deepseek_v2.py

# Copyright 2023-2024 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.
# ==============================================================================

# Adapted from:
# https://github.com/vllm-project/vllm/blob/fb6af8bc086328ca6659e72d11ffd4309ce4de22/vllm/model_executor/models/deepseek_v2.py
"""Inference-only DeepseekV2 model."""

import concurrent.futures
import logging
import os
from enum import IntEnum, auto
from typing import Any, Dict, Iterable, Optional, Tuple

import torch
import torch.nn.functional as F
from torch import nn
from tqdm import tqdm
from transformers import PretrainedConfig

from sglang.srt.distributed import (
    get_tensor_model_parallel_world_size,
    parallel_state,
    tensor_model_parallel_all_reduce,
)
from sglang.srt.eplb.expert_distribution import get_global_expert_distribution_recorder
from sglang.srt.eplb.expert_location import ModelConfigForExpertLocation
from sglang.srt.eplb.expert_location_dispatch import ExpertLocationDispatchInfo
from sglang.srt.layers.activation import SiluAndMul
from sglang.srt.layers.amx_utils import PackWeightMethod
from sglang.srt.layers.communicator import (
    LayerCommunicator,
    LayerScatterModes,
    enable_moe_dense_fully_dp,
)
from sglang.srt.layers.dp_attention import (
    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 (
    ColumnParallelLinear,
    MergedColumnParallelLinear,
    ReplicatedLinear,
    RowParallelLinear,
)
from sglang.srt.layers.logits_processor import LogitsProcessor
from sglang.srt.layers.moe.ep_moe.layer import (
    DeepEPMoE,
    get_moe_impl_class,
    use_flashinfer_trtllm_moe,
)
from sglang.srt.layers.moe.ep_moe.token_dispatcher import DeepEPDispatcher
from sglang.srt.layers.moe.topk import TopK
from sglang.srt.layers.quantization import deep_gemm_wrapper
from sglang.srt.layers.quantization.base_config import QuantizationConfig
from sglang.srt.layers.quantization.fp8_kernel import (
    is_fp8_fnuz,
    per_tensor_quant_mla_fp8,
    per_token_group_quant_mla_deep_gemm_masked_fp8,
)
from sglang.srt.layers.quantization.fp8_utils import (
    block_quant_dequant,
    block_quant_to_tensor_quant,
    channel_quant_to_tensor_quant,
    normalize_e4m3fn_to_e4m3fnuz,
    requant_weight_ue8m0_inplace,
)
from sglang.srt.layers.quantization.int8_utils import (
    block_dequant as int8_block_dequant,
)
from sglang.srt.layers.radix_attention import RadixAttention
from sglang.srt.layers.rotary_embedding import get_rope, get_rope_wrapper
from sglang.srt.layers.utils import is_sm100_supported
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
from sglang.srt.model_loader.weight_utils import default_weight_loader
from sglang.srt.two_batch_overlap import (
    MaybeTboDeepEPDispatcher,
    model_forward_maybe_tbo,
)
from sglang.srt.utils import (
    BumpAllocator,
    DeepEPMode,
    LazyValue,
    add_prefix,
    bind_or_assign,
    cpu_has_amx_support,
    get_bool_env_var,
    get_device_sm,
    get_int_env_var,
    is_cpu,
    is_cuda,
    is_flashinfer_available,
    is_hip,
    is_non_idle_and_non_empty,
    log_info_on_rank0,
    use_intel_amx_backend,
)

_is_hip = is_hip()
_is_cuda = is_cuda()
_is_fp8_fnuz = is_fp8_fnuz()
_use_aiter = get_bool_env_var("SGLANG_USE_AITER") and _is_hip
_is_cpu_amx_available = cpu_has_amx_support()
_is_cpu = is_cpu()
_device_sm = get_device_sm()

if _is_cuda:
    from sgl_kernel import (
        awq_dequantize,
        bmm_fp8,
        dsv3_fused_a_gemm,
        dsv3_router_gemm,
        merge_state_v2,
    )
elif _is_cpu and _is_cpu_amx_available:
    pass
elif _is_hip:
    from sglang.srt.layers.quantization.awq_triton import (
        awq_dequantize_triton as awq_dequantize,
    )
else:
    from vllm._custom_ops import awq_dequantize

if _is_hip:
    from sglang.srt.layers.attention.triton_ops.rocm_mla_decode_rope import (
        decode_attention_fwd_grouped_rope,
    )

_is_flashinfer_available = is_flashinfer_available()
_is_sm100_supported = is_cuda() and is_sm100_supported()


logger = logging.getLogger(__name__)


class AttnForwardMethod(IntEnum):
    # Use multi-head attention
    MHA = auto()

    # Use absorbed multi-latent attention
    MLA = auto()

    # Use multi-head attention, but with KV cache chunked.
    # This method can avoid OOM when prefix lengths are long.
    MHA_CHUNKED_KV = auto()

    # Use MLA but with fused RoPE
    MLA_FUSED_ROPE = auto()

    # Use MLA with fused RoPE kernel for CPU
    MLA_FUSED_ROPE_CPU = auto()


class DeepseekV2MLP(nn.Module):
    def __init__(
        self,
        hidden_size: int,
        intermediate_size: int,
        hidden_act: str,
        quant_config: Optional[QuantizationConfig] = None,
        reduce_results: bool = True,
        prefix: str = "",
        tp_rank: Optional[int] = None,
        tp_size: Optional[int] = None,
    ) -> None:
        super().__init__()
        self.tp_size = tp_size

        self.gate_up_proj = MergedColumnParallelLinear(
            hidden_size,
            [intermediate_size] * 2,
            bias=False,
            quant_config=quant_config,
            prefix=add_prefix("gate_up_proj", prefix),
            tp_rank=tp_rank,
            tp_size=tp_size,
        )
        self.down_proj = RowParallelLinear(
            intermediate_size,
            hidden_size,
            bias=False,
            quant_config=quant_config,
            reduce_results=reduce_results,
            prefix=add_prefix("down_proj", prefix),
            tp_rank=tp_rank,
            tp_size=tp_size,
        )
        if hidden_act != "silu":
            raise ValueError(
                f"Unsupported activation: {hidden_act}. "
                "Only silu is supported for now."
            )
        self.act_fn = SiluAndMul()

    def forward(self, x, forward_batch=None, can_fuse_mlp_allreduce=False):
        if (self.tp_size == 1) and x.shape[0] == 0:
            return x

        gate_up, _ = self.gate_up_proj(x)
        x = self.act_fn(gate_up)
        x, _ = self.down_proj(x, can_fuse_mlp_allreduce=can_fuse_mlp_allreduce)
        return x


class MoEGate(nn.Module):
    def __init__(
        self,
        config,
        prefix: str = "",
        is_nextn: bool = False,
    ):
        super().__init__()
        self.is_nextn = is_nextn
        self.weight = nn.Parameter(
            torch.empty((config.n_routed_experts, config.hidden_size))
        )
        if config.topk_method == "noaux_tc":
            self.e_score_correction_bias = nn.Parameter(
                torch.empty((config.n_routed_experts), dtype=torch.float32)
            )
        else:
            self.e_score_correction_bias = None
        if _is_cpu and _is_cpu_amx_available:
            self.quant_method = PackWeightMethod(weight_names=["weight"])

    def forward(self, hidden_states):
        if use_intel_amx_backend(self):
            return torch.ops.sgl_kernel.weight_packed_linear(
                hidden_states,
                self.weight,
                None,  # bias
                True,  # is_vnni
            )

        # NOTE: For some unknown reason, router_gemm seems degrade accept length.
        if (
            _is_cuda
            and not self.is_nextn
            and hidden_states.shape[0] < 4
            and hidden_states.shape[1] == 7168
            and self.weight.shape[0] == 256
            and _device_sm >= 90
        ):
            # router gemm output float32
            logits = dsv3_router_gemm(hidden_states, self.weight)
        else:
            logits = F.linear(hidden_states, self.weight, None)

        return logits


class DeepseekV2MoE(nn.Module):

    def __init__(
        self,
        config: PretrainedConfig,
        layer_id: int,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
        alt_stream: Optional[torch.cuda.Stream] = None,
        is_nextn: bool = False,
    ):
        super().__init__()
        self.tp_size = get_tensor_model_parallel_world_size()
        self.routed_scaling_factor = config.routed_scaling_factor
        self.n_shared_experts = config.n_shared_experts
        self.num_fused_shared_experts = (
            0
            if global_server_args_dict["disable_shared_experts_fusion"]
            else config.n_shared_experts
        )
        self.config = config
        self.layer_id = layer_id
        self.alt_stream = alt_stream

        if self.tp_size > config.n_routed_experts:
            raise ValueError(
                f"Tensor parallel size {self.tp_size} is greater than "
                f"the number of experts {config.n_routed_experts}."
            )

        if config.hidden_act != "silu":
            raise ValueError(
                f"Unsupported activation: {config.hidden_act}. "
                "Only silu is supported for now."
            )

        self.gate = MoEGate(
            config=config, prefix=add_prefix("gate", prefix), is_nextn=is_nextn
        )

        self.topk = (
            TopK(
                top_k=config.num_experts_per_tok + self.num_fused_shared_experts,
                renormalize=config.norm_topk_prob,
                use_grouped_topk=True,
                num_expert_group=config.n_group,
                num_fused_shared_experts=self.num_fused_shared_experts,
                topk_group=config.topk_group,
                correction_bias=self.gate.e_score_correction_bias,
                routed_scaling_factor=self.routed_scaling_factor,
            )
            if not use_flashinfer_trtllm_moe
            else None
        )

        self.experts = get_moe_impl_class()(
            num_experts=config.n_routed_experts
            + self.num_fused_shared_experts
            + global_server_args_dict["ep_num_redundant_experts"],
            top_k=config.num_experts_per_tok + self.num_fused_shared_experts,
            hidden_size=config.hidden_size,
            intermediate_size=config.moe_intermediate_size,
            layer_id=self.layer_id,
            quant_config=quant_config,
            routed_scaling_factor=self.routed_scaling_factor,
            prefix=add_prefix("experts", prefix),
            **(
                dict(deepep_mode=DeepEPMode[global_server_args_dict["deepep_mode"]])
                if global_server_args_dict["enable_deepep_moe"]
                else {}
            ),
            # Additional args for FusedMoE
            **(
                dict(
                    enable_flashinfer_cutlass_moe=True,
                    enable_ep_moe=global_server_args_dict["enable_ep_moe"],
                )
                if global_server_args_dict["enable_flashinfer_cutlass_moe"]
                else {}
            ),
            **(
                dict(
                    renormalize=config.norm_topk_prob,
                    use_grouped_topk=True,
                    num_expert_group=config.n_group,
                    num_fused_shared_experts=self.num_fused_shared_experts,
                    topk_group=config.topk_group,
                    correction_bias=self.gate.e_score_correction_bias,
                )
                if use_flashinfer_trtllm_moe
                else {}
            ),
        )

