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 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.layers.activation import SiluAndMul
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 get_moe_impl_class
from sglang.srt.layers.moe.ep_moe.token_dispatcher import DeepEPDispatcher
from sglang.srt.layers.moe.topk import select_experts
from sglang.srt.layers.quantization.base_config import QuantizationConfig
from sglang.srt.layers.quantization.deep_gemm import _ENABLE_JIT_DEEPGEMM
from sglang.srt.layers.quantization.fp8_kernel import (
    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,
)
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
from sglang.srt.layers.vocab_parallel_embedding import (
    ParallelLMHead,
    VocabParallelEmbedding,
)
from sglang.srt.managers.expert_distribution import (
    ExpertDistributionRecorder,
    get_global_expert_distribution_recorder,
)
from sglang.srt.managers.expert_location import ModelConfigForExpertLocation
from sglang.srt.managers.expert_location_dispatch import ExpertLocationDispatchInfo
from sglang.srt.managers.schedule_batch import global_server_args_dict
from sglang.srt.model_executor.forward_batch_info import ForwardBatch, ForwardMode
from sglang.srt.model_loader.weight_utils import default_weight_loader
from sglang.srt.operations import execute_operations
from sglang.srt.operations_strategy import compute_layer_operations
from sglang.srt.utils import (
    BumpAllocator,
    DeepEPMode,
    add_prefix,
    get_bool_env_var,
    get_int_env_var,
    is_cuda,
    is_hip,
    log_info_on_rank0,
)

_is_hip = is_hip()
_is_cuda = is_cuda()

if _is_cuda:
    from sgl_kernel import awq_dequantize, bmm_fp8, merge_state_v2

    from sglang.srt.layers.quantization.deep_gemm import (
        grouped_gemm_nt_f8f8bf16_masked as deep_gemm_grouped_gemm_nt_f8f8bf16_masked,
    )
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,
    )

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()


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):
        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)
        return x


class MoEGate(nn.Module):
    def __init__(
        self,
        config,
        prefix: str = "",
    ):
        super().__init__()
        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))
            )
        else:
            self.e_score_correction_bias = None

    def forward(self, hidden_states):
        logits = F.linear(hidden_states, self.weight, None)
        return logits


def is_non_idle_and_non_empty(forward_mode, hidden_states):
    return (
        (forward_mode is not None)
        and not forward_mode.is_idle()
        and hidden_states.shape[0] > 0
    )


class DeepseekV2MoE(nn.Module):

    def __init__(
        self,
        config: PretrainedConfig,
        layer_id: int,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ):
        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.n_share_experts_fusion = global_server_args_dict["n_share_experts_fusion"]
        self.layer_id = layer_id

        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))

        self.experts = get_moe_impl_class()(
            num_experts=config.n_routed_experts + self.n_share_experts_fusion,
            top_k=config.num_experts_per_tok + min(self.n_share_experts_fusion, 1),
            hidden_size=config.hidden_size,
            intermediate_size=config.moe_intermediate_size,
            layer_id=self.layer_id,
            renormalize=config.norm_topk_prob,
            quant_config=quant_config,
            use_grouped_topk=True,
            num_expert_group=config.n_group,
            topk_group=config.topk_group,
            correction_bias=self.gate.e_score_correction_bias,
            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 {}
            ),
        )

        if config.n_shared_experts is not None and self.n_share_experts_fusion == 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 {}
                ),
            )

        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
            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 = DeepEPDispatcher(
                group=parallel_state.get_tp_group().device_group,
                router_topk=self.top_k,
                permute_fusion=True,
                num_experts=config.n_routed_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,  # TODO
                return_recv_hook=True,
            )

    @property
    def _enable_deepep_moe(self):
        return global_server_args_dict["enable_deepep_moe"]

    def op_gate(self, state):
        if (not self._enable_deepep_moe) or 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):
        if (self.n_share_experts_fusion == 0) and (
            (not self._enable_deepep_moe)
            or is_non_idle_and_non_empty(
                state.forward_batch.forward_mode, state.hidden_states_mlp_input
            )
        ):
            state.shared_output = self.shared_experts(state.hidden_states_mlp_input)
        else:
            state.shared_output = None

