raise RuntimeError(f"Unsupported weight with name={name}")
+            pad_func = lambda x: torch.cat(
+                [x.unflatten(0, (-1, head_dim)), x.new_zeros(dummy_shape)], dim=0
+            ).flatten(0, 1)
+            wq, wk, wv = pad_func(wq), pad_func(wk), pad_func(wv)
+            loaded_weight = torch.cat([wq, wk, wv], dim=0)
+        if "attn.proj.weight" in name:
+            padded_weight = loaded_weight.new_zeros(
+                loaded_weight.shape[0], head_dim * num_dummy_heads
+            )
+            loaded_weight = torch.cat([loaded_weight, padded_weight], dim=-1)
+        if "attn.q_norm.weight" in name or "attn.k_norm.weight" in name:
+            padded_weight = loaded_weight.new_zeros(head_dim * num_dummy_heads)
+            loaded_weight = torch.cat([loaded_weight, padded_weight], dim=0)
+        return loaded_weight
+
+    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
+        """Load weights for the model, separating vision and language weights"""
+        weights = list(weights)
+
+        # Separate vision tower weights and language model weights
+        vision_weights = []
+        language_weights = []
+
+        for name, loaded_weight in weights:
+            if name.startswith("vision_tower."):
+                vision_name = name.replace(r"attn.qkv.", r"attn.qkv_proj.")
+                vision_weights.append((vision_name, loaded_weight))
+            else:
+                # All other weights go to language model
+                language_weights.append((name, loaded_weight))
+
+        # Load vision tower weights
+        vision_state_dict = dict(vision_weights)
+        params_dict = dict(self.named_parameters(remove_duplicate=False))
+        for name, loaded_weight in vision_state_dict.items():
+            if name not in params_dict:
+                raise ValueError(f"Weight {name} not found in params_dict")
+            param = params_dict[name]
+            weight_loader = getattr(param, "weight_loader", default_weight_loader)
+            loaded_weight = self._pad_vit_attn_dummy_heads(name, loaded_weight)
+            weight_loader(param, loaded_weight)
+
+        # Load language model weights
+        if language_weights:
+            self.language_model.load_weights(language_weights)
+
+    @classmethod
+    def get_model_config_for_expert_location(cls, config):
+        return DeepseekV2ForCausalLM.get_model_config_for_expert_location(config)
+
+    def pad_input_ids(self, input_ids: List[int], mm_inputs: MultimodalInputs):
+        """Pad input_ids with multimodal tokens"""
+        # Get image token ID for padding pattern
+        pattern = MultiModalityDataPaddingPatternMultimodalTokens()
+        padded_input_ids = pattern.pad_input_tokens(input_ids, mm_inputs)
+        return padded_input_ids
+
+    def get_image_feature(self, items: List[MultimodalDataItem]) -> torch.Tensor:
+        # Extract pixel values and grid information (following reference pattern)
+        pixel_values = torch.cat([item.feature for item in items], dim=0).type(
+            self.vision_tower.dtype
+        )
+        image_grid_thw = torch.concat(
+            [item.image_grid_thw for item in items], dim=0
+        ).to(self.vision_tower.device)
+
+        # Add dimension checks like in reference code
+        assert pixel_values.dim() == 2, f"{pixel_values.dim()=}"
+        assert image_grid_thw.dim() == 2, f"{image_grid_thw.dim()=}"
+
+        # Process through vision tower
+        image_embeds = self.vision_tower(pixel_values, image_grid_thw)
+
+        # Ensure consistent dtype for FlashInfer compatibility
+        # Force bfloat16 to match model's expected dtype
+        if image_embeds.dtype != torch.bfloat16 and hasattr(
+            self.language_model.model, "embed_tokens"
+        ):
+            target_dtype = self.language_model.model.embed_tokens.weight.dtype
+            image_embeds = image_embeds.to(target_dtype)
+
+        return image_embeds
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        forward_batch: ForwardBatch,
+        **kwargs: object,
+    ) -> torch.Tensor:
+        hidden_states = general_mm_embed_routine(
+            input_ids=input_ids,
