deep seek 소스 python

Python으로 작성된 많은 프로젝트들이 비교적 짧은 코드로도 강력한 기능을 구현할 수 있는 이유는 여러 가지가 있습니다.

높은 추상화 수준:
Python은 고수준 언어로, 복잡한 작업들을 간결한 코드로 표현할 수 있도록 설계되었습니다. 이는 반복적이고 보일러플레이트 코드 없이도 핵심 로직에 집중할 수 있게 해줍니다.
풍부한 라이브러리와 프레임워크 활용:
Deep Seek 같은 프로젝트에서는 외부 라이브러리나 프레임워크(예: 딥러닝 프레임워크, 데이터 처리 라이브러리 등)를 많이 활용하는 경우가 많습니다. 이러한 라이브러리들은 내부적으로 C/C++ 등 최적화된 언어로 구현되어 있어, Python 코드에서는 간단한 인터페이스만 호출하면 복잡한 기능을 사용할 수 있습니다.
모듈화와 재사용:
Python의 모듈 시스템과 객체 지향 설계 덕분에 코드가 잘 구조화되어 있다면, 반복되는 기능을 재사용할 수 있어 전체 코드베이스의 크기가 작아질 수 있습니다.

즉, Deep Seek의 소스 코드가 상대적으로 작아 보이는 것은 Python의 간결한 문법과 외부 라이브러리의 강력한 기능 덕분이라고 볼 수 있습니다.


import math
from dataclasses import dataclass
from typing import Tuple, Optional, Literal

import torch
from torch import nn
import torch.nn.functional as F
import torch.distributed as dist

from kernel import act_quant, weight_dequant, fp8_gemm


world_size = 1
rank = 0
block_size = 128
gemm_impl: Literal["bf16", "fp8"] = "bf16"
attn_impl: Literal["naive", "absorb"] = "absorb"

@dataclass
class ModelArgs:
    """
    Data class for defining model arguments and hyperparameters.

    Attributes:
        max_batch_size (int): Maximum batch size.
        max_seq_len (int): Maximum sequence length.
        dtype (Literal["bf16", "fp8"]): Data type for computations.
        vocab_size (int): Vocabulary size.
        dim (int): Model dimension.
        inter_dim (int): Intermediate dimension for MLP layers.
        moe_inter_dim (int): Intermediate dimension for MoE layers.
        n_layers (int): Number of transformer layers.
        n_dense_layers (int): Number of dense layers in the model.
        n_heads (int): Number of attention heads.
        n_routed_experts (int): Number of routed experts for MoE layers.
        n_shared_experts (int): Number of shared experts for MoE layers.
        n_activated_experts (int): Number of activated experts in MoE layers.
        n_expert_groups (int): Number of expert groups.
        n_limited_groups (int): Number of limited groups for MoE routing.
        score_func (Literal["softmax", "sigmoid"]): Scoring function for MoE routing.
        route_scale (float): Scaling factor for routing scores.
        q_lora_rank (int): LoRA rank for query projections.
        kv_lora_rank (int): LoRA rank for key-value projections.
        qk_nope_head_dim (int): Dimension for query-key projections without positional embeddings.
        qk_rope_head_dim (int): Dimension for query-key projections with rotary embeddings.
        v_head_dim (int): Dimension for value projections.
        original_seq_len (int): Original sequence length.
        rope_theta (float): Base for rotary positional encoding.
        rope_factor (float): Scaling factor for extended sequence lengths.
        beta_fast (int): Fast beta correction factor.
        beta_slow (int): Slow beta correction factor.
        mscale (float): Scaling factor for extended attention.
    """
    max_batch_size: int = 8
    max_seq_len: int = 4096 * 4
    dtype: Literal["bf16", "fp8"] = "bf16"
    vocab_size: int = 102400
    dim: int = 2048
    inter_dim: int = 10944
    moe_inter_dim: int = 1408
    n_layers: int = 27
    n_dense_layers: int = 1
    n_heads: int = 16
    # moe
    n_routed_experts: int = 64
    n_shared_experts: int = 2
    n_activated_experts: int = 6
    n_expert_groups: int = 1
    n_limited_groups: int = 1
    score_func: Literal["softmax", "sigmoid"] = "softmax"
    route_scale: float = 1.
    # mla
    q_lora_rank: int = 0
    kv_lora_rank: int = 512
    qk_nope_head_dim: int = 128
    qk_rope_head_dim: int = 64
    v_head_dim: int = 128
    # yarn
    original_seq_len: int = 4096
    rope_theta: float = 10000.0
    rope_factor: float = 40
    beta_fast: int = 32
    beta_slow: int = 1
    mscale: float = 1.


class ParallelEmbedding(nn.Module):
    """
    Embedding layer with parallelism support across distributed processes.

    Args:
        vocab_size (int): Vocabulary size.
        dim (int): Embedding dimension.
    """
    def __init__(self, vocab_size: int, dim: int):
        super().__init__()
        self.vocab_size = vocab_size
        self.dim = dim
        assert vocab_size % world_size == 0, f"Vocabulary size must be divisible by world size (world_size={world_size})"
        self.part_vocab_size = (vocab_size // world_size)
        self.vocab_start_idx = rank * self.part_vocab_size
        self.vocab_end_idx = self.vocab_start_idx + self.part_vocab_size
        self.weight = nn.Parameter(torch.empty(self.part_vocab_size, self.dim))

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        """
        Forward pass for parallel embedding layer.

        Args:
            x (torch.Tensor): Input tensor containing token indices.

        Returns:
            torch.Tensor: Embedded representations.

        Raises:
            ValueError: If `world_size` is not defined.
        """
        if world_size > 1:
            mask = (x < self.vocab_start_idx) | (x >= self.vocab_end_idx)
            x = x - self.vocab_start_idx
            x[mask] = 0
        y = F.embedding(x, self.weight)
        if world_size > 1:
            y[mask] = 0
            dist.all_reduce(y)
        return y


def linear(x: torch.Tensor, weight: torch.Tensor, bias: Optional[torch.Tensor] = None) -> torch.Tensor:
    """
    Applies a linear transformation to the incoming data: y = xA^T + b.
    This function supports specialized implementations based on quantization
    and tensor formats.

    Args:
        x (torch.Tensor): The input tensor.
        weight (torch.Tensor): The weight tensor. It may be quantized and 
            requires dequantization for certain cases.
        bias (Optional[torch.Tensor]): The bias tensor to be added. Default is None.

    Returns:
        torch.Tensor: The result of the linear transformation, which may involve 
        quantization-aware computations depending on the input parameters.

    Notes:
        - If `weight` is quantized (e.g., `element_size() == 1`), a dequantized version 
          is used for computation.
        - If `gemm_impl == "bf16"`, dequantization and a `bf16` GEMM operation are applied.
        - For other cases, the function applies quantization to `x` and uses `fp8_gemm` for computation.
    """
    if weight.element_size() > 1:
        return F.linear(x, weight, bias)
    elif gemm_impl == "bf16":
        weight = weight_dequant(weight, weight.scale)
        return F.linear(x, weight, bias)
    else:
        x, scale = act_quant(x, block_size)
        y = fp8_gemm(x, scale, weight, weight.scale)
        if bias is not None:
            y += bias
        return y


class Linear(nn.Module):
    """
    Custom linear layer with support for quantized weights and optional bias.

    Args:
        in_features (int): Number of input features.
        out_features (int): Number of output features.
        bias (bool): Whether to include a bias term. Defaults to False.
        dtype (optional): Data type for the layer. Defaults to `torch.bfloat16`.
    """
    dtype = torch.bfloat16

    def __init__(self, in_features: int, out_features: int, bias: bool = False, dtype = None):
        super().__init__()
        self.in_features = in_features
        self.out_features = out_features
        self.weight = nn.Parameter(torch.empty(out_features, in_features, dtype=dtype or Linear.dtype))
        if self.weight.element_size() == 1:
            scale_out_features = (out_features + block_size - 1) // block_size
            scale_in_features = (in_features + block_size - 1) // block_size
            self.weight.scale = self.scale = nn.Parameter(torch.empty(scale_out_features, scale_in_features, dtype=torch.float32))
        else:
            self.register_parameter("scale", None)
        if bias:
            self.bias = nn.Parameter(torch.empty(out_features))
        else:
            self.register_parameter("bias", None)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        """
        Forward pass for the custom linear layer.

