# 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"]) # Reusable 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 = None # Batched version tracks positions per sample def forward(self, in_idx, cache=None, start_pos=None): B, num_tokens = in_idx.size() tok_embeds = self.tok_emb(in_idx) x = tok_embeds device = x.device if cache is not None: pos_start = start_pos pos_end = pos_start + num_tokens max_len = pos_end.max().item() full_mask = torch.triu( torch.ones(max_len, max_len, device=device, dtype=torch.bool), diagonal=1 ) mask = torch.zeros(B, 1, num_tokens, max_len, device=device, dtype=torch.bool) for i in range(B): ps, pe = pos_start[i].item(), pos_end[i].item() mask[i, 0] = full_mask[ps:pe, :pe] else: pos_start = torch.zeros(B, dtype=torch.long, device=device) mask = torch.triu( torch.ones(num_tokens, num_tokens, device=device, dtype=torch.bool), diagonal=1 )[None, None, :, :] for i, block in enumerate(self.trf_blocks): blk_cache = [cache.get(i, b_idx) for b_idx in range(B)] if cache is not None else None x, new_blk_cache = block(x, mask, self.cos, self.sin, start_pos=pos_start, cache=blk_cache) if cache is not None: for b_idx in range(B): cache.update(i, b_idx, new_blk_cache[b_idx]) x = self.final_norm(x) logits = self.out_head(x.to(self.cfg["dtype"])) return logits def reset_kv_cache(self, batch_size, device=None): device = device or next(self.parameters()).device self.current_pos = torch.zeros(batch_size, dtype=torch.long, device=device) 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, 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, 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, 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 = keys.view(b, num_tokens, self.num_kv_groups, self.head_dim).transpose(1, 2) values = 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 = self.k_norm(keys) # Apply RoPE queries = apply_rope(queries, cos, sin, offset=start_pos) keys = apply_rope(keys, cos, sin, offset=start_pos) # KV caching next_cache = [] for i in range(b): prev = cache[i] if cache else None if prev is None: k_cat = keys[i:i+1] v_cat = values[i:i+1] else: prev_k, prev_v = prev k_cat = torch.cat([prev_k, keys[i:i+1]], dim=2) v_cat = torch.cat([prev_v, values[i:i+1]], dim=2) next_cache.append((k_cat, v_cat)) keys = torch.cat([k for k, _ in next_cache], dim=0) values = torch.cat([v for _, v in next_cache], dim=0) # 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) # PyTorch fails to do the implicit casting, so we have to be intentional with the types scale = torch.tensor(self.head_dim**0.5, dtype=queries.dtype, device=queries.device) attn_weights = torch.softmax(attn_scores / scale, dim=-1).to(values.dtype) 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): # x: (batch_size, num_heads, seq_len, head_dim) bsz, n_heads, seq_len, head_dim = x.shape assert head_dim % 2 == 0, "Head dimension must be even" assert offset.shape[0] == bsz, "Offset must have one value per batch item" # Prepare cos/sin: (seq_len, head_dim) cos = cos[:cos.shape[0], :].unsqueeze(0).unsqueeze(0) # (1, 1, total_seq_len, head_dim) sin = sin[:sin.shape[0], :].unsqueeze(0).unsqueeze(0) # Build position indices per batch item position_ids = torch.arange(seq_len, device=offset.device).unsqueeze(0) + offset.unsqueeze(1) # (bsz, seq_len) position_ids = position_ids.clamp(max=cos.shape[2] - 1) # Gather cos/sin for each position cos = cos[0, 0, position_ids, :] # (bsz, seq_len, head_dim) sin = sin[0, 0, position_ids, :] # Expand for multi-heads cos = cos.unsqueeze(1) # (bsz, 1, seq_len, head_dim) sin = sin.unsqueeze(1) x1 = x[..., :head_dim // 2] x2 = x[..., head_dim // 2:] rotated = torch.cat((-x2, x1), dim=-1) x_rotated = (x * cos) + (rotated * sin) return x_rotated 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)