vllm-cpu-amxbf16 0.9.1__cp312-cp312-manylinux_2_17_x86_64.whl

vllm/model_executor/models/medusa.py

@@ -0,0 +1,219 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+from collections.abc import Iterable
+from typing import Optional
+
+import torch
+import torch.nn as nn
+
+from vllm.config import VllmConfig
+from vllm.model_executor.layers.logits_processor import LogitsProcessor
+from vllm.model_executor.layers.sampler import SamplerOutput
+from vllm.model_executor.layers.vocab_parallel_embedding import (
+    DEFAULT_VOCAB_PADDING_SIZE, ParallelLMHead)
+from vllm.model_executor.model_loader.weight_utils import default_weight_loader
+from vllm.model_executor.sampling_metadata import SamplingMetadata
+
+
+class ResidualBlock(nn.Module):
+
+    def __init__(self, config: VllmConfig, hidden_size: int,
+                 num_layers: int) -> None:
+        super().__init__()
+
+        self.layers = nn.ModuleList([
+            nn.Linear(hidden_size,
+                      hidden_size,
+                      bias=getattr(config, "medusa_fc_bias", False))
+            for _ in range(num_layers)
+        ])
+        self.act = nn.SiLU()
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        for layer in self.layers:
+            x = x + self.act(layer(x))
+        return x
+
+
+class Medusa(nn.Module):
+    """This class implements the Medusa draft model from the paper: https://arxiv.org/abs/2401.10774
+    Reference implementation: https://github.com/FasterDecoding/Medusa
+    
+    Differences from reference implementation:
+    1. Currently this only supports generating proposals from top-1 tokens.
+    2. We have an optional token_map which reduces draft vocab to most 
+       frequently used tokens to give some additional speed-up by reducing 
+       sampling overhead. This is disabled unless the checkpoint file has 
+       explicit token_map tensor and config has an optional attribute 
+       truncated_vocab_size < vocab_size. To use this technique, one has to find
+       the top-k most frequent tokens in target dataset and add that as a tensor
+       in the draft checkpoint (using key token_map). Also, the draft config
+       needs to have truncated_vocab_size (=k) as an attribute."""
+
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = "") -> None:
+        config = vllm_config.speculative_config.draft_model_config.hf_config
+        super().__init__()
+        self.config = config
+        self.blocks = nn.ModuleList([
+            ResidualBlock(config=config,
+                          hidden_size=self.config.hidden_size,
+                          num_layers=self.config.num_hidden_layers)
+            for _ in range(self.config.num_heads)
+        ])
+        self.orig_vocab_size = config.vocab_size
+        self.truncated_vocab_size = config.truncated_vocab_size
+        self.unpadded_vocab_size = self.truncated_vocab_size
+
+        if getattr(config, "original_lm_head", False):
+            self.lm_head = ParallelLMHead(
+                self.unpadded_vocab_size,
+                config.hidden_size,
+                org_num_embeddings=self.truncated_vocab_size,
+                padding_size=DEFAULT_VOCAB_PADDING_SIZE,
+            )
+            self.lm_heads = [
+                self.lm_head for _ in range(self.config.num_heads)
+            ]
+        else:
+            self.lm_heads = nn.ModuleList([
+                ParallelLMHead(
+                    self.unpadded_vocab_size,
+                    config.hidden_size,
+                    org_num_embeddings=self.truncated_vocab_size,
+                    padding_size=DEFAULT_VOCAB_PADDING_SIZE,
+                ) for _ in range(self.config.num_heads)
+            ])
+
+        logit_scale = getattr(config, "logit_scale", 1.0)
+        self.logits_processor = LogitsProcessor(self.unpadded_vocab_size,
+                                                self.truncated_vocab_size,
+                                                logit_scale)
+
+        # Token map is a idx to token mapping to reduce the vocab size for
+        # the draft model. Using smaller vocab size for draft, containing
+        # only most frequent tokens reduces the speculation overhead. This
+        # doesn't affect the acceptance rate much and thus gives more speed
+        # -up. By default, this is disabled and is only used if the EAGLE
+        # checkpoint file has token_map tensor.
