inference/modeling_openpangu.py

# coding=utf-8
# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
# Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved.
#
# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
# and OPT implementations in this library. It has been modified from its
# original forms to accommodate minor architectural differences compared
# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
#
# 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.

from collections.abc import Iterable
from typing import Any, Optional, Union, Callable

import torch
from torch import nn
import torch_npu

from vllm.attention import Attention, AttentionType, AttentionMetadata
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, VllmConfig, get_current_vllm_config
from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
from vllm.model_executor.layers.activation import SiluAndMul
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (MergedColumnParallelLinear,
                                               QKVParallelLinear,
                                               RowParallelLinear)
from vllm.model_executor.layers.logits_processor import LogitsProcessor
from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.layers.rotary_embedding import get_rope
from vllm.model_executor.layers.vocab_parallel_embedding import (
    DEFAULT_VOCAB_PADDING_SIZE, ParallelLMHead, VocabParallelEmbedding)
from vllm.model_executor.model_loader.weight_utils import (
    default_weight_loader, sharded_weight_loader, row_parallel_weight_loader, maybe_remap_kv_scale_name)
from vllm.sequence import IntermediateTensors

from vllm.model_executor.models.interfaces import SupportsLoRA, SupportsPP
from vllm.model_executor.models.utils import (AutoWeightsLoader, PPMissingLayer, extract_layer_index,
                    is_pp_missing_parameter,
                    make_empty_intermediate_tensors_factory, make_layers,
                    maybe_prefix)
from vllm.forward_context import ForwardContext, get_forward_context
from vllm.utils import direct_register_custom_op

from configuration_openpangu_dense import PanguEmbeddedConfig


def aggregate_hiddden(
        hidden_states: torch.Tensor, 
        cache_states: torch.Tensor, 
        cache_length: torch.Tensor, 
        fn_name: str, 
        aggre_output: torch.Tensor
        ) -> torch.Tensor:
    """
    input_hidden.shape = (S, H) or (B, H)

    conv(H, S) or (B, H, 1)
            ^            ^
    return.shape = (S, H) or (B, H)
    """
    forward_context: ForwardContext = get_forward_context()
    attn_metadata = forward_context.attn_metadata
    if attn_metadata is None: #dummy run
        return hidden_states

    aggregate_fn = forward_context.no_compile_layers[fn_name]
    num_tokens, hidden_dim = hidden_states.shape

    cache_slot_id = forward_context.cache_slot_id
    query_start_loc = forward_context.query_start_loc

    if forward_context.with_prefill:
        is_first_chunk = forward_context.is_first_chunk
        for i, q_start in enumerate(query_start_loc[:-1]):
            slot_id = cache_slot_id[i]
            q_end = query_start_loc[i+1]
            aggre_input = torch.empty(
                (cache_length + q_end - q_start, hidden_dim),
                device=hidden_states.device, dtype=hidden_states.dtype
            )

            if is_first_chunk[i]:
                aggre_input[:cache_length].fill_(0)
            else:
                aggre_input[:cache_length].copy_(cache_states[slot_id, :cache_length])
            aggre_input[cache_length:].copy_(hidden_states[q_start:q_end])

            aggre_input[cache_length:].copy_(hidden_states[q_start:q_end])
            output = aggregate_fn(aggre_input.permute(1, 0))
            aggre_output[q_start:q_end].copy_(output.permute(1, 0))
            cache_states[slot_id, :cache_length].copy_(aggre_input[-cache_length:])
        return aggre_output
    else:
        # decode stage
        num_tokens = query_start_loc[-1]
        cache_slot_id_t = cache_slot_id.unsqueeze(0).permute(1, 0)
        torch_npu.npu_scatter_nd_update_(cache_states[:, -1, :], cache_slot_id_t, hidden_states[:num_tokens])
        aggre_input = cache_states[cache_slot_id].permute(0, 2, 1)
        aggre_output[:num_tokens] = aggregate_fn(aggre_input).squeeze(2)
        torch_npu.npu_scatter_nd_update_(cache_states[:, :cache_length, :], cache_slot_id_t,
                                         cache_states[cache_slot_id, -cache_length:, :])
        return aggre_output

