Skip to content

Refactor KVCacheManager: Simplify token availability calculation and … #6134

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

Merged
merged 1 commit into from
Jul 17, 2025
Merged

Refactor KVCacheManager: Simplify token availability calculation and … #6134

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
14 changes: 4 additions & 10 deletions tensorrt_llm/_torch/pyexecutor/resource_manager.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -536,16 +536,8 @@ def get_num_kv_blocks(self, num_tokens: int) -> int:
return (num_tokens + self.tokens_per_block - 1) // self.tokens_per_block

def get_num_available_tokens(self, max_num_draft_tokens: int = 0) -> int:
if self.max_attention_window_vec and len(
self.max_attention_window_vec) > 1:
# VSWA case, the available tokens should the the minimum of the available tokens for each window size
min_free_blocks = min(self.impl.get_kv_cache_stats().
num_free_blocks_per_window_size.values())
res = min_free_blocks * self.tokens_per_block - self.num_extra_kv_tokens - max_num_draft_tokens
else:
res = (self.get_num_free_blocks() * self.tokens_per_block -
self.num_extra_kv_tokens - max_num_draft_tokens)
return res
return (self.get_num_free_blocks() * self.tokens_per_block -
self.num_extra_kv_tokens - max_num_draft_tokens)

def get_buffers(self, layer_idx: int) -> Optional[torch.Tensor]:
layer_offset = self.layer_offsets[layer_idx]
Expand Down Expand Up @@ -732,6 +724,8 @@ def calculate_max_num_blocks_from_cpp(

# VSWA on Torch backend has not supported the cross attention.
is_cross_attention = False
# check model config
assert model_config.layer_types is not None, "layer_types have to be set correctly for VSWA"

# Construct WorldConfig from self.mapping
world_config_cpp = WorldConfig(
Expand Down