+        self.token_map = None
+
+    def forward(self, hidden_states: torch.Tensor) -> list[torch.Tensor]:
+        return [block(hidden_states) for block in self.blocks]
+
+    def compute_logits(
+            self, hidden_states: list[torch.Tensor],
+            sampling_metadata: SamplingMetadata) -> list[torch.Tensor]:
+        logits_lst: list[torch.Tensor] = []
+
+        for hs, lm_head in zip(hidden_states, self.lm_heads):
+            _logits = self.logits_processor(lm_head, hs, sampling_metadata)
+
+            if _logits is None:
+                # _logits should only be None on rank > 0, in which case
+                # it should remain true for every lm_head
+                assert len(logits_lst) == 0
+                continue
+
+            if self.token_map is None:
+                logits_lst.append(_logits)
+            else:
+                logits_lst.append(-torch.inf * torch.ones(
+                    size=(*_logits.shape[:-1], self.orig_vocab_size),
+                    device=_logits.device,
+                    dtype=_logits.dtype))
+
+                logits_lst[-1][..., self.token_map] = _logits
+
+        return logits_lst
+
+    def sample(
+        self,
+        logits: list[torch.Tensor],
+        sampling_metadata: SamplingMetadata,
+    ) -> list[SamplerOutput]:
+        logits = torch.stack(logits, dim=0).float()
+        logprobs = torch.log_softmax(logits, dim=-1)
+        token_ids = logits.argmax(-1)  # support only top-1 for now
+        probs = torch.softmax(logits, dim=-1)
+
+        token_id_list = []
+        token_prob_list = []
+        token_logprob_list = []
+
+        for idx, seq_group in enumerate(sampling_metadata.seq_groups):
+            token_id_list.append(token_ids[:, seq_group.sample_indices])
+            token_prob_list.append(probs[:, seq_group.sample_indices])
+            token_logprob_list.append(logprobs[:, seq_group.sample_indices])
+
+        outputs: list[Optional[SamplerOutput]] = []
+        for idx in range(len(sampling_metadata.seq_groups)):
+            outputs.append(
+                SamplerOutput(
+                    outputs=None,
+                    sampled_token_probs=token_prob_list[idx].squeeze(1),
+                    logprobs=token_logprob_list[idx].squeeze(1),
+                    sampled_token_ids=token_id_list[idx].squeeze(1),
+                ))
+
+        return outputs
+
+    def generate_proposals(
+        self,
+        previous_hidden_states: torch.Tensor,
+        sampling_metadata: SamplingMetadata,
+    ) -> Optional[list[SamplerOutput]]:
+        # During preemption, we may receive an empty tensor (batch_size=0)
+        if previous_hidden_states.size(0) == 0:
+            # Return None to signal the Top1Proposer that no proposals
+            # were generated for this batch, allowing it to handle this
+            # special case appropriately
+            return None
+
+        return self.sample(
+            logits=self.compute_logits(
+                hidden_states=self.forward(previous_hidden_states),
+                sampling_metadata=sampling_metadata,
+            ),
+            sampling_metadata=sampling_metadata,
+        )
+
+    def load_weights(self, weights: Iterable[tuple[str,
+                                                   torch.Tensor]]) -> set[str]:
+        params_dict = dict(self.named_parameters())
+        loaded_params: set[str] = set()
+
+        weights_map = {}
+
+        for name, loaded_weight in weights:
+            name = name.replace("medusa_heads.", "")
+
+            if name == "token_map":
+                if self.truncated_vocab_size < self.orig_vocab_size:
+                    self.token_map = nn.Parameter(loaded_weight,
+                                                  requires_grad=False)
+            elif name in params_dict:
+                weights_map[name] = loaded_weight
+            elif (getattr(self.config, "original_lm_head", False)
+                  and name == "lm_heads.0.weight"):
+                weights_map["lm_head.weight"] = loaded_weight
+
+        for name, loaded_weight in weights_map.items():
+            if "lm_head" in name and self.token_map is not None and\
+                loaded_weight.shape[0] > self.token_map.shape[0]:
+
+                loaded_weight = loaded_weight[self.token_map]
+
+            param = params_dict[name]
+            weight_loader = getattr(param, "weight_loader",
+                                    default_weight_loader)
+            weight_loader(param, loaded_weight)
+            loaded_params.add(name)
+
+        if self.token_map is not None:
+            self.token_map.to(device=self.lm_heads[0].weight.device)
+
+        assert (self.truncated_vocab_size
+                == self.orig_vocab_size) or (self.token_map is not None)
+
+        return loaded_params

vllm/model_executor/models/mimo.py

@@ -0,0 +1,192 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+# Adapted from
+# https://github.com/huggingface/transformers/blob/v4.28.0/src/transformers/models/qwen2/modeling_qwen2.py
+# Copyright 2025 Xiaomi Corporation.