def aggregate_hiddden_fake(
        hidden_states: torch.Tensor, 
        cache_states: torch.Tensor, 
        cache_length: torch.Tensor, 
        fn_name: str, 
        aggre_output: torch.Tensor
        ) -> torch.Tensor:
    return hidden_states

direct_register_custom_op(
    op_name="aggregate_hiddden",
    op_func=aggregate_hiddden,
    mutates_args=["cache_states", "aggre_output"],
    fake_impl=aggregate_hiddden_fake,
)


class PanguEmbeddedMLP(nn.Module):

    def __init__(
        self,
        hidden_size: int,
        intermediate_size: int,
        hidden_act: str,
        quant_config: Optional[QuantizationConfig] = None,
        bias: bool = False,
        prefix: str = "",
        reduce_results: bool = True,
    ) -> None:
        super().__init__()
        self.gate_up_proj = MergedColumnParallelLinear(
            input_size=hidden_size,
            output_sizes=[intermediate_size] * 2,
            bias=bias,
            quant_config=quant_config,
            prefix=f"{prefix}.gate_up_proj",
        )
        self.down_proj = RowParallelLinear(
            input_size=intermediate_size,
            output_size=hidden_size,
            bias=bias,
            quant_config=quant_config,
            reduce_results=reduce_results,
            prefix=f"{prefix}.down_proj",
        )
        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):
        x, _ = self.gate_up_proj(x)
        x = self.act_fn(x)
        x, _ = self.down_proj(x)
        return x


class PanguEmbeddedAttention(nn.Module):

    def __init__(
        self,
        config: PanguEmbeddedConfig,
        hidden_size: int,
        num_heads: int,
        num_kv_heads: int,
        rope_theta: float = 10000,
        rope_scaling: Optional[dict[str, Any]] = None,
        max_position_embeddings: int = 8192,
        quant_config: Optional[QuantizationConfig] = None,
        bias: bool = False,
        bias_o_proj: bool = False,
        cache_config: Optional[CacheConfig] = None,
        prefix: str = "",
        attn_type: str = AttentionType.DECODER,
    ) -> None:
        super().__init__()
        layer_idx = extract_layer_index(prefix)
        self.hidden_size = hidden_size
        tp_size = get_tensor_model_parallel_world_size()
        self.total_num_heads = num_heads
        self.num_heads = self.total_num_heads // tp_size
        self.total_num_kv_heads = num_kv_heads
        self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
        # MistralConfig has an optional head_dim introduced by Mistral-Nemo
        head_dim = getattr(config, "head_dim", None)
        if head_dim is None:
            head_dim = self.hidden_size // self.total_num_heads
        self.head_dim = head_dim
        # Phi models introduced a partial_rotary_factor parameter in the config
        self.partial_rotary_factor = getattr(config, "partial_rotary_factor", 1)
        self.q_size = self.num_heads * self.head_dim
        self.kv_size = self.num_kv_heads * self.head_dim
        self.scaling = self.head_dim**-0.5
        self.rope_theta = rope_theta
        self.rotary_dim = getattr(config, "qk_rope_dim", head_dim)
        self.max_position_embeddings = max_position_embeddings
        self.v_channels = getattr(config, "v_channels", None)

        self.qkv_proj = QKVParallelLinear(
            hidden_size=hidden_size,
            head_size=self.head_dim,
            total_num_heads=self.total_num_heads,
            total_num_kv_heads=self.total_num_kv_heads,
            bias=bias,
            quant_config=quant_config,
            prefix=f"{prefix}.qkv_proj",
        )

        self.o_proj = RowParallelLinear(
            input_size=self.total_num_heads * self.head_dim,
            output_size=hidden_size,
            bias=bias_o_proj,
            quant_config=quant_config,
            prefix=f"{prefix}.o_proj",
        )

        self._init_rotary_emb(config,
                              rope_scaling=rope_scaling,
                              quant_config=quant_config)