+            positions=positions,
+            forward_batch=forward_batch,
+            multimodal_model=self,
+            language_model=self.language_model,
+        )
+        return hidden_states
+
+
+EntryClass = [DotsVLMForCausalLM]

sglang/srt/models/dots_vlm_vit.py

@@ -0,0 +1,337 @@
+import logging
+from typing import Optional
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.utils.checkpoint
+from torch.nn import LayerNorm
+from transformers.modeling_utils import PreTrainedModel
+
+from sglang.srt.configs.dots_vlm import DotsVisionConfig
+from sglang.srt.distributed import parallel_state
+from sglang.srt.layers.attention.vision import VisionAttention
+from sglang.srt.layers.quantization import QuantizationConfig
+from sglang.srt.utils import add_prefix
+
+logger = logging.getLogger(__name__)
+
+
+class VisionRotaryEmbedding(nn.Module):
+    def __init__(self, dim: int, theta: float = 10000.0) -> None:
+        super().__init__()
+        inv_freq = 1.0 / (theta ** (torch.arange(0, dim, 2, dtype=torch.float) / dim))
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+
+    def forward(self, seqlen: int) -> torch.Tensor:
+        seq = torch.arange(
+            seqlen, device=self.inv_freq.device, dtype=self.inv_freq.dtype
+        )
+        freqs = torch.outer(seq, self.inv_freq)
+        return freqs
+
+
+class PatchMerger(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        context_dim: int,
+        spatial_merge_size: int = 2,
+        pre_norm="layernorm",
+        init_merger_std=None,
+        quant_config: Optional[QuantizationConfig] = None,
+    ) -> None:
+        super().__init__()
+        self.hidden_size = context_dim * (spatial_merge_size**2)
+        self.pre_norm = pre_norm
+        if self.pre_norm == "layernorm":
+            self.ln_q = LayerNorm(context_dim, eps=1e-6)
+        elif self.pre_norm == "rmsnorm":
+            self.ln_q = RMSNorm(context_dim, eps=1e-6)
+        else:
+            logger.warning(f"no norm in patch merger: {self.pre_norm}")
+
+        self.mlp = nn.Sequential(
+            nn.Linear(self.hidden_size, self.hidden_size),
+            nn.GELU(),
+            nn.Linear(self.hidden_size, dim),
+        )
+
+        if init_merger_std is not None:
+            nn.init.normal_(self.mlp[0].weight, mean=0.0, std=init_merger_std)
+            nn.init.zeros_(self.mlp[0].bias)
+            nn.init.normal_(self.mlp[2].weight, mean=0.0, std=init_merger_std)
+            nn.init.zeros_(self.mlp[2].bias)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        if self.pre_norm:
+            x = self.mlp(self.ln_q(x).view(-1, self.hidden_size))
+        else:
+            x = self.mlp(x.view(-1, self.hidden_size))
+        return x
+
+
+class RMSNorm(nn.Module):
+    def __init__(self, dim: int, eps: float = 1e-6):
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(dim))
+        self.eps = eps
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        output = self._norm(x.float()).type_as(x)
+        return output * self.weight
+
+    def extra_repr(self) -> str:
+        return f"{tuple(self.weight.shape)}, eps={self.eps}"
+
+    def _norm(self, x: torch.Tensor) -> torch.Tensor:
+        return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
+
+
+class DotsSwiGLUFFN(nn.Module):
+    def __init__(self, config, quant_config: Optional[QuantizationConfig] = None):
+        super().__init__()
+        hidden_features = config.intermediate_size
+        in_features = config.embed_dim
+        bias = config.use_bias
+
+        self.fc1 = nn.Linear(in_features, hidden_features, bias=bias)
+        self.fc2 = nn.Linear(hidden_features, in_features, bias=bias)
+        self.fc3 = nn.Linear(in_features, hidden_features, bias=bias)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = F.silu(self.fc1(x)) * self.fc3(x)
+        x = self.fc2(x)
+        return x
+
+
+class DotsPatchEmbed(nn.Module):
+    def __init__(self, config, quant_config: Optional[QuantizationConfig] = None):
+        super().__init__()
+        self.num_channels = config.num_channels
+        self.patch_size = config.patch_size
+        self.temporal_patch_size = config.temporal_patch_size