        self.shared_experts_is_int8 = False
        self.shared_experts_is_fp8 = False
        self.shared_experts_weight_block_size = None
        if config.n_shared_experts is not None and self.num_fused_shared_experts == 0:
            intermediate_size = config.moe_intermediate_size * config.n_shared_experts
            # disable tp for shared experts when enable deepep moe
            self.shared_experts = DeepseekV2MLP(
                hidden_size=config.hidden_size,
                intermediate_size=intermediate_size,
                hidden_act=config.hidden_act,
                quant_config=quant_config,
                reduce_results=False,
                prefix=add_prefix("shared_experts", prefix),
                **(
                    dict(tp_rank=0, tp_size=1)
                    if global_server_args_dict["enable_deepep_moe"]
                    else {}
                ),
            )
            is_packed_weight = hasattr(
                self.shared_experts.gate_up_proj.quant_method, "quant_config"
            ) and self.shared_experts.gate_up_proj.quant_method.quant_config.get_name() in {
                "awq",
                "awq_marlin",
                "moe_wna16",
            }
            self.shared_experts_is_int8 = (
                not is_packed_weight
                and self.shared_experts.gate_up_proj.weight.dtype == torch.int8
            )
            self.shared_experts_is_fp8 = (
                not is_packed_weight
                and self.shared_experts.gate_up_proj.weight.dtype == torch.float8_e4m3fn
            )
            if self.shared_experts_is_fp8:
                assert (
                    self.shared_experts.gate_up_proj.quant_method.quant_config.weight_block_size
                    == self.shared_experts.down_proj.quant_method.quant_config.weight_block_size
                )
                self.shared_experts_weight_block_size = (
                    self.shared_experts.gate_up_proj.quant_method.quant_config.weight_block_size
                )

        self.top_k = config.num_experts_per_tok

        if global_server_args_dict["enable_deepep_moe"]:
            # TODO: we will support tp < ep in the future
            self.ep_size = get_tensor_model_parallel_world_size()
            self.num_experts = (
                config.n_routed_experts
                + global_server_args_dict["ep_num_redundant_experts"]
            )
            self.renormalize = config.norm_topk_prob
            self.topk_group = config.topk_group
            self.num_expert_group = config.n_group
            self.correction_bias = (
                self.gate.e_score_correction_bias.data
                if self.gate.e_score_correction_bias is not None
                else None
            )

            self.deepep_dispatcher = MaybeTboDeepEPDispatcher(
                group=parallel_state.get_tp_group().device_group,
                router_topk=self.top_k,
                permute_fusion=True,
                num_experts=self.num_experts,
                num_local_experts=config.n_routed_experts // self.tp_size,
                hidden_size=config.hidden_size,
                params_dtype=config.torch_dtype,
                deepep_mode=DeepEPMode[global_server_args_dict["deepep_mode"]],
                async_finish=True,
                return_recv_hook=True,
            )

        self._enable_deepep_moe = global_server_args_dict["enable_deepep_moe"]

    def get_moe_weights(self):
        return [
            x.data
            for name, x in self.experts.named_parameters()
            if name not in ["correction_bias"]
        ]

    def forward(
        self,
        hidden_states: torch.Tensor,
        forward_batch: Optional[ForwardBatch] = None,
        can_fuse_mlp_allreduce: bool = False,
    ) -> torch.Tensor:
        if not self._enable_deepep_moe:
            DUAL_STREAM_TOKEN_THRESHOLD = 1024
            if (
                self.alt_stream is not None
                and self.num_fused_shared_experts == 0
                and hidden_states.shape[0] <= DUAL_STREAM_TOKEN_THRESHOLD
            ):
                return self.forward_normal_dual_stream(
                    hidden_states, can_fuse_mlp_allreduce
                )
            else:
                return self.forward_normal(hidden_states, can_fuse_mlp_allreduce)
        else:
            return self.forward_deepep(hidden_states, forward_batch)

    def forward_normal_dual_stream(
        self, hidden_states: torch.Tensor, can_fuse_mlp_allreduce: bool = False
    ) -> torch.Tensor:

        current_stream = torch.cuda.current_stream()
        self.alt_stream.wait_stream(current_stream)
        shared_output = self._forward_shared_experts(hidden_states)

        with torch.cuda.stream(self.alt_stream):
            # router_logits: (num_tokens, n_experts)
            router_logits = self.gate(hidden_states)
            kwargs = {"hidden_states": hidden_states}
            if self.topk is not None:
                kwargs["topk_output"] = self.topk(hidden_states, router_logits)
            else:
                kwargs["router_logits"] = router_logits
            final_hidden_states = self.experts(**kwargs)
            if not _is_cuda:
                final_hidden_states *= self.routed_scaling_factor
        current_stream.wait_stream(self.alt_stream)
        final_hidden_states += shared_output
        if self.tp_size > 1 and not can_fuse_mlp_allreduce:
            final_hidden_states = tensor_model_parallel_all_reduce(final_hidden_states)
        return final_hidden_states

    def forward_normal(
        self, hidden_states: torch.Tensor, can_fuse_mlp_allreduce: bool = False
    ) -> torch.Tensor:
        if hasattr(self, "shared_experts") and use_intel_amx_backend(
            self.shared_experts.gate_up_proj
        ):
            return self.forward_cpu(hidden_states, can_fuse_mlp_allreduce)

        shared_output = self._forward_shared_experts(hidden_states)
        # router_logits: (num_tokens, n_experts)
        router_logits = self.gate(hidden_states)
        kwargs = {"hidden_states": hidden_states}
        if self.topk is not None:
            kwargs["topk_output"] = self.topk(hidden_states, router_logits)
        else:
            kwargs["router_logits"] = router_logits
        final_hidden_states = self.experts(**kwargs)
        if not _is_cuda and not _use_aiter:
            # fused in biased_grouped_topk so we can skip here
            final_hidden_states *= self.routed_scaling_factor
        if shared_output is not None:
            final_hidden_states = final_hidden_states + shared_output
        if self.tp_size > 1 and not can_fuse_mlp_allreduce:
            final_hidden_states = tensor_model_parallel_all_reduce(final_hidden_states)
        return final_hidden_states

    def forward_cpu(
        self, hidden_states: torch.Tensor, can_fuse_mlp_allreduce: bool = False
    ) -> torch.Tensor:
        # router_logits: (num_tokens, n_experts)
        router_logits = self.gate(hidden_states)
        topk_output = self.topk(hidden_states, router_logits)
        fused_experts_out = self.experts(
            hidden_states=hidden_states, topk_output=topk_output
        )

        assert use_intel_amx_backend(
            self.shared_experts.gate_up_proj
        ) == use_intel_amx_backend(self.shared_experts.down_proj)
        # [Note] inplace should be False in fused_experts.
        # If inplace is True in fused_experts (self.experts), hidden_states will be changed after fused_experts
        # While hidden_states is still needed in shared_expert.
        final_hidden_states = torch.ops.sgl_kernel.shared_expert_cpu(
            hidden_states,
            self.shared_experts.gate_up_proj.weight,
            self.shared_experts.down_proj.weight,
            fused_experts_out,
            self.routed_scaling_factor,
            True,  # inplace
            self.shared_experts_is_int8,  # use_int8_w8a8
            self.shared_experts_is_fp8,  # use_fp8_w8a16
            (
                self.shared_experts.gate_up_proj.weight_scale
                if self.shared_experts_is_int8
                else (
                    self.shared_experts.gate_up_proj.weight_scale_inv
                    if self.shared_experts_is_fp8
                    else None
                )
            ),  # w1_scale
            (
                self.shared_experts.down_proj.weight_scale
                if self.shared_experts_is_int8
                else (
                    self.shared_experts.down_proj.weight_scale_inv
                    if self.shared_experts_is_fp8
                    else None
                )
            ),  # w2_scale
            (
                self.shared_experts_weight_block_size
                if self.shared_experts_is_fp8
                else None
            ),  # block_size
            None,  # a1_scale
            None,  # a2_scale
            True,  # is_vnni
        )
        if self.tp_size > 1 and not can_fuse_mlp_allreduce:
            final_hidden_states = tensor_model_parallel_all_reduce(final_hidden_states)
        return final_hidden_states

    def forward_deepep(
        self, hidden_states: torch.Tensor, forward_batch: ForwardBatch
    ) -> torch.Tensor:
        shared_output = None
        if hidden_states.shape[0] > 0:
            # router_logits: (num_tokens, n_experts)
            router_logits = self.gate(hidden_states)
            shared_output = self._forward_shared_experts(hidden_states)
            topk_weights, topk_idx, _ = self.topk(
                hidden_states,
                router_logits,
                num_token_non_padded=forward_batch.num_token_non_padded,
                expert_location_dispatch_info=ExpertLocationDispatchInfo.init_new(
                    layer_id=self.layer_id,
                ),
            )
        else:
            topk_idx = torch.full(
                (0, self.top_k), -1, dtype=torch.int, device=hidden_states.device
            )
            topk_weights = torch.empty(
                (0, self.top_k), dtype=torch.float32, device=hidden_states.device
            )
        if self.ep_size > 1:
            # TODO(ch-wan): allow users to set num_max_dispatch_tokens_per_rank value
            (
                hidden_states,
                topk_idx,
                topk_weights,
                reorder_topk_ids,
                num_recv_tokens_per_expert,
                seg_indptr,
                masked_m,
                expected_m,
            ) = self.deepep_dispatcher.dispatch(
                hidden_states=hidden_states,
                topk_idx=topk_idx,
                topk_weights=topk_weights,
                forward_batch=forward_batch,
            )
        final_hidden_states = self.experts(
            hidden_states=hidden_states,
            topk_idx=topk_idx,
            topk_weights=topk_weights,
            reorder_topk_ids=reorder_topk_ids,
            seg_indptr=seg_indptr,
            masked_m=masked_m,
            expected_m=expected_m,
            num_recv_tokens_per_expert=num_recv_tokens_per_expert,
            forward_batch=forward_batch,
        )
        if self.ep_size > 1:
            final_hidden_states = self.deepep_dispatcher.combine(
                hidden_states=final_hidden_states,
                topk_idx=topk_idx,
                topk_weights=topk_weights,
                forward_batch=forward_batch,
            )

        if shared_output is not None:
            x = shared_output
            x.add_(final_hidden_states, alpha=self.routed_scaling_factor)
            final_hidden_states = x
        else:
            final_hidden_states *= self.routed_scaling_factor

        return final_hidden_states

    def _forward_shared_experts(self, hidden_states):
        if self.num_fused_shared_experts == 0:
            return self.shared_experts(hidden_states)
        else:
            return None

    def op_gate(self, state):
        if is_non_idle_and_non_empty(
            state.forward_batch.forward_mode, state.hidden_states_mlp_input
        ):
            # router_logits: (num_tokens, n_experts)
            state.router_logits = self.gate(state.hidden_states_mlp_input)
        else:
            state.router_logits = None

    def op_shared_experts(self, state):
        hidden_states_mlp_input = state.pop("hidden_states_mlp_input")
        if (self.num_fused_shared_experts == 0) and is_non_idle_and_non_empty(
            state.forward_batch.forward_mode, hidden_states_mlp_input
        ):
            state.shared_output = self.shared_experts(hidden_states_mlp_input)
        else:
            state.shared_output = None

    def op_select_experts(self, state):
        router_logits = state.pop("router_logits")
        hidden_states = state.hidden_states_mlp_input

        if router_logits is not None:
            with get_global_expert_distribution_recorder().with_current_layer(
                self.layer_id
            ):
                state.topk_weights_local, state.topk_idx_local, _ = self.topk(
                    hidden_states=hidden_states,
                    router_logits=router_logits,
                    num_token_non_padded=state.forward_batch.num_token_non_padded,
                    expert_location_dispatch_info=ExpertLocationDispatchInfo.init_new(
                        layer_id=self.layer_id,
                    ),
                )
        else:
            state.topk_idx_local = torch.full(
                (0, self.top_k), -1, dtype=torch.int, device=hidden_states.device
            )
            state.topk_weights_local = torch.empty(
                (0, self.top_k), dtype=torch.float32, device=hidden_states.device
            )

    def op_dispatch_a(self, state):
        if self.ep_size > 1:
            # TODO(ch-wan): allow users to set num_max_dispatch_tokens_per_rank value
            self.deepep_dispatcher.dispatch_a(
                hidden_states=state.hidden_states_mlp_input,
                topk_idx=state.pop("topk_idx_local"),
                topk_weights=state.pop("topk_weights_local"),
                forward_batch=state.forward_batch,
                tbo_subbatch_index=state.get("tbo_subbatch_index"),
            )

