lightseekorg--tokenspeed/python/tokenspeed/runtime/models/qwen3_moe.py

# Copyright (c) 2026 LightSeek Foundation
#
# Permission is hereby granted, free of charge, to any person obtaining a copy
# of this software and associated documentation files (the "Software"), to deal
# in the Software without restriction, including without limitation the rights
# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
# copies of the Software, and to permit persons to whom the Software is
# furnished to do so, subject to the following conditions:
#
# The above copyright notice and this permission notice shall be included in
# all copies or substantial portions of the Software.
#
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
# SOFTWARE.

"""Inference-only Qwen3 MoE model compatible with HuggingFace weights."""

from __future__ import annotations

from collections.abc import Iterable

import torch

from tokenspeed.runtime.configs.qwen3_moe_config import Qwen3MoeConfig
from tokenspeed.runtime.distributed.comm_manager import CommManager
from tokenspeed.runtime.distributed.mapping import Mapping
from tokenspeed.runtime.execution.context import ForwardContext
from tokenspeed.runtime.layers.moe import (
    ExpertCheckpointSchema,
    build_moe_checkpoint_loader,
)
from tokenspeed.runtime.layers.quantization.base_config import QuantizationConfig
from tokenspeed.runtime.layers.utils import get_layer_id
from tokenspeed.runtime.model_loader.weight_utils import default_weight_loader
from tokenspeed.runtime.models.qwen3 import (
    Qwen3DecoderLayer,
    Qwen3ForCausalLM,
    Qwen3MLP,
    Qwen3Model,
)
from tokenspeed.runtime.models.qwen3_5_moe import (
    Qwen3_5MoeMLP,
    Qwen3_5MoeSparseMoeBlock,
    _is_moe_layer,
)
from tokenspeed.runtime.utils import add_prefix


class Qwen3MoeDecoderLayer(Qwen3DecoderLayer):
    def __init__(
        self,
        config: Qwen3MoeConfig,
        mapping: Mapping,
        layer_id: int = 0,
        quant_config: QuantizationConfig | None = None,
        prefix: str = "",
    ) -> None:
        super().__init__(
            config=config,
            mapping=mapping,
            layer_id=layer_id,
            quant_config=quant_config,
            prefix=prefix,
        )
        if _is_moe_layer(layer_id, config):
            self.mlp = Qwen3_5MoeSparseMoeBlock(
                config=config,
                mapping=self.mapping,
                quant_config=quant_config,
                layer_index=layer_id,
                prefix=add_prefix("mlp", prefix),
            )
        elif isinstance(self.mlp, Qwen3MLP):
            self.mlp = Qwen3_5MoeMLP(
                hidden_size=config.hidden_size,
                intermediate_size=config.intermediate_size,
                hidden_act=config.hidden_act,
                mapping=self.mapping,
                quant_config=quant_config,
                reduce_results=False,
                prefix=add_prefix("mlp", prefix),
            )
        is_moe = isinstance(self.mlp, Qwen3_5MoeSparseMoeBlock)
        self.comm_manager = CommManager(
            mapping=self.mapping,
            layer_id=layer_id,
            is_moe=is_moe,
            prev_is_moe=_is_moe_layer(layer_id - 1, config),
            input_layernorm=self.input_layernorm,
            post_attn_layernorm=self.post_attention_layernorm,
        )

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        ctx: ForwardContext,
        out_cache_loc: torch.Tensor,
        residual: torch.Tensor | None,
        cos_sin: tuple[torch.Tensor, torch.Tensor] | None,
    ) -> tuple[torch.Tensor, torch.Tensor]:
        if not ctx.forward_mode.is_idle():
            hidden_states, residual = self.comm_manager.input_reduce_norm(
                hidden_states, residual
            )
            hidden_states = self.comm_manager.pre_attn_comm(hidden_states, ctx)
            hidden_states = self.self_attn(
                positions=positions,
                hidden_states=hidden_states,
                ctx=ctx,
                out_cache_loc=out_cache_loc,
                cos_sin=cos_sin,
            )
            hidden_states, residual = self.comm_manager.post_attn_reduce_norm(
                hidden_states, residual, ctx
            )

        hidden_states, residual = self.forward_mlp(hidden_states, residual, ctx)
        return hidden_states, residual

    def forward_mlp(
        self,
        hidden_states: torch.Tensor,
        residual: torch.Tensor,
        ctx: ForwardContext,
    ) -> tuple[torch.Tensor, torch.Tensor]:
        if isinstance(self.mlp, Qwen3_5MoeSparseMoeBlock):
            num_global_tokens, max_num_tokens_per_gpu = (
                self.mlp.comm_manager.get_num_tokens(ctx)
            )
            hidden_states = self.mlp(
                hidden_states,
                num_global_tokens,
                max_num_tokens_per_gpu,
                ctx,
            )
            return hidden_states, residual

