biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Feature] support sequence parallelism using compilation pass (#16155)

Browse Source 
Signed-off-by: cascade812 <cascade812@outlook.com>
Signed-off-by: Tyler Michael Smith <tyler@neuralmagic.com>
Co-authored-by: Tyler Michael Smith <tyler@neuralmagic.com>
This commit is contained in:
cascade
2025-04-27 06:29:35 -07:00
committed by GitHub
parent ed7a29d9f8
commit 690fe019f0
21 changed files with 1072 additions and 44 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
vllm/v1/worker/gpu_model_runner.py
View File

@@ -1027,7 +1027,15 @@ class GPUModelRunner(LoRAModelRunnerMixin):
num_scheduled_tokens)
else:
# Eager mode.
num_input_tokens = num_scheduled_tokens
# Pad tokens to multiple of tensor_parallel_size when
# enabled collective fusion for SP
tp_size = self.vllm_config.parallel_config.tensor_parallel_size
if self.vllm_config.compilation_config.pass_config. \
enable_sequence_parallelism and tp_size > 1:
from vllm.utils import round_up
num_input_tokens = round_up(num_scheduled_tokens, tp_size)
else:
num_input_tokens = num_scheduled_tokens
attn_metadata.num_input_tokens = num_input_tokens
# _prepare_inputs may reorder the batch, so we must gather multi