    def op_dispatch_b(self, state):
        if self.ep_size > 1:
            with get_global_expert_distribution_recorder().with_current_layer(
                self.layer_id
            ):
                (
                    state.hidden_states_experts_input,
                    state.topk_idx_dispatched,
                    state.topk_weights_dispatched,
                    state.reorder_topk_ids,
                    state.num_recv_tokens_per_expert,
                    state.seg_indptr,
                    state.masked_m,
                    state.expected_m,
                ) = self.deepep_dispatcher.dispatch_b(
                    tbo_subbatch_index=state.get("tbo_subbatch_index"),
                )

    def op_experts(self, state):
        state.hidden_states_experts_output = self.experts(
            hidden_states=state.pop("hidden_states_experts_input"),
            topk_idx=state.topk_idx_dispatched,
            topk_weights=state.topk_weights_dispatched,
            reorder_topk_ids=state.pop("reorder_topk_ids"),
            seg_indptr=state.pop("seg_indptr"),
            masked_m=state.pop("masked_m"),
            expected_m=state.pop("expected_m"),
            num_recv_tokens_per_expert=state.pop("num_recv_tokens_per_expert"),
            forward_batch=state.forward_batch,
        )

    def op_combine_a(self, state):
        if self.ep_size > 1:
            self.deepep_dispatcher.combine_a(
                hidden_states=state.pop("hidden_states_experts_output"),
                topk_idx=state.pop("topk_idx_dispatched"),
                topk_weights=state.pop("topk_weights_dispatched"),
                forward_batch=state.forward_batch,
                tbo_subbatch_index=state.get("tbo_subbatch_index"),
            )

    def op_combine_b(self, state):
        if self.ep_size > 1:
            state.hidden_states_after_combine = self.deepep_dispatcher.combine_b(
                tbo_subbatch_index=state.get("tbo_subbatch_index"),
            )

    def op_output(self, state):
        final_hidden_states = state.pop("hidden_states_after_combine")

        if (shared_output := state.pop("shared_output")) is not None:
            x = shared_output
            x.add_(final_hidden_states, alpha=self.routed_scaling_factor)
            final_hidden_states = x
        else:
            final_hidden_states *= self.routed_scaling_factor

        state.hidden_states_mlp_output = final_hidden_states


def yarn_get_mscale(scale: float = 1, mscale: float = 1) -> float:
    import math

    if scale <= 1:
        return 1.0
    return 0.1 * mscale * math.log(scale) + 1.0


class DeepseekV2AttentionMLA(nn.Module):

    def __init__(
        self,
        config: PretrainedConfig,
        hidden_size: int,
        num_heads: int,
        qk_nope_head_dim: int,
        qk_rope_head_dim: int,
        v_head_dim: int,
        q_lora_rank: int,
        kv_lora_rank: int,
        rope_theta: float = 10000,
        rope_scaling: Optional[Dict[str, Any]] = None,
        max_position_embeddings: int = 8192,
        quant_config: Optional[QuantizationConfig] = None,
        reduce_results: bool = True,
        layer_id: int = None,
        prefix: str = "",
        alt_stream: Optional[torch.cuda.Stream] = None,
    ) -> None:
        super().__init__()
        self.layer_id = layer_id
        self.hidden_size = hidden_size
        self.qk_nope_head_dim = qk_nope_head_dim
        self.qk_rope_head_dim = qk_rope_head_dim
        self.qk_head_dim = qk_nope_head_dim + qk_rope_head_dim
        self.v_head_dim = v_head_dim
        self.q_lora_rank = q_lora_rank
        self.kv_lora_rank = kv_lora_rank
        attn_tp_rank = get_attention_tp_rank()
        attn_tp_size = get_attention_tp_size()

        self.num_heads = num_heads
        assert num_heads % attn_tp_size == 0
        self.num_local_heads = num_heads // attn_tp_size
        self.scaling = self.qk_head_dim**-0.5
        self.rope_theta = rope_theta
        self.max_position_embeddings = max_position_embeddings

        # For tensor parallel attention
        if self.q_lora_rank is not None:
            self.fused_qkv_a_proj_with_mqa = ReplicatedLinear(
                self.hidden_size,
                self.q_lora_rank + self.kv_lora_rank + self.qk_rope_head_dim,
                bias=False,
                quant_config=quant_config,
                prefix=add_prefix("fused_qkv_a_proj_with_mqa", prefix),
            )
            self.q_a_layernorm = RMSNorm(self.q_lora_rank, eps=config.rms_norm_eps)
            self.q_b_proj = ColumnParallelLinear(
                q_lora_rank,
                self.num_heads * self.qk_head_dim,
                bias=False,
                quant_config=quant_config,
                prefix=add_prefix("q_b_proj", prefix),
                tp_rank=attn_tp_rank,
                tp_size=attn_tp_size,
            )
        else:
            self.q_proj = ColumnParallelLinear(
                self.hidden_size,
                self.num_heads * self.qk_head_dim,
                bias=False,
                quant_config=quant_config,
                prefix=add_prefix("q_proj", prefix),
                tp_rank=attn_tp_rank,
                tp_size=attn_tp_size,
            )
            self.kv_a_proj_with_mqa = ReplicatedLinear(
                self.hidden_size,
                self.kv_lora_rank + self.qk_rope_head_dim,
                bias=False,
                quant_config=quant_config,
                prefix=add_prefix("kv_a_proj_with_mqa", prefix),
            )

        self.kv_b_proj = ColumnParallelLinear(
            self.kv_lora_rank,
            self.num_heads * (self.qk_nope_head_dim + self.v_head_dim),
            bias=False,
            quant_config=quant_config,
            prefix=add_prefix("kv_b_proj", prefix),
            tp_rank=attn_tp_rank,
            tp_size=attn_tp_size,
        )
        # O projection.
        self.o_proj = RowParallelLinear(
            self.num_heads * self.v_head_dim,
            self.hidden_size,
            bias=False,
            quant_config=quant_config,
            reduce_results=reduce_results,
            prefix=add_prefix("o_proj", prefix),
            tp_rank=attn_tp_rank,
            tp_size=attn_tp_size,
        )
        self.kv_a_layernorm = RMSNorm(self.kv_lora_rank, eps=config.rms_norm_eps)

        if rope_scaling:
            rope_scaling["rope_type"] = "deepseek_yarn"

        self.rotary_emb = get_rope_wrapper(
            qk_rope_head_dim,
            rotary_dim=qk_rope_head_dim,
            max_position=max_position_embeddings,
            base=rope_theta,
            rope_scaling=rope_scaling,
            is_neox_style=False,
            device=global_server_args_dict["device"],
        )

        if rope_scaling:
            mscale_all_dim = rope_scaling.get("mscale_all_dim", False)
            scaling_factor = rope_scaling["factor"]
            mscale = yarn_get_mscale(scaling_factor, float(mscale_all_dim))
            self.scaling = self.scaling * mscale * mscale
        else:
            self.rotary_emb.forward = self.rotary_emb.forward_native

        self.attn_mqa = RadixAttention(
            self.num_local_heads,
            self.kv_lora_rank + self.qk_rope_head_dim,
            self.scaling,
            num_kv_heads=1,
            layer_id=layer_id,
            v_head_dim=self.kv_lora_rank,
            quant_config=quant_config,
            prefix=add_prefix("attn_mqa", prefix),
        )

        self.attn_mha = RadixAttention(
            self.num_local_heads,
            self.qk_nope_head_dim + self.qk_rope_head_dim,
            self.scaling,
            num_kv_heads=self.num_local_heads,
            layer_id=layer_id,
            v_head_dim=self.v_head_dim,
            quant_config=quant_config,
            prefix=add_prefix("attn_mha", prefix),
        )

        self.alt_stream = alt_stream
        self.attn_mha.kv_b_proj = None

        self.w_kc = None
        self.w_vc = None
        self.w_scale = 1.0

        self.w_scale_k = None
        self.w_scale_v = None
        self.use_deep_gemm_bmm = False

        self.flashinfer_mla_disable_ragged = global_server_args_dict[
            "flashinfer_mla_disable_ragged"
        ]
        self.disable_chunked_prefix_cache = global_server_args_dict[
            "disable_chunked_prefix_cache"
        ]

        self.current_attention_backend = (
            None  # Attention backend used by current forward batch
        )
        self.rocm_fused_decode_mla = get_bool_env_var(
            "SGLANG_ROCM_FUSED_DECODE_MLA", "false"
        )

        # TODO: Design a finer way to determine the threshold
        self.chunked_prefix_cache_threshold = get_int_env_var(
            "SGL_CHUNKED_PREFIX_CACHE_THRESHOLD", 8192
        )

        # If we have self.fused_qkv_a_proj_with_mqa and we're running on CPU, we will choose the torch.ops.sgl_kernel.qkv_proj_with_rope_fused_weight kernel
        # which requires self.w_kc and self.w_vc to be packed.
        # If not, we will use torch.bmm and weight shouldn't be packed in this case
        has_fused_proj = hasattr(self, "fused_qkv_a_proj_with_mqa")
        if has_fused_proj and _is_cpu and _is_cpu_amx_available:
            self.quant_method = PackWeightMethod(
                weight_names=["w_kc", "w_vc"], transpose_dims=[[1, 2], [1, 2]]
            )

        is_packed_weight = (
            has_fused_proj
            and hasattr(self.fused_qkv_a_proj_with_mqa.quant_method, "quant_config")
            and self.fused_qkv_a_proj_with_mqa.quant_method.quant_config.get_name()
            in {"awq", "awq_marlin", "moe_wna16"}
        )
        self.use_min_latency_fused_a_gemm = (
            has_fused_proj
            and not is_packed_weight
            and self.fused_qkv_a_proj_with_mqa.weight.dtype == torch.bfloat16
            and self.fused_qkv_a_proj_with_mqa.weight.shape[0] == 2112
            and self.fused_qkv_a_proj_with_mqa.weight.shape[1] == 7168
            and _is_cuda
            and _device_sm >= 90
        )

        self.qkv_proj_with_rope_is_int8 = (
            has_fused_proj
            and not is_packed_weight
            and self.fused_qkv_a_proj_with_mqa.weight.dtype == torch.int8
        )
        self.qkv_proj_with_rope_is_fp8 = (
            has_fused_proj
            and not is_packed_weight
            and self.fused_qkv_a_proj_with_mqa.weight.dtype == torch.float8_e4m3fn
        )

        self.weight_block_size = None
        if self.qkv_proj_with_rope_is_fp8 and _is_cpu and _is_cpu_amx_available:
            assert getattr(
                self.fused_qkv_a_proj_with_mqa.quant_method, "block_quant", False
            ) == getattr(self.q_b_proj.quant_method, "block_quant", False)
            use_block_quant = getattr(
                self.fused_qkv_a_proj_with_mqa.quant_method, "block_quant", False
            )

            if use_block_quant:
                assert (
                    self.fused_qkv_a_proj_with_mqa.quant_method.quant_config.weight_block_size
                    == self.q_b_proj.quant_method.quant_config.weight_block_size
                )
                self.weight_block_size = (
                    self.fused_qkv_a_proj_with_mqa.quant_method.quant_config.weight_block_size
                )

    def dispatch_attn_forward_method(
        self, forward_batch: ForwardBatch
    ) -> AttnForwardMethod:
        def _dispatch_mla_subtype():
            if _is_hip:
                if (
                    self.rocm_fused_decode_mla
                    and forward_batch.forward_mode.is_decode()
                ):
                    return AttnForwardMethod.MLA_FUSED_ROPE
                else:
                    return AttnForwardMethod.MLA
            else:
                if hasattr(self, "fused_qkv_a_proj_with_mqa") and use_intel_amx_backend(
                    self
                ):
                    return AttnForwardMethod.MLA_FUSED_ROPE_CPU
                else:
                    return AttnForwardMethod.MLA

