Flash attention v2 forward #10484
Merged
Flash attention v2 forward #10484
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
…nto flash_attn_v2 for updates.
cccddd77
requested review from
jackalcooper,
hjchen2 and
liujuncheng
as code owners
MARD1NO
reviewed
Apr 12, 2024
| const auto batch_size = query->shape()->At(0); | ||
| const auto seqlen_q = query->shape()->At(2); | ||
| const auto num_heads = query->shape()->At(1); | ||
| // const auto max_seqlen_batch_q = query->shape()->At(2); |
| const int64_t& seed = 0) const { | ||
| const auto og_size = query->shape()->At(3); | ||
| const auto batch_size = query->shape()->At(0); | ||
| const auto seqlen_q = query->shape()->At(2); |
There was a problem hiding this comment.
这里应该也要检查下 q k v三者shape在bsz，headsize，numhead一不一致
| const auto num_heads = query->shape()->At(1); | ||
| // const auto max_seqlen_batch_q = query->shape()->At(2); | ||
| const auto max_seqlen_batch_k = key->shape()->At(2); | ||
| const auto max_seqlen_batch_v = value->shape()->At(2); |
There was a problem hiding this comment.
k和v的seqlen是不是一定相等，如果是应该也要check
| // Query (Batch x Num_heads x Q_seq_len x Dim_per_head) | ||
| // Key (Batch x Num_heads x KV_seq_len x Dim_per_head) | ||
| // Value (Batch x Num_heads x KV_seq_len x Dim_per_head) | ||
| auto q_padded = JUST(pad_last_dim<8>(query)); |
There was a problem hiding this comment.
如果last dim size满足被8整除，这个是不是可以去掉，应该需要if else
| JUST(OpInterpUtil::Dispatch<one::Tensor>(*op_, {q_, k_, v_}, ctx)); | ||
|
|
||
| auto output_padded = JUST(functional::Transpose(output_, {0, 2, 1, 3})); | ||
| return JUST(functional::Slice(output_padded, {0, 0, 0, 0}, |
|
|
||
| } // namespace impl | ||
|
|
||
| ONEFLOW_FUNCTION_LIBRARY(m) { |
There was a problem hiding this comment.
整体我觉得应该可以放到nn_functor.cpp里
Closed
MARD1NO
reviewed
Apr 12, 2024
|
|
||
| } // namespace oneflow | ||
|
|
||
| #endif // ONEFLOW_USER_KERNELS_FLASH_ATTENTION_KERNEL_H_ |
| return Maybe<void>::Ok(); | ||
| } | ||
|
|
||
| } // namespace oneflow |
| const Shape& q_shape = ctx->InputShape("query", 0); | ||
| const Shape& k_shape = ctx->InputShape("key", 0); | ||
| const Shape& v_shape = ctx->InputShape("value", 0); | ||
|
|
There was a problem hiding this comment.
如果有输入alibi，应该要检查下是否等于 numhead / batch * numhead shape
| #endif | ||
| } | ||
|
|
||
| // void set_params_dgrad(Flash_bwd_params ¶ms, |
There was a problem hiding this comment.
这部分删除，后续单独拉出来一个 scaled_dot_product_attention_kernel_grad.cu ，后向的东西都放这
|
|
||
| void set_params_alibi(Flash_fwd_params& params, const Tensor* alibi_slopes_, int batch_size, | ||
| int num_heads) { | ||
| params.alibi_slopes_ptr = nullptr; |
There was a problem hiding this comment.
一个小tips是如果为了临时Debug/跑通绕过的代码，最好加个 check, Log
TODO(Author): Need Support ALibi params
if (alibi_slopes != nullptr) {
// throw / log...
}
如果后面有人以为这个函数已经支持，传入参数发现结果不对，这样提示更加明确。后续想要增加支持的时候，直接全局搜索 TODO(Author): 也很快能找到需要修改的代码处
|
|
||
| const int arch = cuda_stream->cuda_arch() / 10; | ||
| const bool is_supported_arch = (arch == 80 || arch == 86 || arch == 89 || arch == 90); | ||
| CHECK(is_supported_arch); |
There was a problem hiding this comment.
应该加点log信息，表明只支持80 86 89 90
| } | ||
|
|
||
| Maybe<void> ScaledDotProductFlashAttentionOp::GetSbp(user_op::SbpContext* ctx) { | ||
| ctx->NewBuilder() |
There was a problem hiding this comment.
- SBP应该要考虑Optional输入alibi
- 考虑模型并行情况下，应该是对numhead维度进行split，需要添加这个SBP
|
Oneflow编译应该会注入一个宏 CUDA_VERSION，可以根据这个来控制只有11.6以上才编译： 例如：#if CUDA_VERSION >= 11000 |
|
Code got formatted by CI. Please request CI again if you still want to have this PR merged. If the PR is from a forked repo, please download the patch files from the GitHub Actions web page and apply them locally. |
cccddd77
requested review from
oneflow-ci-bot
and removed request for
oneflow-ci-bot
|
View latest API docs preview at: https://oneflow-staging.oss-cn-beijing.aliyuncs.com/docs/Oneflow-Inc/oneflow/pr/10484/ |
Speed stats: |
cccddd77
requested review from
oneflow-ci-bot
and removed request for
oneflow-ci-bot
cccddd77
force-pushed
the
flash_attention_v2_forward
branch
from
3e028a3
to
ffe3f74
Compare
|
Code got formatted by CI. Please request CI again if you still want to have this PR merged. If the PR is from a forked repo, please download the patch files from the GitHub Actions web page and apply them locally. |
cccddd77
requested review from
oneflow-ci-bot
and removed request for
oneflow-ci-bot
|
View latest API docs preview at: https://oneflow-staging.oss-cn-beijing.aliyuncs.com/docs/Oneflow-Inc/oneflow/pr/10484/ |
Speed stats: |
MARD1NO
approved these changes
Apr 18, 2024
mosout
reviewed
Apr 22, 2024
| @@ -0,0 +1,39 @@ | |||
| include(ExternalProject) | |||
There was a problem hiding this comment.
这个第三方库 之后最好把它挪到external目录下去 原则上新添加的第三方依赖都放external下
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.
集成了flash attn v2 forward算子