        Args:
            x (torch.Tensor): Input tensor.

        Returns:
            torch.Tensor: Transformed tensor after linear computation.
        """
        return linear(x, self.weight, self.bias)


class ColumnParallelLinear(Linear):
    """
    Linear layer with column parallelism, splitting output features across distributed processes.

    Args:
        in_features (int): Number of input features.
        out_features (int): Total number of output features.
        bias (bool): Whether to include a bias term. Defaults to False.
        dtype (optional): Data type for the layer. Defaults to `torch.bfloat16`.
    """
    def __init__(self, in_features: int, out_features: int, bias: bool = False, dtype = None):
        assert out_features % world_size == 0, f"Output features must be divisible by world size (world_size={world_size})"
        self.part_out_features = out_features // world_size
        super().__init__(in_features, self.part_out_features, bias, dtype)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        """
        Forward pass for column parallel linear layer.

        Args:
            x (torch.Tensor): Input tensor.

        Returns:
            torch.Tensor: Transformed tensor with column-parallel computation.
        """
        y = linear(x, self.weight, self.bias)
        return y


class RowParallelLinear(Linear):
    """
    Linear layer with row parallelism, splitting input features across distributed processes.

    Args:
        in_features (int): Total number of input features.
        out_features (int): Number of output features.
        bias (bool): Whether to include a bias term. Defaults to False.
        dtype (optional): Data type for the layer. Defaults to `torch.bfloat16`.
    """
    def __init__(self, in_features: int, out_features: int, bias: bool = False, dtype = None):
        assert in_features % world_size == 0, f"Input features must be divisible by world size (world_size={world_size})"
        self.part_in_features = in_features // world_size
        super().__init__(self.part_in_features, out_features, bias, dtype)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        """
        Forward pass for row parallel linear layer.

        Args:
            x (torch.Tensor): Input tensor.

        Returns:
            torch.Tensor: Transformed tensor with row-parallel computation.
        """
        y = linear(x, self.weight)
        if world_size > 1:
            dist.all_reduce(y)
        if self.bias is not None:
            y += self.bias
        return y


class RMSNorm(nn.Module):
    """
    Root Mean Square Layer Normalization (RMSNorm).

    Args:
        dim (int): Dimension of the input tensor.
        eps (float): Epsilon value for numerical stability. Defaults to 1e-6.
    """
    def __init__(self, dim: int, eps: float = 1e-6):
        super().__init__()
        self.dim = dim
        self.eps = eps
        self.weight = nn.Parameter(torch.ones(dim))

    def forward(self, x: torch.Tensor):
        """
        Forward pass for RMSNorm.

        Args:
            x (torch.Tensor): Input tensor.

        Returns:
            torch.Tensor: Normalized tensor with the same shape as input.
        """
        return F.rms_norm(x, (self.dim,), self.weight, self.eps)


def precompute_freqs_cis(args: ModelArgs) -> torch.Tensor:
    """
    Precomputes frequency-based complex exponential values for rotary positional embeddings.

    Args:
        args (ModelArgs): Model arguments containing positional embedding parameters.

    Returns:
        torch.Tensor: Precomputed complex exponential values for positional embeddings.
    """
    dim = args.qk_rope_head_dim
    seqlen = args.max_seq_len
    beta_fast = args.beta_fast
    beta_slow = args.beta_slow
    base = args.rope_theta
    factor = args.rope_factor

    def find_correction_dim(num_rotations, dim, base, max_seq_len):
        """
        Computes the correction dimension for a given number of rotations in the rotary positional embedding.

        Args:
            num_rotations (float): Number of rotations to compute the correction for.
            dim (int): Dimensionality of the embedding space.
            base (float): Base value for the exponential computation.
            max_seq_len (int): Maximum sequence length.

        Returns:
            float: The correction dimension based on the input parameters.
        """
        return dim * math.log(max_seq_len / (num_rotations * 2 * math.pi)) / (2 * math.log(base))

    def find_correction_range(low_rot, high_rot, dim, base, max_seq_len):
        """
        Computes the range of correction dimensions for rotary positional embeddings.

        Args:
            low_rot (float): Lower bound for the number of rotations.
            high_rot (float): Upper bound for the number of rotations.
            dim (int): Dimensionality of the embedding space.
            base (float): Base value for the exponential computation.
            max_seq_len (int): Maximum sequence length.

        Returns:
            Tuple[int, int]: The range of correction dimensions (low, high), clamped to valid indices.
        """
        low = math.floor(find_correction_dim(low_rot, dim, base, max_seq_len))
        high = math.ceil(find_correction_dim(high_rot, dim, base, max_seq_len))
        return max(low, 0), min(high, dim-1)

    def linear_ramp_factor(min, max, dim):
        """
        Computes a linear ramp function used to smooth values between a minimum and maximum range.

        Args:
            min (float): Minimum value for the ramp function.
            max (float): Maximum value for the ramp function.
            dim (int): Dimensionality of the ramp tensor.

        Returns:
            torch.Tensor: A tensor of shape (dim,) with values linearly interpolated between 0 and 1,
                clamped to the range [0, 1].
        """
        if min == max:
            max += 0.001
        linear_func = (torch.arange(dim, dtype=torch.float32) - min) / (max - min)
        ramp_func = torch.clamp(linear_func, 0, 1)
        return ramp_func

    freqs = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim))
    if seqlen > args.original_seq_len:
        low, high = find_correction_range(beta_fast, beta_slow, dim, base, args.original_seq_len)
        smooth = 1 - linear_ramp_factor(low, high, dim // 2)
        freqs = freqs / factor * (1 - smooth) + freqs * smooth

    t = torch.arange(seqlen)
    freqs = torch.outer(t, freqs)
    freqs_cis = torch.polar(torch.ones_like(freqs), freqs)
    return freqs_cis


def apply_rotary_emb(x: torch.Tensor, freqs_cis: torch.Tensor) -> torch.Tensor:
    """
    Applies rotary positional embeddings to the input tensor.

    Args:
        x (torch.Tensor): Input tensor with positional embeddings to be applied.
        freqs_cis (torch.Tensor): Precomputed complex exponential values for positional embeddings.

    Returns:
        torch.Tensor: Tensor with rotary embeddings applied.
    """
    dtype = x.dtype
    x = torch.view_as_complex(x.float().view(*x.shape[:-1], -1, 2))
    freqs_cis = freqs_cis.view(1, x.size(1), 1, x.size(-1))
    y = torch.view_as_real(x * freqs_cis).flatten(3)
    return y.to(dtype)


class MLA(nn.Module):
    """
    Multi-Headed Attention Layer (MLA).

    Attributes:
        dim (int): Dimensionality of the input features.
        n_heads (int): Number of attention heads.
        n_local_heads (int): Number of local attention heads for distributed systems.
        q_lora_rank (int): Rank for low-rank query projection.
        kv_lora_rank (int): Rank for low-rank key/value projection.
        qk_nope_head_dim (int): Dimensionality of non-positional query/key projections.
        qk_rope_head_dim (int): Dimensionality of rotary-positional query/key projections.
        qk_head_dim (int): Total dimensionality of query/key projections.
        v_head_dim (int): Dimensionality of value projections.
        softmax_scale (float): Scaling factor for softmax in attention computation.
    """
    def __init__(self, args: ModelArgs):
        super().__init__()
        self.dim = args.dim
        self.n_heads = args.n_heads
        self.n_local_heads = args.n_heads // world_size
        self.q_lora_rank = args.q_lora_rank
        self.kv_lora_rank = args.kv_lora_rank
        self.qk_nope_head_dim = args.qk_nope_head_dim
        self.qk_rope_head_dim = args.qk_rope_head_dim
        self.qk_head_dim = args.qk_nope_head_dim + args.qk_rope_head_dim
        self.v_head_dim = args.v_head_dim

