iHore/vLLM
8
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity

[CI] Allowlist-shrink batch 2: wire 9 tests/models/ files into Buildkite, fix 2 stale bugs found (#59050)

Browse Source 
Signed-off-by: Turner Jabbour <doubleujabbour@gmail.com>
Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com>
This commit is contained in:
Turner Jabbour
2026-10-06 21:16:53 -07:00
committed by GitHub
co-authored by Claude Sonnet 5
parent ef014bdf06
commit aecb717f10
7 changed files with 143 additions and 141 deletions
Download Patch File Download Diff File Expand all files Collapse all files
.buildkite/test_areas/models_basic.yaml
+18 -4
View File
@@ -68,8 +68,10 @@ steps:
- tests/models/test_deepseek_v4_vl_rocm.py
- tests/models/test_deepseek_v41_decoder_replay_layers.py
- tests/models/test_deepseek_v41_replay_batch.py
- tests/models/transformers/test_layer_registry_e2e.py
- tests/models/test_deepseek_v41_replay_start.py
commands:
- pytest -v -s models/test_terratorch.py models/transformers/test_backend.py models/test_registry.py models/test_deepseek_v4_vl_rocm.py models/test_deepseek_v41_decoder_replay_layers.py models/test_deepseek_v41_replay_batch.py
- pytest -v -s models/test_terratorch.py models/transformers/test_backend.py models/test_registry.py models/test_deepseek_v4_vl_rocm.py models/transformers/test_layer_registry_e2e.py models/test_deepseek_v41_replay_start.py models/test_deepseek_v41_decoder_replay_layers.py models/test_deepseek_v41_replay_batch.py
mirror:
amd:
label: ":amd: (MI355) Basic Models (Other)"
@@ -81,8 +83,15 @@ steps:
source_file_dependencies:
- tests/models/test_deepseek_v4_vl_rocm.py
- tests/models/test_hyv4_rocm.py
- tests/models/transformers/test_layer_registry_e2e.py
- tests/models/test_deepseek_v41_replay_start.py
- tests/models/test_deepseek_v41_decoder_replay_layers.py
- tests/models/test_deepseek_v41_replay_batch.py
- tests/models/test_deepseek_v4_dspark_rocm.py
- tests/models/test_deepseek_v4_rocm_wo_a.py
- tests/models/test_deepseek_v4_rocm_compressor_gemm_fusion.py
commands:
- pytest -v -s models/test_terratorch.py models/transformers/test_backend.py models/test_registry.py models/test_deepseek_v4_vl_rocm.py
- pytest -v -s models/test_terratorch.py models/transformers/test_backend.py models/test_registry.py models/test_deepseek_v4_vl_rocm.py models/transformers/test_layer_registry_e2e.py models/test_deepseek_v41_replay_start.py models/test_deepseek_v41_decoder_replay_layers.py models/test_deepseek_v41_replay_batch.py models/test_deepseek_v4_dspark_rocm.py models/test_deepseek_v4_rocm_wo_a.py models/test_deepseek_v4_rocm_compressor_gemm_fusion.py
- VLLM_ROCM_USE_AITER=1 pytest -v -s models/test_hyv4_rocm.py -m 'not distributed'
- label: ":nvidia: (B200) Inkling"
@@ -229,7 +238,7 @@ steps:
key: basic-models-test-other-cpu
depends_on:
- image-build-cpu
timeout_in_minutes: 20
timeout_in_minutes: 30
source_file_dependencies:
- vllm/
- "!vllm/distributed/kv_transfer/"
@@ -239,6 +248,11 @@ steps:
- tests/models/test_qwen3_5_mtp_config.py
- tests/models/transformers/fusers/
- tests/model_executor/layers/test_activation.py
- tests/models/test_language_model_cache_is_weak.py
- tests/models/deepseek_v32/test_sequence_parallel.py
- tests/models/test_deepseek_v4_fi_moe_ep.py
- tests/models/test_dspark_mla.py
- tests/models/transformers/test_layer_registry.py
device: cpu-small
commands:
- pytest -v -s models/test_utils.py models/test_vision.py models/test_adapters.py models/test_qwen3_5_mtp_config.py models/transformers/fusers/ model_executor/layers/test_activation.py
- pytest -v -s models/test_utils.py models/test_vision.py models/test_adapters.py models/test_qwen3_5_mtp_config.py models/transformers/fusers/ model_executor/layers/test_activation.py models/test_language_model_cache_is_weak.py models/deepseek_v32/test_sequence_parallel.py models/test_deepseek_v4_fi_moe_ep.py models/test_dspark_mla.py models/transformers/test_layer_registry.py
tests/models/test_dspark_mla.py
+1
View File
@@ -85,6 +85,7 @@ def test_dspark_markov_head_is_replicated(
assert head.markov_w2.tp_size == 1
assert head.markov_w1.weight.shape == (128, 8)
assert head.markov_w2.weight.shape == (128, 8)
head.markov_w2.quant_method.process_weights_after_loading(head.markov_w2)
def fail_collective(*args, **kwargs):
raise AssertionError("replicated Markov head must not invoke TP collectives")
tests/models/transformers/test_layer_registry.py
+9 -106
View File
@@ -7,6 +7,9 @@
is set and the symbol exists, and otherwise falls back to
`vllm.model_executor.layers.<module>`. These tests pin that contract and the
logging that reports which source was used.
