ayo/ipex-llm
1
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions 1

Combine apply_rotary_pos_emb for gpt-neox (#9074)

Browse source
This commit is contained in:
Kai Huang 2023-10-07 16:27:46 +08:00 committed by GitHub
parent 0b40ef8261
commit 78ea7ddb1c
2 changed files with 2 additions and 10 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

2
python/llm/dev/benchmark/all-in-one/README.md
View file

@ -5,7 +5,7 @@ Before running, make sure to have [bigdl-llm](../../../README.md) and [bigdl-nan
## Dependencies ## Dependencies
```bash ```bash
pip install omageconfig pip install omegaconf
pip install pandas pip install pandas
``` ```

10
python/llm/src/bigdl/llm/transformers/models/utils.py
View file

@ -71,7 +71,7 @@ def rotate_every_two(x):
def apply_rotary_pos_emb(q, k, cos, sin, position_ids, model_family): def apply_rotary_pos_emb(q, k, cos, sin, position_ids, model_family):
if model_family in ["llama", "baichuan", "internlm", "aquila"]: if model_family in ["llama", "baichuan", "internlm", "aquila", "gpt_neox"]:
# The first two dimensions of cos and sin are always 1, so we can `squeeze` them. # The first two dimensions of cos and sin are always 1, so we can `squeeze` them.
cos = cos.squeeze(1).squeeze(0) # [seq_len, dim] cos = cos.squeeze(1).squeeze(0) # [seq_len, dim]
sin = sin.squeeze(1).squeeze(0) # [seq_len, dim] sin = sin.squeeze(1).squeeze(0) # [seq_len, dim]
@ -86,14 +86,6 @@ def apply_rotary_pos_emb(q, k, cos, sin, position_ids, model_family):
q_embed = (q * cos) + (rotate_every_two(q) * sin) q_embed = (q * cos) + (rotate_every_two(q) * sin)
k_embed = (k * cos) + (rotate_every_two(k) * sin) k_embed = (k * cos) + (rotate_every_two(k) * sin)
return q_embed, k_embed return q_embed, k_embed
elif model_family == "gpt_neox":
gather_indices = position_ids[:, None, :, None] # [bs, 1, seq_len, 1]
gather_indices = gather_indices.repeat(1, cos.shape[1], 1, cos.shape[3])
cos = torch.gather(cos.repeat(gather_indices.shape[0], 1, 1, 1), 2, gather_indices)
sin = torch.gather(sin.repeat(gather_indices.shape[0], 1, 1, 1), 2, gather_indices)
q_embed = (q * cos) + (rotate_half(q) * sin)
k_embed = (k * cos) + (rotate_half(k) * sin)
return q_embed, k_embed
else: else:
invalidInputError(False, invalidInputError(False,
f"{model_family} is not supported.") f"{model_family} is not supported.")