+# Copyright 2024 The Qwen team.
+# Copyright 2023 The vLLM team.
+# 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.
+"""Inference-only MiMo model compatible with HuggingFace weights."""
+from collections.abc import Iterable
+from typing import Optional, Union
+
+import torch
+import torch.nn as nn
+
+from vllm.compilation.decorators import support_torch_compile
+from vllm.config import VllmConfig
+from vllm.distributed import get_pp_group
+from vllm.logger import init_logger
+from vllm.model_executor.layers.logits_processor import LogitsProcessor
+from vllm.model_executor.layers.sampler import get_sampler
+from vllm.model_executor.layers.vocab_parallel_embedding import ParallelLMHead
+from vllm.model_executor.model_loader.weight_utils import (
+    default_weight_loader, maybe_remap_kv_scale_name)
+from vllm.model_executor.models.qwen2 import Qwen2ForCausalLM, Qwen2Model
+from vllm.model_executor.sampling_metadata import SamplingMetadata
+from vllm.sequence import IntermediateTensors
+
+from .utils import PPMissingLayer, is_pp_missing_parameter, maybe_prefix
+
+logger = init_logger(__name__)
+
+
+@support_torch_compile(
+    dynamic_arg_dims={
+        "input_ids": 0,
+        "positions": -1,
+        "intermediate_tensors": 0,
+        "inputs_embeds": 0,
+    })
+class MiMoModel(Qwen2Model):
+
+    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]:
+        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:
+            assert intermediate_tensors is not None
+            hidden_states = intermediate_tensors["hidden_states"]
+            residual = intermediate_tensors["residual"]
+        for layer in self.layers[self.start_layer:self.end_layer]:
+            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 = hidden_states + residual
+        return hidden_states
+
+    def load_weights(self, weights: Iterable[tuple[str,
+                                                   torch.Tensor]]) -> set[str]:
+        stacked_params_mapping = [
+            ("qkv_proj", "q_proj", "q"),
+            ("qkv_proj", "k_proj", "k"),
+            ("qkv_proj", "v_proj", "v"),
+            ("gate_up_proj", "gate_proj", 0),
+            ("gate_up_proj", "up_proj", 1),
+        ]
+        params_dict = dict(self.named_parameters(remove_duplicate=False))
+        loaded_params: set[str] = set()
+        for name, loaded_weight in weights:
+            if "mtp_layers" in name:
+                continue
+            if "rotary_emb.inv_freq" in name:
+                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
+            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
+                # Remapping the name of FP8 kv-scale.
+                name = maybe_remap_kv_scale_name(name, params_dict)
+                if name is None:
+                    continue
+                if is_pp_missing_parameter(name, self):
+                    continue
+                param = params_dict[name]
+                weight_loader = getattr(param, "weight_loader",
+                                        default_weight_loader)
+                weight_loader(param, loaded_weight)
+            loaded_params.add(name)
+        return loaded_params
+
+
+class MiMoForCausalLM(Qwen2ForCausalLM, nn.Module):
+
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
+        nn.Module.__init__(self)
+        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.quant_config = quant_config
+
+        self.model = MiMoModel(vllm_config=vllm_config,
+                               prefix=maybe_prefix(prefix, "model"))
+
+        if get_pp_group().is_last_rank:
+            if config.tie_word_embeddings:
+                self.lm_head = self.model.embed_tokens
+            else:
+                self.lm_head = ParallelLMHead(config.vocab_size,
+                                              config.hidden_size,
+                                              quant_config=quant_config,
+                                              prefix=maybe_prefix(
+                                                  prefix, "lm_head"))
+        else:
+            self.lm_head = PPMissingLayer()
+
+        self.logits_processor = LogitsProcessor(config.vocab_size)
+        self.sampler = get_sampler()
+
+        self.make_empty_intermediate_tensors = (
+            self.model.make_empty_intermediate_tensors)
+
+    def compute_logits(
+        self,
+        hidden_states: torch.Tensor,
+        sampling_metadata: SamplingMetadata,
+    ) -> Optional[torch.Tensor]:
+        hidden_states = self.model.norm(hidden_states)
+        logits = self.logits_processor(self.lm_head, hidden_states,
+                                       sampling_metadata)
+        return logits

vllm/model_executor/models/mimo_mtp.py

@@ -0,0 +1,285 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+# Adapted from
+# https://github.com/vllm-project/vllm/blob/v0.7.3/vllm/model_executor/models/deepseek_mtp.py
+# Copyright 2025 Xiaomi Corporation.
