LLMs-from-scratch/pkg/llms_from_scratch/kv_cache/qwen3.py

# Copyright (c) Sebastian Raschka under Apache License 2.0 (see LICENSE.txt).
# Source for "Build a Large Language Model From Scratch"
#   - https://www.manning.com/books/build-a-large-language-model-from-scratch
# Code: https://github.com/rasbt/LLMs-from-scratch

from .utils import KVCache   # noqa: F401
from ..qwen3 import (   # noqa: F401
    QWEN_CONFIG_06_B, QWEN3_CONFIG_1_7B, QWEN3_CONFIG_4B,
    QWEN3_CONFIG_8B, QWEN3_CONFIG_14B, QWEN3_CONFIG_32B,
    Qwen3Tokenizer, load_weights_into_qwen,
    download_from_huggingface,
    download_from_huggingface_from_snapshots
)

import torch
import torch.nn as nn


class Qwen3Model(nn.Module):
    def __init__(self, cfg):
        super().__init__()

        # Main model parameters
        self.tok_emb = nn.Embedding(cfg["vocab_size"], cfg["emb_dim"], dtype=cfg["dtype"])

        self.trf_blocks = nn.ModuleList(  # ModuleList since Sequential can only accept one input, and we need `x, mask, cos, sin`
            [TransformerBlock(cfg) for _ in range(cfg["n_layers"])]
        )
        self.final_norm = RMSNorm(cfg["emb_dim"])
        self.out_head = nn.Linear(cfg["emb_dim"], cfg["vocab_size"], bias=False, dtype=cfg["dtype"])

        # Reusuable utilities
        if cfg["head_dim"] is None:
            head_dim = cfg["emb_dim"] // cfg["n_heads"]
        else:
            head_dim = cfg["head_dim"]
        cos, sin = compute_rope_params(
            head_dim=head_dim,
            theta_base=cfg["rope_base"],
            context_length=cfg["context_length"]
        )
        self.register_buffer("cos", cos, persistent=False)
        self.register_buffer("sin", sin, persistent=False)
        self.cfg = cfg
        self.current_pos = 0  # Track current position in KV cache

    def forward(self, in_idx, use_cache=False, cache=None):
        # Forward pass
        tok_embeds = self.tok_emb(in_idx)
        x = tok_embeds

        num_tokens = x.shape[1]
        if use_cache:
            pos_start = self.current_pos
            pos_end = pos_start + num_tokens
            self.current_pos = pos_end
            mask = torch.triu(
                torch.ones(pos_end, pos_end, device=x.device, dtype=torch.bool), diagonal=1
            )[pos_start:pos_end, :pos_end]
        else:
            pos_start = 0  # Not strictly necessary but helps torch.compile
            mask = torch.triu(
                torch.ones(num_tokens, num_tokens, device=x.device, dtype=torch.bool), diagonal=1
            )
        # Shape (1, 1, num_tokens, num_tokens) to broadcast across batch and heads
        mask = mask[None, None, :, :]

        next_cache = []
        for i, block in enumerate(self.trf_blocks):
            blk_cache = cache.get(i) if cache else None
            x, new_blk_cache = block(x, mask, self.cos, self.sin,
                                     use_cache=use_cache,
                                     start_pos=pos_start,
                                     cache=blk_cache)
            if cache:
                cache.update(i, new_blk_cache)
            next_cache.append(new_blk_cache)

        x = self.final_norm(x)
        logits = self.out_head(x.to(self.cfg["dtype"]))
        return logits

    def reset_kv_cache(self):
        self.current_pos = 0


class TransformerBlock(nn.Module):
    def __init__(self, cfg):
        super().__init__()
        self.att = GroupedQueryAttention(
            d_in=cfg["emb_dim"],
            num_heads=cfg["n_heads"],
            head_dim=cfg["head_dim"],
            num_kv_groups=cfg["n_kv_groups"],
            qk_norm=cfg["qk_norm"],
            dtype=cfg["dtype"]
        )
        self.ff = FeedForward(cfg)
        self.norm1 = RMSNorm(cfg["emb_dim"], eps=1e-6)
        self.norm2 = RMSNorm(cfg["emb_dim"], eps=1e-6)

    def forward(self, x, mask, cos, sin, use_cache=False, start_pos=0, cache=None):
        # Shortcut connection for attention block
        shortcut = x
        x = self.norm1(x)
        x, next_cache = self.att(x, mask, cos, sin, use_cache=use_cache, start_pos=start_pos, cache=cache)  # Shape [batch_size, num_tokens, emb_size]
        x = x + shortcut  # Add the original input back

        # Shortcut connection for feed-forward block
        shortcut = x
        x = self.norm2(x)
        x = self.ff(x)
        x = x + shortcut  # Add the original input back

        return x, next_cache


class FeedForward(nn.Module):
    def __init__(self, cfg):
        super().__init__()
        self.fc1 = nn.Linear(cfg["emb_dim"], cfg["hidden_dim"], dtype=cfg["dtype"], bias=False)
        self.fc2 = nn.Linear(cfg["emb_dim"], cfg["hidden_dim"], dtype=cfg["dtype"], bias=False)
        self.fc3 = nn.Linear(cfg["hidden_dim"], cfg["emb_dim"], dtype=cfg["dtype"], bias=False)

    def forward(self, x):
        x_fc1 = self.fc1(x)
        x_fc2 = self.fc2(x)
        x = nn.functional.silu(x_fc1) * x_fc2
        return self.fc3(x)


class GroupedQueryAttention(nn.Module):
    def __init__(
        self, d_in, num_heads, num_kv_groups, head_dim=None, qk_norm=False, dtype=None
    ):
        super().__init__()
        assert num_heads % num_kv_groups == 0, "num_heads must be divisible by num_kv_groups"