        if hasattr(config, "interleaved_sliding_window"):
            interleaved_sliding_window = config.interleaved_sliding_window
            if isinstance(interleaved_sliding_window, int):
                sliding_window = interleaved_sliding_window
            elif isinstance(interleaved_sliding_window, list):
                sw_idx = layer_idx % len(interleaved_sliding_window)
                sliding_window = interleaved_sliding_window[sw_idx]
            else:
                raise ValueError(
                    f"{type(interleaved_sliding_window)} is not supported.")
        else:
            sliding_window = None

        self.attn = Attention(
            self.num_heads,
            self.head_dim,
            self.scaling,
            num_kv_heads=self.num_kv_heads,
            cache_config=cache_config,
            quant_config=quant_config,
            per_layer_sliding_window=sliding_window,
            attn_type=attn_type,
            prefix=f"{prefix}.attn",
            sinks={}
        )
        # Patch for Sink
        param_sink_number = getattr(config, 'param_sink_number', 0)
        param_sink_with_value = getattr(config, 'param_sink_with_value', False)
        if param_sink_number > 0:
            self.enable_sink = True
            self.param_sink_query = torch.zeros((
                param_sink_number,
                self.num_heads,
                self.head_dim),
                dtype=config.torch_dtype
                )
            self.param_sink_key = torch.nn.Parameter(
                torch.empty((
                    param_sink_number,
                    self.num_kv_heads,
                    self.head_dim),
                    dtype=config.torch_dtype
                )
            )
            if param_sink_with_value:
                self.param_sink_value = torch.nn.Parameter(
                    torch.empty((
                        param_sink_number,
                        self.num_kv_heads,
                        self.v_channels),
                        dtype=config.torch_dtype
                    )
                )
            else:
                self.param_sink_value = torch.zeros((
                    param_sink_number,
                    self.num_kv_heads,
                    self.v_channels),
                    dtype=config.torch_dtype
                )
        else:
            self.enable_sink = False

        attn_groupnorm = getattr(config, 'attn_groupnorm', False)
        if attn_groupnorm:
            self.groupnorm = RMSNorm(hidden_size=self.head_dim, eps=config.rms_norm_eps)
        else:
            self.groupnorm = None

        attn_elementwise_gate = getattr(config, 'attn_elementwise_gate', False)
        if attn_elementwise_gate:
            self.attention_gate = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
        else:
            self.attention_gate = None

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
        qkv, _ = self.qkv_proj(hidden_states)
        q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
        q, k = self.rotary_emb(positions, q, k)
        attn_output = self.attn(
            q, k, v,
            ** (dict(
                sink_query=self.param_sink_query,
                sink_key=self.param_sink_key,
                sink_value=self.param_sink_value,
                v_head_size=self.v_channels
            ) if self.enable_sink else {})
        )
        # groupnorm (s, h, d)
        if self.groupnorm is not None:
            num_tokens, hidden_dim = attn_output.shape
            attn_norm = attn_output.view(num_tokens, self.num_heads, self.head_dim)
            attn_norm = self.groupnorm(attn_norm)
            attn_output = attn_norm.view(num_tokens, hidden_dim)
        # gate (s, h*d)
        if self.attention_gate is not None:
            gate_score = self.attention_gate(hidden_states)
            attn_output = attn_output * torch.sigmoid(gate_score)
        output, _ = self.o_proj(attn_output)
        return output

    def _init_rotary_emb(self, config: PanguEmbeddedConfig,
                         rope_scaling: Optional[dict[str, Any]],
                         quant_config: Optional[QuantizationConfig]) -> None:
        is_neox_style = True
        is_gguf = quant_config and quant_config.get_name() == "gguf"
        if is_gguf and config.model_type == "Pangu":
            is_neox_style = False

        self.rotary_emb = get_rope(
            self.head_dim,
            rotary_dim=self.rotary_dim,
            max_position=self.max_position_embeddings,
            base=self.rope_theta,
            rope_scaling=rope_scaling,
            is_neox_style=is_neox_style,
        )


class PanguEmbeddedDecoderLayer(nn.Module):

