-
Notifications
You must be signed in to change notification settings - Fork 182
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
Updates to support new sdpa function #458
base: main
Are you sure you want to change the base?
Conversation
For my understanding, is it possible for SDPA to detect that the input key/value shapes are targeting GQA so that we do not need to pass In this case, the key/value shapes are not broadcastable, so there should not be any uncertainty for semantics. |
@awgu Had much discussion on this one with different folks around the org. TLDR yes its completely possible to recognize when users want to do this and enable it if the shapes work. This however doesnt fit naturally into the existing broadcasting semantics. So in theory it is possible for users to "make a mistake" pass in mishaped inputs and not get an error. The consensus was to add this extra check so that users give strong single of their intention. |
### Approach: Using the current function declaration **Constraint:** Q_Heads % KV_Heads == 0 **Major change:** - Added a new argument enable_gqa: bool to sdpa function call - It adds a meaning to the last third dimension. Sample use cases this would enable: LLama3 ``` # LLama3 8b call to SDPA query = torch.rand(batch, 32, seq_len_q, D) key = torch.rand(batch, 8, seq_len_kv, D) value = torch.rand(batch, 8, seq_len_kv, D) output = scaled_dot_product_attention(query, key, value, is_causal=True, enable_gqa=True) # Output Shape (batch, 32, seq_len_q, D) ``` ### Design Choice: - Check if Query.size(-3) == Key.size(-3) == Value.size(-3) or, Query.size(-3) % Key.size(-3) == 0 - The function adjusts the key and value tensors to match the query tensor's head dimension by using repeat_interleave if their number of heads are not equal, facilitating correct and efficient computation in attention mechanisms. - By default the enable_gqa flag is set to False, which ensures that regular sdpa functionality remains unchanged. ### Benchmarks: - **sdpa.py: #130634** For different batch sizes enable_gqa=True shows a substansial improvement in the run_time of sdpa | batch_size | q_num_heads | kv_num_heads | q_seq_len | kv_seq_len | embed_dim | forward_time when enable_gqa=True | forward_time when enable_gqa=False | | ------------ | ------------- | -------------- | ----------- | ------------ | ----------- | ----------- | ---------------- | | 1 | 32 | 8 | 2048 | 2048 | 2048 | 100.71 | 119.70 | | 8 | 32 | 8 | 2048 | 2048 | 2048 | 539.78 | 628.83 | | 16 | 32 | 8 | 2048 | 2048 | 2048 | 1056.81 | 1225.48 | | 32 | 32 | 8 | 2048 | 2048 | 2048 | 2099.54 | 2440.45 | ![Screenshot 2024-07-25 at 9 07 40 PM](https://github.com/user-attachments/assets/a3e5f716-c39f-4096-9e6c-82a735e57b7b) - **TorchTitan: pytorch/torchtitan#458 Pull Request resolved: #128898 Approved by: https://github.com/drisspg
### Approach: Using the current function declaration **Constraint:** Q_Heads % KV_Heads == 0 **Major change:** - Added a new argument enable_gqa: bool to sdpa function call - It adds a meaning to the last third dimension. Sample use cases this would enable: LLama3 ``` # LLama3 8b call to SDPA query = torch.rand(batch, 32, seq_len_q, D) key = torch.rand(batch, 8, seq_len_kv, D) value = torch.rand(batch, 8, seq_len_kv, D) output = scaled_dot_product_attention(query, key, value, is_causal=True, enable_gqa=True) # Output Shape (batch, 32, seq_len_q, D) ``` ### Design Choice: - Check if Query.size(-3) == Key.size(-3) == Value.size(-3) or, Query.size(-3) % Key.size(-3) == 0 - The function adjusts the key and value tensors to match the query tensor's head dimension by using repeat_interleave if their number of heads are not equal, facilitating correct and efficient computation in attention mechanisms. - By default the enable_gqa flag is set to False, which ensures that regular sdpa functionality remains unchanged. ### Benchmarks: - **sdpa.py: #130634** For different batch sizes enable_gqa=True shows a substansial improvement in the run_time of sdpa | batch_size | q_num_heads | kv_num_heads | q_seq_len | kv_seq_len | embed_dim | forward_time when enable_gqa=True | forward_time when enable_gqa=False | | ------------ | ------------- | -------------- | ----------- | ------------ | ----------- | ----------- | ---------------- | | 1 | 32 | 8 | 2048 | 2048 | 2048 | 100.71 | 119.70 | | 8 | 32 | 8 | 2048 | 2048 | 2048 | 539.78 | 628.83 | | 16 | 32 | 8 | 2048 | 2048 | 2048 | 1056.81 | 1225.48 | | 32 | 32 | 8 | 2048 | 2048 | 2048 | 2099.54 | 2440.45 | ![Screenshot 2024-07-25 at 9 07 40 PM](https://github.com/user-attachments/assets/a3e5f716-c39f-4096-9e6c-82a735e57b7b) - **TorchTitan: pytorch/torchtitan#458 Pull Request resolved: #128898 Approved by: https://github.com/drisspg
### Approach: Using the current function declaration **Constraint:** Q_Heads % KV_Heads == 0 **Major change:** - Added a new argument enable_gqa: bool to sdpa function call - It adds a meaning to the last third dimension. Sample use cases this would enable: LLama3 ``` # LLama3 8b call to SDPA query = torch.rand(batch, 32, seq_len_q, D) key = torch.rand(batch, 8, seq_len_kv, D) value = torch.rand(batch, 8, seq_len_kv, D) output = scaled_dot_product_attention(query, key, value, is_causal=True, enable_gqa=True) # Output Shape (batch, 32, seq_len_q, D) ``` ### Design Choice: - Check if Query.size(-3) == Key.size(-3) == Value.size(-3) or, Query.size(-3) % Key.size(-3) == 0 - The function adjusts the key and value tensors to match the query tensor's head dimension by using repeat_interleave if their number of heads are not equal, facilitating correct and efficient computation in attention mechanisms. - By default the enable_gqa flag is set to False, which ensures that regular sdpa functionality remains unchanged. ### Benchmarks: - **sdpa.py: #130634** For different batch sizes enable_gqa=True shows a substansial improvement in the run_time of sdpa | batch_size | q_num_heads | kv_num_heads | q_seq_len | kv_seq_len | embed_dim | forward_time when enable_gqa=True | forward_time when enable_gqa=False | | ------------ | ------------- | -------------- | ----------- | ------------ | ----------- | ----------- | ---------------- | | 1 | 32 | 8 | 2048 | 2048 | 2048 | 100.71 | 119.70 | | 8 | 32 | 8 | 2048 | 2048 | 2048 | 539.78 | 628.83 | | 16 | 32 | 8 | 2048 | 2048 | 2048 | 1056.81 | 1225.48 | | 32 | 32 | 8 | 2048 | 2048 | 2048 | 2099.54 | 2440.45 | ![Screenshot 2024-07-25 at 9 07 40 PM](https://github.com/user-attachments/assets/a3e5f716-c39f-4096-9e6c-82a735e57b7b) - **TorchTitan: pytorch/torchtitan#458 Differential Revision: D60772086 Pull Request resolved: #132689 Approved by: https://github.com/drisspg
No description provided.