Support the case where query_len>context_len in the multi-queries paged attention. #8356
Conversation
This file contains hidden or 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
| Hi @WoosukKwon , could you take a look at the test and the modified ref impl when you get a chance? |
WoosukKwon approved these changes Nov 8, 2024
WoosukKwon reviewed Nov 8, 2024
| @vanbasten23 If possible, can you merge the PR today so that I can use this features in the nightly wheel tomorrow? |
| @vanbasten23 can you fix the linter? |
Sorry for the late reply as I was OOO last Friday. Let me try to merge it today. |
| hm, after I take into account the effective_q_lens in the kernel, I found the results from the ref impl and the kernel mismatch. Need to look into it. |
vanbasten23 changed the title | Ok, I fixed the kernel and the test. cc @WoosukKwon If the test |
vanbasten23 commented Nov 13, 2024
| self.fail(f'Unsupported dtype: {dtype}') | ||
| for b in range(batch_size): | ||
| # N.B. For the output ([batch_size, query_len, num_q_heads, head_dim]) at query_len dim, all the value after the effective_q_len will be thrown away due to we padding the query seq len. The values after the effective_q_len may differ between the kernel and the ref impl because of the causal mask. | ||
| effective_q_len = effective_q_lens[b] |
There was a problem hiding this comment.
@WoosukKwon , the kernel makes sure that the results up to effective_q_len are correct and the kernel makes no guarantee on the results beyond effective_q_len
There was a problem hiding this comment.
Makes sense. Thanks for the clarification!
vanbasten23 commented Nov 13, 2024
| In this PR, the test fails on TPU v4 (but succeeded TPU v5e): Notice the failure comes from the reference impl (pure jax, not Pallas kernel related) |
| As Jack mentioned, it might be due to OOM in the kernel but the error is not surfaced until later because the kernel execution is async. |
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
4 participants
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.
This PR support the case where query_len>context len. In reality, vLLM may pad the query_len (but not for the kv seq len). That means the query_len may be longer than the kv_len. This PR
Test plans: