xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-13 15:55:48 +08:00
Code Issues Packages Projects Releases Wiki Activity

[V1] Avoid list creation in input preparation (#12457)

Browse Source 
Signed-off-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>
This commit is contained in:
Woosuk Kwon 2025-01-26 19:57:56 -08:00 committed by GitHub
parent 582cf78798
commit 28e0750847
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 12 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

17
vllm/v1/worker/gpu_model_runner.py
View File

@ -171,7 +171,8 @@ class GPUModelRunner:
# OPTIMIZATION: Cache the tensors rather than creating them every step. # OPTIMIZATION: Cache the tensors rather than creating them every step.
self.arange_np = np.arange(max(self.max_num_reqs + 1, 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) dtype=np.int32)
# NOTE(woosuk): These tensors are "stateless", i.e., they are literally # NOTE(woosuk): These tensors are "stateless", i.e., they are literally
# a faster version of creating a new tensor every time. Thus, we should # a faster version of creating a new tensor every time. Thus, we should
@ -358,8 +359,15 @@ class GPUModelRunner:
# Get batched arange. # Get batched arange.
# E.g., [2, 5, 3] -> [0, 1, 0, 1, 2, 3, 4, 0, 1, 2] # E.g., [2, 5, 3] -> [0, 1, 0, 1, 2, 3, 4, 0, 1, 2]
arange = np.concatenate( # Equivalent to but faster than:
[self.arange_np[:n] for n in num_scheduled_tokens]) # 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. # Get positions.
positions_np = self.positions_np[:total_num_scheduled_tokens] positions_np = self.positions_np[:total_num_scheduled_tokens]
@ -406,8 +414,7 @@ class GPUModelRunner:
# Prepare the attention metadata. # Prepare the attention metadata.
self.query_start_loc_np[0] = 0 self.query_start_loc_np[0] = 0
np.cumsum(num_scheduled_tokens, self.query_start_loc_np[1:num_reqs + 1] = cu_num_tokens
out=self.query_start_loc_np[1:num_reqs + 1])
self.seq_lens_np[:num_reqs] = ( self.seq_lens_np[:num_reqs] = (
self.input_batch.num_computed_tokens_cpu[:num_reqs] + self.input_batch.num_computed_tokens_cpu[:num_reqs] +