granitemoe.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 GraniteMoe model."""

from collections.abc import Iterable
from itertools import islice
from typing import Any, Optional

import torch
from torch import nn

from vllm.attention import Attention
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, VllmConfig
from vllm.distributed import (
    get_pp_group,
    get_tensor_model_parallel_world_size,
    tensor_model_parallel_all_gather,
)
from vllm.model_executor.layers.fused_moe import FusedMoE
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (
    QKVParallelLinear,
    ReplicatedLinear,
    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,
    maybe_remap_kv_scale_name,
)
from vllm.model_executor.models.utils import sequence_parallel_chunk
from vllm.sequence import IntermediateTensors

from .interfaces import SupportsLoRA, SupportsPP
from .utils import AutoWeightsLoader, is_pp_missing_parameter, make_layers, maybe_prefix


class GraniteMoeMoE(nn.Module):
    """A tensor-parallel MoE implementation for GraniteMoe that shards each
    expert across all ranks.
    Each expert's weights are sharded across all ranks and a fused MoE
    kernel is used for the forward pass, and finally we reduce the outputs
    across ranks.
    """

    def __init__(
        self,
        num_experts: int,
        top_k: int,
        hidden_size: int,
        intermediate_size: int,
        params_dtype: Optional[torch.dtype] = None,
        quant_config: Optional[QuantizationConfig] = None,
        tp_size: Optional[int] = None,
        is_sequence_parallel=False,
        prefix: str = "",
    ):
        super().__init__()
        self.hidden_size = hidden_size
        self.is_sequence_parallel = is_sequence_parallel

        # Gate always runs at half / full precision for now.
        self.gate = ReplicatedLinear(
            hidden_size,
            num_experts,
            bias=False,
            params_dtype=params_dtype,
            quant_config=None,
            prefix=f"{prefix}.gate",
        )

        self.experts = FusedMoE(
            num_experts=num_experts,
            top_k=top_k,
            hidden_size=hidden_size,
            intermediate_size=intermediate_size,
            params_dtype=params_dtype,
            reduce_results=True,
            renormalize=True,
            quant_config=quant_config,
            tp_size=tp_size,
            prefix=f"{prefix}.experts",
            is_sequence_parallel=self.is_sequence_parallel,
        )

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        # NOTE: hidden_states can have either 1D or 2D shape.
        orig_shape = hidden_states.shape
        hidden_states = hidden_states.view(-1, self.hidden_size)

        if self.is_sequence_parallel:
            hidden_states = sequence_parallel_chunk(hidden_states)

        # router_logits: (num_tokens, n_experts)
        router_logits, _ = self.gate(hidden_states)
        final_hidden_states = self.experts(hidden_states, router_logits)

        if self.is_sequence_parallel:
            final_hidden_states = tensor_model_parallel_all_gather(
                final_hidden_states, 0
            )
            num_tokens = orig_shape[0]
            final_hidden_states = final_hidden_states[:num_tokens]

        return final_hidden_states.view(orig_shape)


class GraniteMoeAttention(nn.Module):
    def __init__(
        self,
        hidden_size: int,
        num_heads: int,
        num_kv_heads: int,
        max_position: int = 4096 * 32,
        rope_theta: float = 10000,
        rope_scaling: Optional[dict[str, Any]] = None,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        attention_multiplier: Optional[float] = None,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.hidden_size = hidden_size
        tp_size = get_tensor_model_parallel_world_size()
        self.total_num_heads = num_heads
        assert self.total_num_heads % tp_size == 0
        self.num_heads = self.total_num_heads // tp_size
        self.total_num_kv_heads = num_kv_heads
        if self.total_num_kv_heads >= tp_size:
            # Number of KV heads is greater than TP size, so we partition
            # the KV heads across multiple tensor parallel GPUs.
            assert self.total_num_kv_heads % tp_size == 0
        else:
            # Number of KV heads is less than TP size, so we replicate
            # the KV heads across multiple tensor parallel GPUs.
            assert tp_size % self.total_num_kv_heads == 0
        self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
        self.head_dim = hidden_size // self.total_num_heads
        self.q_size = self.num_heads * self.head_dim
        self.kv_size = self.num_kv_heads * self.head_dim
        self.scaling = (
            attention_multiplier
            if attention_multiplier is not None
            else self.head_dim**-1
        )
        self.rope_theta = rope_theta

