[ROCm][CI] Fix nixl tests (#45219)
Signed-off-by: Andreas Karatzas <akaratza@amd.com>
This commit is contained in:
@@ -6,6 +6,7 @@ import json
|
||||
from pathlib import Path
|
||||
|
||||
import pytest
|
||||
from transformers import PretrainedConfig
|
||||
|
||||
from vllm.config import ModelConfig, ParallelConfig, SpeculativeConfig
|
||||
from vllm.transformers_utils.model_arch_config_convertor import (
|
||||
@@ -114,6 +115,22 @@ def _assert_model_config_methods(
|
||||
assert model_config.get_head_size() == expected["head_size"]
|
||||
|
||||
|
||||
def test_head_size_falls_back_when_head_dim_is_zero():
|
||||
"""Regression test for configs that materialize missing head_dim as 0."""
|
||||
hf_config = PretrainedConfig(
|
||||
model_type="deepseek_vl_v2",
|
||||
hidden_size=1280,
|
||||
num_attention_heads=10,
|
||||
num_key_value_heads=10,
|
||||
head_dim=0,
|
||||
kv_lora_rank=None,
|
||||
)
|
||||
|
||||
convertor = ModelArchConfigConvertorBase(hf_config, hf_config)
|
||||
|
||||
assert convertor.get_head_size() == 128
|
||||
|
||||
|
||||
@pytest.mark.parametrize("model", BASE_MODELS_TO_TEST)
|
||||
def test_base_model_arch_config(model: str):
|
||||
"""Test model architecture config for base models."""
|
||||
|
||||
@@ -13,13 +13,13 @@ tp_configs=(
|
||||
"GPU_MEMORY_UTILIZATION=0.6 PREFILLER_TP_SIZE=2 DECODER_TP_SIZE=2"
|
||||
"GPU_MEMORY_UTILIZATION=0.6 PREFILLER_TP_SIZE=1 DECODER_TP_SIZE=2"
|
||||
"GPU_MEMORY_UTILIZATION=0.6 PREFILLER_TP_SIZE=2 DECODER_TP_SIZE=1"
|
||||
"GPU_MEMORY_UTILIZATION=0.8 MODEL_NAMES=deepseek-ai/deepseek-vl2-tiny" # MLA case
|
||||
"GPU_MEMORY_UTILIZATION=0.8 MODEL_NAMES=deepseek-ai/deepseek-vl2-tiny"
|
||||
"GPU_MEMORY_UTILIZATION=0.8 PREFILLER_TP_SIZE=1 DECODER_TP_SIZE=2 MODEL_NAMES=deepseek-ai/deepseek-vl2-tiny"
|
||||
"GPU_MEMORY_UTILIZATION=0.8 PREFILLER_TP_SIZE=2 DECODER_TP_SIZE=1 MODEL_NAMES=deepseek-ai/deepseek-vl2-tiny"
|
||||
)
|
||||
dp_ep_configs=(
|
||||
"DP_EP=1 GPU_MEMORY_UTILIZATION=0.8 PREFILLER_TP_SIZE=1 DECODER_TP_SIZE=2 MODEL_NAMES=deepseek-ai/deepseek-vl2-tiny" # MLA+P-TP1, D-DPEP=2 (TP=1)
|
||||
"DP_EP=1 GPU_MEMORY_UTILIZATION=0.8 PREFILLER_TP_SIZE=2 DECODER_TP_SIZE=2 MODEL_NAMES=deepseek-ai/deepseek-vl2-tiny" # MLA+P-TP2, D-DPEP=2 (TP=1)
|
||||
"DP_EP=1 GPU_MEMORY_UTILIZATION=0.8 PREFILLER_TP_SIZE=1 DECODER_TP_SIZE=2 MODEL_NAMES=deepseek-ai/deepseek-vl2-tiny" # P-TP1, D-DPEP=2 (TP=1)
|
||||
"DP_EP=1 GPU_MEMORY_UTILIZATION=0.8 PREFILLER_TP_SIZE=2 DECODER_TP_SIZE=2 MODEL_NAMES=deepseek-ai/deepseek-vl2-tiny" # P-TP2, D-DPEP=2 (TP=1)
|
||||
)
|
||||
# We assume HMA enabled by default.
|
||||
hybrid_ssm_configs=(
|
||||
|
||||
@@ -4,9 +4,9 @@
|
||||
|
||||
import importlib
|
||||
import importlib.metadata as metadata
|
||||
import pathlib
|
||||
import subprocess
|
||||
import sys
|
||||
import types
|
||||
|
||||
import pytest
|
||||
import torch
|
||||
@@ -20,6 +20,34 @@ def _print_distribution_version(package_name: str) -> None:
|
||||
print(f"{package_name}: {version}")
|
||||
|
||||
|
||||
def _import_nixl_ep_cpp(nixl_ep: types.ModuleType) -> types.ModuleType:
|
||||
candidate_module_names = []
|
||||
|
||||
config_module_name = getattr(getattr(nixl_ep, "Config", None), "__module__", None)
|
||||
if config_module_name and config_module_name.endswith("nixl_ep_cpp"):
|
||||
candidate_module_names.append(config_module_name)
|
||||
|
||||
if torch.version.cuda is not None:
|
||||
cuda_major = torch.version.cuda.split(".", maxsplit=1)[0]
|
||||
candidate_module_names.append(f"nixl_ep_cu{cuda_major}.nixl_ep_cpp")
