xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-03-16 10:17:13 +08:00
Code Issues Packages Projects Releases Wiki Activity

minimize fill_

Browse Source 
Signed-off-by: Tyler Michael Smith <tyler@neuralmagic.com>
This commit is contained in:
Tyler Michael Smith 2025-02-04 22:03:51 +00:00
parent 34fb0cbbd0
commit 1244c25908
1 changed files with 2 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

9
vllm/v1/worker/gpu_model_runner.py
View File

@ -457,15 +457,10 @@ class GPUModelRunner:
self.slot_mapping_cpu[:total_num_scheduled_tokens],
non_blocking=True)
# TODO: Which of these are needed?
# Fill unused with -1. Needed for reshape_and_cache
self.slot_mapping[total_num_scheduled_tokens:].fill_(-1)
self.seq_lens[num_reqs:].fill_(0)
self.query_start_loc[num_reqs + 1:].fill_(-1)
self.positions[total_num_scheduled_tokens:].fill_(0)
self.input_batch.block_table.get_device_tensor()[num_reqs:].fill_(-1)
# Fill with -1s -- needed for reshape_and_cache
self.slot_mapping[total_num_scheduled_tokens:].fill_(
-1) # Definitely needed
# Prepare for cascade attention if needed.
common_prefix_len = (scheduler_output.num_common_prefix_blocks *