Merge branch 'main' into patches
Signed-off-by: MOVZX <movzx@yahoo.com>
This commit is contained in:
@@ -70,20 +70,21 @@ def test_moe_sp_rejects_pipeline_parallel() -> None:
|
||||
),
|
||||
)
|
||||
with (
|
||||
patch("vllm.platforms.current_platform.is_cuda", return_value=True),
|
||||
patch("vllm.platforms.current_platform.is_cuda_alike", return_value=True),
|
||||
pytest.raises(ValueError, match="MoE SP requires PP=1"),
|
||||
):
|
||||
Qwen4ExpForConditionalGenerationConfig.verify_and_update_config(config)
|
||||
|
||||
|
||||
@spawn_new_process_for_each_test
|
||||
@pytest.mark.parametrize("backend", ["amd", "nvidia"])
|
||||
@pytest.mark.parametrize("use_sp", [False, True])
|
||||
def test_qwen4_exp_mtp_returns_sample_and_multi_streams(use_sp: bool) -> None:
|
||||
def test_qwen4_exp_mtp_returns_sample_and_multi_streams(
|
||||
backend: str, use_sp: bool
|
||||
) -> None:
|
||||
"""MTP returns complete, contiguous outputs after gathering padded SP shards."""
|
||||
from vllm.models.qwen4_exp.nvidia.mtp import (
|
||||
Qwen4ExpMultiTokenPredictor,
|
||||
)
|
||||
|
||||
model = object.__new__(Qwen4ExpMultiTokenPredictor)
|
||||
mtp_module = f"vllm.models.qwen4_exp.{backend}.mtp"
|
||||
model = object.__new__(import_module(mtp_module).Qwen4ExpMultiTokenPredictor)
|
||||
model.use_hc_sequence_parallel = use_sp
|
||||
model.use_sequence_parallel = False
|
||||
torch.nn.Module.__init__(model)
|
||||
@@ -98,7 +99,9 @@ def test_qwen4_exp_mtp_returns_sample_and_multi_streams(use_sp: bool) -> None:
|
||||
def decoder(**kwargs):
|
||||
"""Check that the embedding residual shares the hidden state's token slice."""
|
||||
hidden = kwargs["hidden_states"]
|
||||
assert kwargs["prev_block_output"].shape == (hidden.shape[0], model.hidden_size)
|
||||
if backend == "nvidia":
|
||||
prev_block_output = kwargs["prev_block_output"]
|
||||
assert prev_block_output.shape == (hidden.shape[0], model.hidden_size)
|
||||
return hidden, hidden, torch.zeros(hidden.shape[0], model.hc_count)
|
||||
|
||||
model.layers = [decoder]
|
||||
@@ -120,9 +123,9 @@ def test_qwen4_exp_mtp_returns_sample_and_multi_streams(use_sp: bool) -> None:
|
||||
return torch.cat([local, remote, torch.full_like(remote, float("nan"))])
|
||||
|
||||
with (
|
||||
patch("vllm.models.qwen4_exp.nvidia.mtp.get_pp_group", return_value=pp_group),
|
||||
patch("vllm.models.qwen4_exp.nvidia.mtp.sp_shard", side_effect=lambda x: x[:2]),
|
||||
patch("vllm.models.qwen4_exp.nvidia.mtp.sp_all_gather", side_effect=gather),
|
||||
patch(f"{mtp_module}.get_pp_group", return_value=pp_group),
|
||||
patch(f"{mtp_module}.sp_shard", side_effect=lambda x: x[:2]),
|
||||
patch(f"{mtp_module}.sp_all_gather", side_effect=gather),
|
||||
):
|
||||
sample_hidden, returned_multi_hidden = model.forward(
|
||||
input_ids=None,
|
||||
|
||||
Reference in New Issue
Block a user