    def __init__(
        self,
        config: PanguEmbeddedConfig,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ) -> None:
        super().__init__()
        torch_npu.npu.config.allow_internal_format = False
        self.hidden_size = config.hidden_size
        rope_theta = getattr(config, "rope_theta", 10000)
        rope_scaling = getattr(config, "rope_scaling", None)
        if rope_scaling is not None and getattr(
                config, "original_max_position_embeddings", None):
            rope_scaling["original_max_position_embeddings"] = (
                config.original_max_position_embeddings)
        max_position_embeddings = getattr(config, "max_position_embeddings",
                                          8192)
        # Support abacusai/Smaug-72B-v0.1 with attention_bias
        # Support internlm/internlm-7b with bias
        attention_bias = getattr(config, "attention_bias", False) or getattr(
            config, "bias", False)
        bias_o_proj = attention_bias
        # support internlm/internlm3-8b with qkv_bias
        if hasattr(config, 'qkv_bias'):
            attention_bias = config.qkv_bias

        # By default, PanguEmbedded uses causal attention as it is a decoder-only model.
        # You can override the HF config with `is_causal=False` to enable
        # bidirectional attention, which is used in some embedding models
        # (e.g. parasail-ai/GritLM-7B-vllm)
        if getattr(config, "is_causal", True):
            attn_type = AttentionType.DECODER
        else:
            attn_type = AttentionType.ENCODER_ONLY

        self.self_attn = PanguEmbeddedAttention(
            config=config,
            hidden_size=self.hidden_size,
            num_heads=config.num_attention_heads,
            num_kv_heads=getattr(config, "num_key_value_heads",
                                 config.num_attention_heads),
            rope_theta=rope_theta,
            rope_scaling=rope_scaling,
            max_position_embeddings=max_position_embeddings,
            quant_config=quant_config,
            bias=attention_bias,
            bias_o_proj=bias_o_proj,
            cache_config=cache_config,
            prefix=f"{prefix}.self_attn",
            attn_type=attn_type,
        )
        self.mlp = PanguEmbeddedMLP(
            hidden_size=self.hidden_size,
            intermediate_size=config.intermediate_size,
            hidden_act=config.hidden_act,
            quant_config=quant_config,
            bias=getattr(config, "mlp_bias", False),
            prefix=f"{prefix}.mlp",
        )
        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)

        # merge_conv
        layer_idx = extract_layer_index(prefix)
        self.router_sliding_window = getattr(config, 'router_sliding_window', 0)
        if self.router_sliding_window > 1 and layer_idx in [0, config.num_hidden_layers - 1]:
            self.merge_conv = torch.nn.Conv1d(
                in_channels=config.hidden_size,
                out_channels=config.hidden_size,
                kernel_size=self.router_sliding_window,
                groups=config.hidden_size,
                bias=False,
            )
            vllm_config = get_current_vllm_config()
            self.max_num_seqs = vllm_config.scheduler_config.max_num_seqs
            self.cache_states = \
                torch.zeros((self.max_num_seqs, self.router_sliding_window, config.hidden_size), device='npu')
            self.cache_length = torch.tensor(self.router_sliding_window - 1).npu()
            # add conv to static_forward_context
            self.conv_name = f"{prefix}.conv"
            vllm_config.compilation_config.static_forward_context[self.conv_name] = self.merge_conv

        else:
            self.merge_conv = None
            self.cache_states = None

    def aggregate(self, hidden_states: torch.Tensor) -> torch.Tensor:
        aggre_output = torch.zeros((hidden_states.shape), dtype=hidden_states.dtype, device=hidden_states.device)
        torch.ops.vllm.aggregate_hiddden(
                        hidden_states=hidden_states, 
                        cache_states=self.cache_states, 
                        cache_length=self.cache_length,
                        fn_name=self.conv_name, 
                        aggre_output=aggre_output
                        )
        return aggre_output

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        residual: Optional[torch.Tensor] = None,
    ) -> tuple[torch.Tensor, torch.Tensor]:
        # Self Attention
        if residual is None:
            residual = hidden_states
            hidden_states = self.input_layernorm(hidden_states)
        else:
            hidden_states, residual = self.input_layernorm(
                hidden_states, residual)
        hidden_states = self.self_attn(positions=positions,
                                       hidden_states=hidden_states)

        # Add
        hidden_states = residual + hidden_states
        residual = hidden_states
        # Conv
        if self.merge_conv is not None:
            hidden_states = self.aggregate(hidden_states=hidden_states)

