deepseek_mtp.py

# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from collections.abc import Iterable

import torch
import torch.nn as nn
from transformers import PretrainedConfig

from vllm.compilation.decorators import support_torch_compile
from vllm.config import VllmConfig
from vllm.logger import init_logger
from vllm.model_executor.layers.fused_moe import FusedMoE
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.vocab_parallel_embedding import (
    ParallelLMHead,
    VocabParallelEmbedding,
)
from vllm.model_executor.model_loader.weight_utils import (
    default_weight_loader,
    maybe_remap_kv_scale_name,
)
from vllm.platforms import current_platform
from vllm.sequence import IntermediateTensors

from .deepseek_v2 import (
    DeepseekV2DecoderLayer,
    DeepseekV2MixtureOfExperts,
    DeepseekV2MoE,
    get_spec_layer_idx_from_weight_name,
)
from .interfaces import SupportsPP
from .utils import maybe_prefix

logger = init_logger(__name__)


class SharedHead(nn.Module):
    def __init__(
        self,
        config: PretrainedConfig,
        prefix: str,
        quant_config: QuantizationConfig | None = None,
    ) -> None:
        super().__init__()
        self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.head = ParallelLMHead(
            config.vocab_size,
            config.hidden_size,
            quant_config=quant_config,
            prefix=maybe_prefix(prefix, "head"),
        )

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        return self.norm(hidden_states)


class DeepSeekMultiTokenPredictorLayer(nn.Module):
    def __init__(self, vllm_config: VllmConfig, prefix: str) -> None:
        super().__init__()

        config = vllm_config.speculative_config.draft_model_config.hf_config
        self.config = config
        quant_config = vllm_config.quant_config

        self.enorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.hnorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.eh_proj = nn.Linear(config.hidden_size * 2, config.hidden_size, bias=False)

        self.device = current_platform.device_type

        self.is_v32 = hasattr(config, "index_topk")
        if self.is_v32:
            topk_tokens = config.index_topk
            topk_indices_buffer = torch.empty(
                vllm_config.scheduler_config.max_num_batched_tokens,
                topk_tokens,
                dtype=torch.int32,
                device=self.device,
            )
        else:
            topk_indices_buffer = None

        self.shared_head = SharedHead(
            config=config, prefix=prefix, quant_config=quant_config
        )
        self.mtp_block = DeepseekV2DecoderLayer(
            vllm_config,
            prefix,
            config=self.config,
            topk_indices_buffer=topk_indices_buffer,
        )

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        previous_hidden_states: torch.Tensor,
        inputs_embeds: torch.Tensor | None = None,
        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 = torch.where(positions.unsqueeze(-1) == 0, 0, inputs_embeds)
        inputs_embeds = self.enorm(inputs_embeds)
        previous_hidden_states = self.hnorm(previous_hidden_states)

        hidden_states = self.eh_proj(
            torch.cat([inputs_embeds, previous_hidden_states], dim=-1)
        )

        hidden_states, residual = self.mtp_block(
            positions=positions, hidden_states=hidden_states, residual=None
        )
        hidden_states = residual + hidden_states
        return hidden_states


class DeepSeekMultiTokenPredictor(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
        # to map the exact layer index from weights

        self.layers = torch.nn.ModuleDict(
            {
                str(idx): DeepSeekMultiTokenPredictorLayer(
                    vllm_config, f"{prefix}.layers.{idx}"
                )
                for idx in range(
                    self.mtp_start_layer_idx,
                    self.mtp_start_layer_idx + self.num_mtp_layers,
                )
            }
        )
        self.embed_tokens = VocabParallelEmbedding(
            config.vocab_size,
            config.hidden_size,
        )
        self.logits_processor = LogitsProcessor(config.vocab_size)