        if self.q_lora_rank == 0:
            self.wq = ColumnParallelLinear(self.dim, self.n_heads * self.qk_head_dim)
        else:
            self.wq_a = Linear(self.dim, self.q_lora_rank)
            self.q_norm = RMSNorm(self.q_lora_rank)
            self.wq_b = ColumnParallelLinear(self.q_lora_rank, self.n_heads * self.qk_head_dim)
        self.wkv_a = Linear(self.dim, self.kv_lora_rank + self.qk_rope_head_dim)
        self.kv_norm = RMSNorm(self.kv_lora_rank)
        self.wkv_b = ColumnParallelLinear(self.kv_lora_rank, self.n_heads * (self.qk_nope_head_dim + self.v_head_dim))
        self.wo = RowParallelLinear(self.n_heads * self.v_head_dim, self.dim)
        self.softmax_scale = self.qk_head_dim ** -0.5
        if args.max_seq_len > args.original_seq_len:
            mscale = 0.1 * args.mscale * math.log(args.rope_factor) + 1.0
            self.softmax_scale = self.softmax_scale * mscale * mscale

        if attn_impl == "naive":
            self.register_buffer("k_cache", torch.zeros(args.max_batch_size, args.max_seq_len, self.n_local_heads, self.qk_head_dim), persistent=False)
            self.register_buffer("v_cache", torch.zeros(args.max_batch_size, args.max_seq_len, self.n_local_heads, self.v_head_dim), persistent=False)
        else:
            self.register_buffer("kv_cache", torch.zeros(args.max_batch_size, args.max_seq_len, self.kv_lora_rank), persistent=False)
            self.register_buffer("pe_cache", torch.zeros(args.max_batch_size, args.max_seq_len, self.qk_rope_head_dim), persistent=False)

    def forward(self, x: torch.Tensor, start_pos: int, freqs_cis: torch.Tensor, mask: Optional[torch.Tensor]):
        """
        Forward pass for the Multi-Headed Attention Layer (MLA).

        Args:
            x (torch.Tensor): Input tensor of shape (batch_size, seq_len, dim).
            start_pos (int): Starting position in the sequence for caching.
            freqs_cis (torch.Tensor): Precomputed complex exponential values for rotary embeddings.
            mask (Optional[torch.Tensor]): Mask tensor to exclude certain positions from attention.

        Returns:
            torch.Tensor: Output tensor with the same shape as the input.
        """
        bsz, seqlen, _ = x.size()
        end_pos = start_pos + seqlen
        if self.q_lora_rank == 0:
            q = self.wq(x)
        else:
            q = self.wq_b(self.q_norm(self.wq_a(x)))
        q = q.view(bsz, seqlen, self.n_local_heads, self.qk_head_dim)
        q_nope, q_pe = torch.split(q, [self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1)
        q_pe = apply_rotary_emb(q_pe, freqs_cis)
        kv = self.wkv_a(x)
        kv, k_pe = torch.split(kv, [self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)
        k_pe = apply_rotary_emb(k_pe.unsqueeze(2), freqs_cis)
        if attn_impl == "naive":
            q = torch.cat([q_nope, q_pe], dim=-1)
            kv = self.wkv_b(self.kv_norm(kv))
            kv = kv.view(bsz, seqlen, self.n_local_heads, self.qk_nope_head_dim + self.v_head_dim)
            k_nope, v = torch.split(kv, [self.qk_nope_head_dim, self.v_head_dim], dim=-1)
            k = torch.cat([k_nope, k_pe.expand(-1, -1, self.n_local_heads, -1)], dim=-1)
            self.k_cache[:bsz, start_pos:end_pos] = k
            self.v_cache[:bsz, start_pos:end_pos] = v
            scores = torch.einsum("bshd,bthd->bsht", q, self.k_cache[:bsz, :end_pos]) * self.softmax_scale
        else:
            wkv_b = self.wkv_b.weight if self.wkv_b.scale is None else weight_dequant(self.wkv_b.weight, self.wkv_b.scale, block_size) 
            wkv_b = wkv_b.view(self.n_local_heads, -1, self.kv_lora_rank)
            q_nope = torch.einsum("bshd,hdc->bshc", q_nope, wkv_b[:, :self.qk_nope_head_dim])
            self.kv_cache[:bsz, start_pos:end_pos] = self.kv_norm(kv)
            self.pe_cache[:bsz, start_pos:end_pos] = k_pe.squeeze(2)
            scores = (torch.einsum("bshc,btc->bsht", q_nope, self.kv_cache[:bsz, :end_pos]) +
                      torch.einsum("bshr,btr->bsht", q_pe, self.pe_cache[:bsz, :end_pos])) * self.softmax_scale
        if mask is not None:
            scores += mask.unsqueeze(1)
        scores = scores.softmax(dim=-1, dtype=torch.float32).type_as(x)
        if attn_impl == "naive":
            x = torch.einsum("bsht,bthd->bshd", scores, self.v_cache[:bsz, :end_pos])
        else:
            x = torch.einsum("bsht,btc->bshc", scores, self.kv_cache[:bsz, :end_pos])
            x = torch.einsum("bshc,hdc->bshd", x, wkv_b[:, -self.v_head_dim:])
        x = self.wo(x.flatten(2))
        return x


class MLP(nn.Module):
    """
    Multi-Layer Perceptron (MLP) used as a feed-forward layer.

    Attributes:
        w1 (nn.Module): Linear layer for input-to-hidden transformation.
        w2 (nn.Module): Linear layer for hidden-to-output transformation.
        w3 (nn.Module): Additional linear layer for feature transformation.
    """
    def __init__(self, dim: int, inter_dim: int):
        """
        Initializes the MLP layer.

        Args:
            dim (int): Input and output dimensionality.
            inter_dim (int): Hidden layer dimensionality.
        """
        super().__init__()
        self.w1 = ColumnParallelLinear(dim, inter_dim)
        self.w2 = RowParallelLinear(inter_dim, dim)
        self.w3 = ColumnParallelLinear(dim, inter_dim)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        """
        Forward pass for the MLP layer.

        Args:
            x (torch.Tensor): Input tensor.

        Returns:
            torch.Tensor: Output tensor after MLP computation.
        """
        return self.w2(F.silu(self.w1(x)) * self.w3(x))


class Gate(nn.Module):
    """
    Gating mechanism for routing inputs in a mixture-of-experts (MoE) model.

    Attributes:
        dim (int): Dimensionality of input features.
        topk (int): Number of top experts activated for each input.
        n_groups (int): Number of groups for routing.
        topk_groups (int): Number of groups to route inputs to.
        score_func (str): Scoring function ('softmax' or 'sigmoid').
        route_scale (float): Scaling factor for routing weights.
        weight (torch.nn.Parameter): Learnable weights for the gate.
        bias (Optional[torch.nn.Parameter]): Optional bias term for the gate.
    """
    def __init__(self, args: ModelArgs):
        """
        Initializes the Gate module.

        Args:
            args (ModelArgs): Model arguments containing gating parameters.
        """
        super().__init__()
        self.dim = args.dim
        self.topk = args.n_activated_experts
        self.n_groups = args.n_expert_groups
        self.topk_groups = args.n_limited_groups
        self.score_func = args.score_func
        self.route_scale = args.route_scale
        self.weight = nn.Parameter(torch.empty(args.n_routed_experts, args.dim))
        self.bias = nn.Parameter(torch.empty(args.n_routed_experts)) if self.dim == 7168 else None

    def forward(self, x: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
        """
        Forward pass for the gating mechanism.

        Args:
            x (torch.Tensor): Input tensor.

        Returns:
            Tuple[torch.Tensor, torch.Tensor]: Routing weights and selected expert indices.
        """
        scores = linear(x, self.weight)
        if self.score_func == "softmax":
            scores = scores.softmax(dim=-1, dtype=torch.float32)
        else:
            scores = scores.sigmoid()
        original_scores = scores
        if self.bias is not None:
            scores = scores + self.bias
        if self.n_groups > 1:
            scores = scores.view(x.size(0), self.n_groups, -1)
            if self.bias is None:
                group_scores = scores.amax(dim=-1)
            else:
                group_scores = scores.topk(2, dim=-1)[0].sum(dim=-1)
            indices = group_scores.topk(self.topk_groups, dim=-1)[1]
            mask = scores.new_ones(x.size(0), self.n_groups, dtype=bool).scatter_(1, indices, False)
            scores = scores.masked_fill_(mask.unsqueeze(-1), float("-inf")).flatten(1)
        indices = torch.topk(scores, self.topk, dim=-1)[1]
        weights = original_scores.gather(1, indices)
        if self.score_func == "sigmoid":
            weights /= weights.sum(dim=-1, keepdim=True)
        weights *= self.route_scale
        return weights.type_as(x), indices


class Expert(nn.Module):
    """
    Expert layer for Mixture-of-Experts (MoE) models.