        # Determine attention backend used by current forward batch
        if forward_batch.forward_mode.is_decode_or_idle():
            attention_backend = global_server_args_dict["decode_attention_backend"]
        else:
            attention_backend = global_server_args_dict["prefill_attention_backend"]
        self.current_attention_backend = attention_backend

        if attention_backend == "ascend":
            return AttnForwardMethod.MLA
        elif attention_backend == "flashinfer":
            # Flashinfer MLA: Do not absorb when enabling ragged prefill
            if (
                not self.flashinfer_mla_disable_ragged
                and forward_batch.forward_mode.is_extend()
                and not forward_batch.forward_mode.is_target_verify()
                and not forward_batch.forward_mode.is_draft_extend()
                and sum(forward_batch.extend_prefix_lens_cpu) == 0
            ):
                return AttnForwardMethod.MHA
            else:
                return _dispatch_mla_subtype()
        elif attention_backend == "fa3":
            # Flash Attention: Use MHA with chunked KV cache when prefilling on long sequences.
            if forward_batch.extend_prefix_lens_cpu is not None:
                sum_extend_prefix_lens = sum(forward_batch.extend_prefix_lens_cpu)
            if (
                forward_batch.forward_mode.is_extend()
                and not self.disable_chunked_prefix_cache
                and not forward_batch.forward_mode.is_target_verify()
                and not forward_batch.forward_mode.is_draft_extend()
                and (
                    sum_extend_prefix_lens >= self.chunked_prefix_cache_threshold
                    or sum_extend_prefix_lens == 0
                )
            ):
                return AttnForwardMethod.MHA_CHUNKED_KV
            else:
                return _dispatch_mla_subtype()
        elif attention_backend == "aiter":
            if (
                forward_batch.forward_mode.is_extend()
                and not forward_batch.forward_mode.is_target_verify()
                and not forward_batch.forward_mode.is_draft_extend()
            ):
                return AttnForwardMethod.MHA
            else:
                return AttnForwardMethod.MLA
        else:
            # Triton: Use normal computation for prefill and use weight absorption for extend/decode
            if (
                forward_batch.forward_mode.is_extend()
                and not forward_batch.forward_mode.is_target_verify()
                and not forward_batch.forward_mode.is_draft_extend()
                and sum(forward_batch.extend_prefix_lens_cpu) == 0
            ):
                return AttnForwardMethod.MHA
            else:
                return _dispatch_mla_subtype()

    def op_prepare(self, state):
        state.attn_intermediate_state = self.forward_prepare(
            positions=state.positions,
            hidden_states=state.pop("hidden_states_after_comm_pre_attn"),
            forward_batch=state.forward_batch,
            zero_allocator=state.zero_allocator,
        )

    def op_core(self, state):
        state.hidden_states_after_attn = self.forward_core(
            state.pop("attn_intermediate_state")
        )

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        forward_batch: ForwardBatch,
        zero_allocator: BumpAllocator,
    ):
        s = self.forward_prepare(
            positions=positions,
            hidden_states=hidden_states,
            forward_batch=forward_batch,
            zero_allocator=zero_allocator,
        )
        return self.forward_core(s)

    def forward_prepare(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        forward_batch: ForwardBatch,
        zero_allocator: BumpAllocator,
    ):
        if self.attn_mha.kv_b_proj is None:
            self.attn_mha.kv_b_proj = self.kv_b_proj

        if hidden_states.shape[0] == 0:
            assert (
                not self.o_proj.reduce_results
            ), "short-circuiting allreduce will lead to hangs"
            return hidden_states, None, forward_batch, None

        attn_forward_method = self.dispatch_attn_forward_method(forward_batch)

        if attn_forward_method == AttnForwardMethod.MHA:
            inner_state = self.forward_normal_prepare(
                positions, hidden_states, forward_batch, zero_allocator
            )
        elif attn_forward_method == AttnForwardMethod.MHA_CHUNKED_KV:
            inner_state = self.forward_normal_chunked_kv_prepare(
                positions, hidden_states, forward_batch, zero_allocator
            )
        elif attn_forward_method == AttnForwardMethod.MLA:
            inner_state = self.forward_absorb_prepare(
                positions, hidden_states, forward_batch, zero_allocator
            )
        elif attn_forward_method == AttnForwardMethod.MLA_FUSED_ROPE:
            inner_state = self.forward_absorb_fused_mla_rope_prepare(
                positions, hidden_states, forward_batch, zero_allocator
            )
        elif attn_forward_method == AttnForwardMethod.MLA_FUSED_ROPE_CPU:
            inner_state = self.forward_absorb_fused_mla_rope_cpu_prepare(
                positions, hidden_states, forward_batch, zero_allocator
            )
        else:
            raise NotImplementedError
        return None, attn_forward_method, forward_batch, inner_state

    def forward_core(self, intermediate_state):
        hidden_states, attn_forward_method, forward_batch, inner_state = (
            intermediate_state
        )
        if inner_state is None:
            return hidden_states

        if attn_forward_method == AttnForwardMethod.MHA:
            return self.forward_normal_core(*inner_state)
        elif attn_forward_method == AttnForwardMethod.MHA_CHUNKED_KV:
            return self.forward_normal_chunked_kv_core(*inner_state)
        elif attn_forward_method == AttnForwardMethod.MLA:
            return self.forward_absorb_core(*inner_state)
        elif attn_forward_method == AttnForwardMethod.MLA_FUSED_ROPE:
            return self.forward_absorb_fused_mla_rope_core(*inner_state)
        elif attn_forward_method == AttnForwardMethod.MLA_FUSED_ROPE_CPU:
            return self.forward_absorb_fused_mla_rope_cpu_core(*inner_state)
        else:
            raise NotImplementedError

    def forward_normal_prepare(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        forward_batch: ForwardBatch,
        zero_allocator: BumpAllocator,
    ):
        if self.q_lora_rank is not None:
            q, latent_cache = self.fused_qkv_a_proj_with_mqa(hidden_states)[0].split(
                [self.q_lora_rank, self.kv_lora_rank + self.qk_rope_head_dim], dim=-1
            )
            q = self.q_a_layernorm(q)
            q = self.q_b_proj(q)[0].view(-1, self.num_local_heads, self.qk_head_dim)
        else:
            q = self.q_proj(hidden_states)[0].view(
                -1, self.num_local_heads, self.qk_head_dim
            )
            latent_cache = self.kv_a_proj_with_mqa(hidden_states)[0]

        _, q_pe = q.split([self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1)
        kv_a, _ = latent_cache.split([self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)
        latent_cache = latent_cache.unsqueeze(1)
        kv_a = self.kv_a_layernorm(kv_a)
        kv = self.kv_b_proj(kv_a)[0]
        kv = kv.view(-1, self.num_local_heads, self.qk_nope_head_dim + self.v_head_dim)
        k_nope = kv[..., : self.qk_nope_head_dim]
        v = kv[..., self.qk_nope_head_dim :]
        k_pe = latent_cache[:, :, self.kv_lora_rank :]
        q_pe, k_pe = self.rotary_emb(positions, q_pe, k_pe)
        q[..., self.qk_nope_head_dim :] = q_pe
        k = torch.empty_like(q)
        k[..., : self.qk_nope_head_dim] = k_nope
        k[..., self.qk_nope_head_dim :] = k_pe

        latent_cache[:, :, : self.kv_lora_rank] = kv_a.unsqueeze(1)
        latent_cache[:, :, self.kv_lora_rank :] = k_pe

        # Save latent cache
        forward_batch.token_to_kv_pool.set_kv_buffer(
            self.attn_mha, forward_batch.out_cache_loc, latent_cache, None
        )

        return q, k, v, forward_batch

    def forward_normal_core(self, q, k, v, forward_batch):
        attn_output = self.attn_mha(q, k, v, forward_batch, save_kv_cache=False)
        attn_output = attn_output.reshape(-1, self.num_local_heads * self.v_head_dim)
        output, _ = self.o_proj(attn_output)
        return output

    def forward_absorb_prepare(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        forward_batch: ForwardBatch,
        zero_allocator: BumpAllocator,
    ):
        from sglang.srt.model_executor.cuda_graph_runner import get_is_capture_mode

        if self.q_lora_rank is not None:
            if hidden_states.shape[0] <= 16 and self.use_min_latency_fused_a_gemm:
                fused_qkv_a_proj_out = dsv3_fused_a_gemm(
                    hidden_states, self.fused_qkv_a_proj_with_mqa.weight.T
                )
            else:
                fused_qkv_a_proj_out = self.fused_qkv_a_proj_with_mqa(hidden_states)[0]
            q, latent_cache = fused_qkv_a_proj_out.split(
                [self.q_lora_rank, self.kv_lora_rank + self.qk_rope_head_dim], dim=-1
            )
            k_nope = latent_cache[..., : self.kv_lora_rank]

            # overlap qk norm
            if self.alt_stream is not None and get_is_capture_mode():
                current_stream = torch.cuda.current_stream()
                self.alt_stream.wait_stream(current_stream)
                q = self.q_a_layernorm(q)
                with torch.cuda.stream(self.alt_stream):
                    k_nope = self.kv_a_layernorm(k_nope)
                current_stream.wait_stream(self.alt_stream)
            else:
                q = self.q_a_layernorm(q)
                k_nope = self.kv_a_layernorm(k_nope)

            k_nope = k_nope.unsqueeze(1)
            q = self.q_b_proj(q)[0].view(-1, self.num_local_heads, self.qk_head_dim)
        else:
            q = self.q_proj(hidden_states)[0].view(
                -1, self.num_local_heads, self.qk_head_dim
            )
            latent_cache = self.kv_a_proj_with_mqa(hidden_states)[0]
            k_nope = latent_cache[..., : self.kv_lora_rank]
            k_nope = self.kv_a_layernorm(k_nope).unsqueeze(1)

        q_nope, q_pe = q.split([self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1)
        k_pe = latent_cache[..., self.kv_lora_rank :].unsqueeze(1)

        if self.use_deep_gemm_bmm:
            q_nope_val, q_nope_scale, masked_m, expected_m, aligned_m = (
                per_token_group_quant_mla_deep_gemm_masked_fp8(q_nope.transpose(0, 1))
            )
            q_nope_out = q_nope.new_empty(
                (self.num_local_heads, aligned_m, self.kv_lora_rank)
            )
            deep_gemm_wrapper.grouped_gemm_nt_f8f8bf16_masked(
                (q_nope_val, q_nope_scale),
                (self.w_kc, self.w_scale_k),
                q_nope_out,
                masked_m,
                expected_m,
            )
            q_nope_out = q_nope_out[:, :expected_m, :]
        elif _is_hip:
            # TODO(haishaw): add bmm_fp8 to ROCm
            q_nope_out = torch.bmm(
                q_nope.to(torch.bfloat16).transpose(0, 1),
                self.w_kc.to(torch.bfloat16) * self.w_scale,
            )
        elif self.w_kc.dtype == torch.float8_e4m3fn:
            q_nope_val, q_nope_scale = per_tensor_quant_mla_fp8(
                q_nope.transpose(0, 1),
                zero_allocator.allocate(1),
            )
            q_nope_out = bmm_fp8(
                q_nope_val, self.w_kc, q_nope_scale, self.w_scale, torch.bfloat16
            )
        else:
            q_nope_out = torch.bmm(q_nope.transpose(0, 1), self.w_kc)

        q_nope_out = q_nope_out.transpose(0, 1)
        q_pe, k_pe = self.rotary_emb(positions, q_pe, k_pe)

