[Test] Unify L2/L3 test layout, Buildkite steps, and test helpers (#2556)

Signed-off-by: wangyu <410167048@qq.com>
Signed-off-by: Huang, Zeyu <11222265+fhfuih@users.noreply.github.com>
Signed-off-by: wangyu <53896905+yenuo26@users.noreply.github.com>
Co-authored-by: Huang, Zeyu <11222265+fhfuih@users.noreply.github.com>
This commit is contained in:
wangyuandHuang, Zeyu authored and GitHub committed 2026-05-08 14:12:23 +08:00
1 parent b85833eeec
commit 6f2ad7b403
134 files changed
+4904 -4185

No files matched your search

@@ -8,16 +8,33 @@ produced by generator2tokenizer_async_chunk and cut leading context samples
from decoded audio arrays.
"""
import functools
import pytest
import torch
from vllm_omni.model_executor.models.voxtral_tts.voxtral_tts import (
apply_ctx_frames_cutting,
parse_batched_audio_input,
)
pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
@functools.lru_cache(maxsize=1)
def _voxtral_tts_parsers():
# ``vllm_omni...voxtral_tts`` does ``from vllm.model_executor.models.utils import ...``.
# If that module was first pulled in under a different import context in the
# same process, ``direct_register_custom_op`` for ``sequence_parallel_chunk_impl``
# can run twice. Match ``tests.model_executor.helpers`` + pin ``utils`` first.
from tests.model_executor.helpers import bootstrap_vllm_layer_custom_op_modules
bootstrap_vllm_layer_custom_op_modules()
import vllm.model_executor.models.utils # noqa: F401
from vllm_omni.model_executor.models.voxtral_tts.voxtral_tts import (
apply_ctx_frames_cutting,
parse_batched_audio_input,
)
return parse_batched_audio_input, apply_ctx_frames_cutting
NUM_CODEBOOKS = 37
@@ -37,6 +54,7 @@ def _make_input_ids(*requests: tuple[int, int, torch.Tensor]) -> torch.Tensor:
def test_parse_single_request():
"""Single request is parsed correctly."""
parse_batched_audio_input, _ = _voxtral_tts_parsers()
ctx_frames = 3
context_length = 5
total_frames = ctx_frames + context_length
@@ -54,6 +72,7 @@ def test_parse_single_request():
def test_parse_multiple_requests():
"""Multiple requests in a batch are parsed correctly."""
parse_batched_audio_input, _ = _voxtral_tts_parsers()
req1 = (2, 4, torch.ones((6, NUM_CODEBOOKS), dtype=torch.long))
req2 = (0, 3, torch.full((3, NUM_CODEBOOKS), 7, dtype=torch.long))
input_ids = _make_input_ids(req1, req2)
@@ -70,6 +89,7 @@ def test_parse_multiple_requests():
def test_parse_zero_ctx_frames():
"""ctx_frames=0 means no context, just chunk data."""
parse_batched_audio_input, _ = _voxtral_tts_parsers()
ctx_frames = 0
context_length = 10
tokens = torch.randint(0, 100, (context_length, NUM_CODEBOOKS))
@@ -83,6 +103,7 @@ def test_parse_zero_ctx_frames():
def test_parse_fails_on_misaligned_input():
"""Input not divisible by num_codebooks should raise AssertionError."""
parse_batched_audio_input, _ = _voxtral_tts_parsers()
# 2 header elements + 5 data elements (not divisible by 37)
input_ids = torch.tensor([0, 1] + [0] * 5)
with pytest.raises(AssertionError, match="divisible by"):
@@ -91,6 +112,7 @@ def test_parse_fails_on_misaligned_input():
def test_parse_custom_num_codebooks():
"""Works with a non-default num_codebooks value."""
parse_batched_audio_input, _ = _voxtral_tts_parsers()
cb = 4
ctx_frames = 1
context_length = 2
@@ -111,6 +133,7 @@ def test_parse_custom_num_codebooks():
def test_cut_removes_leading_samples():
"""Context frames are removed from the front of the audio array."""
_, apply_ctx_frames_cutting = _voxtral_tts_parsers()
downsample_factor = 240
ctx_frames = 5
total_samples = 2400 # 10 frames * 240
@@ -126,6 +149,7 @@ def test_cut_removes_leading_samples():
def test_cut_zero_ctx_frames_unchanged():
"""ctx_frames=0 leaves the audio unchanged."""
_, apply_ctx_frames_cutting = _voxtral_tts_parsers()
downsample_factor = 240