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

        if self._enable_deepep_moe:
            if router_logits is not None:
                state.topk_weights_local, state.topk_idx_local = select_experts(
                    hidden_states=hidden_states,
                    router_logits=router_logits,
                    top_k=self.top_k,
                    use_grouped_topk=True,
                    renormalize=self.renormalize,
                    topk_group=self.topk_group,
                    num_expert_group=self.num_expert_group,
                    correction_bias=self.correction_bias,
                    routed_scaling_factor=self.routed_scaling_factor,
                    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._enable_deepep_moe and (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.pop("hidden_states_mlp_input"),
                topk_idx=state.pop("topk_idx_local"),
                topk_weights=state.pop("topk_weights_local"),
                forward_mode=state.forward_batch.forward_mode,
            )

    def op_dispatch_b(self, state):
        if self._enable_deepep_moe and (self.ep_size > 1):
            (
                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()

    def op_experts(self, state):
        if self._enable_deepep_moe:
            state.pop("router_logits")
            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_mode=state.forward_batch.forward_mode,
            )
        else:
            state.hidden_states_experts_output = self.experts(
                hidden_states=state.pop("hidden_states_mlp_input"),
                router_logits=state.pop("router_logits"),
            )

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

    def op_combine_b(self, state):
        if self._enable_deepep_moe and (self.ep_size > 1):
            state.hidden_states_after_combine = self.deepep_dispatcher.combine_b()

    def op_output(self, state):
        final_hidden_states = (
            state.pop("hidden_states_after_combine")
            if self._enable_deepep_moe
            else state.pop("hidden_states_experts_output")
        )

        final_hidden_states *= self.routed_scaling_factor

        if (s := state.pop("shared_output")) is not None:
            final_hidden_states = final_hidden_states + s

        if (not self._enable_deepep_moe) and (self.tp_size > 1):
            final_hidden_states = tensor_model_parallel_all_reduce(final_hidden_states)

        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(
            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,
        )

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

        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.attention_backend = global_server_args_dict["attention_backend"]
        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
        )

    def dispatch_attn_forward_method(
        self, forward_batch: ForwardBatch
    ) -> AttnForwardMethod:
        if self.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 AttnForwardMethod.MLA
        elif self.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 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 AttnForwardMethod.MLA

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        forward_batch: ForwardBatch,
        zero_allocator: BumpAllocator,
    ) -> torch.Tensor:
        if hidden_states.shape[0] == 0:
            assert (
                not self.o_proj.reduce_results
            ), "short-circuiting allreduce will lead to hangs"
            return hidden_states

        attn_forward_method = self.dispatch_attn_forward_method(forward_batch)

        if attn_forward_method == AttnForwardMethod.MHA:
            return self.forward_normal(positions, hidden_states, forward_batch)
        elif attn_forward_method == AttnForwardMethod.MHA_CHUNKED_KV:
            return self.forward_normal_chunked_kv(
                positions, hidden_states, forward_batch
            )
        else:
            if _is_hip:
                if (
                    self.rocm_fused_decode_mla
                    and forward_batch.forward_mode.is_decode()
                ):
                    return self.forward_absorb_fused_mla_rope(
                        positions, hidden_states, forward_batch
                    )
                else:
                    return self.forward_absorb(
                        positions, hidden_states, forward_batch, zero_allocator
                    )
            else:
                return self.forward_absorb(
                    positions, hidden_states, forward_batch, zero_allocator
                )

    def forward_normal(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        forward_batch: ForwardBatch,
    ) -> torch.Tensor:
        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.contiguous())
        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
        )
        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(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        forward_batch: ForwardBatch,
        zero_allocator: BumpAllocator,
    ) -> torch.Tensor:
        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
            )
            k_nope = latent_cache[..., : self.kv_lora_rank]

