[Rebase] Rebase to vLLM 0.29.0 (#7230)

Signed-off-by: tzhouam <tzhouam@connect.ust.hk>
Signed-off-by: Zhou Taichang <tzhouam@connect.ust.hk>
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
This commit is contained in:
Zhou TaichangandClaude Opus 5 authored and GitHub committed 2026-09-10 12:08:28 +08:00
1 parent 333f14e97e
commit aff7d64948
120 files changed
+2958 -1690

No files matched your search

@@ -59,7 +59,7 @@ def _force_default_gemm(monkeypatch):
monkeypatch.setattr(
"vllm.model_executor.layers.linear.dispatch_unquantized_gemm",
lambda: default_unquantized_gemm,
lambda *_args, **_kwargs: default_unquantized_gemm,
)
@@ -1,5 +1,5 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
"""Equivalence test for DreamZero self-attn QKV fusion.
@@ -52,7 +52,7 @@ def _force_default_gemm(monkeypatch):
monkeypatch.setattr(
"vllm.model_executor.layers.linear.dispatch_unquantized_gemm",
lambda: default_unquantized_gemm,
lambda *_args, **_kwargs: default_unquantized_gemm,
)
@@ -1,5 +1,5 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
"""Unit tests for GLM-Image quantization support (W4A16/AutoRound).
These tests verify that the GLM-Image DiT transformer correctly accepts and uses
@@ -69,7 +69,7 @@ def _force_default_gemm(monkeypatch):
monkeypatch.setattr(
"vllm.model_executor.layers.linear.dispatch_unquantized_gemm",
lambda: default_unquantized_gemm,
lambda *_args, **_kwargs: default_unquantized_gemm,
)
@@ -1,5 +1,5 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
from types import SimpleNamespace
import pytest
@@ -64,6 +64,13 @@ def setup_tp_group(monkeypatch, mocker):
lambda: 0,
)
# vLLM 0.29 added a CUDA-only fused embedding kernel that engages whenever
# tp_size > 1; these tests drive CPU tensors, so keep the portable masked path.
monkeypatch.setattr(
"vllm.model_executor.layers.vocab_parallel_embedding.current_platform.is_cuda",
lambda: False,
)
mock_tp_group = mocker.MagicMock()
mock_tp_group.world_size = 2
mocker.patch("vllm.distributed.parallel_state.get_tp_group", return_value=mock_tp_group)
@@ -94,7 +101,7 @@ def setup_tp_group(monkeypatch, mocker):
monkeypatch.setattr(
"vllm.model_executor.layers.linear.dispatch_unquantized_gemm",
lambda: default_unquantized_gemm,
lambda *_args, **_kwargs: default_unquantized_gemm,
)
with set_current_vllm_config(VllmConfig(device_config=device_config)):
@@ -1,12 +1,39 @@
import pytest
import torch
from vllm_omni.diffusion.models.qwen_image.qwen_image_transformer import QwenEmbedRope
from vllm_omni.diffusion.models.qwen_image.qwen_image_transformer import (
QwenEmbedRope,
_apply_qwen_image_rotary_emb,
)
from vllm_omni.diffusion.models.qwen_image.rope_utils import txt_seq_lens_from_embeds
pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
@pytest.mark.parametrize("dtype", [torch.float32, torch.float16, torch.bfloat16])
@pytest.mark.parametrize("noncontiguous", [False, True])
def test_qwen_rotary_preserves_frequency_precision(dtype, noncontiguous):
from diffusers.models.transformers.transformer_qwenimage import apply_rotary_emb_qwen
generator = torch.Generator().manual_seed(42)
x = torch.randn(2, 17, 3, 128, generator=generator).to(dtype)
if noncontiguous:
x = x.transpose(1, 2).contiguous().transpose(1, 2)
angles = torch.randn(17, 64, generator=generator)
freqs = torch.polar(torch.ones_like(angles), angles)
expected = apply_rotary_emb_qwen(x, freqs, use_real=False)
