nemotron_nas.py

# 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/llama/modeling_llama.py
# 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 deci model compatible with HuggingFace weights."""

from collections.abc import Iterable
from itertools import islice

import torch
from torch import nn
from transformers import LlamaConfig

from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, VllmConfig
from vllm.distributed import get_pp_group
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.rotary_embedding import get_rope
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.model_executor.models.llama import LlamaAttention, LlamaMLP
from vllm.sequence import IntermediateTensors
from vllm.v1.attention.backend import AttentionType

from .interfaces import HasNoOps, SupportsLoRA, SupportsPP
from .utils import (
    AutoWeightsLoader,
    PPMissingLayer,
    is_pp_missing_parameter,
    make_empty_intermediate_tensors_factory,
    make_layers,
    maybe_prefix,
)


def _ffn_mult_to_intermediate_size(ffn_mult: float, n_embd: int) -> int:
    # DeciLM-specific code
    intermediate_size = int(2 * ffn_mult * n_embd / 3)
    return _find_multiple(intermediate_size, 256)


def _find_multiple(n: int, k: int) -> int:
    # DeciLM-specific code
    if n % k == 0:
        return n
    return n + k - (n % k)


class DeciLMAttention(LlamaAttention):
    def __init__(
        self,
        config: LlamaConfig,
        hidden_size: int,
        num_heads: int,
        num_kv_heads: int,
        max_position_embeddings: int = 8192,
        quant_config: QuantizationConfig | None = None,
        bias: bool = False,
        bias_o_proj: bool = False,
        cache_config: CacheConfig | None = None,
        prefix: str = "",
        attn_type: str = AttentionType.DECODER,
    ) -> None:
        super().__init__(
            config,
            hidden_size,
            num_heads,
            num_kv_heads,
            max_position_embeddings,
            quant_config,
            bias,
            bias_o_proj,
            cache_config,
            prefix,
            attn_type,
        )

    def _init_rotary_emb(
        self,
        config,
        quant_config: QuantizationConfig | None,
    ) -> None:
        # Enables YARN for Mistral and LLaMA4 derivatives.
        is_neox_style = True
        if hasattr(config, "position_embedding_type"):
            is_neox_style = config.position_embedding_type not in [
                "mistral_yarn",
                "rope_llama4",
            ]

        self.rotary_emb = get_rope(
            self.head_dim,
            max_position=self.max_position_embeddings,
            rope_parameters=config.rope_parameters,
            is_neox_style=is_neox_style,
        )


class DeciLMDecoderLayer(nn.Module):
    def __init__(
        self,
        config: LlamaConfig,
        layer_idx: int,
        cache_config: CacheConfig | None = None,
        quant_config: QuantizationConfig | None = None,
        prefix: str = "",
    ) -> None:
        super().__init__()
        block_config = config.block_configs[layer_idx]
        self._is_no_op_attention = block_config.attention.no_op
        self._is_no_op_ffn = block_config.ffn.no_op

        self.hidden_size = config.hidden_size
        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

        if not self._is_no_op_attention:
            num_kv_heads = (
                config.num_attention_heads // block_config.attention.n_heads_in_group
            )
            self.self_attn = DeciLMAttention(
                config=config,
                hidden_size=self.hidden_size,
                num_heads=config.num_attention_heads,
                num_kv_heads=num_kv_heads,
                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",
            )
            self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)

        if not self._is_no_op_ffn:
            if hasattr(block_config.ffn, "ffn_mult"):
                ffn_mult = block_config.ffn.ffn_mult
                intermediate_size = _ffn_mult_to_intermediate_size(
                    ffn_mult, config.hidden_size
                )
            else:
                intermediate_size = block_config.ffn.intermediate_size

            self.mlp = LlamaMLP(
                hidden_size=self.hidden_size,
                intermediate_size=intermediate_size,
                hidden_act=config.hidden_act,
                quant_config=quant_config,
                bias=getattr(config, "mlp_bias", False),
                prefix=f"{prefix}.mlp",
            )
            self.post_attention_layernorm = RMSNorm(
                config.hidden_size, eps=config.rms_norm_eps
            )

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        residual: torch.Tensor | None,
    ) -> tuple[torch.Tensor, torch.Tensor]:
        # Self Attention

        if self._is_no_op_attention:
            pass
        else:
            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,
            )

        # Fully Connected
        if not self._is_no_op_ffn:
            hidden_states, residual = self.post_attention_layernorm(
                hidden_states, residual
            )
            hidden_states = self.mlp(hidden_states)
        return hidden_states, residual


@support_torch_compile
class DeciModel(nn.Module):
    def __init__(
        self,
        *,
        vllm_config: VllmConfig,
        prefix: str = "",
        layer_type: type[DeciLMDecoderLayer] = DeciLMDecoderLayer,
    ):
        super().__init__()

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

        self.config = config
        self.quant_config = quant_config
        self.padding_idx = config.pad_token_id

        self.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,
                quant_config=quant_config,
            )
        else:
            self.embed_tokens = PPMissingLayer()

        def get_layer(prefix: str):
            layer_idx = int(prefix.rsplit(".", 1)[1])
            return layer_type(
                config,
                layer_idx,
                cache_config,
                quant_config=quant_config,
                prefix=prefix,
            )

        self.start_layer, self.end_layer, self.layers = make_layers(
            config.num_hidden_layers,
            get_layer,
            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.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory(
            ["hidden_states", "residual"], config.hidden_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 | None,
        positions: torch.Tensor,
        intermediate_tensors: IntermediateTensors | None,
        inputs_embeds: torch.Tensor | None = None,
    ) -> torch.Tensor | IntermediateTensors:
        if get_pp_group().is_first_rank:
            if inputs_embeds is not None:
                hidden_states = inputs_embeds
            else:
                hidden_states = self.embed_input_ids(input_ids)
            residual = None
        else:
            assert intermediate_tensors is not None
            hidden_states = intermediate_tensors["hidden_states"]
            residual = intermediate_tensors["residual"]

        kv_cache_index = 0
        for layer in islice(self.layers, self.start_layer, self.end_layer):
            if not layer._is_no_op_attention:
                hidden_states, residual = layer(positions, hidden_states, residual)
                kv_cache_index += 1
            else:
                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)
        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),
        ]
        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
            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 or "zero_point" 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]
                weight_loader = getattr(param, "weight_loader", default_weight_loader)
                weight_loader(param, loaded_weight)
            loaded_params.add(name)
        return loaded_params


class DeciLMForCausalLM(nn.Module, SupportsLoRA, SupportsPP, HasNoOps):
    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",
    }

    # Mistral/Llama models can also be loaded with --load-format mistral
    # from consolidated.safetensors checkpoints
    mistral_mapping = {
        "layers": "model.layers",
        "attention": "self_attn",
        "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 = ""):
        super().__init__()
        config = vllm_config.model_config.hf_config
        quant_config = vllm_config.quant_config

        self.config = config

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

        if get_pp_group().is_last_rank:
            self.lm_head = ParallelLMHead(
                config.vocab_size,
                config.hidden_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(
                config.vocab_size, scale=logit_scale
            )
        else:
            self.lm_head = PPMissingLayer()

        self.make_empty_intermediate_tensors = (
            self.model.make_empty_intermediate_tensors
        )

    def _init_model(self, vllm_config: VllmConfig, prefix: str = ""):
        return DeciModel(vllm_config=vllm_config, prefix=prefix)

    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 | None,
        positions: torch.Tensor,
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
    ) -> 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,
    ) -> torch.Tensor | None:
        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(weights)