[Neuron][Kernel] NKI Flash PagedAttention with BlockSparse Execution Plan #13249
Conversation
Signed-off-by: Lingfan Yu <[email protected]>
|
👋 Hi! Thank you for contributing to the vLLM project. 💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels. Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging. To run CI, PR reviewers can either: Add 🚀 |
lingfanyu
changed the title
Signed-off-by: Lingfan Yu <[email protected]>
Signed-off-by: Lingfan Yu <[email protected]>
Signed-off-by: Lingfan Yu <[email protected]>
Signed-off-by: Lingfan Yu <[email protected]>
|
This pull request has merge conflicts that must be resolved before it can be |
| self.tile_size_kv = tile_size_kv | ||
| self.block_size = block_size | ||
|
|
||
| def plan(self): |
There was a problem hiding this comment.
Questions:
- I suppose this is meant to be amortized for entire forward pass?
- The logic is quite involved, so I wonder what is the range of wallclock time for this function call for reasonable batch size (e.g. 128) - is it in us range?
| def get_active_block_tables(block_tables, query_lens, seq_lens, block_size, | ||
| num_blocks): | ||
| context_lens = seq_lens - query_lens | ||
| blocks_per_seq = (context_lens + block_size - 1) // block_size |
There was a problem hiding this comment.
| blocks_per_seq = (context_lens + block_size - 1) // block_size | |
| blocks_per_seq = (context_lens + (block_size - 1)) // block_size |
Reduce vectorized (e.g. torch) op dispatch
2 participants
When performing attention computation on RaggedTensor (i.e. a batch of variable-length sequences), we view the batch as one flattened sequence in which original sequences are concatenated along sequence dimension. When performing attention on this flattened sequence, attention that belongs to original sequences happens along the block-diagonal.
Therefore, we have developed a BlockSparse version of flash paged attention to minimize wasted computation. This PR contains two parts:
This PR depends on PR #13245 and #13455 .
@liangfu