+# Copyright 2023 The vLLM team.
+# Copyright 2024 DeepSeek-AI team.
+
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Inference-only MiMo-MTP model."""
+from collections.abc import Iterable
+from typing import Optional
+
+import torch
+import torch.nn as nn
+from transformers import PretrainedConfig
+
+from vllm.config import CacheConfig, ModelConfig, VllmConfig
+from vllm.model_executor.layers.layernorm import RMSNorm
+from vllm.model_executor.layers.logits_processor import LogitsProcessor
+from vllm.model_executor.layers.quantization import QuantizationConfig
+from vllm.model_executor.layers.sampler import SamplerOutput, get_sampler
+from vllm.model_executor.layers.vocab_parallel_embedding import (
+    ParallelLMHead, VocabParallelEmbedding)
+from vllm.model_executor.model_loader.weight_utils import default_weight_loader
+from vllm.model_executor.models.qwen2 import Qwen2DecoderLayer
+from vllm.model_executor.sampling_metadata import SamplingMetadata
+from vllm.sequence import IntermediateTensors
+
+from .utils import maybe_prefix
+
+
+class MiMoMultiTokenPredictorLayer(nn.Module):
+
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        prefix: str,
+        model_config: ModelConfig,
+        cache_config: Optional[CacheConfig] = None,
+        quant_config: Optional[QuantizationConfig] = None,
+    ) -> None:
+        super().__init__()
+
+        self.token_layernorm = RMSNorm(config.hidden_size,
+                                       eps=config.rms_norm_eps)
+        self.hidden_layernorm = RMSNorm(config.hidden_size,
+                                        eps=config.rms_norm_eps)
+        self.input_proj = nn.Linear(config.hidden_size * 2,
+                                    config.hidden_size,
+                                    bias=False)
+        self.mtp_block = Qwen2DecoderLayer(config=config,
+                                           cache_config=cache_config,
+                                           quant_config=quant_config,
+                                           prefix=prefix)
+        self.final_layernorm = RMSNorm(config.hidden_size,
+                                       eps=config.rms_norm_eps)
+
+    def forward(
+        self,
+        inputs_embeds: torch.Tensor,
+        positions: torch.Tensor,
+        previous_hidden_states: torch.Tensor,
+        spec_step_index: int = 0,
+    ) -> torch.Tensor:
+        assert inputs_embeds is not None
+        # masking inputs at position 0, as not needed by MTP
+        inputs_embeds[positions == 0] = 0
+        inputs_embeds = self.token_layernorm(inputs_embeds)
+        previous_hidden_states = self.hidden_layernorm(previous_hidden_states)
+
+        hidden_states = self.input_proj(
+            torch.cat([previous_hidden_states, inputs_embeds], dim=-1))
+
+        hidden_states, residual = self.mtp_block(positions=positions,
+                                                 hidden_states=hidden_states,
+                                                 residual=None)
+        hidden_states = residual + hidden_states
+        return self.final_layernorm(hidden_states)
+
+
+class MiMoMultiTokenPredictor(nn.Module):
+
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
+        super().__init__()
+
+        config = vllm_config.model_config.hf_config
+        self.mtp_start_layer_idx = config.num_hidden_layers
+        self.num_mtp_layers = config.num_nextn_predict_layers
+
+        self.embed_tokens = VocabParallelEmbedding(
+            config.vocab_size,
+            config.hidden_size,
+        )
+
+        self.mtp_layers = torch.nn.ModuleDict({
+            str(idx):
+            MiMoMultiTokenPredictorLayer(
+                config,
+                f"{prefix}.layers.{idx}",
+                model_config=vllm_config.model_config,
+                cache_config=vllm_config.cache_config,
+                quant_config=vllm_config.quant_config,
+            )
+            for idx in range(self.mtp_start_layer_idx,
+                             self.mtp_start_layer_idx + self.num_mtp_layers)
+        })
+
+        self.logits_processor = LogitsProcessor(config.vocab_size)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        previous_hidden_states: torch.Tensor,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        spec_step_idx: int = 0,
+    ) -> torch.Tensor:
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+        return self.mtp_layers[str(self.mtp_start_layer_idx + spec_step_idx)](
+            inputs_embeds,
+            positions,
+            previous_hidden_states,
+            spec_step_idx,
+        )
+
+    def compute_logits(
+        self,
+        hidden_states: torch.Tensor,
+        lm_head: ParallelLMHead,
+        sampling_metadata: SamplingMetadata,
+        spec_step_idx: int = 0,
+    ) -> torch.Tensor:
