ring_attn_fp8_kv_fusion #787
Conversation
Summary of ChangesHello @xiehaohx, I'm Gemini Code Assist1! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed! This pull request introduces significant optimizations to the Highlights
🧠 New Feature in Public Preview: You can now enable Memory to help Gemini Code Assist learn from your team's feedback. This makes future code reviews more consistent and personalized to your project's style. Click here to enable Memory in your admin console. Using Gemini Code AssistThe full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips. Invoking Gemini You can request assistance from Gemini at any point by creating a comment using either
Customization To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a Limitations & Feedback Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for Github and other Google products, sign up here. You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension. Footnotes
|
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Code Review
This pull request introduces support for FP8 communication and key-value tensor fusion within the Ring Attention mechanism. These are valuable optimizations that can enhance performance by reducing communication overhead. The implementation adds use_fp8_comm and use_kv_fusion flags to enable these features. My review has identified a critical bug related to the KV fusion logic in the final attention step, which would result in incorrect outputs. Additionally, there's a significant amount of code duplication that complicates maintenance. I've provided detailed comments and suggestions to address these issues.
|
|
||
| block_out, block_lse = self.ring_attn_sub(q, k, v) | ||
| if use_kv_fusion: | ||
| next_kv = torch.cat((kv, txt_kv), dim=1) |
There was a problem hiding this comment.
There is a critical bug in the logic for the final step of the ring attention when use_kv_fusion is enabled. The result of torch.cat((kv, txt_kv), dim=1) is assigned to next_kv, but the subsequent attention calculation at line 164 still uses the original kv. This means the text tokens (txt_kv) are not included in the attention computation for the final block, leading to incorrect results. The else branch for the non-fused case correctly updates k and v in-place. The same should be done for kv.
kv = torch.cat((kv, txt_kv), dim=1)| for step in range(world_size): | ||
| if step + 1 != world_size: | ||
| next_k = RING_COMM.send_recv(k) | ||
| next_v = RING_COMM.send_recv(v) | ||
| RING_COMM.commit() | ||
| if use_fp8_comm: | ||
| if use_kv_fusion: | ||
| kv_fp8, kv_scale = quant_fp8_vllm(kv.reshape(-1, hidden_dims)) | ||
| kv_fp8 = kv_fp8.reshape(original_shape) | ||
| kv_scale = kv_scale.reshape(original_shape[0], original_shape[1], original_shape[2], 1) | ||
| next_kv_fp8 = RING_COMM.send_recv(kv_fp8) | ||
| next_kv_scale = RING_COMM.send_recv(kv_scale) | ||
| else: | ||
| k_fp8, k_scale = quant_fp8_vllm(k.reshape(-1, hidden_dims)) | ||
| v_fp8, v_scale = quant_fp8_vllm(v.reshape(-1, hidden_dims)) | ||
| k_fp8 = k_fp8.reshape(original_shape) | ||
| v_fp8 = v_fp8.reshape(original_shape) | ||
| k_scale = k_scale.reshape(original_shape[0], original_shape[1], original_shape[2], 1) | ||
| v_scale = v_scale.reshape(original_shape[0], original_shape[1], original_shape[2], 1) | ||
| next_k_fp8 = RING_COMM.send_recv(k_fp8) | ||
| next_k_scale = RING_COMM.send_recv(k_scale) | ||
| next_v_fp8 = RING_COMM.send_recv(v_fp8) | ||
| next_v_scale = RING_COMM.send_recv(v_scale) | ||
| RING_COMM.commit() | ||
| else: | ||
| if use_kv_fusion: | ||
| next_kv = RING_COMM.send_recv(kv) | ||
| else: | ||
| next_k = RING_COMM.send_recv(k) | ||
| next_v = RING_COMM.send_recv(v) | ||
| RING_COMM.commit() | ||
|
|
||
| if step + 1 == world_size: | ||
| k = torch.cat((k, txt_k), dim=1) | ||
| v = torch.cat((v, txt_v), dim=1) | ||
|
|
||
| block_out, block_lse = self.ring_attn_sub(q, k, v) | ||
| if use_kv_fusion: | ||
| next_kv = torch.cat((kv, txt_kv), dim=1) | ||
| else: | ||
| k = torch.cat((k, txt_k), dim=1) | ||
| v = torch.cat((v, txt_v), dim=1) | ||
|
|
||
| if use_kv_fusion: | ||
| block_out, block_lse = self.ring_attn_sub_kv_fusion(q, kv) | ||
| else: | ||
| block_out, block_lse = self.ring_attn_sub(q, k, v) | ||
| out, lse = self.update_out_and_lse(out, lse, block_out, block_lse) | ||
|
|
||
| if step + 1 != world_size: | ||
| RING_COMM.wait() | ||
| k = next_k | ||
| v = next_v | ||
| if use_fp8_comm: | ||
| if use_kv_fusion: | ||
| kv = dequant_fp8_vllm(next_kv_fp8, next_kv_scale, original_dtype) | ||
| else: | ||
| k = dequant_fp8_vllm(next_k_fp8, next_k_scale, original_dtype) | ||
| v = dequant_fp8_vllm(next_v_fp8, next_v_scale, original_dtype) | ||
| else: | ||
| if use_kv_fusion: | ||
| kv = next_kv | ||
| else: | ||
| k = next_k | ||
| v = next_v |
There was a problem hiding this comment.
The logic within the for loop contains significant code duplication, particularly in the handling of use_fp8_comm with and without use_kv_fusion. The quantization, communication, and dequantization steps are nearly identical for k, v, and the fused kv tensor. This repetition makes the code harder to read, maintain, and debug.
Consider refactoring this logic into smaller helper functions. For instance, you could have a helper for quantization and communication:
def _quant_and_send(tensor, original_shape, hidden_dims, comm):
tensor_fp8, tensor_scale = quant_fp8_vllm(tensor.reshape(-1, hidden_dims))
tensor_fp8 = tensor_fp8.reshape(original_shape)
tensor_scale = tensor_scale.reshape(original_shape[0], original_shape[1], original_shape[2], 1)
next_tensor_fp8 = comm.send_recv(tensor_fp8)
next_tensor_scale = comm.send_recv(tensor_scale)
return next_tensor_fp8, next_tensor_scaleA similar helper could be created for dequantization. This would abstract away the repeated logic and simplify the main loop's structure.
| cur_rank = dist.get_rank(seq_p_group) | ||
| world_size = dist.get_world_size(seq_p_group) | ||
|
|
||
| img_qkv_len = slice_qkv_len |
There was a problem hiding this comment.
The parameter img_qkv_len was renamed to slice_qkv_len in the method signature, which is a good change for generality. However, it's immediately reassigned to a local variable img_qkv_len. For consistency and to improve code clarity, it would be better to use slice_qkv_len throughout the function's body, removing this redundant assignment.
# Consider removing this line and replacing all instances of `img_qkv_len`
# in this function with `slice_qkv_len` for better clarity and consistency.
img_qkv_len = slice_qkv_lenThere was a problem hiding this comment.
To adapt to the project requirements, the input parameter must be named slice_qkv_len. In subsequent code, img_qkv_len used to differentiate between image and text components.
1 participant
ring_attn: fp8_comm & kv_fusion