Skip to content

Fix calculating used blocks in unified attn #232

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

Open
wants to merge 3 commits into
base: main
Choose a base branch
from
+7 −5
Open

Fix calculating used blocks in unified attn #232

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
fbd0268
Fix calculating used blocks
madamczyk-intel Sep 22, 2025
f7a6c97
Merge branch 'main' into fix_unified_acc
madamczyk-intel Sep 25, 2025
5ae2d11
Merge branch 'main' into fix_unified_acc
madamczyk-intel Sep 26, 2025
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
12 changes: 7 additions & 5 deletions vllm_gaudi/extension/unified.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -349,16 +349,16 @@ class Context:
block_usages: torch.tensor

@staticmethod
def create(num_computed_tokens: torch.tensor, block_table: torch.tensor, block_size: int) -> Self:
def create(total_tokens: torch.tensor, block_table: torch.tensor, block_size: int) -> Self:
""" Create a new Context obj """
num_ctx_blocks = (num_computed_tokens + block_size - 1) // block_size
num_ctx_blocks = (total_tokens + block_size - 1) // block_size
if num_ctx_blocks.sum() <= 0:
return None

group_ids, group_offsets = indices_and_offsets(num_ctx_blocks)
block_ids = fetch_2d(block_table, group_ids, group_offsets)
block_usages = torch.clamp(
num_computed_tokens.index_select(0, group_ids) - group_offsets * block_size + 1, 1, block_size)
total_tokens.index_select(0, group_ids) - group_offsets * block_size + 1, 1, block_size)

ctx = Context(group_ids, group_offsets, block_ids, block_usages)
all_shapes = [v.shape for v in ctx._values() if torch.is_tensor(v)]
Expand Down Expand Up @@ -408,7 +408,9 @@ def create_unified_batch(req_ids: list[str], all_token_ids: torch.tensor, num_co
""" Calculate all necessary tensors needed for batch scheduling """
total_tokens = num_computed_tokens + num_scheduled_tokens
query_len = num_scheduled_tokens.sum().item()
contains_prompts = torch.any(total_tokens <= num_prompt_tokens).item()
is_prompt = total_tokens <= num_prompt_tokens
cached_tokens = num_computed_tokens + torch.where(is_prompt, 0, num_scheduled_tokens)
contains_prompts = torch.any(is_prompt).item()
num_output_tokens = total_tokens - num_prompt_tokens + 1
num_output_tokens = torch.clamp(num_output_tokens, torch.zeros_like(num_scheduled_tokens), num_scheduled_tokens)
group_starts = torch.cumsum(num_scheduled_tokens, dim=0) - num_scheduled_tokens
Expand Down Expand Up @@ -440,7 +442,7 @@ def first_dim(t: Optional[torch.tensor]) -> int:
if contains_prompts:
causal_bias = create_causal_bias(token_groups, token_positions, dtype)

ctx = Context.create(num_computed_tokens, block_table, block_size)
ctx = Context.create(cached_tokens, block_table, block_size)
if ctx:
shared_ctx, unique_ctx = ctx.split(num_scheduled_tokens)
if shared_ctx:
Expand Down