sgl-project--sglang/python/sglang/srt/models/qwen3_vl_moe.py

# Copyright 2025 Qwen Team
# Copyright 2025 SGLang Team
# 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 Qwen3-VL model compatible with HuggingFace weights."""

import logging
import re
from functools import lru_cache
from typing import Iterable, Optional, Tuple, Union

import torch
import torch.nn as nn

from sglang.srt.configs.qwen3_vl import Qwen3VLMoeConfig, Qwen3VLMoeTextConfig
from sglang.srt.eplb.expert_location import ModelConfigForExpertLocation
from sglang.srt.layers.moe.fused_moe_triton.layer import FusedMoE
from sglang.srt.layers.quantization.base_config import QuantizationConfig
from sglang.srt.layers.utils import get_layer_id
from sglang.srt.model_executor.forward_batch_info import ForwardBatch, PPProxyTensors
from sglang.srt.model_loader.weight_utils import default_weight_loader
from sglang.srt.models.qwen3_moe import Qwen3MoeDecoderLayer, Qwen3MoeModel
from sglang.srt.models.qwen3_vl import Qwen3VLForConditionalGeneration
from sglang.srt.utils.hf_transformers_utils import get_processor

logger = logging.getLogger(__name__)

cached_get_processor = lru_cache(get_processor)


class Qwen3MoeLLMModel(Qwen3MoeModel):
    def __init__(
        self,
        *,
        config: Qwen3VLMoeTextConfig,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
        decoder_layer_type=Qwen3MoeDecoderLayer,
    ):
        super().__init__(
            config=config,
            quant_config=quant_config,
            prefix=prefix,
            decoder_layer_type=decoder_layer_type,
        )
        self.hidden_size = config.hidden_size
        # Currently, we use 3 as len(config.vision_config.deepstack_visual_indexes) is not directly accessible here.
        # This approach follows the original implementation.
        # TODO: make config of type Qwen3VLMoeConfig, so that we can directly obtain deepstack_visual_indexes.
        self.deepstack_embed_to_decoder_layer = range(3)

    def get_input_embeddings(self) -> nn.Embedding:
        return self.embed_tokens

    def get_deepstack_embeds(
        self, layer_idx: int, input_deepstack_embeds: Optional[torch.Tensor]
    ) -> Optional[torch.Tensor]:
        """Get deepstack embeddings for a given layer index, or None if not applicable."""
        if (
            input_deepstack_embeds is None
            or layer_idx not in self.deepstack_embed_to_decoder_layer
        ):
            return None
        sep = self.hidden_size * layer_idx
        return input_deepstack_embeds[:, sep : sep + self.hidden_size]

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        forward_batch: ForwardBatch,
        input_embeds: torch.Tensor = None,
        pp_proxy_tensors: Optional[PPProxyTensors] = None,
        input_deepstack_embeds: Optional[torch.Tensor] = None,
    ) -> Union[torch.Tensor, PPProxyTensors]:
        if self.pp_group.is_first_rank:
            if input_embeds is None:
                hidden_states = self.embed_tokens(input_ids)
            else:
                hidden_states = input_embeds
            residual = None
        else:
            assert pp_proxy_tensors is not None
            hidden_states = pp_proxy_tensors["hidden_states"]
            residual = pp_proxy_tensors["residual"]

        aux_hidden_states = []
        for layer_idx, layer in enumerate(
            self.layers[self.start_layer : self.end_layer]
        ):
            layer_idx += self.start_layer
            if layer_idx in self.layers_to_capture:
                aux_hidden_states.append(
                    hidden_states + residual if residual is not None else hidden_states
                )

            # SGLang applies residual at the START of the next layer, not at the END like HuggingFace.
            # See: https://github.com/huggingface/transformers/blob/v5.0.0rc0/src/transformers/models/qwen3_vl/modeling_qwen3_vl.py#L549
            # To match HF behavior, deepstack must be added AFTER residual: (hidden_states + residual) + deepstack
            # The order matters because addition with different tensors is not associative in practice.
            # Deepstack for prev_layer is applied at the start of current layer via post_residual_addition.
            deepstack_embeds = self.get_deepstack_embeds(
                layer_idx - 1, input_deepstack_embeds
            )
            hidden_states, residual = layer(
                positions,
                hidden_states,
                forward_batch,
                residual,
                post_residual_addition=deepstack_embeds,
            )