    Attributes:
        w1 (nn.Module): Linear layer for input-to-hidden transformation.
        w2 (nn.Module): Linear layer for hidden-to-output transformation.
        w3 (nn.Module): Additional linear layer for feature transformation.
    """
    def __init__(self, dim: int, inter_dim: int):
        """
        Initializes the Expert layer.

        Args:
            dim (int): Input and output dimensionality.
            inter_dim (int): Hidden layer dimensionality.
        """
        super().__init__()
        self.w1 = Linear(dim, inter_dim)
        self.w2 = Linear(inter_dim, dim)
        self.w3 = Linear(dim, inter_dim)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        """
        Forward pass for the Expert layer.

        Args:
            x (torch.Tensor): Input tensor.

        Returns:
            torch.Tensor: Output tensor after expert computation.
        """
        return self.w2(F.silu(self.w1(x)) * self.w3(x))


class MoE(nn.Module):
    """
    Mixture-of-Experts (MoE) module.

    Attributes:
        dim (int): Dimensionality of input features.
        n_routed_experts (int): Total number of experts in the model.
        n_local_experts (int): Number of experts handled locally in distributed systems.
        n_activated_experts (int): Number of experts activated for each input.
        gate (nn.Module): Gating mechanism to route inputs to experts.
        experts (nn.ModuleList): List of expert modules.
        shared_experts (nn.Module): Shared experts applied to all inputs.
    """
    def __init__(self, args: ModelArgs):
        """
        Initializes the MoE module.

        Args:
            args (ModelArgs): Model arguments containing MoE parameters.
        """
        super().__init__()
        self.dim = args.dim
        assert args.n_routed_experts % world_size == 0, f"Number of experts must be divisible by world size (world_size={world_size})"
        self.n_routed_experts = args.n_routed_experts
        self.n_local_experts = args.n_routed_experts // world_size
        self.n_activated_experts = args.n_activated_experts
        self.experts_start_idx = rank * self.n_local_experts
        self.experts_end_idx = self.experts_start_idx + self.n_local_experts
        self.gate = Gate(args)
        self.experts = nn.ModuleList([Expert(args.dim, args.moe_inter_dim) if self.experts_start_idx <= i < self.experts_end_idx else None
                                      for i in range(self.n_routed_experts)])
        self.shared_experts = MLP(args.dim, args.n_shared_experts * args.moe_inter_dim)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        """
        Forward pass for the MoE module.

        Args:
            x (torch.Tensor): Input tensor.

        Returns:
            torch.Tensor: Output tensor after expert routing and computation.
        """
        shape = x.size()
        x = x.view(-1, self.dim)
        weights, indices = self.gate(x)
        y = torch.zeros_like(x)
        counts = torch.bincount(indices.flatten(), minlength=self.n_routed_experts).tolist()
        for i in range(self.experts_start_idx, self.experts_end_idx):
            if counts[i] == 0:
                continue
            expert = self.experts[i]
            idx, top = torch.where(indices == i)
            y[idx] += expert(x[idx]) * weights[idx, top, None]
        z = self.shared_experts(x)
        if world_size > 1:
            dist.all_reduce(y)
        return (y + z).view(shape)


class Block(nn.Module):
    """
    Transformer block combining attention and feed-forward layers.

    Attributes:
        attn (nn.Module): Attention layer (MLA).
        ffn (nn.Module): Feed-forward network (MLP or MoE).
        attn_norm (nn.Module): Layer normalization for attention.
        ffn_norm (nn.Module): Layer normalization for feed-forward network.
    """
    def __init__(self, layer_id: int, args: ModelArgs):
        """
        Initializes the Transformer block.

        Args:
            layer_id (int): Layer index in the transformer.
            args (ModelArgs): Model arguments containing block parameters.
        """
        super().__init__()
        self.attn = MLA(args)
        self.ffn = MLP(args.dim, args.inter_dim) if layer_id < args.n_dense_layers else MoE(args)
        self.attn_norm = RMSNorm(args.dim)
        self.ffn_norm = RMSNorm(args.dim)

    def forward(self, x: torch.Tensor, start_pos: int, freqs_cis: torch.Tensor, mask: Optional[torch.Tensor]) -> torch.Tensor:
        """
        Forward pass for the Transformer block.

        Args:
            x (torch.Tensor): Input tensor.
            start_pos (int): Starting position in the sequence.
            freqs_cis (torch.Tensor): Precomputed complex exponential values for rotary embeddings.
            mask (Optional[torch.Tensor]): Mask tensor to exclude certain positions from attention.

        Returns:
            torch.Tensor: Output tensor after block computation.
        """
        x = x + self.attn(self.attn_norm(x), start_pos, freqs_cis, mask)
        x = x + self.ffn(self.ffn_norm(x))
        return x


class Transformer(nn.Module):
    """
    Transformer model with positional embeddings, multiple layers, and output projection.

    Attributes:
        max_seq_len (int): Maximum sequence length for the transformer.
        embed (nn.Module): Embedding layer for input tokens.
        layers (torch.nn.ModuleList): List of transformer blocks.
        norm (nn.Module): Layer normalization applied after all blocks.
        head (nn.Module): Output projection layer mapping to vocabulary size.
        freqs_cis (torch.Tensor): Precomputed complex exponential values for rotary embeddings.
    """
    def __init__(self, args: ModelArgs):
        """
        Initializes the Transformer model.

        Args:
            args (ModelArgs): Model arguments containing transformer parameters.
        """
        global world_size, rank
        world_size = dist.get_world_size() if dist.is_initialized() else 1
        rank = dist.get_rank() if dist.is_initialized() else 0
        Linear.dtype = torch.float8_e4m3fn if args.dtype == "fp8" else torch.bfloat16
        super().__init__()
        self.max_seq_len = args.max_seq_len
        self.embed = ParallelEmbedding(args.vocab_size, args.dim)
        self.layers = torch.nn.ModuleList()
        for layer_id in range(args.n_layers):
            self.layers.append(Block(layer_id, args))
        self.norm = RMSNorm(args.dim)
        self.head = ColumnParallelLinear(args.dim, args.vocab_size, dtype=torch.get_default_dtype())
        self.register_buffer("freqs_cis", precompute_freqs_cis(args), persistent=False)

    @torch.inference_mode()
    def forward(self, tokens: torch.Tensor, start_pos: int = 0):
        """
        Forward pass for the Transformer model.

        Args:
            tokens (torch.Tensor): Input tensor of token IDs with shape (batch_size, seq_len).
            start_pos (int, optional): Starting position in the sequence for rotary embeddings. Defaults to 0.

        Returns:
            torch.Tensor: Logits tensor of shape (batch_size, vocab_size).
        """
        seqlen = tokens.size(1)
        h = self.embed(tokens)
        freqs_cis = self.freqs_cis[start_pos:start_pos+seqlen]
        mask = None
        if seqlen > 1:
            mask = torch.full((seqlen, seqlen), float("-inf"), device=tokens.device).triu_(1)
        for layer in self.layers:
            h = layer(h, start_pos, freqs_cis, mask)
        h = self.norm(h)[:, -1]
        logits = self.head(h)
        if world_size > 1:
            all_logits = [torch.empty_like(logits) for _ in range(world_size)]
            dist.all_gather(all_logits, logits)
            logits = torch.cat(all_logits, dim=-1)
        return logits


if __name__ == "__main__":
    torch.set_default_dtype(torch.bfloat16)
    torch.set_default_device("cuda")
    torch.manual_seed(0)
    args = ModelArgs()
    x = torch.randint(0, args.vocab_size, (2, 128))
    model = Transformer(args)
    print(model(x).size())