        # Fully Connected
        hidden_states = self.post_attention_layernorm(hidden_states)
        hidden_states = self.mlp(hidden_states)

        return hidden_states, residual


@support_torch_compile
class PanguEmbeddedModel(nn.Module):

    def __init__(self,
                 *,
                 vllm_config: VllmConfig,
                 prefix: str = "",
                 layer_type: type[nn.Module] = PanguEmbeddedDecoderLayer):
        super().__init__()

        config = vllm_config.model_config.hf_config
        cache_config = vllm_config.cache_config
        quant_config = vllm_config.quant_config
        lora_config = vllm_config.lora_config

        self.config = config
        self.quant_config = quant_config
        lora_vocab = (lora_config.lora_extra_vocab_size *
                      (lora_config.max_loras or 1)) if lora_config else 0
        self.vocab_size = config.vocab_size + lora_vocab
        self.org_vocab_size = config.vocab_size
        if get_pp_group().is_first_rank or (config.tie_word_embeddings
                                            and get_pp_group().is_last_rank):
            self.embed_tokens = VocabParallelEmbedding(
                self.vocab_size,
                config.hidden_size,
                org_num_embeddings=config.vocab_size,
                quant_config=quant_config,
                prefix=f"{prefix}.embed_tokens",
            )
        else:
            self.embed_tokens = PPMissingLayer()
        self.start_layer, self.end_layer, self.layers = make_layers(
            config.num_hidden_layers,
            lambda prefix: layer_type(config=config,
                                      cache_config=cache_config,
                                      quant_config=quant_config,
                                      prefix=prefix),
            prefix=f"{prefix}.layers",
        )
        if get_pp_group().is_last_rank:
            self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        else:
            self.norm = PPMissingLayer()

        self.aux_hidden_state_layers: tuple[int] = tuple()

        self.make_empty_intermediate_tensors = (
            make_empty_intermediate_tensors_factory(
                ["hidden_states", "residual"], config.hidden_size))

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

    def forward(
        self,
        input_ids: Optional[torch.Tensor],
        positions: torch.Tensor,
        intermediate_tensors: Optional[IntermediateTensors],
        inputs_embeds: Optional[torch.Tensor] = None,
    ) -> Union[torch.Tensor, IntermediateTensors, tuple[torch.Tensor,
                                                        list[torch.Tensor]]]:
        if get_pp_group().is_first_rank:
            if inputs_embeds is not None:
                hidden_states = inputs_embeds
            else:
                hidden_states = self.get_input_embeddings(input_ids)
            residual = None
        else:
            hidden_states = intermediate_tensors["hidden_states"]
            residual = intermediate_tensors["residual"]

        aux_hidden_states = []
        for idx, layer in enumerate(
                self.layers[self.start_layer:self.end_layer]):
            if idx in self.aux_hidden_state_layers:
                aux_hidden_states.append(hidden_states + residual)
            hidden_states, residual = layer(positions, hidden_states, residual)

        if not get_pp_group().is_last_rank:
            return IntermediateTensors({
                "hidden_states": hidden_states,
                "residual": residual
            })

        hidden_states, _ = self.norm(hidden_states, residual)

        if len(aux_hidden_states) > 0:
            return hidden_states, aux_hidden_states
        return hidden_states

    def load_weights(self, weights: Iterable[tuple[str,
                                                   torch.Tensor]]) -> set[str]:
        stacked_params_mapping = [
            # (param_name, shard_name, shard_id)
            (".qkv_proj", ".q_proj", "q"),
            (".qkv_proj", ".k_proj", "k"),
            (".qkv_proj", ".v_proj", "v"),
            (".gate_up_proj", ".gate_proj", 0),
            (".gate_up_proj", ".up_proj", 1),
        ]
        # skip second norms.1.weights
        skip_unneeded_norm = (not isinstance(self.norm, nn.ModuleList))