        # Handle deepstack for the last processed layer if it exists.
        last_deepstack = self.get_deepstack_embeds(
            self.end_layer - 1, input_deepstack_embeds
        )

        if not self.pp_group.is_last_rank:
            return PPProxyTensors(
                {
                    "hidden_states": hidden_states,
                    "residual": residual,
                }
            )
        else:
            if hidden_states.shape[0] != 0:
                if residual is None:
                    hidden_states = self.norm(hidden_states)
                else:
                    hidden_states, _ = self.norm(
                        hidden_states, residual, post_residual_addition=last_deepstack
                    )

        if len(aux_hidden_states) == 0:
            return hidden_states

        return hidden_states, aux_hidden_states


def load_fused_expert_weights(
    name: str,
    params_dict: dict,
    loaded_weight: torch.Tensor,
    shard_id: str,
    num_experts: int,
):
    param = params_dict[name]
    # weight_loader = typing.cast(Callable[..., bool], param.weight_loader)
    weight_loader = param.weight_loader
    # let ep moe layer to gracefully handle expert_ids that do not belong to local moe rank
    for expert_id in range(num_experts):
        curr_expert_weight = loaded_weight[expert_id]
        weight_loader(
            param,
            curr_expert_weight,
            name,
            shard_id,
            expert_id,
        )
    return True


class Qwen3VLMoeForConditionalGeneration(Qwen3VLForConditionalGeneration):
    def __init__(
        self,
        config: Qwen3VLMoeConfig,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
        language_model_cls=Qwen3MoeLLMModel,
    ):
        super().__init__(config, quant_config, prefix, language_model_cls)

    _lora_pattern_moe = re.compile(
        r"^(?:model\.layers\.(\d+)\.(?:self_attn\.(?:qkv_proj|o_proj)|mlp\.experts)|lm_head|model\.embed_tokens)$"
    )

    def should_apply_lora(self, module_name: str) -> bool:
        return bool(self._lora_pattern_moe.match(module_name))

    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),
        ]

        expert_params_mapping = FusedMoE.make_expert_params_mapping(
            ckpt_gate_proj_name="gate_proj",
            ckpt_down_proj_name="down_proj",
            ckpt_up_proj_name="up_proj",
            num_experts=self.config.num_experts,
        )

        # Skip loading extra parameters for GPTQ/modelopt models.
        ignore_suffixes = (
            ".bias",
            "_bias",
            ".k_scale",
            "_k_scale",
            ".v_scale",
            "_v_scale",
            ".weight_scale",
            "_weight_scale",
            ".input_scale",
            "_input_scale",
        )

        is_fused_expert = False
        fused_expert_params_mapping = [
            ("experts.w13_weight", "experts.gate_up_proj", 0, "w1"),
            ("experts.w2_weight", "experts.down_proj", 0, "w2"),
        ]

        num_experts = self.config.num_experts

        # Pre-define `params_dict` to avoid repeated expensive traversal of model parameters.
        params_dict = dict(self.named_parameters())

        for name, loaded_weight in weights:
            name = name.replace(r"model.language_model.", r"model.")
            layer_id = get_layer_id(name)
            if (
                "visual" not in name
                and layer_id is not None
                and hasattr(self, "model")
                and hasattr(self.model, "start_layer")
                and (
                    layer_id < self.model.start_layer
                    or layer_id >= self.model.end_layer
                )
            ):
                continue

            for param_name, weight_name, shard_id in stacked_params_mapping:
                if "experts.gate_up_proj" in name or "experts.down_proj" in name:
                    is_fused_expert = True
                    expert_params_mapping = fused_expert_params_mapping

                # Skip non-stacked layers and experts (experts handled below).
                if weight_name not in name:
                    continue
                if "visual" in name:
                    continue

