[Hardware] Replace torch.cuda.device_count/current_device/set_device API (#36145)
Signed-off-by: Kunshang Ji <jikunshang95@gmail.com> Signed-off-by: Kunshang Ji <kunshang.ji@intel.com>
This commit is contained in:
@@ -300,7 +300,7 @@ def async_tp_pass_on_test_model(
|
||||
set_random_seed(0)
|
||||
|
||||
device = torch.device(f"cuda:{local_rank}")
|
||||
torch.cuda.set_device(device)
|
||||
torch.accelerator.set_device_index(device)
|
||||
torch.set_default_device(device)
|
||||
torch.set_default_dtype(dtype)
|
||||
|
||||
|
||||
@@ -262,7 +262,7 @@ def all_reduce_fusion_pass_on_test_model(
|
||||
set_random_seed(0)
|
||||
|
||||
device = torch.device(f"cuda:{local_rank}")
|
||||
torch.cuda.set_device(device)
|
||||
torch.accelerator.set_device_index(device)
|
||||
torch.set_default_device(device)
|
||||
torch.set_default_dtype(dtype)
|
||||
|
||||
|
||||
@@ -228,7 +228,7 @@ def sequence_parallelism_pass_on_test_model(
|
||||
set_random_seed(0)
|
||||
|
||||
device = torch.device(f"cuda:{local_rank}")
|
||||
torch.cuda.set_device(device)
|
||||
torch.accelerator.set_device_index(device)
|
||||
torch.set_default_device(device)
|
||||
torch.set_default_dtype(dtype)
|
||||
|
||||
|
||||
+2
-2
@@ -428,7 +428,7 @@ class HfRunner:
|
||||
)
|
||||
|
||||
# don't put this import at the top level
|
||||
# it will call torch.cuda.device_count()
|
||||
# it will call torch.accelerator.device_count()
|
||||
from transformers import AutoProcessor
|
||||
|
||||
self.processor = AutoProcessor.from_pretrained(
|
||||
@@ -1535,7 +1535,7 @@ def clean_gpu_memory_between_tests():
|
||||
|
||||
from tests.utils import wait_for_gpu_memory_to_clear
|
||||
|
||||
num_gpus = torch.cuda.device_count()
|
||||
num_gpus = torch.accelerator.device_count()
|
||||
if num_gpus > 0:
|
||||
try:
|
||||
wait_for_gpu_memory_to_clear(
|
||||
|
||||
@@ -14,7 +14,7 @@ import torch # noqa: E402
|
||||
from vllm.platforms import current_platform # noqa: F401, E402
|
||||
|
||||
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
|
||||
count = torch.cuda.device_count()
|
||||
count = torch.accelerator.device_count()
|
||||
|
||||
if count == 0:
|
||||
sys.exit(0) # Skip: no GPUs available
|
||||
|
||||
@@ -42,7 +42,7 @@ def set_env_vars_and_device(env: dict[str, str]) -> None:
|
||||
update_environment_variables(env)
|
||||
local_rank = os.environ["LOCAL_RANK"]
|
||||
device = torch.device(f"cuda:{local_rank}")
|
||||
torch.cuda.set_device(device)
|
||||
torch.accelerator.set_device_index(device)
|
||||
|
||||
# Create a minimal vllm config for init_distributed_environment
|
||||
vllm_config = VllmConfig()
|
||||
|
||||
@@ -43,7 +43,7 @@ def all_reduce_test_worker(
|
||||
monkeypatch.delenv("CUDA_VISIBLE_DEVICES", raising=False)
|
||||
|
||||
device = torch.device(f"cuda:{rank}")
|
||||
torch.cuda.set_device(device)
|
||||
torch.accelerator.set_device_index(device)
|
||||
init_test_distributed_environment(tp_size, pp_size, rank, distributed_init_port)
|
||||
num_elements = 8
|
||||
all_tensors = [
|
||||
@@ -69,7 +69,7 @@ def reduce_scatter_test_worker(
|
||||
# they will be able to set the device to the correct GPU
|
||||
monkeypatch.delenv("CUDA_VISIBLE_DEVICES", raising=False)
|
||||
device = torch.device(f"cuda:{rank}")
|
||||
torch.cuda.set_device(device)
|
||||
torch.accelerator.set_device_index(device)
|
||||
init_test_distributed_environment(tp_size, pp_size, rank, distributed_init_port)
|
||||
|
||||
num_elements = 8
|
||||
@@ -100,7 +100,7 @@ def all_gather_test_worker(
|
||||
# they will be able to set the device to the correct GPU
|
||||
monkeypatch.delenv("CUDA_VISIBLE_DEVICES", raising=False)
|
||||
device = torch.device(f"cuda:{rank}")
|
||||
torch.cuda.set_device(device)
|
||||
torch.accelerator.set_device_index(device)
|
||||
init_test_distributed_environment(tp_size, pp_size, rank, distributed_init_port)
|
||||
num_dimensions = 3
|
||||
tensor_size = list(range(2, num_dimensions + 2))
|
||||
@@ -134,7 +134,7 @@ def broadcast_tensor_dict_test_worker(
|
||||
# they will be able to set the device to the correct GPU
|
||||
monkeypatch.delenv("CUDA_VISIBLE_DEVICES", raising=False)
|
||||
device = torch.device(f"cuda:{rank}")
|
||||
torch.cuda.set_device(device)
|
||||
torch.accelerator.set_device_index(device)
|
||||
init_test_distributed_environment(tp_size, pp_size, rank, distributed_init_port)
|
||||
test_dict = {
|
||||
# device tensor
|
||||
@@ -171,7 +171,7 @@ def send_recv_tensor_dict_test_worker(
|
||||
):
|
||||
monkeypatch.delenv("CUDA_VISIBLE_DEVICES", raising=False)