        return q_pe, k_pe, q_nope_out, k_nope, forward_batch, zero_allocator

    def forward_absorb_core(
        self, q_pe, k_pe, q_nope_out, k_nope, forward_batch, zero_allocator
    ):
        if (
            self.current_attention_backend == "fa3"
            or self.current_attention_backend == "flashinfer"
            or self.current_attention_backend == "cutlass_mla"
        ):
            attn_output = self.attn_mqa(
                q_nope_out, k_nope, k_nope, forward_batch, q_rope=q_pe, k_rope=k_pe
            )
        else:
            q = torch.cat([q_nope_out, q_pe], dim=-1)
            k = torch.cat([k_nope, k_pe], dim=-1)
            attn_output = self.attn_mqa(q, k, k_nope, forward_batch)
        attn_output = attn_output.view(-1, self.num_local_heads, self.kv_lora_rank)

        if self.use_deep_gemm_bmm:
            attn_output_val, attn_output_scale, masked_m, expected_m, aligned_m = (
                per_token_group_quant_mla_deep_gemm_masked_fp8(
                    attn_output.transpose(0, 1)
                )
            )
            attn_bmm_output = attn_output.new_empty(
                (self.num_local_heads, aligned_m, self.v_head_dim)
            )
            deep_gemm_wrapper.grouped_gemm_nt_f8f8bf16_masked(
                (attn_output_val, attn_output_scale),
                (self.w_vc, self.w_scale_v),
                attn_bmm_output,
                masked_m,
                expected_m,
            )
            attn_bmm_output = (
                attn_bmm_output[:, :expected_m, :].transpose(0, 1).flatten(1, 2)
            )
        elif _is_hip:
            # TODO(haishaw): add bmm_fp8 to ROCm
            attn_bmm_output = torch.bmm(
                attn_output.to(torch.bfloat16).transpose(0, 1),
                self.w_vc.to(torch.bfloat16) * self.w_scale,
            )
            attn_bmm_output = attn_bmm_output.transpose(0, 1).flatten(1, 2)
        elif self.w_vc.dtype == torch.float8_e4m3fn:
            attn_output_val, attn_output_scale = per_tensor_quant_mla_fp8(
                attn_output.transpose(0, 1),
                zero_allocator.allocate(1),
            )
            attn_bmm_output = bmm_fp8(
                attn_output_val,
                self.w_vc,
                attn_output_scale,
                self.w_scale,
                torch.bfloat16,
            )
            attn_bmm_output = attn_bmm_output.transpose(0, 1).flatten(1, 2)
        else:
            attn_bmm_output = torch.empty(
                (attn_output.shape[0], self.num_local_heads * self.v_head_dim),
                dtype=attn_output.dtype,
                device=attn_output.device,
            )
            torch.bmm(
                attn_output.transpose(0, 1),
                self.w_vc,
                out=attn_bmm_output.view(
                    -1, self.num_local_heads, self.v_head_dim
                ).transpose(0, 1),
            )
        output, _ = self.o_proj(attn_bmm_output)

        return output

    def forward_absorb_fused_mla_rope_prepare(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        forward_batch: ForwardBatch,
        zero_allocator: BumpAllocator,
    ):
        enable_rope_fusion = (
            os.getenv("SGLANG_FUSED_MLA_ENABLE_ROPE_FUSION", "1") == "1"
        )
        q_len = hidden_states.shape[0]
        q_input = hidden_states.new_empty(
            q_len, self.num_local_heads, self.kv_lora_rank + self.qk_rope_head_dim
        )
        if self.q_lora_rank is not None:
            q, latent_cache = self.fused_qkv_a_proj_with_mqa(hidden_states)[0].split(
                [self.q_lora_rank, self.kv_lora_rank + self.qk_rope_head_dim], dim=-1
            )
            q = self.q_a_layernorm(q)
            q = self.q_b_proj(q)[0].view(-1, self.num_local_heads, self.qk_head_dim)
        else:
            q = self.q_proj(hidden_states)[0].view(
                -1, self.num_local_heads, self.qk_head_dim
            )
            latent_cache = self.kv_a_proj_with_mqa(hidden_states)[0]
        q_nope, q_pe = q.split([self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1)

        if _is_hip:
            # TODO(haishaw): add bmm_fp8 to ROCm
            q_nope_out = torch.bmm(
                q_nope.to(torch.bfloat16).transpose(0, 1),
                self.w_kc.to(torch.bfloat16) * self.w_scale,
            )
        elif self.w_kc.dtype == torch.float8_e4m3fn:
            q_nope_val, q_nope_scale = per_tensor_quant_mla_fp8(
                q_nope.transpose(0, 1),
                zero_allocator.allocate(1),
                dtype=torch.float8_e4m3fn,
            )
            q_nope_out = bmm_fp8(
                q_nope_val, self.w_kc, q_nope_scale, self.w_scale, torch.bfloat16
            )
        else:
            q_nope_out = torch.bmm(q_nope.transpose(0, 1), self.w_kc)
        q_input[..., : self.kv_lora_rank] = q_nope_out.transpose(0, 1)
        v_input = latent_cache[..., : self.kv_lora_rank]
        v_input = self.kv_a_layernorm(v_input.contiguous()).unsqueeze(1)
        k_input = latent_cache.unsqueeze(1)
        k_input[..., : self.kv_lora_rank] = v_input

        if not enable_rope_fusion:
            k_pe = k_input[..., self.kv_lora_rank :]
            q_pe, k_pe = self.rotary_emb(positions, q_pe, k_pe)
            q_input[..., self.kv_lora_rank :] = q_pe
            k_input[..., self.kv_lora_rank :] = k_pe
            k_pe_output = None
        else:
            k_pe_output = torch.empty_like(k_input[..., self.kv_lora_rank :])

        q_input[..., self.kv_lora_rank :] = q_pe

        # attn_output = self.attn_mqa(q_input, k_input, v_input, forward_batch)
        # Use Fused ROPE with use_rope=OFF.
        attn_output = torch.empty(
            (q_len, self.num_local_heads, self.kv_lora_rank),
            dtype=q.dtype,
            device=q.device,
        )
        attn_logits, _, kv_indptr, kv_indices, _, _, _ = (
            forward_batch.attn_backend.forward_metadata
        )
        cos_sin_cache = self.rotary_emb.cos_sin_cache
        num_kv_split = forward_batch.attn_backend.num_kv_splits
        sm_scale = self.attn_mqa.scaling
        if attn_logits is None:
            attn_logits = torch.empty(
                (
                    forward_batch.batch_size,
                    self.num_local_heads,
                    num_kv_split,
                    self.kv_lora_rank + 1,
                ),
                dtype=torch.float32,
                device=q.device,
            )

        # save current latent cache.
        forward_batch.token_to_kv_pool.set_kv_buffer(
            self.attn_mqa, forward_batch.out_cache_loc, k_input, None
        )
        key_cache_buf = forward_batch.token_to_kv_pool.get_key_buffer(
            self.attn_mqa.layer_id
        )
        val_cache_buf = key_cache_buf[..., : self.kv_lora_rank]

        return (
            q_input,
            key_cache_buf,
            val_cache_buf,
            attn_output,
            kv_indptr,
            kv_indices,
            k_pe_output,
            cos_sin_cache,
            positions,
            attn_logits,
            num_kv_split,
            sm_scale,
            enable_rope_fusion,
            k_input,
            forward_batch,
            zero_allocator,
        )

    def forward_absorb_fused_mla_rope_cpu_prepare(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        forward_batch: ForwardBatch,
        zero_allocator: BumpAllocator,
    ):
        assert self.q_lora_rank is not None and use_intel_amx_backend(
            self
        ), "forward_absorb_fused_mla_rope_cpu_prepare requires q_lora_rank is not None and use_intel_amx_backend"

        q_input, k_input, v_input = (
            torch.ops.sgl_kernel.qkv_proj_with_rope_fused_weight(
                hidden_states,
                self.fused_qkv_a_proj_with_mqa.weight,
                self.q_b_proj.weight,
                self.w_kc,
                self.q_a_layernorm.weight,
                self.kv_a_layernorm.weight,
                positions,
                self.rotary_emb.cos_sin_cache,
                self.kv_a_layernorm.variance_epsilon,
                self.qkv_proj_with_rope_is_int8,
                self.qkv_proj_with_rope_is_fp8,
                (
                    self.fused_qkv_a_proj_with_mqa.weight_scale
                    if self.qkv_proj_with_rope_is_int8
                    else (
                        self.fused_qkv_a_proj_with_mqa.weight_scale_inv
                        if self.qkv_proj_with_rope_is_fp8
                        else None
                    )
                ),
                (
                    self.q_b_proj.weight_scale
                    if self.qkv_proj_with_rope_is_int8
                    else (
                        self.q_b_proj.weight_scale_inv
                        if self.qkv_proj_with_rope_is_fp8
                        else None
                    )
                ),
                True,  # is_vnni
                self.weight_block_size,
                self.q_lora_rank,
                self.kv_lora_rank,
                self.qk_rope_head_dim,
            )
        )
        return (q_input, k_input, v_input, forward_batch, zero_allocator)

    def forward_absorb_fused_mla_rope_core(
        self,
        q_input,
        key_cache_buf,
        val_cache_buf,
        attn_output,
        kv_indptr,
        kv_indices,
        k_pe_output,
        cos_sin_cache,
        positions,
        attn_logits,
        num_kv_split,
        sm_scale,
        enable_rope_fusion,
        k_input,
        forward_batch,
        zero_allocator,
    ):
        decode_attention_fwd_grouped_rope(
            q_input,
            key_cache_buf,
            val_cache_buf,
            attn_output,
            kv_indptr,
            kv_indices,
            k_pe_output,
            self.kv_lora_rank,
            self.rotary_emb.rotary_dim,
            cos_sin_cache,
            positions,
            attn_logits,
            num_kv_split,
            sm_scale,
            logit_cap=self.attn_mqa.logit_cap,
            use_rope=enable_rope_fusion,
            is_neox_style=self.rotary_emb.is_neox_style,
        )

        if enable_rope_fusion:
            k_input[..., self.kv_lora_rank :] = k_pe_output
            forward_batch.token_to_kv_pool.set_kv_buffer(
                self.attn_mqa, forward_batch.out_cache_loc, k_input, None
            )

        attn_output = attn_output.view(-1, self.num_local_heads, self.kv_lora_rank)

        if _is_hip:
            # TODO(haishaw): add bmm_fp8 to ROCm
            attn_bmm_output = torch.bmm(
                attn_output.to(torch.bfloat16).transpose(0, 1),
                self.w_vc.to(torch.bfloat16) * self.w_scale,
            )
        elif self.w_vc.dtype == torch.float8_e4m3fn:
            attn_output_val, attn_output_scale = per_tensor_quant_mla_fp8(
                attn_output.transpose(0, 1),
                zero_allocator.allocate(1),
                dtype=torch.float8_e4m3fn,
            )
            attn_bmm_output = bmm_fp8(
                attn_output_val,
                self.w_vc,
                attn_output_scale,
                self.w_scale,
                torch.bfloat16,
            )
        else:
            attn_bmm_output = torch.bmm(attn_output.transpose(0, 1), self.w_vc)
        attn_output = attn_bmm_output.transpose(0, 1).flatten(1, 2)
        output, _ = self.o_proj(attn_output)

        return output

    def forward_absorb_fused_mla_rope_cpu_core(
        self, q_input, k_input, v_input, forward_batch, zero_allocator
    ):
        assert self.q_lora_rank is not None and use_intel_amx_backend(
            self
        ), "forward_absorb_fused_mla_rope_cpu_core requires q_lora_rank is not None and use_intel_amx_backend"

        attn_output = self.attn_mqa(q_input, k_input, v_input, forward_batch)
        attn_output = attn_output.view(-1, self.num_local_heads, self.kv_lora_rank)

