[LLM INFER] Append attn #9244
Merged
[LLM INFER] Append attn #9244
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Change-Id: Ibe8920ba41ea9775e676b05b12dc01cb9da95b5e
|
Thanks for your contribution! |
Codecov ReportAttention: Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## develop #9244 +/- ##
========================================
Coverage 52.73% 52.74%
========================================
Files 661 661
Lines 107422 107371 -51
========================================
- Hits 56653 56630 -23
+ Misses 50769 50741 -28 ☔ View full report in Codecov by Sentry. |
yuanlehome
force-pushed
the
append_attn
branch
from
4d69d01 to
4a4a4b4
Compare
| static_cast<uint8_t>(quant_value2 + 128.0f); | ||
| } | ||
| // write k | ||
| // 大分块 lane_id / 4 / 2 |
yuanlehome
force-pushed
the
append_attn
branch
from
50a48de to
3789175
Compare
yuanlehome
force-pushed
the
append_attn
branch
from
ed9da7a to
84a6864
Compare
qingqing01
approved these changes
Oct 22, 2024
ZHUI
approved these changes
Oct 23, 2024
| @@ -871,7 +873,7 @@ def set_state_dict(self, state_dict): | |||
| weight_scales_loader = EmptyWeightScale( | |||
| weight_scale_map_dict, | |||
| num_of_layers=self.config.num_hidden_layers, | |||
| num_head=self.num_attention_heads, | |||
There was a problem hiding this comment.
这里是调试时改的，仅仅改下命名，不会有什么影响
| @@ -835,7 +838,10 @@ def set_state_dict(self, state_dict): | |||
|
|
|||
| for k, v in cache_scales_loader.scale.items(): | |||
| for i_layer, weight_scale in enumerate(v): | |||
| weight_scale = weight_scale.astype("float32") | |||
| if self.config.append_attn: | |||
| weight_scale = paddle.to_tensor(weight_scale).cast(paddle.get_default_dtype()) | |||
There was a problem hiding this comment.
为什么 append_attn 下 可以不同 fp32？
There was a problem hiding this comment.
因为kernel实现里是要求half精度的，访存 不同
| @@ -1684,8 +1692,8 @@ def benchmark(predictor, predictor_args, model_args): | |||
| batch_benchmark_texts = batchfy_text(benchmark_texts, predictor_args.batch_size) | |||
| print("***********Start Benchmark**********") | |||
|
|
|||
| warmup_time = 10 | |||
| test_time = 100 | |||
There was a problem hiding this comment.
这个修改没啥影响，没注意到提到commit上了
| super().__init__(config) | ||
| self.max_seq_len = config.max_seq_len | ||
| self.block_size = config.block_size | ||
|
|
||
| def set_transformer_block(self, transformer_config): | ||
| if self.use_weight_only: | ||
| self.transformer_block = FusedBlockMultiTransformerWeightOnly(transformer_config) | ||
| elif "a8w8" in self.quant_type: |
There was a problem hiding this comment.
因为并没有支持，之前同学参考其他代码时一并copy过来了，我这里顺便给删掉
lvdongyi
pushed a commit
to lvdongyi/PaddleNLP
that referenced
this pull request
Oct 23, 2024
* refine paddle::empty(), fix memory error, support multi_stream for attention * fix and rename attention as append_attention * rename file --------- Co-authored-by: lizhenyun <lizhenyun@baidu.com> Co-authored-by: lizhenyun01 <1500424927@qq.com>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
PR types
New features
PR changes
Others
Description
大模型推理attention组网重构，新的append_attn方案相比旧方案有10%到90%的性能提升。
目前已支持了llama/qwen/qwen-moe/mixtral的推理。
使用方式，原推理脚本的 --block_attn选项改为--append_attn即可。
TODO：