            # overlap qk norm
            if self.alt_stream is not None and torch.cuda.is_current_stream_capturing():
                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_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 self.w_kc.dtype == torch.float8_e4m3fnuz:
            # TODO(kernel): add bmm_fp8 for torch.float8_e4m3fnuz
            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)

        if self.attention_backend == "fa3" or self.attention_backend == "flashinfer":
            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_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, :]
        elif self.w_vc.dtype == torch.float8_e4m3fnuz:
            # TODO(kernel): add bmm_fp8 for torch.float8_e4m3fnuz
            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),
            )
            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(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        forward_batch: ForwardBatch,
        zero_allocator: BumpAllocator,
    ) -> torch.Tensor:
        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 self.w_kc.dtype == torch.float8_e4m3fnuz:
            # TODO(kernel): add bmm_fp8 for torch.float8_e4m3fnuz
            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]

        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 self.w_vc.dtype == torch.float8_e4m3fnuz:
            # TODO(kernel): add bmm_fp8 for torch.float8_e4m3fnuz
            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 _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(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        forward_batch: ForwardBatch,
    ) -> torch.Tensor:
        # 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.contiguous())
        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
        )

        # 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.layer_id = layer_id
        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=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,
            )
        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 forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        forward_batch: ForwardBatch,
        residual: Optional[torch.Tensor],
        zero_allocator: BumpAllocator,
    ) -> torch.Tensor:
        return execute_operations(
            inputs=dict(
                positions=positions,
                hidden_states=hidden_states,
                forward_batch=forward_batch,
                residual=residual,
                zero_allocator=zero_allocator,
            ),
            operations=compute_layer_operations(self),
        )

    def op_comm_prepare_attn(
        self,
        state,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        forward_batch: ForwardBatch,
        residual: Optional[torch.Tensor],
        zero_allocator: BumpAllocator,
    ):
        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,
            )
        )

    def op_attn(self, state):
        state.hidden_states_after_attn = self.self_attn(
            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_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.forward_mode
            )
        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,
        )

        state.clear(expect_keys={"positions", "forward_batch", "zero_allocator"})
        return hidden_states, residual


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.embed_tokens = VocabParallelEmbedding(
            config.vocab_size,
            config.hidden_size,
            enable_tp=not global_server_args_dict["enable_dp_attention"],
        )
        # TODO(haishaw): multi-stream performance on ROCm
        self.alt_stream = None if _is_hip else torch.cuda.Stream()
        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)

        self.dp_size = get_local_attention_dp_size()

    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:
        zero_allocator = BumpAllocator(
            # TODO for two-batch-overlap, we need a larger buffer size
            buffer_size=len(self.layers) * 2,
            dtype=torch.float32,
            device=(
                input_embeds.device if input_embeds is not None else input_ids.device
            ),
        )

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

        residual = None
        for i in range(len(self.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 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_n_share_experts_fusion()
        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.dp_size = get_local_attention_dp_size()

    def determine_n_share_experts_fusion(
        self, architecture: str = "DeepseekV3ForCausalLM"
    ):
        self.n_share_experts_fusion = global_server_args_dict["n_share_experts_fusion"]
        if self.n_share_experts_fusion > 0:
            # Only Deepseek V3/R1 can use shared experts fusion optimization now.
            if (
                not _is_cuda
                or self.config.architectures[0] != architecture
                or self.config.n_routed_experts != 256
            ):
                self.n_share_experts_fusion = 0
                global_server_args_dict["n_share_experts_fusion"] = 0
                log_info_on_rank0(
                    logger,
                    "Only Deepseek V3/R1 on NV-platform can use shared experts fusion optimization. Shared experts fusion optimization is disabled.",
                )
            else:
                assert (
                    self.n_share_experts_fusion == self.tp_size
                ), f"Shared experts fusion optimization is enabled in DeepSeek V3/R1, set it to {self.tp_size} can get best optimized performance."
        elif self.n_share_experts_fusion == 0:
            if (
                _is_cuda
                and torch.cuda.get_device_capability("cuda") >= (9, 0)
                and self.config.architectures[0] == architecture
                and self.config.n_routed_experts == 256
                and (not global_server_args_dict["enable_deepep_moe"])
            ):
                self.n_share_experts_fusion = self.tp_size
                global_server_args_dict["n_share_experts_fusion"] = self.tp_size
                log_info_on_rank0(
                    logger,
                    "Deepseek V3/R1 with fp8 can use shared experts fusion optimization when SM version >=90. Shared experts fusion optimization is enabled.",
                )