        # [Note] Align shapes of bmm inputs.
        # Shapes of inputs:
        #   q_nope: [M, B, K]
        #   original self.w_kc: [B, K, N]
        #   current self.w_kc (which has been converted in PackWeightMethod): [B, N, K]

        # Shapes of inputs to sgl_kernel.cpu.bmm:
        #   out: [B, M, N]
        #   mat1: [B, M, K]
        #   mat2: [B, N, K]
        B = self.w_vc.size(0)
        N = self.w_vc.size(1)
        M = attn_output.size(0)
        output = torch.empty([M, int(B * N)], dtype=attn_output.dtype)
        attn_bmm_output = output.view([M, B, N]).transpose_(0, 1)
        torch.ops.sgl_kernel.bmm_cpu(
            attn_bmm_output,
            attn_output.transpose(0, 1),
            self.w_vc,
            True,  # is_vnni
            None,  # scale
        )
        attn_output = output
        output, _ = self.o_proj(attn_output)

        return output

    def _chunked_prefix_attn_mha(
        self,
        q: torch.Tensor,
        accum_output: torch.Tensor,
        accum_lse: torch.Tensor,
        forward_batch: ForwardBatch,
    ) -> torch.Tensor:

        assert forward_batch.num_prefix_chunks is not None
        for i in range(forward_batch.num_prefix_chunks):
            forward_batch.set_prefix_chunk_idx(i)

            # Fetch latent cache from memory pool with precomputed chunked kv indices
            latent_cache_buf = forward_batch.token_to_kv_pool.get_key_buffer(
                self.attn_mha.layer_id
            )
            latent_cache = latent_cache_buf[
                forward_batch.prefix_chunk_kv_indices[i]
            ].contiguous()

            kv_a_normed, k_pe = latent_cache.split(
                [self.kv_lora_rank, self.qk_rope_head_dim], dim=-1
            )
            kv_a_normed = kv_a_normed.squeeze(1).contiguous()
            kv = self.kv_b_proj(kv_a_normed)[0]
            kv = kv.view(
                -1, self.num_local_heads, self.qk_nope_head_dim + self.v_head_dim
            )
            v = kv[..., self.qk_nope_head_dim :]
            k_nope = kv[..., : self.qk_nope_head_dim]

            k = torch.empty(
                (
                    k_nope.shape[0],
                    self.num_local_heads,
                    self.qk_nope_head_dim + self.qk_rope_head_dim,
                ),
                dtype=v.dtype,
                device=v.device,
            )
            k[..., : self.qk_nope_head_dim] = k_nope
            k[..., self.qk_nope_head_dim :] = k_pe

            output, lse = self.attn_mha(q, k, v, forward_batch, save_kv_cache=False)
            lse = torch.transpose(lse, 0, 1).contiguous()
            tmp_output = torch.empty_like(accum_output)
            tmp_lse = torch.empty_like(accum_lse)
            merge_state_v2(output, lse, accum_output, accum_lse, tmp_output, tmp_lse)
            accum_output, accum_lse = tmp_output, tmp_lse

        return accum_output

    def forward_normal_chunked_kv_prepare(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        forward_batch: ForwardBatch,
        zero_allocator: BumpAllocator,
    ):
        # In normal mha, the k and v tensors will become overly large when the prefix length is long.
        # To avoid this, we split the kv cache into chunks and process them one after another.
        # Since mha is compute friendly, the for loop induced here will not introduce significant overhead.
        # The top comments in https://github.com/vllm-project/vllm/blob/main/vllm/v1/attention/backends/mla/common.py
        # will be helpful for understanding the purpose of this function.

        # First do normal mha forward to get output for extended part
        if self.q_lora_rank is not None:
            q, latent_cache = self.fused_qkv_a_proj_with_mqa(hidden_states)[0].split(
                [self.q_lora_rank, self.kv_lora_rank + self.qk_rope_head_dim], dim=-1
            )
            q = self.q_a_layernorm(q)
            q = self.q_b_proj(q)[0].view(-1, self.num_local_heads, self.qk_head_dim)
        else:
            q = self.q_proj(hidden_states)[0].view(
                -1, self.num_local_heads, self.qk_head_dim
            )
            latent_cache = self.kv_a_proj_with_mqa(hidden_states)[0]
        _, q_pe = q.split([self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1)
        kv_a, _ = latent_cache.split([self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)
        latent_cache = latent_cache.unsqueeze(1)
        kv_a = self.kv_a_layernorm(kv_a)
        kv = self.kv_b_proj(kv_a)[0]
        kv = kv.view(-1, self.num_local_heads, self.qk_nope_head_dim + self.v_head_dim)
        k_nope = kv[..., : self.qk_nope_head_dim]
        v = kv[..., self.qk_nope_head_dim :]
        k_pe = latent_cache[:, :, self.kv_lora_rank :]

        q_pe, k_pe = self.rotary_emb(positions, q_pe, k_pe)
        q[..., self.qk_nope_head_dim :] = q_pe
        k = torch.empty_like(q)
        k[..., : self.qk_nope_head_dim] = k_nope
        k[..., self.qk_nope_head_dim :] = k_pe

        latent_cache[:, :, : self.kv_lora_rank] = kv_a.unsqueeze(1)
        latent_cache[:, :, self.kv_lora_rank :] = k_pe

        # Save latent cache
        forward_batch.token_to_kv_pool.set_kv_buffer(
            self.attn_mha, forward_batch.out_cache_loc, latent_cache, None
        )

        return q, k, v, forward_batch

    def forward_normal_chunked_kv_core(self, q, k, v, forward_batch):
        # Do mha for extended part without prefix
        forward_batch.set_attn_attend_prefix_cache(False)
        attn_output, lse = self.attn_mha(q, k, v, forward_batch, save_kv_cache=False)
        lse = torch.transpose(lse, 0, 1).contiguous()

        # Do mha attention with chunked prefix cache if there are any sequence with prefix
        if any(forward_batch.extend_prefix_lens_cpu):
            # Only initialize the info once
            if forward_batch.num_prefix_chunks is None:
                forward_batch.prepare_chunked_prefix_cache_info(q.device)

            forward_batch.set_attn_attend_prefix_cache(True)
            attn_output = self._chunked_prefix_attn_mha(
                q=q,
                accum_output=attn_output,
                accum_lse=lse,
                forward_batch=forward_batch,
            )

        attn_output = attn_output.reshape(-1, self.num_local_heads * self.v_head_dim)
        output, _ = self.o_proj(attn_output)
        return output


class DeepseekV2DecoderLayer(nn.Module):

    def __init__(
        self,
        config: PretrainedConfig,
        layer_id: int,
        quant_config: Optional[QuantizationConfig] = None,
        is_nextn: bool = False,
        prefix: str = "",
        alt_stream: Optional[torch.cuda.Stream] = None,
    ) -> None:
        super().__init__()
        self.hidden_size = config.hidden_size
        self.config = config
        rope_theta = getattr(config, "rope_theta", 10000)
        rope_scaling = getattr(config, "rope_scaling", None)
        max_position_embeddings = getattr(config, "max_position_embeddings", 8192)
        self.enable_dp_attention = global_server_args_dict["enable_dp_attention"]
        self.speculative_algorithm = global_server_args_dict["speculative_algorithm"]
        self.layer_id = layer_id
        self.is_nextn = is_nextn
        self.self_attn = DeepseekV2AttentionMLA(
            config=config,
            hidden_size=self.hidden_size,
            num_heads=config.num_attention_heads,
            qk_nope_head_dim=config.qk_nope_head_dim,
            qk_rope_head_dim=config.qk_rope_head_dim,
            v_head_dim=config.v_head_dim,
            q_lora_rank=(
                config.q_lora_rank if hasattr(config, "q_lora_rank") else None
            ),
            kv_lora_rank=config.kv_lora_rank,
            rope_theta=rope_theta,
            rope_scaling=rope_scaling,
            max_position_embeddings=max_position_embeddings,
            quant_config=quant_config,
            layer_id=layer_id,
            reduce_results=False,
            prefix=add_prefix("self_attn", prefix),
            alt_stream=alt_stream,
        )

        self.is_layer_sparse = self._is_layer_sparse(layer_id, is_nextn=is_nextn)
        is_previous_layer_sparse = self._is_layer_sparse(layer_id - 1, is_nextn=False)

        self.layer_scatter_modes = LayerScatterModes.init_new(
            layer_id=layer_id,
            num_layers=1 if is_nextn else config.num_hidden_layers,
            is_layer_sparse=self.is_layer_sparse,
            is_previous_layer_sparse=is_previous_layer_sparse,
        )

        if self.is_layer_sparse:
            self.mlp = DeepseekV2MoE(
                config=config,
                quant_config=quant_config,
                prefix=add_prefix("mlp", prefix),
                layer_id=self.layer_id,
                alt_stream=alt_stream,
                is_nextn=is_nextn,
            )
        else:
            if enable_moe_dense_fully_dp():
                mlp_tp_rank, mlp_tp_size = 0, 1
            else:
                mlp_tp_rank, mlp_tp_size = None, None
            self.mlp = DeepseekV2MLP(
                hidden_size=config.hidden_size,
                intermediate_size=config.intermediate_size,
                hidden_act=config.hidden_act,
                quant_config=quant_config,
                prefix=add_prefix("mlp", prefix),
                tp_rank=mlp_tp_rank,
                tp_size=mlp_tp_size,
            )

        self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.post_attention_layernorm = RMSNorm(
            config.hidden_size, eps=config.rms_norm_eps
        )

        self.layer_communicator = LayerCommunicator(
            layer_scatter_modes=self.layer_scatter_modes,
            input_layernorm=self.input_layernorm,
            post_attention_layernorm=self.post_attention_layernorm,
        )

    def _is_layer_sparse(self, layer_id: int, is_nextn: bool) -> bool:
        return is_nextn or (
            self.config.n_routed_experts is not None
            and layer_id >= self.config.first_k_dense_replace
            and layer_id % self.config.moe_layer_freq == 0
        )

    def _should_fuse_mlp_allreduce_with_next_layer(self, forward_batch) -> bool:
        """Check if MLP allreduce can be fused with next layer's add_rmsnorm"""

        if (
            self.layer_id == self.config.num_hidden_layers - 1
            or get_tensor_model_parallel_world_size() <= 1
        ):
            return False

        if not global_server_args_dict.get("enable_flashinfer_allreduce_fusion", False):
            return False

        if not _is_sm100_supported or not _is_flashinfer_available:
            return False

        if hasattr(forward_batch, "input_ids") and (
            forward_batch.input_ids.shape[0] == 0
            or forward_batch.input_ids.shape[0] > 128
        ):
            return False

        return True

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        forward_batch: ForwardBatch,
        residual: Optional[torch.Tensor],
        zero_allocator: BumpAllocator,
    ) -> torch.Tensor:

        hidden_states, residual = self.layer_communicator.prepare_attn(
            hidden_states, residual, forward_batch
        )

        hidden_states = self.self_attn(
            positions=positions,
            hidden_states=hidden_states,
            forward_batch=forward_batch,
            zero_allocator=zero_allocator,
        )

        hidden_states, residual = self.layer_communicator.prepare_mlp(
            hidden_states, residual, forward_batch
        )

        can_fuse_mlp_allreduce = (
            self._should_fuse_mlp_allreduce_with_next_layer(forward_batch)
            and not (self.enable_dp_attention and self.speculative_algorithm.is_eagle())
            and not self.is_nextn
        )

        hidden_states = self.mlp(hidden_states, forward_batch, can_fuse_mlp_allreduce)

        if can_fuse_mlp_allreduce:
            hidden_states._sglang_needs_allreduce_fusion = True

        if not can_fuse_mlp_allreduce:
            hidden_states, residual = self.layer_communicator.postprocess_layer(
                hidden_states, residual, forward_batch
            )