+        self.embed_dim = config.embed_dim
+        self.config = config
+        self.proj = nn.Conv2d(
+            config.num_channels,
+            config.embed_dim,
+            kernel_size=(config.patch_size, config.patch_size),
+            stride=(config.patch_size, config.patch_size),
+        )
+        self.norm = RMSNorm(config.embed_dim, eps=config.rms_norm_eps)
+
+    def forward(self, x: torch.Tensor, grid_thw=None) -> torch.Tensor:
+        x = x.view(
+            -1,
+            self.num_channels,
+            self.temporal_patch_size,
+            self.patch_size,
+            self.patch_size,
+        )[:, :, 0]
+        x = self.proj(x).view(-1, self.embed_dim)
+        x = self.norm(x)
+        return x
+
+
+class DotsViTPreprocessor(nn.Module):
+    def __init__(self, config, quant_config: Optional[QuantizationConfig] = None):
+        super().__init__()
+        self.patch_h = config.patch_size
+        self.patch_w = config.patch_size
+        self.embed_dim = config.embed_dim
+        self.config = config
+        self.patchifier = DotsPatchEmbed(config, quant_config)
+
+    def forward(self, x: torch.Tensor, grid_thw=None) -> torch.Tensor:
+        tokens = self.patchifier(x, grid_thw)
+        return tokens
+
+
+class DotsVisionBlock(nn.Module):
+    def __init__(
+        self,
+        config: DotsVisionConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+        attn_implementation: str = "flash_attention_2",
+    ):
+        super().__init__()
+        if attn_implementation == "flash_attention_2":
+            qkv_backend = "fa3"
+            softmax_in_single_precision = False
+        else:
+            raise RuntimeError("Unimplemented")
+        self.attn = VisionAttention(
+            embed_dim=config.embed_dim,
+            num_heads=config.num_attention_heads,
+            projection_size=config.embed_dim,
+            use_qkv_parallel=True,
+            qkv_backend=qkv_backend,
+            softmax_in_single_precision=softmax_in_single_precision,
+            flatten_batch=True,
+            quant_config=quant_config,
+            prefix=add_prefix("attn", prefix),
+            num_dummy_heads=config.num_dummy_heads,
+            qkv_bias=config.use_bias,
+            proj_bias=config.use_bias,
+        )
+        self.norm1 = RMSNorm(config.embed_dim, eps=config.rms_norm_eps)
+        self.mlp = DotsSwiGLUFFN(config, quant_config)
+        self.norm2 = RMSNorm(config.embed_dim, eps=config.rms_norm_eps)
+
+    def forward(self, hidden_states, cu_seqlens, rotary_pos_emb) -> torch.Tensor:
+        hidden_states = hidden_states + self.attn(
+            self.norm1(hidden_states),
+            cu_seqlens=cu_seqlens,
+            position_embeddings=rotary_pos_emb,
+        )
+        hidden_states = hidden_states + self.mlp(self.norm2(hidden_states))
+        return hidden_states
+
+
+class DotsVisionTransformer(PreTrainedModel):
+    def __init__(
+        self,
+        config: DotsVisionConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+    ) -> None:
+        super().__init__(config)
+        self.config = config
+        self._update_vision_config()
+        self.spatial_merge_size = config.spatial_merge_size
+
+        self.patch_embed = DotsViTPreprocessor(config, quant_config)
+        self._init_weights(self.patch_embed.patchifier.proj)
+
+        head_dim = config.embed_dim // config.num_attention_heads
+
+        self.rotary_pos_emb = VisionRotaryEmbedding(head_dim // 2)
+
+        _num_hidden_layers = config.num_hidden_layers
+        self.blocks = nn.ModuleList(
+            [
+                DotsVisionBlock(
+                    config, quant_config, f"blocks.{i}", config.attn_implementation
+                )
+                for i in range(_num_hidden_layers)
+            ]
+        )
+
+        if self.config.post_norm:
+            self.post_trunk_norm = RMSNorm(config.embed_dim, eps=config.rms_norm_eps)
+
+        self.merger = PatchMerger(
+            dim=config.hidden_size,
+            context_dim=config.embed_dim,
+            spatial_merge_size=config.spatial_merge_size,
+            init_merger_std=self.config.init_merger_std,
