Rope imbedding kernel to use avx2 #23694
Open
+240
−3
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
Signed-off-by: Liqun Fu <liqun.fu@microsoft.com>
Signed-off-by: liqunfu <liqun.fu@microsoft.com>
![github-actions[bot]](https://avatars.githubusercontent.com/in/15368?s=60&v=4){kind=small}
Comment on lines
235
to
238
| auto ret = ApplyAttention(q_rotary, packed_qkv ? nullptr : k_rotary, packed_qkv ? nullptr : V.Get<Tensor>().Data<T>(), | ||
| past_key, past_value, output, present_k, present_v, | ||
| seqlens_k, parameters, allocator, context); | ||
| if (profiler_->IsEnabled()) { |
There was a problem hiding this comment.
Suggested change
| auto ret = ApplyAttention(q_rotary, packed_qkv ? nullptr : k_rotary, packed_qkv ? nullptr : V.Get<Tensor>().Data<T>(), | |
| past_key, past_value, output, present_k, present_v, | |
| seqlens_k, parameters, allocator, context); | |
| if (profiler_->IsEnabled()) { | |
| auto ret = ApplyAttention(q_rotary, packed_qkv ? nullptr : k_rotary, packed_qkv ? nullptr : V.Get<Tensor>().Data<T>(), | |
| past_key, past_value, output, present_k, present_v, | |
| seqlens_k, parameters, allocator, context); | |
| if (profiler_->IsEnabled()) { |
Signed-off-by: liqunfu <liqun.fu@microsoft.com>
Comment on lines
51
to
53
| Status GroupQueryAttention<T>::Compute(OpKernelContext* context) const { | ||
|
|
||
| const std::string node_name = this->Node().Name(); |
There was a problem hiding this comment.
Suggested change
| Status GroupQueryAttention<T>::Compute(OpKernelContext* context) const { | |
| const std::string node_name = this->Node().Name(); | |
| Status GroupQueryAttention<T>::Compute(OpKernelContext* context) const { | |
| const std::string node_name = this->Node().Name(); |
Comment on lines
189
to
191
| domain="com.microsoft", | ||
|
|
||
| ), |
There was a problem hiding this comment.
Suggested change
| domain="com.microsoft", | |
| ), | |
| domain="com.microsoft", | |
| ), |
Comment on lines
1943
to
1950
| node_name = ( | ||
| ("packed_" if packed else "") + | ||
| ("rotary_" if rotary else "") + | ||
| ("rotary_interleaved_" if rotary_interleaved else "") + | ||
| "softcap_" + str(softcap) + "_" + | ||
| "smooth_softmax_" + str(use_smooth_softmax) + "_" + | ||
| "b_" + str(b) + "_sq_" + str(sq) + "_skv_" + str(skv) + "_n_" + str(n) + "_n2_" + str(n2) + "_h_" + str(h) | ||
| ) |
There was a problem hiding this comment.
Suggested change
| node_name = ( | |
| ("packed_" if packed else "") + | |
| ("rotary_" if rotary else "") + | |
| ("rotary_interleaved_" if rotary_interleaved else "") + | |
| "softcap_" + str(softcap) + "_" + | |
| "smooth_softmax_" + str(use_smooth_softmax) + "_" + | |
| "b_" + str(b) + "_sq_" + str(sq) + "_skv_" + str(skv) + "_n_" + str(n) + "_n2_" + str(n2) + "_h_" + str(h) | |
| ) | |
| node_name = ( | |
| ("packed_" if packed else "") | |
| + ("rotary_" if rotary else "") | |
| + ("rotary_interleaved_" if rotary_interleaved else "") | |
| + "softcap_" | |
| + str(softcap) | |
| + "_" | |
| + "smooth_softmax_" | |
| + str(use_smooth_softmax) | |
| + "_" | |
| + "b_" | |
| + str(b) | |
| + "_sq_" | |
| + str(sq) | |
| + "_skv_" | |
| + str(skv) | |
| + "_n_" | |
| + str(n) | |
| + "_n2_" | |
| + str(n2) | |
| + "_h_" | |
| + str(h) | |
| ) |
snnn
reviewed
Feb 14, 2025
| const std::string node_name = this->Node().Name(); | ||
|
|
||
| // Initialize the profiler_ with a unique log file based on the node name | ||
| profiler_ = new onnxruntime::profiling::Profiler(); |
There was a problem hiding this comment.
it is used to profile GQA. I have removed it because we need the code to be in for the release.
Signed-off-by: liqunfu <liqun.fu@microsoft.com>
Comment on lines
196
to
199
| auto ret = ApplyAttention(q_rotary, packed_qkv ? nullptr : k_rotary, packed_qkv ? nullptr : V.Get<Tensor>().Data<T>(), | ||
| past_key, past_value, output, present_k, present_v, | ||
| seqlens_k, parameters, allocator, context); | ||
| return ret; |
There was a problem hiding this comment.
Suggested change
| auto ret = ApplyAttention(q_rotary, packed_qkv ? nullptr : k_rotary, packed_qkv ? nullptr : V.Get<Tensor>().Data<T>(), | |
| past_key, past_value, output, present_k, present_v, | |
| seqlens_k, parameters, allocator, context); | |
| return ret; | |
| auto ret = ApplyAttention(q_rotary, packed_qkv ? nullptr : k_rotary, packed_qkv ? nullptr : V.Get<Tensor>().Data<T>(), | |
| past_key, past_value, output, present_k, present_v, | |
| seqlens_k, parameters, allocator, context); | |
| return ret; |
Comment on lines
92
to
94
| sequence_length, seqlen_past_kv_cache, seqlen_present_kv_cache, head_size, past_key_data, | ||
| present_key_data, past_present_share_buffer, packed_qkv, is_prompt, tp, allocator); | ||
|
|
||
| present_key_data, past_present_share_buffer, packed_qkv, is_prompt, tp, allocator); | ||
| // Compute the attentionScore * Value: out(B, N, S, H_v) = attention_probs(B, N, S, T) x V(B, N, T, H_v) |
There was a problem hiding this comment.
Suggested change
| sequence_length, seqlen_past_kv_cache, seqlen_present_kv_cache, head_size, past_key_data, | |
| present_key_data, past_present_share_buffer, packed_qkv, is_prompt, tp, allocator); | |
| present_key_data, past_present_share_buffer, packed_qkv, is_prompt, tp, allocator); | |
| // Compute the attentionScore * Value: out(B, N, S, H_v) = attention_probs(B, N, S, T) x V(B, N, T, H_v) | |
| sequence_length, seqlen_past_kv_cache, seqlen_present_kv_cache, head_size, past_key_data, | |
| present_key_data, past_present_share_buffer, packed_qkv, is_prompt, tp, allocator); | |
| // Compute the attentionScore * Value: out(B, N, S, H_v) = attention_probs(B, N, S, T) x V(B, N, T, H_v) |
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.
Description
Credit to chethanpk who provided with Rope Embedding in a patch. The patch is in the first commit of this PR.
Motivation and Context
improve GQP performance on intel x64.