100

101

102

103

104

105

106

107

108

109

110

111

112

113

114

115

116

117

118

119

120

121

122

123

124

125

126

127

128

129

130

131

132

133

134

135

136

137

138

139

140

141

142

143

144

145

146

147

148

149

150

151

152

153

154

155

156

157

158

159

160

161

162

163

164

165

166

167

168

169

170

171

172

173

174

175

176

177

178

179

180

181

182

183

184

185

186

187

188

189

190

191

192

193

194

195

196

197

198

199

200

201

202

203

204

205

206

207

208

209

210

211

212

213

214

215

216

217

218

219

220

221

222

223

224

225

226

227

228

229

230

231

232

233

234

235

236

237

238

239

240

241

242

243

244

245

246

247

248

249

250

251

252

253

254

255

256

257

258

259

260

261

262

263

264

265

266

267

268

269

270

271

272

273

274

275

276

277

278

279

280

281

282

283

284

285

286

287

288

289

290

291

292

293

294

295

296

297

298

299

300

301

302

303

304

305

306

307

308

309

310

311

312

313

314

315

316

317

318

319

320

321

322

323

324

325

326

327

328

329

330

331

332

333

334

335

336

337

338

339

340

341

342

343

344

345

346

347

348

349

350

351

352

353

354

355

356

357

358

359

360

361

362

363

364

365

366

367

368

369

370

371

372

373

374

375

376

377

378

379

380

381

382

383

384

385

386

387

388

389

390

391

392

393

394

395

396

397

398

399

400

401

402

403

404

405

406

407

408

409

410

411

412

413

414

415

416

417

418

419

420

421

422

423

424

425

426

427

428

429

430

431

432

433

434

435

436

437

438

439

440

441

442

443

444

445

446

447

448

449

450

451

452

453

454

455

456

457

458

459

460

461

462

463

464

465

466

467

468

469

470

471

472

473

474

475

476

477

478

479

480

481

482

483

484

485

486

487

488

489

490

491

492

493

494

495

496

497

498

499

500

501

502

503

504

505

506

507

508

509

510

511

512

513

514

515

516

517

518

519

520

521

522

523

524

525

526

527

528

529

530

531

532

533

534

535

536

537

538

539

540

541

542

543

544

545

546

547

548

549

550

551

552

553

554

555

556

557

558

559

560

561

562

563

564

565

566

567

568

569

570

571

572

573

574

575

576

577

578

579

580

581

582

583

584

585

586

587

588

589

590

591

592

593

594

595

596

597

598

599

600

601

602

603

604

605

606

607

608

609

610

611

612

613

614

615

616

617

618

619

620

621

622

623

624

625

626

627

628

629

630

631

632

633

634

635

636

637

638

639

640

641

642

643

644

645

646

647

648

649

650

651

652

653

654

655

656

657

658

659

660

661

662

663

664

665

666

667

668

669

670

671

672

673

674

675

676

677

678

679

680

681

682

683

684

685

686

687

688

689

690

691

692

693

694

695

696

697

698

699

700

701

702

703

704

705

706

707

708

709

710

711

712

713

714

715

716

717

718

719

720

721

722

723

724

725

726

727

728

729

730

731

732

733

734

735

736

737

738

739

740

741

742

743

744

745

746

747

748

749

750

751

752

753

754

755

756

757

758

759

760

761

762

763

764

765

766

767

768

769

770

771

772

773

774

775

776

777

778

779

780

781

782

783

784

785

786

787

788

789

790

791

792

793

794

795

796

797

798

799

800

801

802

803

804

805

806

import math

from dataclasses import dataclass

from typing import Tuple, Optional, Literal

import torch

from torch import nn

import torch.nn.functional as F

import torch.distributed as dist

from kernel import act_quant, weight_dequant, fp8_gemm

world_size = 1

rank = 0

block_size = 128

gemm_impl: Literal["bf16", "fp8"] = "bf16"

attn_impl: Literal["naive", "absorb"] = "absorb"

@dataclass

class ModelArgs:

"""

Data class for defining model arguments and hyperparameters.

Attributes:

max_batch_size (int): Maximum batch size.

max_seq_len (int): Maximum sequence length.

dtype (Literal["bf16", "fp8"]): Data type for computations.

vocab_size (int): Vocabulary size.

dim (int): Model dimension.

inter_dim (int): Intermediate dimension for MLP layers.

moe_inter_dim (int): Intermediate dimension for MoE layers.

n_layers (int): Number of transformer layers.

n_dense_layers (int): Number of dense layers in the model.

n_heads (int): Number of attention heads.

n_routed_experts (int): Number of routed experts for MoE layers.

n_shared_experts (int): Number of shared experts for MoE layers.

n_activated_experts (int): Number of activated experts in MoE layers.

n_expert_groups (int): Number of expert groups.

n_limited_groups (int): Number of limited groups for MoE routing.

score_func (Literal["softmax", "sigmoid"]): Scoring function for MoE routing.

route_scale (float): Scaling factor for routing scores.

q_lora_rank (int): LoRA rank for query projections.

kv_lora_rank (int): LoRA rank for key-value projections.

qk_nope_head_dim (int): Dimension for query-key projections without positional embeddings.

qk_rope_head_dim (int): Dimension for query-key projections with rotary embeddings.

v_head_dim (int): Dimension for value projections.

original_seq_len (int): Original sequence length.

rope_theta (float): Base for rotary positional encoding.

rope_factor (float): Scaling factor for extended sequence lengths.

beta_fast (int): Fast beta correction factor.

beta_slow (int): Slow beta correction factor.

mscale (float): Scaling factor for extended attention.

"""

max_batch_size: int = 8

max_seq_len: int = 4096 * 4

dtype: Literal["bf16", "fp8"] = "bf16"

vocab_size: int = 102400

dim: int = 2048

inter_dim: int = 10944

moe_inter_dim: int = 1408

n_layers: int = 27

n_dense_layers: int = 1

n_heads: int = 16

# moe

n_routed_experts: int = 64

n_shared_experts: int = 2

n_activated_experts: int = 6

n_expert_groups: int = 1

n_limited_groups: int = 1

score_func: Literal["softmax", "sigmoid"] = "softmax"

route_scale: float = 1.

# mla

q_lora_rank: int = 0

kv_lora_rank: int = 512

qk_nope_head_dim: int = 128

qk_rope_head_dim: int = 64

v_head_dim: int = 128

# yarn

original_seq_len: int = 4096

rope_theta: float = 10000.0

rope_factor: float = 40

beta_fast: int = 32

beta_slow: int = 1

mscale: float = 1.

class ParallelEmbedding(nn.Module):

"""

Embedding layer with parallelism support across distributed processes.

Args:

vocab_size (int): Vocabulary size.

dim (int): Embedding dimension.

"""

def __init__(self, vocab_size: int, dim: int):

super().__init__()

self.vocab_size = vocab_size

self.dim = dim

assert vocab_size % world_size == 0, f"Vocabulary size must be divisible by world size (world_size={world_size})"

self.part_vocab_size = (vocab_size // world_size)

self.vocab_start_idx = rank * self.part_vocab_size

self.vocab_end_idx = self.vocab_start_idx + self.part_vocab_size

self.weight = nn.Parameter(torch.empty(self.part_vocab_size, self.dim))

def forward(self, x: torch.Tensor) -> torch.Tensor:

"""

Forward pass for parallel embedding layer.

Args:

x (torch.Tensor): Input tensor containing token indices.

Returns:

torch.Tensor: Embedded representations.

Raises:

ValueError: If `world_size` is not defined.

"""

if world_size > 1:

mask = (x < self.vocab_start_idx) | (x >= self.vocab_end_idx)

x = x - self.vocab_start_idx

x[mask] = 0

y = F.embedding(x, self.weight)

if world_size > 1:

y[mask] = 0

dist.all_reduce(y)

return y

def linear(x: torch.Tensor, weight: torch.Tensor, bias: Optional[torch.Tensor] = None) -> torch.Tensor:

"""

Applies a linear transformation to the incoming data: y = xA^T + b.

This function supports specialized implementations based on quantization

and tensor formats.

Args:

x (torch.Tensor): The input tensor.

weight (torch.Tensor): The weight tensor. It may be quantized and

requires dequantization for certain cases.

bias (Optional[torch.Tensor]): The bias tensor to be added. Default is None.

Returns:

torch.Tensor: The result of the linear transformation, which may involve

quantization-aware computations depending on the input parameters.

Notes:

- If `weight` is quantized (e.g., `element_size() == 1`), a dequantized version

is used for computation.

- If `gemm_impl == "bf16"`, dequantization and a `bf16` GEMM operation are applied.

- For other cases, the function applies quantization to `x` and uses `fp8_gemm` for computation.