        return hidden_states, residual

    def op_comm_prepare_attn(
        self,
        state,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        forward_batch: ForwardBatch,
        residual: Optional[torch.Tensor],
        zero_allocator: BumpAllocator,
        tbo_subbatch_index: Optional[int] = None,
    ):
        state.hidden_states_after_comm_pre_attn, state.residual_after_input_ln = (
            self.layer_communicator.prepare_attn(hidden_states, residual, forward_batch)
        )
        state.update(
            dict(
                forward_batch=forward_batch,
                positions=positions,
                zero_allocator=zero_allocator,
                tbo_subbatch_index=tbo_subbatch_index,
            )
        )

    def op_comm_prepare_mlp(self, state):
        state.hidden_states_mlp_input, state.residual_after_comm_pre_mlp = (
            self.layer_communicator.prepare_mlp(
                state.pop("hidden_states_after_attn"),
                state.pop("residual_after_input_ln"),
                state.forward_batch,
            )
        )

    def op_mlp(self, state):
        hidden_states = state.pop("hidden_states_mlp_input")
        if not (
            enable_moe_dense_fully_dp()
            and (not self.is_layer_sparse)
            and hidden_states.shape[0] == 0
        ):
            state.hidden_states_mlp_output = self.mlp(
                hidden_states, state.forward_batch
            )
        else:
            state.hidden_states_mlp_output = hidden_states

    def op_comm_postprocess_layer(self, state):
        hidden_states, residual = self.layer_communicator.postprocess_layer(
            state.pop("hidden_states_mlp_output"),
            state.pop("residual_after_comm_pre_mlp"),
            state.forward_batch,
        )

        output = dict(
            positions=state.positions,
            hidden_states=hidden_states,
            residual=residual,
            forward_batch=state.forward_batch,
            zero_allocator=state.zero_allocator,
            tbo_subbatch_index=state.tbo_subbatch_index,
        )

        state.clear(
            expect_keys={
                "positions",
                "forward_batch",
                "zero_allocator",
                "tbo_subbatch_index",
            }
        )
        return output


class DeepseekV2Model(nn.Module):
    fall_back_to_pt_during_load = False

    def __init__(
        self,
        config: PretrainedConfig,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.padding_id = config.pad_token_id
        self.vocab_size = config.vocab_size
        self.first_k_dense_replace = config.first_k_dense_replace

        self.embed_tokens = VocabParallelEmbedding(
            config.vocab_size,
            config.hidden_size,
            enable_tp=not global_server_args_dict["enable_dp_attention"],
        )
        self.alt_stream = torch.cuda.Stream() if _is_cuda else None
        self.layers = nn.ModuleList(
            [
                DeepseekV2DecoderLayer(
                    config,
                    layer_id,
                    quant_config=quant_config,
                    prefix=add_prefix(f"layers.{layer_id}", prefix),
                    alt_stream=self.alt_stream,
                )
                for layer_id in range(config.num_hidden_layers)
            ]
        )
        self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)

    def get_input_embeddings(self) -> torch.Tensor:
        return self.embed_tokens

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        forward_batch: ForwardBatch,
        input_embeds: torch.Tensor = None,
    ) -> torch.Tensor:
        total_num_layers = len(self.layers)
        device = input_embeds.device if input_embeds is not None else input_ids.device
        zero_allocator = BumpAllocator(
            buffer_size=total_num_layers * 2 * (2 if forward_batch.can_run_tbo else 1),
            dtype=torch.float32,
            device=device,
        )

        if input_embeds is None:
            hidden_states = self.embed_tokens(input_ids)
        else:
            hidden_states = input_embeds

        residual = None

        normal_num_layers = (
            self.first_k_dense_replace
            if forward_batch.can_run_tbo
            else total_num_layers
        )
        for i in range(normal_num_layers):
            with get_global_expert_distribution_recorder().with_current_layer(i):
                layer = self.layers[i]
                hidden_states, residual = layer(
                    positions, hidden_states, forward_batch, residual, zero_allocator
                )

        if normal_num_layers != total_num_layers:
            hidden_states, residual = model_forward_maybe_tbo(
                layers=self.layers[normal_num_layers:],
                enable_tbo=True,
                positions=positions,
                forward_batch=forward_batch,
                hidden_states=hidden_states,
                residual=residual,
                input_data_scatter_mode=self.layers[
                    normal_num_layers - 1
                ].layer_scatter_modes.layer_output_mode,
                zero_allocator=zero_allocator,
            )

        if not forward_batch.forward_mode.is_idle():
            if residual is None:
                hidden_states = self.norm(hidden_states)
            else:
                hidden_states, _ = self.norm(hidden_states, residual)
        return hidden_states


class DeepseekV2ForCausalLM(nn.Module):

    def __init__(
        self,
        config: PretrainedConfig,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.config = config
        self.tp_size = get_tensor_model_parallel_world_size()
        self.quant_config = quant_config
        self.determine_num_fused_shared_experts()
        self.model = DeepseekV2Model(
            config, quant_config, prefix=add_prefix("model", prefix)
        )
        self.lm_head = ParallelLMHead(
            config.vocab_size,
            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._routed_experts_weights_of_layer = LazyValue(
            lambda: {
                layer_id: layer.mlp.get_moe_weights()
                for layer_id, layer in enumerate(self.model.layers)
                if isinstance(layer.mlp, DeepseekV2MoE)
            }
        )

    @property
    def routed_experts_weights_of_layer(self):
        return self._routed_experts_weights_of_layer.value

    def determine_num_fused_shared_experts(
        self, architecture: str = "DeepseekV3ForCausalLM"
    ):
        self.num_fused_shared_experts = 0
        if global_server_args_dict["disable_shared_experts_fusion"]:
            return

        # Only Deepseek V3/R1 can use shared experts fusion optimization now.
        disable_reason = None
        if (
            not _is_cuda
            or torch.cuda.get_device_capability("cuda") < (8, 0)
            or self.config.architectures[0] != architecture
            or self.config.n_routed_experts != 256
            or self.config.n_shared_experts != 1
        ):
            disable_reason = "Only Deepseek V3/R1 on NV-platform with capability >= 80 can use shared experts fusion optimization."
        elif (
            global_server_args_dict["enable_deepep_moe"]
            or global_server_args_dict["enable_ep_moe"]
        ):
            disable_reason = "Deepseek V3/R1 can not use shared experts fusion optimization when in deepep_moe or ep_moe mode."

        if disable_reason is not None:
            global_server_args_dict["disable_shared_experts_fusion"] = True
            log_info_on_rank0(
                logger,
                f"{disable_reason} Shared experts fusion optimization is disabled.",
            )
            return

        self.num_fused_shared_experts = self.config.n_shared_experts

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

    @torch.no_grad()
    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        forward_batch: ForwardBatch,
        input_embeds: torch.Tensor = None,
    ) -> torch.Tensor:
        hidden_states = self.model(input_ids, positions, forward_batch, input_embeds)

        return self.logits_processor(
            input_ids, hidden_states, self.lm_head, forward_batch
        )

    def post_load_weights(self, is_nextn=False, weight_names=None):

        # Perform post-processing after loading weights
        if is_nextn:
            layer_ids = [self.config.num_hidden_layers]
        else:
            if weight_names is None:
                layer_ids = range(self.config.num_hidden_layers)
            else:
                layer_ids = set()
                for name in weight_names:
                    if "kv_b_proj" in name:
                        layer_id = int(name.split(".")[2])
                        if layer_id < self.config.num_hidden_layers:
                            layer_ids.add(layer_id)

        for layer_id in layer_ids:
            self_attn = (
                self.model.layers[layer_id].self_attn
                if not is_nextn
                else self.model.decoder.self_attn
            )
            if hasattr(self_attn.kv_b_proj, "qweight"):
                # AWQ compatible
                if _is_cuda or _is_hip:
                    w = awq_dequantize(
                        self_attn.kv_b_proj.qweight,
                        self_attn.kv_b_proj.scales,
                        self_attn.kv_b_proj.qzeros,
                    ).T
                else:
                    w = awq_dequantize(
                        self_attn.kv_b_proj.qweight,
                        self_attn.kv_b_proj.scales,
                        self_attn.kv_b_proj.qzeros,
                        0,
                        0,
                        0,
                    ).T
            else:
                w = self_attn.kv_b_proj.weight
            # NOTE(HandH1998): Since `bmm_fp8` only supports per-tensor scale, we have to requantize `self_attn.kv_b_proj`.
            # This may affect the accuracy of fp8 model.
            # Fix deepseek v3 blockwise bmm by using deep_gemm
            use_deep_gemm_bmm = False

            if w.dtype in (
                torch.float8_e4m3fn,
                torch.float8_e4m3fnuz,
            ):
                if (
                    hasattr(self.quant_config, "weight_block_size")
                    and self.quant_config.weight_block_size is not None
                ):
                    weight_block_size = self.quant_config.weight_block_size
                    assert hasattr(self_attn.kv_b_proj, "weight_scale_inv")
                    if _is_fp8_fnuz:
                        weight, weight_scale, _ = normalize_e4m3fn_to_e4m3fnuz(
                            weight=w,
                            weight_scale=self_attn.kv_b_proj.weight_scale_inv,
                            input_scale=None,
                        )
                    else:
                        weight = w
                        weight_scale = self_attn.kv_b_proj.weight_scale_inv

                    if (
                        _is_cuda
                        and weight_block_size[0] == 128
                        and weight_block_size[1] == 128
                    ):
                        if (
                            deep_gemm_wrapper.ENABLE_JIT_DEEPGEMM
                            and not deep_gemm_wrapper.DEEPGEMM_BLACKWELL
                            and get_bool_env_var("SGL_USE_DEEPGEMM_BMM", "false")
                        ):
                            block_scale = weight_scale
                            use_deep_gemm_bmm = True
                        else:
                            w = block_quant_dequant(
                                weight,
                                weight_scale,
                                weight_block_size,
                                torch.bfloat16,
                            )
                    else:
                        w, scale = block_quant_to_tensor_quant(
                            weight, weight_scale, weight_block_size
                        )
                        self_attn.w_scale = scale
                else:
                    if _is_fp8_fnuz:
                        weight, weight_scale, _ = normalize_e4m3fn_to_e4m3fnuz(
                            weight=w,
                            weight_scale=self_attn.kv_b_proj.weight_scale,
                            input_scale=None,
                        )
                    else:
                        weight = w
                        weight_scale = self_attn.kv_b_proj.weight_scale

                    w, scale = channel_quant_to_tensor_quant(weight, weight_scale)
                    self_attn.w_scale = scale

            if w.dtype == torch.int8:
                if hasattr(self.quant_config, "weight_block_size"):
                    # block-wise int8 need it
                    weight_block_size = self.quant_config.weight_block_size
                    if weight_block_size is not None:
                        assert hasattr(self_attn.kv_b_proj, "weight_scale_inv")
                        weight = w
                        weight_scale = self_attn.kv_b_proj.weight_scale_inv
                        w = int8_block_dequant(
                            weight, weight_scale, weight_block_size
                        ).to(torch.bfloat16)
                else:
                    # channel-wise int8 need it
                    w = w.to(torch.bfloat16) * self_attn.kv_b_proj.weight_scale.to(
                        torch.bfloat16
                    )

