[torch.compile]: Disable Sequence Parallelism (SP) for piecewise compilation (#38373)

Signed-off-by: SouthWest7 <am1ao@qq.com>
Signed-off-by: Xinan Miao <1403572259@qq.com>
Co-authored-by: SouthWest7 <am1ao@qq.com>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
Co-authored-by: OpenAI Codex <codex@openai.com>
Co-authored-by: Wang Xingran <72983099+wangxingran222@users.noreply.github.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
This commit is contained in:
Xinan Miao
2026-04-26 17:44:42 +00:00
committed by GitHub
co-authored by SouthWest7 gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com> OpenAI Codex Wang Xingran mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
parent b39c266dae
commit 32e45636e3
9 changed files with 222 additions and 79 deletions
@@ -22,6 +22,7 @@ from vllm.config import (
get_current_vllm_config,
set_current_vllm_config,
)
from vllm.config.utils import Range
from vllm.distributed import tensor_model_parallel_all_reduce
from vllm.distributed.parallel_state import (
init_distributed_environment,
@@ -216,6 +217,24 @@ def test_sequence_parallelism_pass(
run_torch_spawn(sequence_parallelism_pass_on_test_model, num_processes)
def test_sequence_parallelism_pass_requires_full_graph_compilation():
vllm_config = VllmConfig()
vllm_config.compilation_config.use_inductor_graph_partition = False
vllm_config.compilation_config.splitting_ops = [
"vllm::unified_attention_with_output"
]
sequence_parallelism_pass = object.__new__(SequenceParallelismPass)
sequence_parallelism_pass.compilation_config = vllm_config.compilation_config
sequence_parallelism_pass.min_token_num = 1
with pytest.raises(
AssertionError,
match="SequenceParallelismPass requires full-graph compilation",
):
sequence_parallelism_pass.is_applicable_for_range(Range(start=8, end=8))
def sequence_parallelism_pass_on_test_model(
local_rank: int,
world_size: int,