CPU-only: no engine is built. The end-to-end counterpart that serves a model
through `vllm_runner` lives in `test_layer_registry_e2e.py` (GPU job).
"""
import logging
@@ -14,7 +17,6 @@ import sys
import types
import pytest
import torch
from vllm.model_executor.models.transformers import layers
@@ -42,16 +44,16 @@ def test_falls_back_to_vllm_when_disabled(monkeypatch, fake_hw_layernorm):
assert layers._resolve("layernorm", "RMSNorm") is VllmRMSNorm
def test_uses_hw_agnostic_when_enabled(monkeypatch, fake_hw_layernorm, caplog):
def test_uses_hw_agnostic_when_enabled(monkeypatch, fake_hw_layernorm, caplog_vllm):
"""Enabled and available: the hw-agnostic class is used and logged."""
monkeypatch.setenv("VLLM_USE_HW_AGNOSTIC", "1")
with caplog.at_level(logging.INFO):
with caplog_vllm.at_level(logging.INFO):
resolved = layers._resolve("layernorm", "RMSNorm")
assert resolved is fake_hw_layernorm.RMSNorm
assert "Using hw-agnostic layer: RMSNorm" in caplog.text
assert "Using hw-agnostic layer: RMSNorm" in caplog_vllm.text
def test_falls_back_when_symbol_missing(monkeypatch, caplog):
def test_falls_back_when_symbol_missing(monkeypatch, caplog_vllm):
"""Enabled but the symbol is not ported: fall back to vLLM and warn."""
monkeypatch.setenv("VLLM_USE_HW_AGNOSTIC", "1")
# A hw-agnostic module without the requested attribute triggers fallback.
@@ -59,10 +61,10 @@ def test_falls_back_when_symbol_missing(monkeypatch, caplog):
monkeypatch.setitem(sys.modules, HW_MODULE, empty)
from vllm.model_executor.layers.layernorm import RMSNorm as VllmRMSNorm
with caplog.at_level(logging.WARNING):
with caplog_vllm.at_level(logging.WARNING):
resolved = layers._resolve("layernorm", "RMSNorm")
assert resolved is VllmRMSNorm
assert "falling back to default" in caplog.text
assert "falling back to default" in caplog_vllm.text
def test_act_and_mul_falls_back_for_unknown_activation(
@@ -76,102 +78,3 @@ def test_act_and_mul_falls_back_for_unknown_activation(
from vllm.model_executor.layers.activation import GeluAndMul
assert isinstance(layers.get_act_and_mul_fn("gelu"), GeluAndMul)
@pytest.fixture(scope="module")
def tiny_llama_path(tmp_path_factory):
"""A randomly-initialized microscopic Llama saved to disk (with an ungated
tokenizer) so vLLM can load it like any local checkpoint."""
from transformers import AutoTokenizer, LlamaConfig, LlamaForCausalLM
tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/llama-tokenizer")
config = LlamaConfig(
vocab_size=tokenizer.vocab_size,
hidden_size=128,
intermediate_size=256,
num_hidden_layers=2,
num_attention_heads=4,
rms_norm_eps=1e-6,
hidden_act="silu",
)
torch.manual_seed(0)
model = LlamaForCausalLM(config)
path = tmp_path_factory.mktemp("tiny_llama")
model.save_pretrained(path)
tokenizer.save_pretrained(path)
return str(path)
# Registered names of the layers the backend can
# currently route to hw-agnostic implementations.
_COVERED_LAYERS = ("rms_norm", "silu_and_mul")
def _layer_providers(model) -> dict[str, str]:
"""Map each covered layer type present in the model to the provider its
implementation came from (``hw_agnostic`` or ``vllm``).