"""

if weight.element_size() > 1:

return F.linear(x, weight, bias)

elif gemm_impl == "bf16":

weight = weight_dequant(weight, weight.scale)

return F.linear(x, weight, bias)

else:

x, scale = act_quant(x, block_size)

y = fp8_gemm(x, scale, weight, weight.scale)

if bias is not None:

y += bias

return y

class Linear(nn.Module):

"""

Custom linear layer with support for quantized weights and optional bias.

Args:

in_features (int): Number of input features.

out_features (int): Number of output features.

bias (bool): Whether to include a bias term. Defaults to False.

dtype (optional): Data type for the layer. Defaults to `torch.bfloat16`.

"""

dtype = torch.bfloat16

def __init__(self, in_features: int, out_features: int, bias: bool = False, dtype = None):

super().__init__()

self.in_features = in_features

self.out_features = out_features

self.weight = nn.Parameter(torch.empty(out_features, in_features, dtype=dtype or Linear.dtype))

if self.weight.element_size() == 1:

scale_out_features = (out_features + block_size - 1) // block_size

scale_in_features = (in_features + block_size - 1) // block_size

self.weight.scale = self.scale = nn.Parameter(torch.empty(scale_out_features, scale_in_features, dtype=torch.float32))

else:

self.register_parameter("scale", None)

if bias:

self.bias = nn.Parameter(torch.empty(out_features))

else:

self.register_parameter("bias", None)

def forward(self, x: torch.Tensor) -> torch.Tensor:

"""

Forward pass for the custom linear layer.

Args:

x (torch.Tensor): Input tensor.

Returns:

torch.Tensor: Transformed tensor after linear computation.

"""

return linear(x, self.weight, self.bias)

class ColumnParallelLinear(Linear):

"""

Linear layer with column parallelism, splitting output features across distributed processes.

Args:

in_features (int): Number of input features.

out_features (int): Total number of output features.

bias (bool): Whether to include a bias term. Defaults to False.

dtype (optional): Data type for the layer. Defaults to `torch.bfloat16`.

"""

def __init__(self, in_features: int, out_features: int, bias: bool = False, dtype = None):

assert out_features % world_size == 0, f"Output features must be divisible by world size (world_size={world_size})"

self.part_out_features = out_features // world_size

super().__init__(in_features, self.part_out_features, bias, dtype)

def forward(self, x: torch.Tensor) -> torch.Tensor:

"""

Forward pass for column parallel linear layer.

Args:

x (torch.Tensor): Input tensor.

Returns:

torch.Tensor: Transformed tensor with column-parallel computation.

"""

y = linear(x, self.weight, self.bias)

return y

class RowParallelLinear(Linear):

"""

Linear layer with row parallelism, splitting input features across distributed processes.

Args:

in_features (int): Total number of input features.

out_features (int): Number of output features.

bias (bool): Whether to include a bias term. Defaults to False.

dtype (optional): Data type for the layer. Defaults to `torch.bfloat16`.

"""

def __init__(self, in_features: int, out_features: int, bias: bool = False, dtype = None):

assert in_features % world_size == 0, f"Input features must be divisible by world size (world_size={world_size})"

self.part_in_features = in_features // world_size

super().__init__(self.part_in_features, out_features, bias, dtype)

def forward(self, x: torch.Tensor) -> torch.Tensor:

"""

Forward pass for row parallel linear layer.

Args:

x (torch.Tensor): Input tensor.

Returns:

torch.Tensor: Transformed tensor with row-parallel computation.

"""

y = linear(x, self.weight)

if world_size > 1:

dist.all_reduce(y)

if self.bias is not None:

y += self.bias

return y

class RMSNorm(nn.Module):

"""

Root Mean Square Layer Normalization (RMSNorm).

Args:

dim (int): Dimension of the input tensor.

eps (float): Epsilon value for numerical stability. Defaults to 1e-6.

"""

def __init__(self, dim: int, eps: float = 1e-6):

super().__init__()

self.dim = dim

self.eps = eps

self.weight = nn.Parameter(torch.ones(dim))

def forward(self, x: torch.Tensor):

"""

Forward pass for RMSNorm.

Args:

x (torch.Tensor): Input tensor.

Returns:

torch.Tensor: Normalized tensor with the same shape as input.

"""

return F.rms_norm(x, (self.dim,), self.weight, self.eps)

def precompute_freqs_cis(args: ModelArgs) -> torch.Tensor:

"""

Precomputes frequency-based complex exponential values for rotary positional embeddings.

Args:

args (ModelArgs): Model arguments containing positional embedding parameters.

Returns:

torch.Tensor: Precomputed complex exponential values for positional embeddings.

"""

dim = args.qk_rope_head_dim

seqlen = args.max_seq_len

beta_fast = args.beta_fast

beta_slow = args.beta_slow

base = args.rope_theta

factor = args.rope_factor

def find_correction_dim(num_rotations, dim, base, max_seq_len):

"""

Computes the correction dimension for a given number of rotations in the rotary positional embedding.

Args:

num_rotations (float): Number of rotations to compute the correction for.

dim (int): Dimensionality of the embedding space.

base (float): Base value for the exponential computation.

max_seq_len (int): Maximum sequence length.

Returns:

float: The correction dimension based on the input parameters.

"""

return dim * math.log(max_seq_len / (num_rotations * 2 * math.pi)) / (2 * math.log(base))

def find_correction_range(low_rot, high_rot, dim, base, max_seq_len):

"""

Computes the range of correction dimensions for rotary positional embeddings.

Args:

low_rot (float): Lower bound for the number of rotations.

high_rot (float): Upper bound for the number of rotations.

dim (int): Dimensionality of the embedding space.

base (float): Base value for the exponential computation.

max_seq_len (int): Maximum sequence length.

Returns:

Tuple[int, int]: The range of correction dimensions (low, high), clamped to valid indices.

"""

low = math.floor(find_correction_dim(low_rot, dim, base, max_seq_len))

high = math.ceil(find_correction_dim(high_rot, dim, base, max_seq_len))

return max(low, 0), min(high, dim-1)

def linear_ramp_factor(min, max, dim):

"""

Computes a linear ramp function used to smooth values between a minimum and maximum range.

Args:

min (float): Minimum value for the ramp function.

max (float): Maximum value for the ramp function.

dim (int): Dimensionality of the ramp tensor.

Returns:

torch.Tensor: A tensor of shape (dim,) with values linearly interpolated between 0 and 1,

clamped to the range [0, 1].

"""

if min == max:

max += 0.001

linear_func = (torch.arange(dim, dtype=torch.float32) - min) / (max - min)

ramp_func = torch.clamp(linear_func, 0, 1)

return ramp_func

freqs = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim))

if seqlen > args.original_seq_len:

low, high = find_correction_range(beta_fast, beta_slow, dim, base, args.original_seq_len)

smooth = 1 - linear_ramp_factor(low, high, dim // 2)

freqs = freqs / factor * (1 - smooth) + freqs * smooth

t = torch.arange(seqlen)

freqs = torch.outer(t, freqs)

freqs_cis = torch.polar(torch.ones_like(freqs), freqs)

return freqs_cis

def apply_rotary_emb(x: torch.Tensor, freqs_cis: torch.Tensor) -> torch.Tensor:

"""

Applies rotary positional embeddings to the input tensor.

Args:

x (torch.Tensor): Input tensor with positional embeddings to be applied.

freqs_cis (torch.Tensor): Precomputed complex exponential values for positional embeddings.

Returns:

torch.Tensor: Tensor with rotary embeddings applied.

"""

dtype = x.dtype

x = torch.view_as_complex(x.float().view(*x.shape[:-1], -1, 2))

freqs_cis = freqs_cis.view(1, x.size(1), 1, x.size(-1))

y = torch.view_as_real(x * freqs_cis).flatten(3)

return y.to(dtype)

class MLA(nn.Module):

"""

Multi-Headed Attention Layer (MLA).

Attributes:

dim (int): Dimensionality of the input features.

n_heads (int): Number of attention heads.

n_local_heads (int): Number of local attention heads for distributed systems.

q_lora_rank (int): Rank for low-rank query projection.

kv_lora_rank (int): Rank for low-rank key/value projection.

qk_nope_head_dim (int): Dimensionality of non-positional query/key projections.

qk_rope_head_dim (int): Dimensionality of rotary-positional query/key projections.

qk_head_dim (int): Total dimensionality of query/key projections.

v_head_dim (int): Dimensionality of value projections.

softmax_scale (float): Scaling factor for softmax in attention computation.