        params_dict = dict(self.named_parameters())
        loaded_params: set[str] = set()
        for name, loaded_weight in weights:
            if valid_name_layer(name, self.end_layer):
                continue
            if skip_unneeded_norm and name.startswith('norms.'):
                norm_idx = int(name.split('norms.')[-1].split('.')[0])
                if norm_idx > 0:
                    continue
                name = name.replace(f"norms.{norm_idx}",
                                    f"norm")

            if "rotary_emb.inv_freq" in name:
                continue
            if ("rotary_emb.cos_cached" in name
                    or "rotary_emb.sin_cached" in name):
                # Models trained using ColossalAI may include these tensors in
                # the checkpoint. Skip them.
                continue
            if (self.quant_config is not None and
                (scale_name := self.quant_config.get_cache_scale(name))):
                # Loading kv cache quantization scales
                param = params_dict[scale_name]
                weight_loader = getattr(param, "weight_loader",
                                        default_weight_loader)
                loaded_weight = (loaded_weight if loaded_weight.dim() == 0 else
                                 loaded_weight[0])
                weight_loader(param, loaded_weight)
                loaded_params.add(scale_name)
                continue
            if "scale" in name:
                # Remapping the name of FP8 kv-scale.
                name = maybe_remap_kv_scale_name(name, params_dict)
                if name is None:
                    continue
            for param_name, weight_name, shard_id in stacked_params_mapping:
                if weight_name not in name:
                    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

                if is_pp_missing_parameter(name, self):
                    continue

                param = params_dict[name]
                weight_loader = param.weight_loader
                weight_loader(param, loaded_weight, shard_id)
                break
            else:
                # Skip loading extra bias for GPTQ models.
                if name.endswith(".bias") and name not in params_dict:
                    continue

                if is_pp_missing_parameter(name, self):
                    continue

                param = params_dict[name]
                if name.endswith("param_sink_key") or name.endswith("param_sink_value"):
                    weight_loader = getattr(param, "weight_loader", sharded_weight_loader(-2)) # [S,N,D]
                elif name.endswith("attention_gate.weight"):
                    weight_loader = getattr(param, "weight_loader", row_parallel_weight_loader)
                else:
                    weight_loader = getattr(param, "weight_loader", default_weight_loader)
                weight_loader(param, loaded_weight)
            loaded_params.add(name)
        return loaded_params


class PanguEmbeddedForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
    packed_modules_mapping = {
        "qkv_proj": ["q_proj", "k_proj", "v_proj"],
        "gate_up_proj": ["gate_proj", "up_proj"]
    }

    # LoRA specific attributes
    embedding_modules = {
        "embed_tokens": "input_embeddings",
        "lm_head": "output_embeddings"
    }
    embedding_padding_modules = ["lm_head"]

    # Mistral/PanguEmbedded models can also be loaded with --load-format mistral
    # from consolidated.safetensors checkpoints
    mistral_mapping = {
        "layers": "model.layers",
        "attention": "self_attn",
        "qscale_act": "input_scale",
        "qscale_weight": "weight_scale",
        "kv_fake_quantizer.qscale_act": "kv_scale",
        "wq": "q_proj",
        "wk": "k_proj",
        "wv": "v_proj",
        "wo": "o_proj",
        "attention_norm": "input_layernorm",
        "feed_forward": "mlp",
        "w1": "gate_proj",
        "w2": "down_proj",
        "w3": "up_proj",
        "ffn_norm": "post_attention_layernorm",
        "tok_embeddings": "model.embed_tokens",
        "output": "lm_head",
        "norm": "model.norm",
    }

    def __init__(self,
                 *,
                 vllm_config: VllmConfig,
                 prefix: str = "",
                 layer_type: type[nn.Module] = PanguEmbeddedDecoderLayer):
        super().__init__()
        config = vllm_config.model_config.hf_config
        quant_config = vllm_config.quant_config
        lora_config = vllm_config.lora_config
        self.config = config
        self.lora_config = lora_config

        self.model = self._init_model(vllm_config=vllm_config,
                                      prefix=maybe_prefix(prefix, "model"),
                                      layer_type=layer_type)

