Rain-Bus
ffd2defdfc
Annotated 16 source files covering the full architecture: engine (scheduler, block manager, model runner), layers (attention, linear, sampler, etc.), model (qwen3), and utils. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
88 lines
3.4 KiB
Python
88 lines
3.4 KiB
Python
import torch
|
|
from torch import nn
|
|
import torch.nn.functional as F
|
|
import torch.distributed as dist
|
|
|
|
from nanovllm.utils.context import get_context
|
|
|
|
|
|
class VocabParallelEmbedding(nn.Module):
|
|
"""词表并行 Embedding：将词表按 TP rank 切分。
|
|
|
|
每个 rank 只存储词表中属于自己的分片。前向计算时：
|
|
1. 只查找属于自己的 token ID，其他位置输出零。
|
|
2. 通过 all-reduce 聚合所有 rank 的结果。
|
|
"""
|
|
|
|
def __init__(
|
|
self,
|
|
num_embeddings: int,
|
|
embedding_dim: int,
|
|
):
|
|
super().__init__()
|
|
self.tp_rank = dist.get_rank()
|
|
self.tp_size = dist.get_world_size()
|
|
assert num_embeddings % self.tp_size == 0
|
|
self.num_embeddings = num_embeddings
|
|
self.num_embeddings_per_partition = self.num_embeddings // self.tp_size
|
|
self.vocab_start_idx = self.num_embeddings_per_partition * self.tp_rank
|
|
self.vocab_end_idx = self.vocab_start_idx + self.num_embeddings_per_partition
|
|
self.weight = nn.Parameter(torch.empty(self.num_embeddings_per_partition, embedding_dim))
|
|
self.weight.weight_loader = self.weight_loader
|
|
|
|
def weight_loader(self, param: nn.Parameter, loaded_weight: torch.Tensor):
|
|
"""加载属于当前 rank 的词表分片。"""
|
|
param_data = param.data
|
|
shard_size = param_data.size(0)
|
|
start_idx = self.tp_rank * shard_size
|
|
loaded_weight = loaded_weight.narrow(0, start_idx, shard_size)
|
|
param_data.copy_(loaded_weight)
|
|
|
|
def forward(self, x: torch.Tensor):
|
|
if self.tp_size > 1:
|
|
# 构造 mask：标记哪些 token ID 属于当前 rank 的范围
|
|
mask = (x >= self.vocab_start_idx) & (x < self.vocab_end_idx)
|
|
# 将全局 token ID 转为局部索引
|
|
x = mask * (x - self.vocab_start_idx)
|
|
y = F.embedding(x, self.weight)
|
|
if self.tp_size > 1:
|
|
# 非当前 rank 范围的 token 输出清零，然后 all-reduce 求和
|
|
y = mask.unsqueeze(1) * y
|
|
dist.all_reduce(y)
|
|
return y
|
|
|
|
|
|
class ParallelLMHead(VocabParallelEmbedding):
|
|
"""并行 LM Head：将隐藏状态映射为词表 logits。
|
|
|
|
与 Embedding 共享权重（如果模型配置了 tie_word_embeddings），
|
|
但前向逻辑不同：使用 F.linear（矩阵乘法）而非 F.embedding（查表）。
|
|
|
|
Prefill 阶段只需要每个序列最后一个 token 的 logits（因为只有最后一个 token 会用于采样），
|
|
所以先用 cu_seqlens 提取最后位置，再做矩阵乘法，减少计算量。
|
|
"""
|
|
|
|
def __init__(
|
|
self,
|
|
num_embeddings: int,
|
|
embedding_dim: int,
|
|
bias: bool = False,
|
|
):
|
|
assert not bias
|
|
super().__init__(num_embeddings, embedding_dim)
|
|
|
|
def forward(self, x: torch.Tensor):
|
|
context = get_context()
|
|
if context.is_prefill:
|
|
# Prefill: 只取每个序列最后一个 token 的隐藏状态用于采样
|
|
last_indices = context.cu_seqlens_q[1:] - 1
|
|
x = x[last_indices].contiguous()
|
|
# 矩阵乘法得到 logits
|
|
logits = F.linear(x, self.weight)
|
|
if self.tp_size > 1:
|
|
# Gather 所有 rank 的 logits 分片到 rank 0
|
|
all_logits = [torch.empty_like(logits) for _ in range(self.tp_size)] if self.tp_rank == 0 else None
|
|
dist.gather(logits, all_logits, 0)
|
|
logits = torch.cat(all_logits, -1) if self.tp_rank == 0 else None
|
|
return logits
|