audio = torch.randn(2400)
@@ -137,6 +161,7 @@ def test_cut_zero_ctx_frames_unchanged():
def test_cut_multiple_requests():
"""Each request in the batch gets its own ctx_frames cut."""
_, apply_ctx_frames_cutting = _voxtral_tts_parsers()
downsample_factor = 100
audio1 = torch.arange(1000, dtype=torch.float32) # 10 frames
audio2 = torch.arange(500, dtype=torch.float32) # 5 frames
@@ -152,6 +177,7 @@ def test_cut_multiple_requests():
def test_cut_all_frames_returns_empty():
"""Cutting all frames returns an empty tensor."""
_, apply_ctx_frames_cutting = _voxtral_tts_parsers()
downsample_factor = 100
ctx_frames = 5
audio = torch.randn(500) # exactly 5 frames
@@ -12,8 +12,7 @@ Euler ODE with CFG. The synthetic model here uses simple linear layers
to exercise the wrapper mechanism while keeping the test lightweight.
"""
import importlib.util
import os
import functools
import pytest
import torch
@@ -32,50 +31,18 @@ N_ACOUSTIC_CODEBOOK = 7
SEMANTIC_CODEBOOK_SIZE = 128
ACOUSTIC_EMBEDDINGS_LEVELS = 1024
# Load CUDAGraphAcousticTransformerWrapper: try package import first, fall back to direct file load
try:
@functools.lru_cache(maxsize=1)
def _voxtral_cudagraph_deps():
"""Load voxtral CUDA graph helpers only when CUDA tests run (avoids re-exec + duplicate vLLM op registration)."""
from vllm_omni.model_executor.models.voxtral_tts.cuda_graph_acoustic_transformer_wrapper import (
CUDAGraphAcousticTransformerWrapper,
)
from vllm_omni.model_executor.models.voxtral_tts.voxtral_tts_audio_generation import (
AudioSpecialTokens,
)
except Exception:
_WRAPPER_PATH = os.path.join(
os.path.dirname(__file__),
os.pardir,
os.pardir,
os.pardir,
os.pardir,
"vllm_omni",
"model_executor",
"models",
"voxtral_tts",
"cuda_graph_acoustic_transformer_wrapper.py",
)
_spec = importlib.util.spec_from_file_location(
"cuda_graph_acoustic_transformer_wrapper", os.path.abspath(_WRAPPER_PATH)
)
_mod = importlib.util.module_from_spec(_spec)
_spec.loader.exec_module(_mod)
CUDAGraphAcousticTransformerWrapper = _mod.CUDAGraphAcousticTransformerWrapper
_AUDIO_GEN_PATH = os.path.join(
os.path.dirname(__file__),
os.pardir,
os.pardir,
os.pardir,
os.pardir,
"vllm_omni",
"model_executor",
"models",
"voxtral_tts",
"voxtral_tts_audio_generation.py",
)
_spec2 = importlib.util.spec_from_file_location("voxtral_tts_audio_generation", os.path.abspath(_AUDIO_GEN_PATH))
_mod2 = importlib.util.module_from_spec(_spec2)
_spec2.loader.exec_module(_mod2)
AudioSpecialTokens = _mod2.AudioSpecialTokens
return CUDAGraphAcousticTransformerWrapper, AudioSpecialTokens
class SyntheticAcousticTransformerArgs:
@@ -102,6 +69,7 @@ class SyntheticAcousticTransformer(nn.Module):
def __init__(self):
super().__init__()
_, AudioSpecialTokens = _voxtral_cudagraph_deps()
self.model_args = SyntheticModelArgs()
self.acoustic_transformer_args = SyntheticAcousticTransformerArgs()
self.acoustic_embeddings_levels = ACOUSTIC_EMBEDDINGS_LEVELS
@@ -128,6 +96,7 @@ class SyntheticModel(nn.Module):
def __init__(self):
super().__init__()
_, AudioSpecialTokens = _voxtral_cudagraph_deps()
self.acoustic_transformer = SyntheticAcousticTransformer()
end_audio_id = AudioSpecialTokens.id(AudioSpecialTokens.end_audio)
empty_audio_id = AudioSpecialTokens.id(AudioSpecialTokens.empty_audio)
@@ -140,6 +109,7 @@ class SyntheticModel(nn.Module):
cfg_alpha: torch.Tensor,
):
"""Eager fallback path: replicate what the wrapper does."""
_, AudioSpecialTokens = _voxtral_cudagraph_deps()
at = self.acoustic_transformer
B = hidden_states.shape[0]
@@ -200,6 +170,7 @@ def model():
@pytest.fixture(scope="module")
def wrapper(model):
"""Create a warmed-up CUDAGraphAcousticTransformerWrapper."""
CUDAGraphAcousticTransformerWrapper, _ = _voxtral_cudagraph_deps()
w = CUDAGraphAcousticTransformerWrapper(
model=model,
capture_sizes=[1, 2, 4, 8, 16, 32],