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

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        previous_hidden_states: torch.Tensor,
        inputs_embeds: torch.Tensor | None = None,
        spec_step_idx: int = 0,
    ) -> torch.Tensor:
        if inputs_embeds is None:
            inputs_embeds = self.embed_tokens(input_ids)
        current_step_idx = spec_step_idx % self.num_mtp_layers
        return self.layers[str(self.mtp_start_layer_idx + current_step_idx)](
            input_ids,
            positions,
            previous_hidden_states,
            inputs_embeds,
            current_step_idx,
        )

    def compute_logits(
        self,
        hidden_states: torch.Tensor,
        spec_step_idx: int = 0,
    ) -> torch.Tensor:
        current_step_idx = spec_step_idx % self.num_mtp_layers
        mtp_layer = self.layers[str(self.mtp_start_layer_idx + current_step_idx)]
        logits = self.logits_processor(
            mtp_layer.shared_head.head, mtp_layer.shared_head(hidden_states)
        )
        return logits


@support_torch_compile
class DeepSeekMTP(nn.Module, SupportsPP, DeepseekV2MixtureOfExperts):
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        super().__init__()
        self.config = vllm_config.model_config.hf_config
        self.model = DeepSeekMultiTokenPredictor(
            vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
        )
        # Set MoE hyperparameters
        self.set_moe_parameters()

    def set_moe_parameters(self):
        self.expert_weights = []
        self.num_moe_layers = self.config.num_nextn_predict_layers
        self.num_expert_groups = self.config.n_group

        self.moe_layers = []
        self.moe_mlp_layers = []
        example_moe = None
        for layer in self.model.layers.values():
            assert isinstance(layer, DeepSeekMultiTokenPredictorLayer)
            layer = layer.mtp_block
            assert isinstance(layer, DeepseekV2DecoderLayer)
            if isinstance(layer.mlp, DeepseekV2MoE):
                example_moe = layer.mlp
                self.moe_mlp_layers.append(layer.mlp)
                self.moe_layers.append(layer.mlp.experts)
        self.extract_moe_parameters(example_moe)

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

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
        spec_step_idx: int = 0,
    ) -> torch.Tensor:
        hidden_states = self.model(
            input_ids, positions, hidden_states, inputs_embeds, spec_step_idx
        )
        return hidden_states

    def compute_logits(
        self,
        hidden_states: torch.Tensor,
        spec_step_idx: int = 0,
    ) -> torch.Tensor | None:
        return self.model.compute_logits(hidden_states, spec_step_idx)

    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
        stacked_params_mapping = [
            ("gate_up_proj", "gate_proj", 0),
            ("gate_up_proj", "up_proj", 1),
            ("fused_qkv_a_proj", "q_a_proj", 0),
            ("fused_qkv_a_proj", "kv_a_proj_with_mqa", 1),
        ]

        expert_params_mapping = FusedMoE.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,
        )

        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
            spec_layer = get_spec_layer_idx_from_weight_name(self.config, name)
            if spec_layer is None:
                continue
            name = self._rewrite_spec_layer_name(spec_layer, 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
                # 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_mapped = name.replace(weight_name, param_name)

                # QKV fusion is optional, fall back to normal
                # weight loading if it's not enabled
                if (
                    param_name == "fused_qkv_a_proj"
                ) and name_mapped not in params_dict:
                    continue
                else:
                    name = name_mapped

                # 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

                    name = maybe_remap_kv_scale_name(name, params_dict)
                    if name is None:
                        continue

                    # According to DeepSeek-V3 Technical Report, MTP modules
                    # shares embedding layer. We only load the first weights.
                    if (
                        spec_layer != self.model.mtp_start_layer_idx
                        and ".layers" 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 _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
        and rename shared layer weights to be top level.
        """
        spec_layer_weight_names = [
            "embed_tokens",
            "enorm",
            "hnorm",
            "eh_proj",
            "shared_head",
        ]
        shared_weight_names = ["embed_tokens"]
        spec_layer_weight = False
        shared_weight = False
        for weight_name in spec_layer_weight_names:
            if weight_name in name:
                spec_layer_weight = True
                if weight_name in shared_weight_names:
                    shared_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."
            )
        elif shared_weight:
            # treat shared weights as top level weights
            name = name.replace(f"model.layers.{spec_layer}.", "model.")
        return name