"""

def __init__(self, args: ModelArgs):

super().__init__()

self.dim = args.dim

self.n_heads = args.n_heads

self.n_local_heads = args.n_heads // world_size

self.q_lora_rank = args.q_lora_rank

self.kv_lora_rank = args.kv_lora_rank

self.qk_nope_head_dim = args.qk_nope_head_dim

self.qk_rope_head_dim = args.qk_rope_head_dim

self.qk_head_dim = args.qk_nope_head_dim + args.qk_rope_head_dim

self.v_head_dim = args.v_head_dim

if self.q_lora_rank == 0:

self.wq = ColumnParallelLinear(self.dim, self.n_heads * self.qk_head_dim)

else:

self.wq_a = Linear(self.dim, self.q_lora_rank)

self.q_norm = RMSNorm(self.q_lora_rank)

self.wq_b = ColumnParallelLinear(self.q_lora_rank, self.n_heads * self.qk_head_dim)

self.wkv_a = Linear(self.dim, self.kv_lora_rank + self.qk_rope_head_dim)

self.kv_norm = RMSNorm(self.kv_lora_rank)

self.wkv_b = ColumnParallelLinear(self.kv_lora_rank, self.n_heads * (self.qk_nope_head_dim + self.v_head_dim))

self.wo = RowParallelLinear(self.n_heads * self.v_head_dim, self.dim)

self.softmax_scale = self.qk_head_dim ** -0.5

if args.max_seq_len > args.original_seq_len:

mscale = 0.1 * args.mscale * math.log(args.rope_factor) + 1.0

self.softmax_scale = self.softmax_scale * mscale * mscale

if attn_impl == "naive":

self.register_buffer("k_cache", torch.zeros(args.max_batch_size, args.max_seq_len, self.n_local_heads, self.qk_head_dim), persistent=False)

self.register_buffer("v_cache", torch.zeros(args.max_batch_size, args.max_seq_len, self.n_local_heads, self.v_head_dim), persistent=False)

else:

self.register_buffer("kv_cache", torch.zeros(args.max_batch_size, args.max_seq_len, self.kv_lora_rank), persistent=False)

self.register_buffer("pe_cache", torch.zeros(args.max_batch_size, args.max_seq_len, self.qk_rope_head_dim), persistent=False)

def forward(self, x: torch.Tensor, start_pos: int, freqs_cis: torch.Tensor, mask: Optional[torch.Tensor]):

"""

Forward pass for the Multi-Headed Attention Layer (MLA).

Args:

x (torch.Tensor): Input tensor of shape (batch_size, seq_len, dim).

start_pos (int): Starting position in the sequence for caching.

freqs_cis (torch.Tensor): Precomputed complex exponential values for rotary embeddings.

mask (Optional[torch.Tensor]): Mask tensor to exclude certain positions from attention.

Returns:

torch.Tensor: Output tensor with the same shape as the input.

"""

bsz, seqlen, _ = x.size()

end_pos = start_pos + seqlen

if self.q_lora_rank == 0:

q = self.wq(x)

else:

q = self.wq_b(self.q_norm(self.wq_a(x)))

q = q.view(bsz, seqlen, self.n_local_heads, self.qk_head_dim)

q_nope, q_pe = torch.split(q, [self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1)

q_pe = apply_rotary_emb(q_pe, freqs_cis)

kv = self.wkv_a(x)

kv, k_pe = torch.split(kv, [self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)

k_pe = apply_rotary_emb(k_pe.unsqueeze(2), freqs_cis)

if attn_impl == "naive":

q = torch.cat([q_nope, q_pe], dim=-1)

kv = self.wkv_b(self.kv_norm(kv))

kv = kv.view(bsz, seqlen, self.n_local_heads, self.qk_nope_head_dim + self.v_head_dim)

k_nope, v = torch.split(kv, [self.qk_nope_head_dim, self.v_head_dim], dim=-1)

k = torch.cat([k_nope, k_pe.expand(-1, -1, self.n_local_heads, -1)], dim=-1)

self.k_cache[:bsz, start_pos:end_pos] = k

self.v_cache[:bsz, start_pos:end_pos] = v

scores = torch.einsum("bshd,bthd->bsht", q, self.k_cache[:bsz, :end_pos]) * self.softmax_scale

else:

wkv_b = self.wkv_b.weight if self.wkv_b.scale is None else weight_dequant(self.wkv_b.weight, self.wkv_b.scale, block_size)

wkv_b = wkv_b.view(self.n_local_heads, -1, self.kv_lora_rank)

q_nope = torch.einsum("bshd,hdc->bshc", q_nope, wkv_b[:, :self.qk_nope_head_dim])

self.kv_cache[:bsz, start_pos:end_pos] = self.kv_norm(kv)

self.pe_cache[:bsz, start_pos:end_pos] = k_pe.squeeze(2)

scores = (torch.einsum("bshc,btc->bsht", q_nope, self.kv_cache[:bsz, :end_pos]) +

torch.einsum("bshr,btr->bsht", q_pe, self.pe_cache[:bsz, :end_pos])) * self.softmax_scale

if mask is not None:

scores += mask.unsqueeze(1)

scores = scores.softmax(dim=-1, dtype=torch.float32).type_as(x)

if attn_impl == "naive":

x = torch.einsum("bsht,bthd->bshd", scores, self.v_cache[:bsz, :end_pos])

else:

x = torch.einsum("bsht,btc->bshc", scores, self.kv_cache[:bsz, :end_pos])

x = torch.einsum("bshc,hdc->bshd", x, wkv_b[:, -self.v_head_dim:])

x = self.wo(x.flatten(2))

return x

class MLP(nn.Module):

"""

Multi-Layer Perceptron (MLP) used as a feed-forward layer.

Attributes:

w1 (nn.Module): Linear layer for input-to-hidden transformation.

w2 (nn.Module): Linear layer for hidden-to-output transformation.

w3 (nn.Module): Additional linear layer for feature transformation.

"""

def __init__(self, dim: int, inter_dim: int):

"""

Initializes the MLP layer.

Args:

dim (int): Input and output dimensionality.

inter_dim (int): Hidden layer dimensionality.

"""

super().__init__()

self.w1 = ColumnParallelLinear(dim, inter_dim)

self.w2 = RowParallelLinear(inter_dim, dim)

self.w3 = ColumnParallelLinear(dim, inter_dim)

def forward(self, x: torch.Tensor) -> torch.Tensor:

"""

Forward pass for the MLP layer.

Args:

x (torch.Tensor): Input tensor.

Returns:

torch.Tensor: Output tensor after MLP computation.

"""

return self.w2(F.silu(self.w1(x)) * self.w3(x))

class Gate(nn.Module):

"""

Gating mechanism for routing inputs in a mixture-of-experts (MoE) model.

Attributes:

dim (int): Dimensionality of input features.

topk (int): Number of top experts activated for each input.

n_groups (int): Number of groups for routing.

topk_groups (int): Number of groups to route inputs to.

score_func (str): Scoring function ('softmax' or 'sigmoid').

route_scale (float): Scaling factor for routing weights.

weight (torch.nn.Parameter): Learnable weights for the gate.

bias (Optional[torch.nn.Parameter]): Optional bias term for the gate.

"""

def __init__(self, args: ModelArgs):

"""

Initializes the Gate module.

Args:

args (ModelArgs): Model arguments containing gating parameters.

"""

super().__init__()

self.dim = args.dim

self.topk = args.n_activated_experts

self.n_groups = args.n_expert_groups

self.topk_groups = args.n_limited_groups

self.score_func = args.score_func

self.route_scale = args.route_scale

self.weight = nn.Parameter(torch.empty(args.n_routed_experts, args.dim))

self.bias = nn.Parameter(torch.empty(args.n_routed_experts)) if self.dim == 7168 else None

def forward(self, x: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:

"""

Forward pass for the gating mechanism.

Args:

x (torch.Tensor): Input tensor.

Returns:

Tuple[torch.Tensor, torch.Tensor]: Routing weights and selected expert indices.