        self.qkv_proj = QKVParallelLinear(
            hidden_size,
            self.head_dim,
            self.total_num_heads,
            self.total_num_kv_heads,
            bias=False,
            quant_config=quant_config,
            prefix=f"{prefix}.qkv_proj",
        )
        self.o_proj = RowParallelLinear(
            self.total_num_heads * self.head_dim,
            hidden_size,
            bias=False,
            quant_config=quant_config,
            prefix=f"{prefix}.o_proj",
        )
        self.rotary_emb = get_rope(
            self.head_dim,
            rotary_dim=self.head_dim,
            max_position=max_position,
            base=int(self.rope_theta),
            is_neox_style=True,
            rope_scaling=rope_scaling,
        )
        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,
            prefix=f"{prefix}.attn",
        )

    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)
        output, _ = self.o_proj(attn_output)
        return output


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

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

        self.hidden_size = config.hidden_size
        # Requires transformers > 4.32.0
        rope_theta = getattr(config, "rope_theta", 10000)
        rope_scaling = getattr(config, "rope_scaling", None)
        self.self_attn = GraniteMoeAttention(
            hidden_size=self.hidden_size,
            num_heads=config.num_attention_heads,
            max_position=config.max_position_embeddings,
            num_kv_heads=config.num_key_value_heads,
            rope_theta=rope_theta,
            rope_scaling=rope_scaling,
            cache_config=cache_config,
            quant_config=quant_config,
            prefix=f"{prefix}.self_attn",
            attention_multiplier=config.attention_multiplier,
        )
        self.block_sparse_moe = GraniteMoeMoE(
            num_experts=config.num_local_experts,
            top_k=config.num_experts_per_tok,
            hidden_size=config.hidden_size,
            intermediate_size=config.intermediate_size,
            quant_config=quant_config,
            is_sequence_parallel=parallel_config.use_sequence_parallel_moe,
            prefix=f"{prefix}.block_sparse_moe",
        )

        self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.post_attention_layernorm = RMSNorm(
            config.hidden_size, eps=config.rms_norm_eps
        )

        self.residual_multiplier = config.residual_multiplier

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
        # Self Attention
        residual = hidden_states
        hidden_states = self.input_layernorm(hidden_states)
        hidden_states = self.self_attn(
            positions=positions,
            hidden_states=hidden_states,
        )
        hidden_states = residual + hidden_states * self.residual_multiplier
        residual = hidden_states
        hidden_states = self.post_attention_layernorm(hidden_states)
        hidden_states = self.block_sparse_moe(hidden_states)
        hidden_states = residual + hidden_states * self.residual_multiplier

        return hidden_states


@support_torch_compile
class GraniteMoeModel(nn.Module):
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        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.quant_config = quant_config  # Required by MixtralModel
        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

        self.embed_tokens = VocabParallelEmbedding(
            self.vocab_size,
            config.hidden_size,
            org_num_embeddings=config.vocab_size,
        )
        self.embedding_multiplier = config.embedding_multiplier

        self.start_layer, self.end_layer, self.layers = make_layers(
            config.num_hidden_layers,
            lambda prefix: GraniteMoeDecoderLayer(vllm_config, prefix=prefix),
            prefix=f"{prefix}.layers",
        )

        self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)

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

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        intermediate_tensors: Optional[IntermediateTensors],
        inputs_embeds: Optional[torch.Tensor] = None,
    ) -> 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)
            hidden_states *= self.embedding_multiplier
        else:
            assert intermediate_tensors is not None
            hidden_states = intermediate_tensors["hidden_states"]
        for layer in islice(self.layers, self.start_layer, self.end_layer):
            hidden_states = layer(positions, hidden_states)
        if not get_pp_group().is_last_rank:
            return IntermediateTensors(
                {
                    "hidden_states": hidden_states,
                }
            )
        hidden_states = self.norm(hidden_states)
        return hidden_states

    def _load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
        """
        This function is copied from `MixtralModel.load_weights`, mainly to
        decouple from mixtral, avoiding impact on support like BNB
        quantization.
        """
        stacked_params_mapping = [
            # (param_name, shard_name, shard_id)
            ("qkv_proj", "q_proj", "q"),
            ("qkv_proj", "k_proj", "k"),
            ("qkv_proj", "v_proj", "v"),
        ]

