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

[V1] Avoid list creation in input preparation #12457

Merged
merged 2 commits into from
Jan 27, 2025
Merged

[V1] Avoid list creation in input preparation #12457

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
ad00877
[V1] Avoid list creation in input preparation
WoosukKwon Jan 26, 2025
31b1a8a
Fix
WoosukKwon Jan 27, 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
17 changes: 12 additions & 5 deletions vllm/v1/worker/gpu_model_runner.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -171,7 +171,8 @@ def __init__(

# OPTIMIZATION: Cache the tensors rather than creating them every step.
self.arange_np = np.arange(max(self.max_num_reqs + 1,
self.max_model_len),
self.max_model_len,
self.max_num_tokens),
dtype=np.int32)
# NOTE(woosuk): These tensors are "stateless", i.e., they are literally
# a faster version of creating a new tensor every time. Thus, we should
Expand Down Expand Up @@ -358,8 +359,15 @@ def _prepare_inputs(self, scheduler_output: "SchedulerOutput"):

# Get batched arange.
# E.g., [2, 5, 3] -> [0, 1, 0, 1, 2, 3, 4, 0, 1, 2]
arange = np.concatenate(
[self.arange_np[:n] for n in num_scheduled_tokens])
# Equivalent to but faster than:
# np.concatenate([np.arange(n) for n in num_scheduled_tokens])
# Step 1. [2, 5, 3] -> [2, 7, 10]
cu_num_tokens = np.cumsum(num_scheduled_tokens)
# Step 2. [2, 7, 10] -> [0, 0, 2, 2, 2, 2, 2, 7, 7, 7]
cumsums_offsets = np.repeat(cu_num_tokens - num_scheduled_tokens,
num_scheduled_tokens)
# Step 3. [0, 1, 0, 1, 2, 3, 4, 0, 1, 2]
arange = self.arange_np[:total_num_scheduled_tokens] - cumsums_offsets

# Get positions.
positions_np = self.positions_np[:total_num_scheduled_tokens]
Expand Down Expand Up @@ -406,8 +414,7 @@ def _prepare_inputs(self, scheduler_output: "SchedulerOutput"):

# Prepare the attention metadata.
self.query_start_loc_np[0] = 0
np.cumsum(num_scheduled_tokens,
out=self.query_start_loc_np[1:num_reqs + 1])
self.query_start_loc_np[1:num_reqs + 1] = cu_num_tokens

self.seq_lens_np[:num_reqs] = (
self.input_batch.num_computed_tokens_cpu[:num_reqs] +
Expand Down
Loading