+            quant_config=quant_config,
+        )
+
+        self.gradient_checkpointing = False
+
+    def _update_vision_config(self):
+        """update vision config to support tp"""
+        world_size = parallel_state.get_tensor_model_parallel_world_size()
+        num_heads = self.config.num_attention_heads
+        head_dim = self.config.embed_dim // num_heads
+        num_dummy_heads = 0
+
+        if num_heads % world_size != 0:
+            num_dummy_heads = (
+                (num_heads + world_size) // world_size
+            ) * world_size - num_heads
+
+        setattr(self.config, "head_dim", head_dim)
+        setattr(self.config, "num_dummy_heads", num_dummy_heads)
+
+    def _init_weights(self, module):
+        std = self.config.initializer_range
+        if isinstance(module, (nn.Linear, nn.Conv2d)):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+
+    @property
+    def dtype(self) -> torch.dtype:
+        return self.blocks[0].mlp.fc2.weight.dtype
+
+    @property
+    def device(self) -> torch.device:
+        return self.blocks[0].mlp.fc2.weight.device
+
+    def get_pos_ids_by_grid(self, grid_thw):
+        pos_ids = []
+        for t, h, w in grid_thw:
+            hpos_ids = torch.arange(h).unsqueeze(1).expand(-1, w)
+            hpos_ids = hpos_ids.reshape(
+                h // self.spatial_merge_size,
+                self.spatial_merge_size,
+                w // self.spatial_merge_size,
+                self.spatial_merge_size,
+            )
+            hpos_ids = hpos_ids.permute(0, 2, 1, 3)
+            hpos_ids = hpos_ids.flatten()
+
+            wpos_ids = torch.arange(w).unsqueeze(0).expand(h, -1)
+            wpos_ids = wpos_ids.reshape(
+                h // self.spatial_merge_size,
+                self.spatial_merge_size,
+                w // self.spatial_merge_size,
+                self.spatial_merge_size,
+            )
+            wpos_ids = wpos_ids.permute(0, 2, 1, 3)
+            wpos_ids = wpos_ids.flatten()
+            pos_ids.append(torch.stack([hpos_ids, wpos_ids], dim=-1).repeat(t, 1))
+
+        return pos_ids
+
+    def rot_pos_emb(self, grid_thw):
+        pos_ids = self.get_pos_ids_by_grid(grid_thw)
+        pos_ids = torch.cat(pos_ids, dim=0)
+        max_grid_size = grid_thw[:, 1:].max()
+        rotary_pos_emb_full = self.rotary_pos_emb(max_grid_size)
+        rotary_pos_emb = rotary_pos_emb_full[pos_ids].flatten(1)
+        return rotary_pos_emb
+
+    def calc_cos_sin(self, rotary_pos_emb):
+        cos = rotary_pos_emb.cos()
+        sin = rotary_pos_emb.sin()
+        cos = cos.unsqueeze(1).repeat(1, 1, 2).unsqueeze(0).float()
+        sin = sin.unsqueeze(1).repeat(1, 1, 2).unsqueeze(0).float()
+        rotary_pos_emb = (cos, sin)
+        return rotary_pos_emb
+
+    def forward(
+        self, hidden_states: torch.Tensor, grid_thw: torch.Tensor, bf16=True
+    ) -> torch.Tensor:
+        if bf16:
+            hidden_states = hidden_states.bfloat16()
+        hidden_states = self.patch_embed(hidden_states, grid_thw)
+
+        rotary_pos_emb = self.rot_pos_emb(grid_thw)
+        rotary_pos_emb = self.calc_cos_sin(rotary_pos_emb)
+
+        cu_seqlens = torch.repeat_interleave(
+            grid_thw[:, 1] * grid_thw[:, 2], grid_thw[:, 0]
+        ).cumsum(
+            dim=0,
+            dtype=grid_thw.dtype if torch.jit.is_tracing() else torch.int32,
+        )
+        cu_seqlens = F.pad(cu_seqlens, (1, 0), value=0)
+
+        for blk in self.blocks:
+            hidden_states = blk(
+                hidden_states, cu_seqlens=cu_seqlens, rotary_pos_emb=rotary_pos_emb
+            )
+
+        if self.config.post_norm:
+            hidden_states = self.post_trunk_norm(hidden_states)
+
+        hidden_states = self.merger(hidden_states)
+        return hidden_states

sglang/srt/models/ernie4.py

@@ -92,7 +92,7 @@ class Ernie4Moe(nn.Module):
             correction_bias=self.gate.e_score_correction_bias,
         )
 
-        self.experts = get_moe_impl_class()(
+        self.experts = get_moe_impl_class(quant_config)(
             num_experts=config.moe_num_experts,
             top_k=config.moe_k,
             hidden_size=config.hidden_size,