|
||||
|
||||
# Keep compatibility with the pre-dispatcher wheel layout.
|
||||
candidate_module_names.append("nixl_ep.nixl_ep_cpp")
|
||||
|
||||
for module_name in dict.fromkeys(candidate_module_names):
|
||||
try:
|
||||
return importlib.import_module(module_name)
|
||||
except ModuleNotFoundError as exc:
|
||||
missing_module = exc.name
|
||||
if missing_module not in (module_name, module_name.split(".", 1)[0]):
|
||||
raise
|
||||
|
||||
raise AssertionError(
|
||||
"No nixl_ep_cpp extension module found; tried "
|
||||
f"{', '.join(dict.fromkeys(candidate_module_names))}"
|
||||
)
|
||||
|
||||
|
||||
@pytest.mark.skipif(torch.version.cuda is None, reason="CUDA NIXL EP canary")
|
||||
def test_nixl_and_nixl_ep_imports() -> None:
|
||||
"""Verify both core NIXL and the NIXL EP extension import successfully."""
|
||||
@@ -38,14 +66,13 @@ def test_nixl_and_nixl_ep_imports() -> None:
|
||||
nixl_ep = importlib.import_module("nixl_ep")
|
||||
print(f"nixl_ep: {nixl_ep.__file__}")
|
||||
|
||||
assert nixl_ep.__file__ is not None
|
||||
extension_dir = pathlib.Path(nixl_ep.__file__).parent
|
||||
extension_files = sorted(extension_dir.glob("nixl_ep_cpp*.so"))
|
||||
assert extension_files, f"No nixl_ep_cpp extension found in {extension_dir}"
|
||||
nixl_ep_cpp = _import_nixl_ep_cpp(nixl_ep)
|
||||
assert nixl_ep_cpp.__file__ is not None
|
||||
extension_file = nixl_ep_cpp.__file__
|
||||
print(f"nixl_ep_cpp: {extension_file}")
|
||||
|
||||
extension_file = extension_files[0]
|
||||
completed = subprocess.run(
|
||||
["ldd", str(extension_file)],
|
||||
["ldd", extension_file],
|
||||
capture_output=True,
|
||||
check=False,
|
||||
text=True,
|
||||
|
||||
@@ -94,6 +94,50 @@ def test_logical_to_kernel_block_ids_with_hma():
|
||||
)
|
||||
|
||||
|
||||
@pytest.mark.cpu_test
|
||||
@pytest.mark.parametrize(
|
||||
"is_rocm,has_mamba,use_host_buffer,done_recving,failed_recving,expected_syncs",
|
||||
[
|
||||
(True, True, False, {"req"}, set(), 1),
|
||||
(False, True, False, {"req"}, set(), 0),
|
||||
(True, False, False, {"req"}, set(), 0),
|
||||
(True, True, True, {"req"}, set(), 0),
|
||||
(True, True, False, set(), set(), 0),
|
||||
(True, True, False, {"req"}, {"req"}, 0),
|
||||
],
|
||||
)
|
||||
def test_sync_device_after_mamba_recv_gates(
|
||||
monkeypatch,
|
||||
is_rocm,
|
||||
has_mamba,
|
||||
use_host_buffer,
|
||||
done_recving,
|
||||
failed_recving,
|
||||
expected_syncs,
|
||||
):
|
||||
"""Only direct-GPU Mamba receives on ROCm need a device fence."""
|
||||
from vllm.distributed.kv_transfer.kv_connector.v1.nixl import base_worker
|
||||
from vllm.distributed.kv_transfer.kv_connector.v1.nixl.worker import (
|
||||
NixlConnectorWorker,
|
||||
)
|
||||
|
||||
worker = object.__new__(NixlConnectorWorker)
|
||||
worker._has_mamba = has_mamba
|
||||
worker.use_host_buffer = use_host_buffer
|
||||
|
||||
sync_calls = []
|
||||
monkeypatch.setattr(base_worker.current_platform, "is_rocm", lambda: is_rocm)
|
||||
monkeypatch.setattr(
|
||||
base_worker.torch.accelerator,
|
||||
"synchronize",
|
||||
lambda: sync_calls.append(True),
|
||||
)
|
||||
|
||||
worker._sync_device_after_mamba_recv(done_recving, failed_recving)
|
||||
|
||||
assert len(sync_calls) == expected_syncs
|
||||
|
||||
|
||||
@pytest.mark.cpu_test
|
||||
@pytest.mark.parametrize(
|
||||
"group_spec_types,remote_physical_per_logical,"
|
||||
|
||||
Reference in New Issue
Block a user