xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-09 22:05:44 +08:00
Code Issues Packages Projects Releases Wiki Activity

Use FP32 in RoPE initialization (#1004)

Browse Source 
Co-authored-by: One <imone@tuta.io>
This commit is contained in:
Woosuk Kwon 2023-09-11 00:26:35 -07:00 committed by GitHub
parent d6770d1f23
commit e67b4f2c2a
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
2 changed files with 7 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
tests/kernels/test_pos_encoding.py
View File

@ -133,9 +133,10 @@ def test_rotary_embedding(
device="cuda")
# Create the rotary embedding.
inv_freq = 1.0 / (base**(torch.arange(0, rotary_dim, 2) / rotary_dim))
inv_freq = 1.0 / (base**(
torch.arange(0, rotary_dim, 2, dtype=torch.float) / rotary_dim))
t = torch.arange(max_position).float()
freqs = torch.einsum("i,j -> ij", t, inv_freq.float())
freqs = torch.einsum("i,j -> ij", t, inv_freq)
cos = freqs.cos()
sin = freqs.sin()
cos_sin_cache = torch.cat((cos, sin), dim=-1)

8
vllm/model_executor/layers/attention.py
View File

@ -264,10 +264,10 @@ class PagedAttentionWithRoPE(PagedAttention):
self.is_neox_style = is_neox_style
# Create the cos and sin cache.
inv_freq = 1.0 / (base**(
torch.arange(0, rotary_dim, 2, device="cuda") / rotary_dim))
t = torch.arange(max_position, device="cuda").float()
freqs = torch.einsum("i,j -> ij", t, inv_freq.float())
inv_freq = 1.0 / (base**(torch.arange(
0, rotary_dim, 2, dtype=torch.float, device="cuda") / rotary_dim))
t = torch.arange(max_position, dtype=torch.float, device="cuda")
freqs = torch.einsum("i,j -> ij", t, inv_freq)
cos = freqs.cos()
sin = freqs.sin()
cache = torch.cat((cos, sin), dim=-1)