+        self.mtp_layers[str(self.mtp_start_layer_idx + spec_step_idx)]
+        logits = self.logits_processor(lm_head, hidden_states,
+                                       sampling_metadata)
+        return logits
+
+
+class MiMoMTP(nn.Module):
+
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
+        super().__init__()
+        self.config = vllm_config.model_config.hf_config
+        self.model = MiMoMultiTokenPredictor(vllm_config=vllm_config,
+                                             prefix=maybe_prefix(
+                                                 prefix, "model"))
+        self.lm_head = ParallelLMHead(self.config.vocab_size,
+                                      self.config.hidden_size)
+
+        self.sampler = get_sampler()
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        previous_hidden_states: torch.Tensor,
+        intermediate_tensors: Optional[IntermediateTensors] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        spec_step_idx: int = 0,
+    ) -> torch.Tensor:
+        assert spec_step_idx == 0, "mimo_mtp only support predict one token now"
+        hidden_states = self.model(input_ids, positions,
+                                   previous_hidden_states, inputs_embeds,
+                                   spec_step_idx)
+        return hidden_states
+
+    def compute_logits(
+        self,
+        hidden_states: torch.Tensor,
+        sampling_metadata: SamplingMetadata,
+        spec_step_idx: int = 0,
+    ) -> Optional[torch.Tensor]:
+        return self.model.compute_logits(hidden_states, self.lm_head,
+                                         sampling_metadata, spec_step_idx)
+
+    def sample(
+        self,
+        logits: torch.Tensor,
+        sampling_metadata: SamplingMetadata,
+    ) -> Optional[SamplerOutput]:
+        next_tokens = self.sampler(logits, sampling_metadata)
+        return next_tokens
+
+    def load_weights(self, weights: Iterable[tuple[str,
+                                                   torch.Tensor]]) -> set[str]:
+        stacked_params_mapping = [
+            ("qkv_proj", "q_proj", "q"),
+            ("qkv_proj", "k_proj", "k"),
+            ("qkv_proj", "v_proj", "v"),
+            ("gate_up_proj", "gate_proj", 0),
+            ("gate_up_proj", "up_proj", 1),
+        ]
+
+        params_dict = dict(self.named_parameters())
+        loaded_params: set[str] = set()
+        for name, loaded_weight in weights:
+
+            if "rotary_emb.inv_freq" in name:
+                continue
+            name = self.map_model_name_to_mtp_param_name(name)
+
+            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
+                if "mtp_layers" not in name:
+                    break
+                # 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:
+                # Skip loading extra bias for GPTQ models.
+                if name.endswith(".bias") and name not in params_dict:
+                    continue
+                if "mtp_layers" not in name and ("embed_tokens" not in name
+                                                 and "lm_head" not in name):
+                    continue
+                param = params_dict[name]
+                weight_loader = getattr(param, "weight_loader",
+                                        default_weight_loader)
+                weight_loader(param, loaded_weight)
+            loaded_params.add(name)
+        return loaded_params
+
+    def map_model_name_to_mtp_param_name(self, name: str) -> str:
+        import regex as re
+        name_without_prefix = [
+            "token_layernorm", "hidden_layernorm", "input_proj",
+            "final_layernorm"
+        ]
+        for sub_name in name_without_prefix:
+            if sub_name in name:
+                return name
+        pattern = r"model.mtp_layers.(\d+)."
+        group = re.match(pattern, name)
+        if group is not None:
+            name = name.replace(group.group(), group.group() + "mtp_block.")
+        return name
+
+    def _rewrite_spec_layer_name(self, spec_layer: int, name: str) -> str:
+        """
+        Rewrite the weight name to match the format of the original model.
+        Add .mtp_block for modules in transformer layer block for spec layer
+        """
+        spec_layer_weight_names = [
+            "embed_tokens", "enorm", "hnorm", "eh_proj", "shared_head"
+        ]
+        spec_layer_weight = False
+        for weight_name in spec_layer_weight_names:
+            if weight_name in name:
+                spec_layer_weight = True
+                break
+        if not spec_layer_weight:
+            # treat rest weights as weights for transformer layer block
+            name = name.replace(f"model.layers.{spec_layer}.",
+                                f"model.layers.{spec_layer}.mtp_block.")
+        return name