    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):

        # Perform post-processing after loading weights
        layer_ids = (
            range(self.config.num_hidden_layers)
            if not is_nextn
            else [self.config.num_hidden_layers]
        )
        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:
                    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
            model_dtype = torch.get_default_dtype()

            if w.dtype in (
                torch.float8_e4m3fn,
                torch.float8_e4m3fnuz,
            ):
                if hasattr(self.quant_config, "weight_block_size"):
                    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")
                        if _is_hip:
                            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
                            and model_dtype == torch.bfloat16
                        ):
                            if _ENABLE_JIT_DEEPGEMM 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,
                                    model_dtype,
                                )
                        else:
                            w, scale = block_quant_to_tensor_quant(
                                weight, weight_scale, weight_block_size
                            )
                            self_attn.w_scale = scale
                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 = w_kc.transpose(1, 2).contiguous().transpose(1, 2)
                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 = self_attn.kv_b_proj.weight_scale
                    if _is_hip:
                        self_attn.w_scale *= 2.0
            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 = ws_kc.transpose(1, 2).contiguous()
                self_attn.w_scale_v = ws_vc.contiguous()
                self_attn.w_kc = w_kc.transpose(1, 2).contiguous()
                self_attn.w_vc = w_vc.contiguous()
                self_attn.use_deep_gemm_bmm = True

    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),
        ]
        if self.n_share_experts_fusion > 0:
            weights_list = list(weights)
            weights_dict = dict(weights_list)
            if self.quant_config is None or self.quant_config.get_name() == "w8a8_int8":
                suffix_list = [
                    "down_proj.weight",
                    "down_proj.weight_scale",
                    "gate_proj.weight",
                    "gate_proj.weight_scale",
                    "up_proj.weight",
                    "up_proj.weight_scale",
                ]
            else:
                suffix_list = [
                    "down_proj.weight",
                    "down_proj.weight_scale_inv",
                    "gate_proj.weight",
                    "gate_proj.weight_scale_inv",
                    "up_proj.weight",
                    "up_proj.weight_scale_inv",
                ]
            names_to_remove = []

            moe_layers = (
                range(
                    self.config.first_k_dense_replace,
                    self.config.num_hidden_layers,
                    self.config.moe_layer_freq,
                )
                if not is_nextn
                else [nextn_layer_id]
            )

            for moe_layer in tqdm(
                moe_layers,
                desc=f"Cloning {self.n_share_experts_fusion} "
                "replicas of the shared expert into MoE",
            ):
                for suffix in suffix_list:
                    shared_expert_weight_name = (
                        f"model.layers.{moe_layer}.mlp.shared_experts.{suffix}"
                    )
                    for num_repeat in range(self.n_share_experts_fusion):
                        weights_list.append(
                            (
                                f"model.layers.{moe_layer}."
                                f"mlp.experts."
                                f"{self.config.n_routed_experts + num_repeat}"
                                f".{suffix}",
                                weights_dict[shared_expert_weight_name],
                            )
                        )
                    names_to_remove += [shared_expert_weight_name]
            weights = [w for w in weights_list if w[0] not in names_to_remove]

        # 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.n_share_experts_fusion,
        )

        # 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",
            ]

        params_dict = dict(self.named_parameters())
        for name, loaded_weight in weights:
            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
                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
                    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]
                            fused_weight = torch.cat(
                                [q_a_proj_weight, kv_a_proj_weight], dim=0
                            )

                            param_name = name.replace(
                                "q_a_proj", "fused_qkv_a_proj_with_mqa"
                            )
                            param = params_dict[param_name]

                            weight_loader = getattr(
                                param, "weight_loader", default_weight_loader
                            )
                            weight_loader(param, fused_weight)
                            cached_a_proj.pop(q_a_proj_name)
                            cached_a_proj.pop(kv_a_proj_name)
                    else:
                        param = params_dict[name]
                        weight_loader = getattr(
                            param, "weight_loader", default_weight_loader
                        )
                        weight_loader(param, loaded_weight)

        self.post_load_weights(is_nextn=is_nextn)

    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]