Skip to content
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.

Sign up for GitHub

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

[WIP][Attention] KV Splits heuristic for MLA #12654

Draft
wants to merge 1 commit into
base: main
Choose a base branch
from
Draft

[WIP][Attention] KV Splits heuristic for MLA #12654

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
21 changes: 17 additions & 4 deletions vllm/attention/backends/triton_mla.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -255,7 +255,7 @@

num_prefill_tokens: int

num_kv_splits: int = 4 # TODO(lucas) add heuristic
num_kv_splits: int = 8 # TODO(lucas) add heuristic
attn_logits: Optional[torch.Tensor] = None
req_idx: Optional[torch.Tensor] = None

Expand Down Expand Up @@ -319,7 +319,8 @@
context_lens_tensor=context_lens_tensor,
block_tables=block_tables,
use_cuda_graph=False,
head_dim=self.head_dim)
head_dim=self.head_dim,
num_kv_splits=self.num_kv_splits)
return self._cached_prefill_metadata

@property
Expand Down Expand Up @@ -366,7 +367,8 @@
block_tables=block_tables,
use_cuda_graph=self.use_cuda_graph,
input_positions=input_positions,
head_dim=self.head_dim)
head_dim=self.head_dim,
num_kv_splits=self.num_kv_splits)
return self._cached_decode_metadata

def advance_step(self,
Expand Down Expand Up @@ -621,6 +623,17 @@
for modality, placeholder_map in
self.multimodal_placeholder_maps.items()
}

# Rough heuristic for num_kv_splits
# not empirically validated
if max_decode_seq_len < 512:
num_kv_splits = 2
elif max_decode_seq_len < 512:
num_kv_splits = 4
elif max_decode_seq_len < 4096:
num_kv_splits = 8
else:
num_kv_splits = 16

Check failure on line 636 in vllm/attention/backends/triton_mla.py

View workflow job for this annotation

GitHub Actions / pre-commit

Ruff (F841) 

vllm/attention/backends/triton_mla.py:636:13: F841 Local variable `num_kv_splits` is assigned to but never used

return TritonMLAMetadata(
num_prefills=self.num_prefills,
Expand All @@ -641,7 +654,7 @@
context_lens_tensor=context_lens_tensor,
block_tables=block_tables,
use_cuda_graph=use_captured_graph,
num_kv_splits=4, # TODO(lucas) add heuristic
num_kv_splits=8, # TODO(lucas) add heuristic
head_dim=self.runner.model_config.get_head_size(),
)

Expand Down
Loading