        self.num_heads = num_heads
        self.num_kv_groups = num_kv_groups
        self.group_size = num_heads // num_kv_groups

        if head_dim is None:
            assert d_in % num_heads == 0, "`d_in` must be divisible by `num_heads` if `head_dim` is not set"
            head_dim = d_in // num_heads

        self.head_dim = head_dim
        self.d_out = num_heads * head_dim

        self.W_query = nn.Linear(d_in, self.d_out, bias=False, dtype=dtype)
        self.W_key = nn.Linear(d_in, num_kv_groups * head_dim, bias=False, dtype=dtype)
        self.W_value = nn.Linear(d_in, num_kv_groups * head_dim, bias=False, dtype=dtype)

        self.out_proj = nn.Linear(self.d_out, d_in, bias=False, dtype=dtype)

        if qk_norm:
            self.q_norm = RMSNorm(head_dim, eps=1e-6)
            self.k_norm = RMSNorm(head_dim, eps=1e-6)
        else:
            self.q_norm = self.k_norm = None

    def forward(self, x, mask, cos, sin, use_cache=False, start_pos=0, cache=None):
        b, num_tokens, _ = x.shape

        # Apply projections
        queries = self.W_query(x)  # (b, num_tokens, num_heads * head_dim)
        keys = self.W_key(x)       # (b, num_tokens, num_kv_groups * head_dim)
        values = self.W_value(x)   # (b, num_tokens, num_kv_groups * head_dim)

        # Reshape
        queries = queries.view(b, num_tokens, self.num_heads, self.head_dim).transpose(1, 2)
        keys_new = keys.view(b, num_tokens, self.num_kv_groups, self.head_dim).transpose(1, 2)
        values_new = values.view(b, num_tokens, self.num_kv_groups, self.head_dim).transpose(1, 2)

        # Optional normalization
        if self.q_norm:
            queries = self.q_norm(queries)
        if self.k_norm:
            keys_new = self.k_norm(keys_new)

        # Apply RoPE
        queries = apply_rope(queries, cos, sin, offset=start_pos)
        keys_new = apply_rope(keys_new, cos, sin, offset=start_pos)

        if use_cache:
            if cache is None:
                keys = keys_new
                values = values_new
            else:
                prev_k, prev_v = cache
                keys = torch.cat([prev_k, keys_new], dim=2)
                values = torch.cat([prev_v, values_new], dim=2)
            next_cache = (keys, values)
        else:
            keys, values = keys_new, values_new
            next_cache = None

        # Expand K and V to match number of heads
        keys = keys.repeat_interleave(self.group_size, dim=1)
        values = values.repeat_interleave(self.group_size, dim=1)

        # Attention
        attn_scores = queries @ keys.transpose(2, 3)
        attn_scores = attn_scores.masked_fill(mask, -torch.inf)
        attn_weights = torch.softmax(attn_scores / self.head_dim**0.5, dim=-1)

        context = (attn_weights @ values).transpose(1, 2).reshape(b, num_tokens, self.d_out)
        return self.out_proj(context), next_cache


def compute_rope_params(head_dim, theta_base=10_000, context_length=4096, dtype=torch.float32):
    assert head_dim % 2 == 0, "Embedding dimension must be even"

    # Compute the inverse frequencies
    inv_freq = 1.0 / (theta_base ** (torch.arange(0, head_dim, 2, dtype=dtype)[: (head_dim // 2)].float() / head_dim))

    # Generate position indices
    positions = torch.arange(context_length, dtype=dtype)

    # Compute the angles
    angles = positions[:, None] * inv_freq[None, :]  # Shape: (context_length, head_dim // 2)

    # Expand angles to match the head_dim
    angles = torch.cat([angles, angles], dim=1)  # Shape: (context_length, head_dim)

    # Precompute sine and cosine
    cos = torch.cos(angles)
    sin = torch.sin(angles)

    return cos, sin


def apply_rope(x, cos, sin, offset=0):
    # x: (batch_size, num_heads, seq_len, head_dim)
    batch_size, num_heads, seq_len, head_dim = x.shape
    assert head_dim % 2 == 0, "Head dimension must be even"

    # Split x into first half and second half
    x1 = x[..., : head_dim // 2]  # First half
    x2 = x[..., head_dim // 2:]  # Second half

    # Adjust sin and cos shapes
    cos = cos[offset:offset + seq_len, :].unsqueeze(0).unsqueeze(0)  # Shape: (1, 1, seq_len, head_dim)
    sin = sin[offset:offset + seq_len, :].unsqueeze(0).unsqueeze(0)

    # Apply the rotary transformation
    rotated = torch.cat((-x2, x1), dim=-1)
    x_rotated = (x * cos) + (rotated * sin)

    # It's ok to use lower-precision after applying cos and sin rotation
    return x_rotated.to(dtype=x.dtype)


class RMSNorm(nn.Module):
    def __init__(self, emb_dim, eps=1e-6, bias=False, qwen3_compatible=True):
        super().__init__()
        self.eps = eps
        self.qwen3_compatible = qwen3_compatible
        self.scale = nn.Parameter(torch.ones(emb_dim))
        self.shift = nn.Parameter(torch.zeros(emb_dim)) if bias else None

    def forward(self, x):
        input_dtype = x.dtype

        if self.qwen3_compatible:
            x = x.to(torch.float32)

        variance = x.pow(2).mean(dim=-1, keepdim=True)
        norm_x = x * torch.rsqrt(variance + self.eps)
        norm_x = norm_x * self.scale

        if self.shift is not None:
            norm_x = norm_x + self.shift

        return norm_x.to(input_dtype)