Add argsort. (#2132)

* Add the argsort cuda kernels.

* CPU version of arg-sort.

* Hook the cuda kernel + rework the cpu bits.

* Add some dedicated test.

* Working cuda kernel.

* Metal kernel.

* Metal adjustments.

* Bugfix.

* Use the fast rope in qwen.

* Rework the expert selection in qwen.
This commit is contained in:
Laurent Mazare
2024-04-27 20:17:35 +02:00
committed by GitHub
parent 6cf82fd7a3
commit 96a48e5cc4
11 changed files with 489 additions and 44 deletions

View File

@ -27,13 +27,6 @@ struct RotaryEmbedding {
cos: Tensor,
}
fn rotate_half(xs: &Tensor) -> Result<Tensor> {
let last_dim = xs.dim(D::Minus1)?;
let xs1 = xs.narrow(D::Minus1, 0, last_dim / 2)?;
let xs2 = xs.narrow(D::Minus1, last_dim / 2, last_dim - last_dim / 2)?;
Tensor::cat(&[&xs2.neg()?, &xs1], D::Minus1)
}
impl RotaryEmbedding {
fn new(dtype: DType, cfg: &Config, dev: &Device) -> Result<Self> {
let dim = cfg.hidden_size / cfg.num_attention_heads;
@ -48,7 +41,6 @@ impl RotaryEmbedding {
.to_dtype(dtype)?
.reshape((max_seq_len, 1))?;
let freqs = t.matmul(&inv_freq)?;
let freqs = Tensor::cat(&[&freqs, &freqs], D::Minus1)?;
Ok(Self {
sin: freqs.sin()?,
cos: freqs.cos()?,
@ -64,10 +56,8 @@ impl RotaryEmbedding {
let (_b_sz, _h, seq_len, _n_embd) = q.dims4()?;
let cos = self.cos.narrow(0, seqlen_offset, seq_len)?;
let sin = self.sin.narrow(0, seqlen_offset, seq_len)?;
let cos = cos.unsqueeze(0)?.unsqueeze(0)?; // (1, 1, seq_len, dim)
let sin = sin.unsqueeze(0)?.unsqueeze(0)?; // (1, 1, seq_len, dim)
let q_embed = (q.broadcast_mul(&cos)? + rotate_half(q)?.broadcast_mul(&sin))?;
let k_embed = (k.broadcast_mul(&cos)? + rotate_half(k)?.broadcast_mul(&sin))?;
let q_embed = candle_nn::rotary_emb::rope(&q.contiguous()?, &cos, &sin)?;
let k_embed = candle_nn::rotary_emb::rope(&k.contiguous()?, &cos, &sin)?;
Ok((q_embed, k_embed))
}
}