"""

scores = linear(x, self.weight)

if self.score_func == "softmax":

scores = scores.softmax(dim=-1, dtype=torch.float32)

else:

scores = scores.sigmoid()

original_scores = scores

if self.bias is not None:

scores = scores + self.bias

if self.n_groups > 1:

scores = scores.view(x.size(0), self.n_groups, -1)

if self.bias is None:

group_scores = scores.amax(dim=-1)

else:

group_scores = scores.topk(2, dim=-1)[0].sum(dim=-1)

indices = group_scores.topk(self.topk_groups, dim=-1)[1]

mask = scores.new_ones(x.size(0), self.n_groups, dtype=bool).scatter_(1, indices, False)

scores = scores.masked_fill_(mask.unsqueeze(-1), float("-inf")).flatten(1)

indices = torch.topk(scores, self.topk, dim=-1)[1]

weights = original_scores.gather(1, indices)

if self.score_func == "sigmoid":

weights /= weights.sum(dim=-1, keepdim=True)

weights *= self.route_scale

return weights.type_as(x), indices

class Expert(nn.Module):

"""

Expert layer for Mixture-of-Experts (MoE) models.

Attributes:

w1 (nn.Module): Linear layer for input-to-hidden transformation.

w2 (nn.Module): Linear layer for hidden-to-output transformation.

w3 (nn.Module): Additional linear layer for feature transformation.

"""

def __init__(self, dim: int, inter_dim: int):

"""

Initializes the Expert layer.

Args:

dim (int): Input and output dimensionality.

inter_dim (int): Hidden layer dimensionality.

"""

super().__init__()

self.w1 = Linear(dim, inter_dim)

self.w2 = Linear(inter_dim, dim)

self.w3 = Linear(dim, inter_dim)

def forward(self, x: torch.Tensor) -> torch.Tensor:

"""

Forward pass for the Expert layer.

Args:

x (torch.Tensor): Input tensor.

Returns:

torch.Tensor: Output tensor after expert computation.

"""

return self.w2(F.silu(self.w1(x)) * self.w3(x))

class MoE(nn.Module):

"""

Mixture-of-Experts (MoE) module.

Attributes:

dim (int): Dimensionality of input features.

n_routed_experts (int): Total number of experts in the model.

n_local_experts (int): Number of experts handled locally in distributed systems.

n_activated_experts (int): Number of experts activated for each input.

gate (nn.Module): Gating mechanism to route inputs to experts.

experts (nn.ModuleList): List of expert modules.

shared_experts (nn.Module): Shared experts applied to all inputs.

"""

def __init__(self, args: ModelArgs):

"""

Initializes the MoE module.

Args:

args (ModelArgs): Model arguments containing MoE parameters.

"""

super().__init__()

self.dim = args.dim

assert args.n_routed_experts % world_size == 0, f"Number of experts must be divisible by world size (world_size={world_size})"

self.n_routed_experts = args.n_routed_experts

self.n_local_experts = args.n_routed_experts // world_size

self.n_activated_experts = args.n_activated_experts

self.experts_start_idx = rank * self.n_local_experts

self.experts_end_idx = self.experts_start_idx + self.n_local_experts

self.gate = Gate(args)

self.experts = nn.ModuleList([Expert(args.dim, args.moe_inter_dim) if self.experts_start_idx <= i < self.experts_end_idx else None

for i in range(self.n_routed_experts)])

self.shared_experts = MLP(args.dim, args.n_shared_experts * args.moe_inter_dim)

def forward(self, x: torch.Tensor) -> torch.Tensor:

"""

Forward pass for the MoE module.

Args:

x (torch.Tensor): Input tensor.

Returns:

torch.Tensor: Output tensor after expert routing and computation.

"""

shape = x.size()

x = x.view(-1, self.dim)

weights, indices = self.gate(x)

y = torch.zeros_like(x)

counts = torch.bincount(indices.flatten(), minlength=self.n_routed_experts).tolist()

for i in range(self.experts_start_idx, self.experts_end_idx):

if counts[i] == 0:

continue

expert = self.experts[i]

idx, top = torch.where(indices == i)

y[idx] += expert(x[idx]) * weights[idx, top, None]

z = self.shared_experts(x)

if world_size > 1:

dist.all_reduce(y)

return (y + z).view(shape)

class Block(nn.Module):

"""

Transformer block combining attention and feed-forward layers.

Attributes:

attn (nn.Module): Attention layer (MLA).

ffn (nn.Module): Feed-forward network (MLP or MoE).

attn_norm (nn.Module): Layer normalization for attention.

ffn_norm (nn.Module): Layer normalization for feed-forward network.

"""

def __init__(self, layer_id: int, args: ModelArgs):

"""

Initializes the Transformer block.

Args:

layer_id (int): Layer index in the transformer.

args (ModelArgs): Model arguments containing block parameters.

"""

super().__init__()

self.attn = MLA(args)

self.ffn = MLP(args.dim, args.inter_dim) if layer_id < args.n_dense_layers else MoE(args)

self.attn_norm = RMSNorm(args.dim)

self.ffn_norm = RMSNorm(args.dim)

def forward(self, x: torch.Tensor, start_pos: int, freqs_cis: torch.Tensor, mask: Optional[torch.Tensor]) -> torch.Tensor:

"""

Forward pass for the Transformer block.

Args:

x (torch.Tensor): Input tensor.

start_pos (int): Starting position in the sequence.

freqs_cis (torch.Tensor): Precomputed complex exponential values for rotary embeddings.

mask (Optional[torch.Tensor]): Mask tensor to exclude certain positions from attention.

Returns:

torch.Tensor: Output tensor after block computation.

"""

x = x + self.attn(self.attn_norm(x), start_pos, freqs_cis, mask)

x = x + self.ffn(self.ffn_norm(x))

return x

class Transformer(nn.Module):

"""

Transformer model with positional embeddings, multiple layers, and output projection.

Attributes:

max_seq_len (int): Maximum sequence length for the transformer.

embed (nn.Module): Embedding layer for input tokens.

layers (torch.nn.ModuleList): List of transformer blocks.

norm (nn.Module): Layer normalization applied after all blocks.

head (nn.Module): Output projection layer mapping to vocabulary size.

freqs_cis (torch.Tensor): Precomputed complex exponential values for rotary embeddings.

"""

def __init__(self, args: ModelArgs):

"""

Initializes the Transformer model.

Args:

args (ModelArgs): Model arguments containing transformer parameters.

"""

global world_size, rank

world_size = dist.get_world_size() if dist.is_initialized() else 1

rank = dist.get_rank() if dist.is_initialized() else 0

Linear.dtype = torch.float8_e4m3fn if args.dtype == "fp8" else torch.bfloat16

super().__init__()

self.max_seq_len = args.max_seq_len

self.embed = ParallelEmbedding(args.vocab_size, args.dim)

self.layers = torch.nn.ModuleList()

for layer_id in range(args.n_layers):

self.layers.append(Block(layer_id, args))

self.norm = RMSNorm(args.dim)

self.head = ColumnParallelLinear(args.dim, args.vocab_size, dtype=torch.get_default_dtype())

self.register_buffer("freqs_cis", precompute_freqs_cis(args), persistent=False)

@torch.inference_mode()

def forward(self, tokens: torch.Tensor, start_pos: int = 0):

"""

Forward pass for the Transformer model.

Args:

tokens (torch.Tensor): Input tensor of token IDs with shape (batch_size, seq_len).

start_pos (int, optional): Starting position in the sequence for rotary embeddings. Defaults to 0.

Returns:

torch.Tensor: Logits tensor of shape (batch_size, vocab_size).

"""

seqlen = tokens.size(1)

h = self.embed(tokens)

freqs_cis = self.freqs_cis[start_pos:start_pos+seqlen]

mask = None

if seqlen > 1:

mask = torch.full((seqlen, seqlen), float("-inf"), device=tokens.device).triu_(1)

for layer in self.layers:

h = layer(h, start_pos, freqs_cis, mask)

h = self.norm(h)[:, -1]

logits = self.head(h)

if world_size > 1:

all_logits = [torch.empty_like(logits) for _ in range(world_size)]

dist.all_gather(all_logits, logits)

logits = torch.cat(all_logits, dim=-1)

return logits

if __name__ == "__main__":

torch.set_default_dtype(torch.bfloat16)

torch.set_default_device("cuda")

torch.manual_seed(0)

args = ModelArgs()

x = torch.randint(0, args.vocab_size, (2, 128))

model = Transformer(args)

print(model(x).size())