        # Params for weights, fp8 weight scales, fp8 activation scales
        # (param_name, weight_name, expert_id, shard_id)
        expert_params_mapping = FusedMoE.make_expert_params_mapping(
            ckpt_gate_proj_name="w1",
            ckpt_down_proj_name="w2",
            ckpt_up_proj_name="w3",
            num_experts=self.config.num_local_experts,
        )

        params_dict = dict(self.named_parameters())
        loaded_params: set[str] = set()
        for name, loaded_weight in weights:
            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") or name.endswith("_bias")
                ) and name not in params_dict:
                    continue
                # Skip layers on other devices.
                if is_pp_missing_parameter(name, self):
                    continue
                if name.endswith("scale"):
                    # Remapping the name of FP8 kv-scale.
                    name = maybe_remap_kv_scale_name(name, params_dict)
                    if name is None:
                        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)
                    # Skip layers on other devices.
                    if is_pp_missing_parameter(name, self):
                        continue
                    if (
                        name.endswith(".bias") or name.endswith("_bias")
                    ) and name not in params_dict:
                        continue
                    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") or name.endswith("_bias")
                    ) and name not in params_dict:
                        continue
                    # Skip layers on other devices.
                    if is_pp_missing_parameter(name, self):
                        continue
                    # Remapping the name of FP8 kv-scale.
                    name = maybe_remap_kv_scale_name(name, params_dict)
                    if name is None:
                        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 load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
        new_weights = {}
        for n, p in weights:
            if n.endswith(".block_sparse_moe.input_linear.weight"):
                for e in range(p.size(0)):
                    w1_name = n.replace(
                        ".block_sparse_moe.input_linear.weight",
                        f".block_sparse_moe.experts.{e}.w1.weight",
                    )
                    w3_name = n.replace(
                        ".block_sparse_moe.input_linear.weight",
                        f".block_sparse_moe.experts.{e}.w3.weight",
                    )
                    w1_param, w3_param = p[e].chunk(2, dim=0)
                    assert w1_name not in new_weights
                    assert w3_name not in new_weights
                    new_weights[w1_name] = w1_param
                    new_weights[w3_name] = w3_param
            elif n.endswith(".block_sparse_moe.output_linear.weight"):
                for e in range(p.size(0)):
                    w2_name = n.replace(
                        ".block_sparse_moe.output_linear.weight",
                        f".block_sparse_moe.experts.{e}.w2.weight",
                    )
                    w2_param = p[e]
                    assert w2_name not in new_weights
                    new_weights[w2_name] = w2_param
            elif n.endswith(".block_sparse_moe.router.layer.weight"):
                gate_name = n.replace(
                    ".block_sparse_moe.router.layer.weight",
                    ".block_sparse_moe.gate.weight",
                )
                assert gate_name not in new_weights
                new_weights[gate_name] = p
            else:
                new_weights[n] = p
        return self._load_weights(new_weights.items())


class GraniteMoeForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
    fall_back_to_pt_during_load = False

    packed_modules_mapping = {
        "qkv_proj": [
            "q_proj",
            "k_proj",
            "v_proj",
        ],
    }

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

    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        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 = GraniteMoeModel(
            vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
        )
        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.weight = self.model.embed_tokens.weight

        self.logits_processor = LogitsProcessor(
            self.unpadded_vocab_size,
            config.vocab_size,
            scale=1 / self.config.logits_scaling,
        )

    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,
    ) -> torch.Tensor:
        hidden_states = self.model(
            input_ids, positions, intermediate_tensors, inputs_embeds
        )
        return hidden_states

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

    def make_empty_intermediate_tensors(
        self, batch_size: int, dtype: torch.dtype, device: torch.device
    ) -> IntermediateTensors:
        return IntermediateTensors(
            {
                "hidden_states": torch.zeros(
                    (batch_size, self.config.hidden_size), dtype=dtype, device=device
                ),
            }
        )

    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)