            w_kc, w_vc = w.unflatten(
                0, (-1, self_attn.qk_nope_head_dim + self_attn.v_head_dim)
            ).split([self_attn.qk_nope_head_dim, self_attn.v_head_dim], dim=1)
            if not use_deep_gemm_bmm:
                self_attn.w_kc = bind_or_assign(
                    self_attn.w_kc, w_kc.transpose(1, 2).contiguous().transpose(1, 2)
                )
                self_attn.w_vc = bind_or_assign(
                    self_attn.w_vc, w_vc.contiguous().transpose(1, 2)
                )
                if (
                    hasattr(self_attn.kv_b_proj, "weight_scale")
                    and self_attn.w_scale is None
                ):
                    self_attn.w_scale = bind_or_assign(
                        self_attn.w_scale, self_attn.kv_b_proj.weight_scale
                    )
                    if _is_hip:
                        self_attn.w_scale *= 2.0
                # TODO: remove this after adding FP8 support in bmm cpu kernel
                if _is_cpu and _is_cpu_amx_available and w.dtype == torch.float8_e4m3fn:
                    self_attn.w_kc = (
                        self_attn.w_kc.to(torch.bfloat16) * self_attn.w_scale
                    )
                    self_attn.w_vc = (
                        self_attn.w_vc.to(torch.bfloat16) * self_attn.w_scale
                    )
            else:
                num_tiles_k = self_attn.qk_nope_head_dim // weight_block_size[1]
                num_tiles_n = self_attn.v_head_dim // weight_block_size[0]
                ws_kc, ws_vc = block_scale.unflatten(
                    0, (-1, (num_tiles_k + num_tiles_n))
                ).split([num_tiles_k, num_tiles_n], dim=1)
                self_attn.w_scale_k = bind_or_assign(
                    self_attn.w_scale_k, ws_kc.transpose(1, 2).contiguous()
                )
                self_attn.w_scale_v = bind_or_assign(
                    self_attn.w_scale_v, ws_vc.contiguous()
                )
                self_attn.w_kc = bind_or_assign(
                    self_attn.w_kc, w_kc.transpose(1, 2).contiguous()
                )
                self_attn.w_vc = bind_or_assign(self_attn.w_vc, w_vc.contiguous())
                self_attn.use_deep_gemm_bmm = True

        if (
            deep_gemm_wrapper.ENABLE_JIT_DEEPGEMM
            and deep_gemm_wrapper.DEEPGEMM_SCALE_UE8M0
            and hasattr(self.quant_config, "weight_block_size")
            and self.quant_config.weight_block_size is not None
        ):
            self._weight_requant_ue8m0(is_nextn)

    def _weight_requant_ue8m0(self, is_nextn=False):
        weight_block_size = self.quant_config.weight_block_size

        moe_layers = list(
            range(
                self.config.first_k_dense_replace,
                self.config.num_hidden_layers,
                self.config.moe_layer_freq,
            )
        )

        num_hidden_layers = 1 if is_nextn else self.config.num_hidden_layers
        for layer_id in range(num_hidden_layers):
            if is_nextn:
                layer = self.model.decoder
            else:
                layer = self.model.layers[layer_id]

            for module in [
                layer.self_attn.fused_qkv_a_proj_with_mqa,
                layer.self_attn.q_b_proj,
                layer.self_attn.kv_b_proj,
                layer.self_attn.o_proj,
            ]:
                requant_weight_ue8m0_inplace(
                    module.weight, module.weight_scale_inv, weight_block_size
                )

            if layer_id in moe_layers or is_nextn:
                shared_experts = getattr(layer.mlp, "shared_experts", None)
                if shared_experts is not None:
                    for module in [
                        shared_experts.gate_up_proj,
                        shared_experts.down_proj,
                    ]:
                        requant_weight_ue8m0_inplace(
                            module.weight, module.weight_scale_inv, weight_block_size
                        )

                experts = layer.mlp.experts
                if isinstance(experts, DeepEPMoE):
                    for w in [
                        experts.w13_weight_fp8,
                        experts.w2_weight_fp8,
                    ]:
                        requant_weight_ue8m0_inplace(w[0], w[1], weight_block_size)
            else:
                mlp = layer.mlp
                assert isinstance(mlp, DeepseekV2MLP)
                for module in [
                    mlp.gate_up_proj,
                    mlp.down_proj,
                ]:
                    requant_weight_ue8m0_inplace(
                        module.weight, module.weight_scale_inv, weight_block_size
                    )

    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]], is_nextn=False):

        if is_nextn:
            if hasattr(self.config, "num_nextn_predict_layers"):
                num_nextn_layers = self.config.num_nextn_predict_layers
                assert num_nextn_layers == 1, "Only 1 nextn layer is supported"
                # compatible with old design
                nextn_layer_id = (
                    0
                    if self.config.num_hidden_layers == 1
                    else self.config.num_hidden_layers
                )
            else:
                raise ValueError("num_nextn_predict_layers is not in the config")

        stacked_params_mapping = [
            # (param_name, shard_name, shard_id)
            ("gate_up_proj", "gate_proj", 0),
            ("gate_up_proj", "up_proj", 1),
        ]

        # Params for weights, fp8 weight scales, fp8 activation scales
        # (param_name, weight_name, expert_id, shard_id)
        expert_params_mapping = get_moe_impl_class().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.n_routed_experts + self.num_fused_shared_experts,
        )
        if self.quant_config and self.quant_config.get_name() == "w4afp8":
            expert_params_mapping += (
                get_moe_impl_class().make_expert_input_scale_params_mapping(
                    num_experts=self.config.n_routed_experts
                )
            )

        # Fuse q_a_proj and kv_a_proj_with_mqa along output dimension when q_lora_rank is not None
        fuse_qkv_a_proj = hasattr(self.config, "q_lora_rank") and (
            self.config.q_lora_rank is not None
        )
        cached_a_proj = {} if fuse_qkv_a_proj else None

        if is_nextn:
            nextn_layer_prefix = f"model.layers.{nextn_layer_id}"
            nextn_spec_weight_names = [
                "shared_head.norm",
                "eh_proj",
                "enorm",
                "hnorm",
            ]

        if self.num_fused_shared_experts > 0:
            assert self.num_fused_shared_experts == 1
            log_info_on_rank0(logger, "Shared experts fusion optimization enabled.")

        with concurrent.futures.ThreadPoolExecutor() as executor:
            futures = []
            params_dict = dict(self.named_parameters())
            weight_names = []
            for name, loaded_weight in weights:
                if self.num_fused_shared_experts > 0 and "mlp.shared_experts" in name:
                    name = name.replace(
                        "mlp.shared_experts",
                        f"mlp.experts.{self.config.n_routed_experts}",
                    )

                weight_names.append(name)

                if not is_nextn:
                    if hasattr(self.config, "num_nextn_predict_layers"):
                        num_nextn_layers = self.config.num_nextn_predict_layers
                        if num_nextn_layers > 0 and name.startswith("model.layers"):
                            name_list = name.split(".")
                            if (
                                len(name_list) >= 3
                                and int(name_list[2]) >= self.config.num_hidden_layers
                            ):
                                continue
                else:
                    if not name.startswith(nextn_layer_prefix):
                        continue

                    # Use shared head and embed weights from target model
                    if "shared_head.head" in name or "embed_tokens" in name:
                        continue

                    is_decoder = True
                    # For nextn specific weights
                    for weight_name in nextn_spec_weight_names:
                        if weight_name in name:
                            name = name.replace(nextn_layer_prefix, "model")
                            is_decoder = False
                            break
                    # For decoder layer weights
                    if is_decoder:
                        name = name.replace(nextn_layer_prefix, "model.decoder")

                if "rotary_emb.inv_freq" in name:
                    continue
                for param_name, weight_name, shard_id in stacked_params_mapping:
                    # Skip non-stacked layers and experts (experts handled below).
                    if weight_name not 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) and name not in params_dict:
                        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
                    futures.append(
                        executor.submit(weight_loader, param, loaded_weight, shard_id)
                    )
                    break
                else:
                    for mapping in expert_params_mapping:
                        param_name, weight_name, expert_id, shard_id = mapping
                        if weight_name not in name:
                            continue
                        name = name.replace(weight_name, param_name)
                        param = params_dict[name]
                        weight_loader = param.weight_loader
                        futures.append(
                            executor.submit(
                                weight_loader,
                                param,
                                loaded_weight,
                                name,
                                shard_id=shard_id,
                                expert_id=expert_id,
                            )
                        )
                        break
                    else:
                        # Skip loading extra bias for GPTQ models.
                        if name.endswith(".bias") and name not in params_dict:
                            continue
                        if fuse_qkv_a_proj and (
                            "q_a_proj" in name or "kv_a_proj_with_mqa" in name
                        ):
                            cached_a_proj[name] = loaded_weight
                            q_a_proj_name = (
                                name
                                if "q_a_proj" in name
                                else name.replace("kv_a_proj_with_mqa", "q_a_proj")
                            )
                            kv_a_proj_name = (
                                name
                                if "kv_a_proj_with_mqa" in name
                                else name.replace("q_a_proj", "kv_a_proj_with_mqa")
                            )

                            # When both q_a_proj and kv_a_proj_with_mqa has been cached, load the fused weight to parameter
                            if (
                                q_a_proj_name in cached_a_proj
                                and kv_a_proj_name in cached_a_proj
                            ):
                                q_a_proj_weight = cached_a_proj[q_a_proj_name]
                                kv_a_proj_weight = cached_a_proj[kv_a_proj_name]
                                cat_dim = 0
                                if self.quant_config is not None and (
                                    self.quant_config.get_name() == "awq"
                                    or self.quant_config.get_name() == "awq_marlin"
                                    or self.quant_config.get_name() == "moe_wna16"
                                ):
                                    cat_dim = 1
                                fused_weight = torch.cat(
                                    [q_a_proj_weight, kv_a_proj_weight], dim=cat_dim
                                )
                                param_name = (
                                    name.replace(
                                        "q_a_proj", "fused_qkv_a_proj_with_mqa"
                                    )
                                    if "q_a_proj" in name
                                    else name.replace(
                                        "kv_a_proj_with_mqa",
                                        "fused_qkv_a_proj_with_mqa",
                                    )
                                )
                                param = params_dict[param_name]

                                weight_loader = getattr(
                                    param, "weight_loader", default_weight_loader
                                )
                                futures.append(
                                    executor.submit(weight_loader, param, fused_weight)
                                )
                                cached_a_proj.pop(q_a_proj_name)
                                cached_a_proj.pop(kv_a_proj_name)
                        else:
                            if (
                                "k_scale" in name or "v_scale" in name
                            ) and name not in params_dict:
                                # modelopt attn kv scale is named differently
                                for scale in ["k_scale", "v_scale"]:
                                    if scale in name:
                                        name = name.replace(
                                            f"{scale[0]}_proj", "attn_mqa"
                                        )
                                        break
                            if name not in params_dict:
                                # modelopt ckpt contains not needed weights for MTP module:
                                # model.decoder.self_attn.attn_mqa.v_scale and
                                # model.decoder.self_attn.attn_mqa.k_scale
                                logger.warning(f"{name} not found in params_dict.")
                                continue
                            param = params_dict[name]
                            weight_loader = getattr(
                                param, "weight_loader", default_weight_loader
                            )
                            futures.append(
                                executor.submit(weight_loader, param, loaded_weight)
                            )

            # Wait for all tasks to complete and raise any exceptions.
            for future in concurrent.futures.as_completed(futures):
                future.result()

        self.post_load_weights(is_nextn=is_nextn, weight_names=weight_names)

    def get_embed_and_head(self):
        return self.model.embed_tokens.weight, self.lm_head.weight

    def set_embed_and_head(self, embed, head):
        del self.model.embed_tokens.weight
        del self.lm_head.weight
        self.model.embed_tokens.weight = embed
        self.lm_head.weight = head
        torch.cuda.empty_cache()
        torch.cuda.synchronize()

    @classmethod
    def get_model_config_for_expert_location(cls, config):
        return ModelConfigForExpertLocation(
            num_layers=config.num_hidden_layers,
            num_logical_experts=config.n_routed_experts,
            num_groups=config.n_group,
        )


class DeepseekV3ForCausalLM(DeepseekV2ForCausalLM):
    pass


EntryClass = [DeepseekV2ForCausalLM, DeepseekV3ForCausalLM]