[CI] Add MTP coverage: Qwen3.5 correctness + no-sync spec decode (#40472)

Signed-off-by: Stefano Castagnetta <scastagnetta@nvidia.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
This commit is contained in:
Stefano Castagnetta
2026-04-30 12:24:09 -07:00
committed by GitHub
co-authored by mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
parent 307b17ce33
commit 92a7c121b6
3 changed files with 53 additions and 1 deletions
+13
View File
@@ -97,3 +97,16 @@ steps:
commands:
- export VLLM_ALLOW_INSECURE_SERIALIZATION=1
- pytest -v -s v1/spec_decode/test_speculators_dflash.py -m slow_test
- label: Spec Decode MTP hybrid (B200)
timeout_in_minutes: 30
device: b200
optional: true
source_file_dependencies:
- vllm/v1/spec_decode/
- vllm/v1/worker/gpu/spec_decode/
- vllm/model_executor/models/qwen3_5.py
- vllm/model_executor/models/qwen3_5_mtp.py
- tests/v1/e2e/spec_decode/
commands:
- pytest -v -s v1/e2e/spec_decode -k "qwen3_5-hybrid"
@@ -82,6 +82,13 @@ SPEC_DECODE_CONFIGS = [
2,
id="eagle-mla-deepseek",
),
pytest.param(
"Qwen/Qwen3.5-0.8B-Base",
"Qwen/Qwen3.5-0.8B-Base",
"mtp",
1,
id="mtp-qwen3_5-hybrid",
),
]
@@ -104,6 +111,14 @@ def test_no_sync_with_spec_decode(
from vllm import LLM, SamplingParams
from vllm.distributed import cleanup_dist_env_and_memory
# Qwen3.5 is a VLM; without this, profile_run runs the ViT warmup
# and peaks well above the 18GB MIG slice used by one of the CI lanes.
# This test only exercises text generation, so the vision tower is
# never needed.
extra_kwargs: dict = {}
if "Qwen3.5" in model:
extra_kwargs["limit_mm_per_prompt"] = {"image": 0, "video": 0}
llm = LLM(
model=model,
max_model_len=256,
@@ -114,6 +129,7 @@ def test_no_sync_with_spec_decode(
},
enforce_eager=True,
async_scheduling=True,
**extra_kwargs,
)
# Assert async scheduling is actually active before running inference.
+24 -1
View File
@@ -1,5 +1,6 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import os
import random
from collections.abc import Iterable
from dataclasses import dataclass
@@ -723,8 +724,13 @@ def test_eagle_correctness_heavy(
[
(("mtp", "XiaomiMiMo/MiMo-7B-Base", 1), False, 0.5), # ref: 65%-70%
(("mtp", "ZixiQi/DeepSeek-V3-4layers-MTP-FP8", 1), False, 0.0), # dummy model
(
("mtp", "Qwen/Qwen3.5-0.8B-Base", 1),
False,
0.20,
), # hybrid + MTP, ref: ~34%-35%
],
ids=["mimo", "deepseek"],
ids=["mimo", "deepseek", "qwen3_5-hybrid"],
)
@single_gpu_only
@large_gpu_mark(min_gb=20)
@@ -750,13 +756,29 @@ def test_mtp_correctness(
method, model_name, tp_size = model_setup
_skip_if_insufficient_gpus_for_tp(tp_size)
if "Qwen3.5" in model_name and os.environ.get("VLLM_USE_V2_MODEL_RUNNER"):
pytest.skip(
"Model Runner V2 does not yet support hybrid models "
"(Qwen3.5 mixes Mamba-style GDN with attention layers)."
)
attn_backend = "TRITON_ATTN" if current_platform.is_rocm() else "auto"
# Qwen3.5 is a VLM; without this, profile_run runs the ViT warmup
# and peaks well above the 18GB MIG slice used by one of the CI
# lanes. This test only exercises text generation, so the vision
# tower is never needed.
extra_kwargs: dict[str, Any] = {}
if "Qwen3.5" in model_name:
extra_kwargs["limit_mm_per_prompt"] = {"image": 0, "video": 0}
ref_llm = LLM(
model=model_name,
max_model_len=2048,
tensor_parallel_size=tp_size,
trust_remote_code=True,
attention_backend=attn_backend,
**extra_kwargs,
)
ref_outputs = ref_llm.chat(test_prompts, sampling_config)
evaluate_llm_for_gsm8k(
@@ -777,6 +799,7 @@ def test_mtp_correctness(
},
max_model_len=2048,
attention_backend=attn_backend,
**extra_kwargs,
)
# MTP supports async scheduling; assert it is active by default.
assert spec_llm.llm_engine.vllm_config.scheduler_config.async_scheduling