actual = _apply_qwen_image_rotary_emb(x, freqs)
assert actual.dtype == dtype
assert actual.shape == x.shape
torch.testing.assert_close(actual, expected, rtol=0, atol=0)
if dtype == torch.bfloat16:
# A regression to early BF16 frequency rounding must be observable.
rounded_freqs = torch.complex(freqs.real.to(dtype).float(), freqs.imag.to(dtype).float())
rounded = _apply_qwen_image_rotary_emb(x, rounded_freqs)
assert not torch.equal(actual, rounded)
def test_txt_seq_lens_from_embeds_uses_padded_width_not_valid_token_count():
prompt_embeds = torch.zeros(2, 64, 8)
prompt_embeds_mask = torch.zeros(2, 64, dtype=torch.bool)
@@ -1,5 +1,5 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
import os
@@ -22,7 +22,7 @@ def _init_distributed(monkeypatch):
monkeypatch.setattr(
"vllm.model_executor.layers.linear.dispatch_unquantized_gemm",
lambda: default_unquantized_gemm,
lambda *_args, **_kwargs: default_unquantized_gemm,
)
os.environ.setdefault("MASTER_ADDR", "localhost")
os.environ.setdefault("MASTER_PORT", "29501")
@@ -41,8 +41,8 @@ def tp1_group():
os.environ.setdefault("MASTER_ADDR", "localhost")
os.environ.setdefault("MASTER_PORT", "29501")
init_distributed_environment(world_size=1, rank=0, local_rank=0, distributed_init_method="env://")
initialize_model_parallel()
init_distributed_environment(world_size=1, rank=0, local_rank=0, distributed_init_method="env://", backend="gloo")
initialize_model_parallel(backend="gloo")
yield
cleanup_dist_env_and_memory()
@@ -54,7 +54,7 @@ def force_default_gemm(monkeypatch):
monkeypatch.setattr(
"vllm.model_executor.layers.linear.dispatch_unquantized_gemm",
lambda: default_unquantized_gemm,
lambda *_args, **_kwargs: default_unquantized_gemm,
)
@@ -482,8 +482,12 @@ class _LockstepP2PDist:
def test_tiny_transformer_sp2_lockstep_matches_dense(tp1_group, force_default_gemm, mocker, task) -> None:
"""Two lockstep rank threads over uneven frame shards must reproduce the
dense tiny-transformer output, and must actually have sharded the work."""
from vllm_omni.diffusion.attention.backends.sdpa import SDPABackend
from vllm_omni.diffusion.models.sana_video import SanaVideoTransformer3DModel
# These CPU tensor tests must not probe a CUDA device during construction.
mocker.patch("vllm_omni.diffusion.attention.selector._cached_get_backend_cls", return_value=SDPABackend)
torch.manual_seed(3)
model = SanaVideoTransformer3DModel(**_TINY_SP_CONFIG).eval()
for _, param in sorted(model.named_parameters()):
@@ -1,5 +1,5 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
"""TP unit tests for SANA-Video: distributed RMSNorm, sharded attention and
sharded GLUMB temporal conv.
@@ -51,7 +51,7 @@ def force_default_gemm(monkeypatch):
monkeypatch.setattr(
"vllm.model_executor.layers.linear.dispatch_unquantized_gemm",
lambda: default_unquantized_gemm,
lambda *_args, **_kwargs: default_unquantized_gemm,
)
@@ -1,5 +1,5 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
import pytest
import torch
from transformers import T5Config
@@ -54,6 +54,13 @@ def setup_tp_group(monkeypatch, mocker):
lambda: 0,
)
# vLLM 0.29 added a CUDA-only fused embedding kernel that engages whenever
# tp_size > 1; these tests drive CPU tensors, so keep the portable masked path.
monkeypatch.setattr(
"vllm.model_executor.layers.vocab_parallel_embedding.current_platform.is_cuda",
lambda: False,
)
# TP group
mock_tp_group = mocker.MagicMock()
mock_tp_group.world_size = 2