-
Notifications
You must be signed in to change notification settings - Fork 2.9k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
[GCU] Support llama for GCU #8445
Conversation
Thanks for your contribution! |
Codecov ReportAttention: Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## develop #8445 +/- ##
===========================================
- Coverage 55.42% 54.29% -1.14%
===========================================
Files 617 617
Lines 96286 96340 +54
===========================================
- Hits 53367 52303 -1064
- Misses 42919 44037 +1118 ☔ View full report in Codecov by Sentry. |
@@ -1528,7 +1535,7 @@ def forward( | |||
attention_mask, (batch_size, seq_length), cache_length, inputs_embeds.dtype | |||
) # [bs, 1, seq_len, seq_len] | |||
is_casual = False | |||
if self.config.use_flash_attention: | |||
if self.config.use_flash_attention and get_env_device() != "gcu": |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
这里在attention mask的处理上,GCU不一样的地方是什么?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
基于 use_flash_attention kernel
的实现,is_casual
情况下也是需要当前与输入相同dtype
的attention_mask
,而不是None
或者bool
类型的mask
。
@@ -297,6 +303,7 @@ def do_generation(): | |||
parser = get_eval_parser() | |||
args = parser.parse_args() | |||
paddle.set_default_dtype(args.dtype) | |||
paddle.set_device(args.device) |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
在训练初始的位置设置set_device,这里再重新设置的原因是什么?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
单独使用的eval.py
测试,没有训练初始位置? 当然不设置默认应该也是当前的device
,可以去除
@@ -934,7 +941,7 @@ def forward( | |||
sin.cast(value_states.dtype) if sin.dtype != value_states.dtype else sin, | |||
) | |||
else: | |||
cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len) | |||
cos, sin, _ = self.rotary_emb(value_states, seq_len=kv_seq_len) |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
一定要加 cos_sin 的优化吗?代码改动很大,而且会导致其他设备性能下降,凭空多了很多开销。
或者你们需要的时候再自己去造一个 cos_sin
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
这里主要是因为算子的实现与paper
或者vllm
一致,使用了与这里不同的sin/cos
。关于其他设备性能开销,一方面,应该大多table
的计算只在初始化阶段,另一方面,我们将按照第一个issue
的建议,在特定设备进行计算,这里仅仅只会多返回一个None
。
1ce5c16
to
0afe3fa
Compare
0afe3fa
to
32d66ef
Compare
https://xly.bce.baidu.com/paddlepaddle/Paddle-NLP/newipipe/detail/10720664/job/26276076 这个PR的 rope 接口改动,貌似导致自动并行代码挂了 |
PR types
New features
PR changes
Models
Description
Support llama for GCU