| """Low-overhead embedding lookup.""" |
|
|
| import torch |
| import torch.nn as nn |
| import triton |
| import triton.language as tl |
|
|
|
|
| @triton.jit |
| def _embedding_chunk_kernel(weight, input_ids, output, |
| embedding_dim: tl.constexpr, |
| block_dim: tl.constexpr, |
| chunks: tl.constexpr): |
| program = tl.program_id(0) |
| row = program // chunks |
| col = (program - row * chunks) * block_dim + tl.arange(0, block_dim) |
| index = tl.load(input_ids + row) |
| mask = col < embedding_dim |
| values = tl.load(weight + index * embedding_dim + col, mask=mask) |
| tl.store(output + row * embedding_dim + col, values, mask=mask) |
|
|
|
|
| @triton.jit |
| def _embedding_64_kernel(weight, input_ids, output, num_rows: tl.constexpr): |
| rows = tl.program_id(0) * 64 + tl.arange(0, 64) |
| cols = tl.arange(0, 64) |
| row_mask = rows < num_rows |
| indices = tl.load(input_ids + rows, mask=row_mask)[:, None] |
| values = tl.load(weight + indices * 64 + cols[None, :], |
| mask=row_mask[:, None]) |
| tl.store(output + rows[:, None] * 64 + cols[None, :], values, |
| mask=row_mask[:, None]) |
|
|
|
|
| class Embedding(nn.Module): |
| def __init__(self, num_embeddings: int, embedding_dim: int, |
| padding_idx: int | None = None): |
| super().__init__() |
| self.emb = nn.Embedding(num_embeddings, embedding_dim, |
| padding_idx=padding_idx) |
| self.embedding_dim = embedding_dim |
|
|
| def forward(self, input_ids): |
| output = torch.empty((*input_ids.shape, self.embedding_dim), |
| device=input_ids.device, dtype=self.emb.weight.dtype) |
| num_rows = input_ids.numel() |
| if self.embedding_dim == 64: |
| _embedding_64_kernel[(triton.cdiv(num_rows, 64),)]( |
| self.emb.weight, input_ids, output, num_rows, num_warps=4) |
| else: |
| block_dim = min(2048, triton.next_power_of_2(self.embedding_dim)) |
| chunks = triton.cdiv(self.embedding_dim, block_dim) |
| _embedding_chunk_kernel[(num_rows * chunks,)]( |
| self.emb.weight, input_ids, output, self.embedding_dim, |
| block_dim, chunks, num_warps=4) |
| return output |
|
|