                # We have mlp.experts[0].gate_proj in the checkpoint.
                # Since we handle the experts below in expert_params_mapping,
                # we need to skip here BEFORE we update the name, otherwise
                # name will be updated to mlp.experts[0].gate_up_proj, which
                # will then be updated below in expert_params_mapping
                # for mlp.experts[0].gate_gate_up_proj, which breaks load.
                if "mlp.experts" in name:
                    continue
                name = name.replace(weight_name, param_name)
                # Skip loading extra parameters for GPTQ/modelopt models.
                if name.endswith(ignore_suffixes) and name not in params_dict:
                    continue
                # [TODO] Skip layers that are on other devices (check if sglang has a similar function)
                # if is_pp_missing_parameter(name, self):
                #     continue

                if name not in params_dict:
                    continue

                param = params_dict[name]
                weight_loader = param.weight_loader
                weight_loader(param, loaded_weight, shard_id)
                break
            else:
                # Track if this is an expert weight to enable early skipping
                is_expert_weight = False

                for mapping in expert_params_mapping:
                    param_name, weight_name, expert_id, shard_id = mapping
                    if weight_name not in name:
                        continue
                    if "visual" in name or self.config.encoder_only:
                        continue
                    # Anyway, this is an expert weight and should not be
                    # attempted to load as other weights later
                    is_expert_weight = True
                    name_mapped = name.replace(weight_name, param_name)
                    if is_fused_expert:
                        loaded_weight = loaded_weight.transpose(-1, -2)  # no bias
                        if "experts.gate_up_proj" in name:
                            loaded_weight = loaded_weight.chunk(2, dim=-2)
                            load_fused_expert_weights(
                                name_mapped,
                                params_dict,
                                loaded_weight[0],
                                "w1",
                                num_experts,
                            )
                            load_fused_expert_weights(
                                name_mapped,
                                params_dict,
                                loaded_weight[1],
                                "w3",
                                num_experts,
                            )
                        else:
                            load_fused_expert_weights(
                                name_mapped,
                                params_dict,
                                loaded_weight,
                                shard_id,
                                num_experts,
                            )
                    else:
                        # Skip loading extra parameters for GPTQ/modelopt models.
                        if (
                            name_mapped.endswith(ignore_suffixes)
                            and name_mapped not in params_dict
                        ):
                            continue
                        param = params_dict[name_mapped]
                        # We should ask the weight loader to return success or
                        # not here since otherwise we may skip experts with
                        # # other available replicas.
                        weight_loader = param.weight_loader
                        weight_loader(
                            param,
                            loaded_weight,
                            name_mapped,
                            shard_id=shard_id,
                            expert_id=expert_id,
                        )
                    name = name_mapped
                    break
                else:
                    if is_expert_weight:
                        # This is an expert weight but not mapped to this rank, skip all remaining processing
                        continue
                    if "visual" in name:
                        # adapt to VisionAttention
                        name = name.replace(r"attn.qkv.", r"attn.qkv_proj.")
                        name = name.replace(r"model.visual.", r"visual.")

                    # Skip loading extra parameters for GPTQ/modelopt models.
                    if name.endswith(ignore_suffixes) and name not in params_dict:
                        continue

                    # Skip loading mm/language parameters
                    if (
                        self.config.encoder_only or self.config.language_only
                    ) and name not in params_dict:
                        continue

                    if name in params_dict.keys():
                        param = params_dict[name]
                        weight_loader = getattr(
                            param, "weight_loader", default_weight_loader
                        )
                        weight_loader(param, loaded_weight)
                    else:
                        logger.warning(f"Parameter {name} not found in params_dict")

        # TODO mimic deepseek
        # Lazy initialization of expert weights cache to avoid slowing down load_weights
        # if not hasattr(self, "routed_experts_weights_of_layer"):
        #     self.routed_experts_weights_of_layer = {
        #         layer_id: self.model.layers[layer_id].mlp.get_moe_weights()
        #         for layer_id in range(self.start_layer, self.end_layer)
        #         if isinstance(self.model.layers[layer_id].mlp, Qwen3MoeSparseMoeBlock)
        #     }

    @classmethod
    def get_model_config_for_expert_location(cls, config):
        return ModelConfigForExpertLocation(
            num_layers=config.text_config.num_hidden_layers,
            num_logical_experts=config.text_config.num_experts,
            num_groups=None,
        )


EntryClass = Qwen3VLMoeForConditionalGeneration