        if get_pp_group().is_last_rank:
            self.unpadded_vocab_size = config.vocab_size
            if lora_config:
                self.unpadded_vocab_size += lora_config.lora_extra_vocab_size
            self.lm_head = ParallelLMHead(
                self.unpadded_vocab_size,
                config.hidden_size,
                org_num_embeddings=config.vocab_size,
                padding_size=(
                    DEFAULT_VOCAB_PADDING_SIZE
                    # We need bigger padding if using lora for kernel
                    # compatibility
                    if not lora_config else
                    lora_config.lora_vocab_padding_size),
                quant_config=quant_config,
                prefix=maybe_prefix(prefix, "lm_head"),
            )
            if config.tie_word_embeddings:
                self.lm_head = self.lm_head.tie_weights(
                    self.model.embed_tokens)

            logit_scale = getattr(config, "logit_scale", 1.0)
            self.logits_processor = LogitsProcessor(self.unpadded_vocab_size,
                                                    config.vocab_size,
                                                    logit_scale)
        else:
            self.lm_head = PPMissingLayer()

        self.make_empty_intermediate_tensors = (
            self.model.make_empty_intermediate_tensors)

    def set_aux_hidden_state_layers(self, layers: tuple[int]) -> None:
        self.model.aux_hidden_state_layers = layers

    def get_eagle3_aux_hidden_state_layers(self) -> tuple[int]:
        num_layers = len(self.model.layers)
        return (2, num_layers // 2, num_layers - 3)

    def _init_model(self,
                    vllm_config: VllmConfig,
                    prefix: str = "",
                    layer_type: type[nn.Module] = PanguEmbeddedDecoderLayer):
        return PanguEmbeddedModel(vllm_config=vllm_config,
                          prefix=prefix,
                          layer_type=layer_type)

    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
        return self.model.get_input_embeddings(input_ids)

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        intermediate_tensors: Optional[IntermediateTensors] = None,
        inputs_embeds: Optional[torch.Tensor] = None,
    ) -> Union[torch.Tensor, IntermediateTensors]:
        model_output = self.model(input_ids, positions, intermediate_tensors,
                                  inputs_embeds)
        return model_output

    def compute_logits(
        self,
        hidden_states: torch.Tensor,
    ) -> Optional[torch.Tensor]:
        logits = self.logits_processor(self.lm_head, hidden_states)
        return logits

    def load_weights(self, weights: Iterable[tuple[str,
                                                   torch.Tensor]]) -> set[str]:
        loader = AutoWeightsLoader(
            self,
            skip_prefixes=(["lm_head."]
                           if self.config.tie_word_embeddings else None),
        )
        return loader.load_weights(
            self.maybe_remap_mistral(name, loaded_weight)
            for name, loaded_weight in weights)

    # This function is used to remap the mistral format as
    # used by Mistral and PanguEmbedded <=2
    def maybe_remap_mistral(
        self,
        name: str,
        loaded_weight: torch.Tensor,
    ) -> tuple[str, torch.Tensor]:

        def permute(w: torch.Tensor, n_heads: int):
            attn_in = self.config.head_dim * n_heads
            attn_out = self.config.hidden_size

            return w.view(n_heads, attn_in // n_heads // 2, 2,
                          attn_out).transpose(1, 2).reshape(attn_in, attn_out)

        mapping = self.mistral_mapping
        modules = name.split(".")

        # rotary embeds should be sliced
        if "wk" in modules and modules[-1] == "weight":
            loaded_weight = permute(loaded_weight,
                                    self.config.num_key_value_heads)
        elif "wq" in modules and modules[-1] == "weight":
            loaded_weight = permute(loaded_weight,
                                    self.config.num_attention_heads)

        num_modules = len(modules)
        for i in range(num_modules):
            item = modules[i]
            next_item = modules[i + 1] if i < num_modules - 1 else None

            combined_item = (f"{item}.{next_item}"
                             if next_item is not None else None)

            if combined_item in mapping:
                name = name.replace(combined_item, mapping[combined_item])
            elif item in mapping and mapping[item] not in name:
                name = name.replace(item, mapping[item])

        return name, loaded_weight

def valid_name_layer(name: str, end_layer: int) -> bool:
    if "layers" in name:
        layer_idx = extract_layer_index(name)
        if layer_idx >= end_layer:
            return True
    return False