"""
def provider_of(module) -> str | None:
for cls in type(module).__mro__:
if "hw_agnostic.layers" in cls.__module__:
return "hw_agnostic"
if ".model_executor.layers." in cls.__module__:
return "vllm"
return None
providers: dict[str, str] = {}
for module in model.modules():
name = getattr(module, "name", None)
if (
name in _COVERED_LAYERS
and name not in providers
and (prov := provider_of(module)) is not None
):
providers[name] = prov
return providers
def _serve(vllm_runner, model_path, prompts):
"""Serve the model through the backend; return (layer_providers, logprobs)."""
with vllm_runner(
model_path,
model_impl="transformers",
max_model_len=64,
enforce_eager=True,
gpu_memory_utilization=0.3,
) as runner:
assert runner.llm.llm_engine.model_config.using_transformers_backend()
providers = runner.apply_model(_layer_providers)[0]
outputs = runner.generate_greedy_logprobs(
prompts, max_tokens=32, num_logprobs=5
)
return providers, outputs
def test_hw_agnostic_matches_vllm_end_to_end(monkeypatch, vllm_runner, tiny_llama_path):
"""Serving the tiny model with hw-agnostic layers matches the vLLM baseline."""
# spawn: worker re-imports layers with the env set (see docstring).
monkeypatch.setenv("VLLM_WORKER_MULTIPROC_METHOD", "spawn")
# apply_model pickles the introspection function.
monkeypatch.setenv("VLLM_ALLOW_INSECURE_SERIALIZATION", "1")
from ..utils import check_logprobs_close
prompts = ["The capital of France is", "vLLM is"]
monkeypatch.setenv("VLLM_USE_HW_AGNOSTIC", "0")
vllm_providers, vllm_outputs = _serve(vllm_runner, tiny_llama_path, prompts)
# Both replaceable layers present in a Llama block must be vLLM's here.
assert vllm_providers == {"rms_norm": "vllm", "silu_and_mul": "vllm"}
monkeypatch.setenv("VLLM_USE_HW_AGNOSTIC", "1")
hw_providers, hw_outputs = _serve(vllm_runner, tiny_llama_path, prompts)
assert hw_providers == {"rms_norm": "hw_agnostic", "silu_and_mul": "hw_agnostic"}
check_logprobs_close(
outputs_0_lst=vllm_outputs,
outputs_1_lst=hw_outputs,
name_0="vllm",
name_1="hw_agnostic",
)
tests/models/transformers/test_layer_registry_e2e.py
+113
View File
@@ -0,0 +1,113 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""End-to-end check of the Transformers backend's hw-agnostic layers.
Serves a tiny Llama through `vllm_runner` with and without
`VLLM_USE_HW_AGNOSTIC` and checks the layer providers and logprobs. Needs a
real engine, so it runs in the GPU job. The CPU-only unit tests for the
resolution logic (`layers._resolve`) are in `test_layer_registry.py`.
"""
import pytest
import torch
@pytest.fixture(scope="module")
def tiny_llama_path(tmp_path_factory):
"""A randomly-initialized microscopic Llama saved to disk (with an ungated
tokenizer) so vLLM can load it like any local checkpoint."""
from transformers import AutoTokenizer, LlamaConfig, LlamaForCausalLM
tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/llama-tokenizer")
config = LlamaConfig(
vocab_size=tokenizer.vocab_size,
hidden_size=128,
intermediate_size=256,
num_hidden_layers=2,
num_attention_heads=4,
rms_norm_eps=1e-6,
hidden_act="silu",
)
torch.manual_seed(0)
model = LlamaForCausalLM(config)
path = tmp_path_factory.mktemp("tiny_llama")
model.save_pretrained(path)
tokenizer.save_pretrained(path)
return str(path)
# Registered names of the layers the backend can
# currently route to hw-agnostic implementations.
_COVERED_LAYERS = ("rms_norm", "silu_and_mul")
def _layer_providers(model) -> dict[str, str]:
"""Map each covered layer type present in the model to the provider its
implementation came from (``hw_agnostic`` or ``vllm``).
"""
def provider_of(module) -> str | None:
for cls in type(module).__mro__:
if "hw_agnostic.layers" in cls.__module__:
return "hw_agnostic"
if ".model_executor.layers." in cls.__module__:
return "vllm"
return None
providers: dict[str, str] = {}
for module in model.modules():
name = getattr(module, "name", None)
if (
name in _COVERED_LAYERS
and name not in providers
and (prov := provider_of(module)) is not None
):
providers[name] = prov
return providers
def _serve(vllm_runner, model_path, prompts):
"""Serve the model through the backend; return (layer_providers, logprobs)."""
with vllm_runner(
model_path,
model_impl="transformers",
max_model_len=64,
enforce_eager=True,
gpu_memory_utilization=0.3,
) as runner:
assert runner.llm.llm_engine.model_config.using_transformers_backend()
providers = runner.apply_model(_layer_providers)[0]
outputs = runner.generate_greedy_logprobs(
prompts, max_tokens=32, num_logprobs=5
)
return providers, outputs
def test_hw_agnostic_matches_vllm_end_to_end(monkeypatch, vllm_runner, tiny_llama_path):
"""Serving the tiny model with hw-agnostic layers matches the vLLM baseline."""