        hidden_states = self.comm_manager.pre_mlp_comm(hidden_states, ctx)
        hidden_states = self.mlp(hidden_states)
        hidden_states, residual = self.comm_manager.post_mlp_fused(
            hidden_states, residual, ctx
        )
        return hidden_states, residual


class Qwen3MoeModel(Qwen3Model):
    def __init__(
        self,
        config: Qwen3MoeConfig,
        mapping: Mapping,
        quant_config: QuantizationConfig | None = None,
        prefix: str = "",
    ) -> None:
        super().__init__(
            config=config,
            mapping=mapping,
            quant_config=quant_config,
            prefix=prefix,
            decoder_layer_type=Qwen3MoeDecoderLayer,
        )

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        ctx: ForwardContext,
        out_cache_loc: torch.Tensor,
        input_embeds: torch.Tensor | None = None,
    ) -> tuple[torch.Tensor, None]:
        if input_embeds is None:
            hidden_states = self.embed_tokens(input_ids)
        else:
            hidden_states = input_embeds
        residual = None

        for layer in self.layers:
            hidden_states, residual = layer(
                positions,
                hidden_states,
                ctx,
                out_cache_loc,
                residual,
                cos_sin=None,
            )

        if not ctx.forward_mode.is_idle():
            hidden_states, _ = layer.comm_manager.final_norm(
                hidden_states, residual, ctx, self.norm
            )
        return hidden_states, None


class Qwen3MoeForCausalLM(Qwen3ForCausalLM):
    model_cls = Qwen3MoeModel

    default_bitsandbytes_target_modules = [
        ".gate_proj.",
        ".down_proj.",
        ".up_proj.",
        ".q_proj.",
        ".k_proj.",
        ".v_proj.",
        ".o_proj.",
    ]
    bitsandbytes_stacked_params_mapping = {
        "q_proj": ("qkv_proj", 0),
        "k_proj": ("qkv_proj", 1),
        "v_proj": ("qkv_proj", 2),
        "gate_proj": ("gate_up_proj", 0),
        "up_proj": ("gate_up_proj", 1),
    }

    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]):
        stacked_params_mapping = [
            ("qkv_proj", "q_proj", "q"),
            ("qkv_proj", "k_proj", "k"),
            ("qkv_proj", "v_proj", "v"),
            ("gate_up_proj", "gate_proj", 0),
            ("gate_up_proj", "up_proj", 1),
        ]
        ignore_suffixes = (
            ".bias",
            "_bias",
            ".k_scale",
            "_k_scale",
            ".v_scale",
            "_v_scale",
            ".weight_scale",
            "_weight_scale",
            ".input_scale",
            "_input_scale",
        )

        params_dict = dict(self.named_parameters(remove_duplicate=False))
        moe_loader = build_moe_checkpoint_loader(
            params_dict=params_dict,
            expert_schema=ExpertCheckpointSchema(
                gate_proj_name="gate_proj",
                down_proj_name="down_proj",
                up_proj_name="up_proj",
            ),
            fused_schema=ExpertCheckpointSchema(
                gate_up_fused_name="gate_up_proj",
                down_proj_name="down_proj",
            ),
            num_experts=self.config.num_experts,
            ep_rank=self.mapping.moe.ep_rank,
            ep_size=self.mapping.moe.ep_size,
        )

        for name, loaded_weight in weights:
            if "Embedding" in self.config.name_or_path:
                name = add_prefix(name, "model")
            layer_id = get_layer_id(name)
            if (
                layer_id is not None
                and hasattr(self.model, "start_layer")
                and (
                    layer_id < self.model.start_layer
                    or layer_id >= self.model.end_layer
                )
            ):
                continue
            if "rotary_emb.inv_freq" in name or "projector" in name:
                continue
            if "rotary_emb.cos_cached" in name or "rotary_emb.sin_cached" in name:
                continue
            if self.config.tie_word_embeddings and "lm_head.weight" in name:
                continue
            if name.startswith("model.vision_tower") and name not in params_dict:
                continue

            for param_name, weight_name, shard_id in stacked_params_mapping:
                if weight_name not in name:
                    continue
                if "mlp.experts" in name:
                    continue
                name = name.replace(weight_name, param_name)
                if name.endswith(ignore_suffixes) and name not in params_dict:
                    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:
                if name.endswith((".bias", "_bias")) and name not in params_dict:
                    continue
                if moe_loader.matches(name):
                    moe_loader.load(name, loaded_weight)
                    continue
                if name.endswith(ignore_suffixes) and name not in params_dict:
                    continue
                if name not in params_dict:
                    continue
                param = params_dict[name]
                weight_loader = getattr(param, "weight_loader", default_weight_loader)
                weight_loader(param, loaded_weight)


EntryClass = Qwen3MoeForCausalLM