# spawn: `layers` resolves its symbols at import time, so the worker must
# re-import it with VLLM_USE_HW_AGNOSTIC already set (fork would inherit
# the parent's already-resolved module).
monkeypatch.setenv("VLLM_WORKER_MULTIPROC_METHOD", "spawn")
# apply_model pickles the introspection function.
monkeypatch.setenv("VLLM_ALLOW_INSECURE_SERIALIZATION", "1")
from ..utils import check_logprobs_close
prompts = ["The capital of France is", "vLLM is"]
monkeypatch.setenv("VLLM_USE_HW_AGNOSTIC", "0")
vllm_providers, vllm_outputs = _serve(vllm_runner, tiny_llama_path, prompts)
# Both replaceable layers present in a Llama block must be vLLM's here.
assert vllm_providers == {"rms_norm": "vllm", "silu_and_mul": "vllm"}
monkeypatch.setenv("VLLM_USE_HW_AGNOSTIC", "1")
hw_providers, hw_outputs = _serve(vllm_runner, tiny_llama_path, prompts)
assert hw_providers == {"rms_norm": "hw_agnostic", "silu_and_mul": "hw_agnostic"}
check_logprobs_close(
outputs_0_lst=vllm_outputs,
outputs_1_lst=hw_outputs,
name_0="vllm",
name_1="hw_agnostic",
)
tests/test_zen_cpu_platform_detection.py
-17
View File
@@ -13,23 +13,6 @@ from vllm.platforms import (
)
@pytest.fixture
def caplog_vllm(caplog):
"""`caplog`, but also captures vLLM's loggers.
vllm/logger.py configures the "vllm" logger with propagate=False, so
records from child loggers like "vllm.platforms" reach vLLM's own
handler but never bubble up to the root logger where plain `caplog`
listens. Attach caplog's handler directly to "vllm" to see them too.
"""
logger = logging.getLogger("vllm")
logger.addHandler(caplog.handler)
try:
yield caplog
finally:
logger.removeHandler(caplog.handler)
def test_is_amd_zen_cpu_detects_amd_with_avx512():
cpuinfo = "vendor_id: AuthenticAMD\nflags: avx avx2 avx512f avx512bw"
with (
tools/pre_commit/test_tethering_allowlist.txt
-11
View File
@@ -48,17 +48,6 @@ tests/kernels/test_shuffle_rows.py # explicitly --ignore'd in kernels.yaml with
tests/lora/test_deepseekv2_tp.py # --ignore'd from sharded LoRA run; no dedicated *_tp job
tests/lora/test_qwen3moe_tp.py # --ignore'd from sharded LoRA run; no dedicated *_tp job
# tests/models/ - the models_*.yaml jobs run specific paths, not all of models/.
tests/models/deepseek_v32/test_sequence_parallel.py # no models_*.yaml job collects this path
tests/models/test_deepseek_v4_dspark_rocm.py # added by #52362 (2026-09-21), no models_*.yaml job collects this path
tests/models/test_deepseek_v4_fi_moe_ep.py # no models_*.yaml job collects this path
tests/models/test_deepseek_v4_rocm_compressor_gemm_fusion.py # no models_*.yaml job collects this path
tests/models/test_deepseek_v4_rocm_wo_a.py # added by #54894 (2026-09-20), no models_*.yaml job collects this path
tests/models/test_deepseek_v41_replay_start.py # added by #56227 (2026-09-18), no models_*.yaml job collects this path
tests/models/test_dspark_mla.py # no models_*.yaml job collects this path
tests/models/test_language_model_cache_is_weak.py # no models_*.yaml job collects this path
tests/models/transformers/test_layer_registry.py # no models_*.yaml job collects this path
# tests/v1/ - the v1 subtree is otherwise covered by many `pytest v1/...` jobs.
tests/v1/ec_connector/integration/test_epd_correctness.py # run_epd_correctness_test.sh exists but no Buildkite job invokes it
tests/v1/ec_connector/integration/test_nixl_failure.py # 2-GPU + NIXL E/PD integration test (#55290); no Buildkite job collects ec_connector/integration/
vllm/triton_utils/importing.py
+2 -3
View File
@@ -3,11 +3,10 @@
import os
import types
from importlib.metadata import version
from importlib.util import find_spec
from vllm.logger import init_logger
from vllm.platforms import current_platform
from vllm.platforms import current_platform, vllm_version_matches_substr
from vllm.utils.math_utils import cdiv
logger = init_logger(__name__)
@@ -35,7 +34,7 @@ if HAS_TRITON:
]
# Check Triton CPU
if "cpu" in version("vllm"):
if vllm_version_matches_substr("cpu"):
if "cpu" in backends:
HAS_TRITON = True
# Suppress following warnings on CPU-only platforms