|
||||
device = torch.device(f"cuda:{rank}")
|
||||
torch.cuda.set_device(device)
|
||||
torch.accelerator.set_device_index(device)
|
||||
init_test_distributed_environment(tp_size, pp_size, rank, distributed_init_port)
|
||||
|
||||
test_dict = {
|
||||
@@ -317,7 +317,7 @@ def send_recv_test_worker(
|
||||
):
|
||||
monkeypatch.delenv("CUDA_VISIBLE_DEVICES", raising=False)
|
||||
device = torch.device(f"cuda:{rank}")
|
||||
torch.cuda.set_device(device)
|
||||
torch.accelerator.set_device_index(device)
|
||||
init_test_distributed_environment(tp_size, pp_size, rank, distributed_init_port)
|
||||
|
||||
size = 64
|
||||
|
||||
@@ -35,7 +35,7 @@ def graph_allreduce(
|
||||
m.delenv("CUDA_VISIBLE_DEVICES", raising=False)
|
||||
m.delenv("HIP_VISIBLE_DEVICES", raising=False)
|
||||
device = torch.device(f"cuda:{rank}")
|
||||
torch.cuda.set_device(device)
|
||||
torch.accelerator.set_device_index(device)
|
||||
init_test_distributed_environment(tp_size, pp_size, rank, distributed_init_port)
|
||||
ensure_model_parallel_initialized(tp_size, pp_size)
|
||||
group = get_tp_group().device_group
|
||||
@@ -62,12 +62,10 @@ def graph_allreduce(
|
||||
for dtype in [torch.float32, torch.float16, torch.bfloat16]:
|
||||
with graph_capture(device=device) as graph_capture_context:
|
||||
# use integers so result matches NCCL exactly
|
||||
inp1 = torch.randint(
|
||||
1, 16, (sz,), dtype=dtype, device=torch.cuda.current_device()
|
||||
)
|
||||
inp2 = torch.randint(
|
||||
1, 16, (sz,), dtype=dtype, device=torch.cuda.current_device()
|
||||
)
|
||||
device_idx = torch.accelerator.current_device_index()
|
||||
inp1 = torch.randint(1, 16, (sz,), dtype=dtype, device=device_idx)
|
||||
inp2 = torch.randint(1, 16, (sz,), dtype=dtype, device=device_idx)
|
||||
|
||||
torch.accelerator.synchronize()
|
||||
graph = torch.cuda.CUDAGraph()
|
||||
with torch.cuda.graph(graph, stream=graph_capture_context.stream):
|
||||
@@ -95,7 +93,7 @@ def eager_allreduce(
|
||||
m.delenv("CUDA_VISIBLE_DEVICES", raising=False)
|
||||
m.delenv("HIP_VISIBLE_DEVICES", raising=False)
|
||||
device = torch.device(f"cuda:{rank}")
|
||||
torch.cuda.set_device(device)
|
||||
torch.accelerator.set_device_index(device)
|
||||
init_test_distributed_environment(tp_size, pp_size, rank, distributed_init_port)
|
||||
|
||||
# we use the first group to communicate once
|
||||
@@ -129,6 +127,6 @@ def test_custom_allreduce(
|
||||
test_target,
|
||||
):
|
||||
world_size = tp_size * pipeline_parallel_size
|
||||
if world_size > torch.cuda.device_count():
|
||||
if world_size > torch.accelerator.device_count():
|
||||
pytest.skip("Not enough GPUs to run the test.")
|
||||
multi_process_parallel(monkeypatch, tp_size, pipeline_parallel_size, test_target)
|
||||
|
||||
@@ -442,7 +442,7 @@ def test_rearrange_expert_weights_with_redundancy(
|
||||
):
|
||||
"""Test the functionality of rearranging expert weights with redundancy."""
|
||||
|
||||
if torch.cuda.device_count() < world_size:
|
||||
if torch.accelerator.device_count() < world_size:
|
||||
pytest.skip(f"Need at least {world_size} GPUs to run the test")
|
||||
distributed_run(
|
||||
_test_rearrange_expert_weights_with_redundancy,
|
||||
@@ -528,7 +528,7 @@ def test_async_transfer_layer_without_mtp(
|
||||
):
|
||||
"""Exercise async EPLB transfer path without MTP/spec decode."""
|
||||
|
||||
if torch.cuda.device_count() < world_size:
|
||||
if torch.accelerator.device_count() < world_size:
|
||||
pytest.skip(f"Need at least {world_size} GPUs to run the test")
|
||||
|
||||
distributed_run(
|
||||
@@ -547,7 +547,7 @@ def test_rearrange_expert_weights_no_change(world_size):
|
||||
unchanged.
|
||||
"""
|
||||
|
||||
if torch.cuda.device_count() < world_size:
|
||||
if torch.accelerator.device_count() < world_size:
|
||||
pytest.skip(f"Need at least {world_size} GPUs to run the test")
|
||||
distributed_run(_test_rearrange_expert_weights_no_change, world_size)
|
||||
|
||||
@@ -623,6 +623,6 @@ def _test_rearrange_expert_weights_profile_mode(env, world_size) -> None:
|
||||
def test_rearrange_expert_weights_profile_mode(world_size):
|
||||
"""Test profile mode (should not copy actual weights)"""
|
||||
|
||||
if torch.cuda.device_count() < world_size:
|
||||
if torch.accelerator.device_count() < world_size:
|
||||
pytest.skip(f"Need at least {world_size} GPUs to run the test")
|
||||
distributed_run(_test_rearrange_expert_weights_profile_mode, world_size)
|
||||
|
||||
@@ -257,7 +257,7 @@ def test_eplb_fml(
|
||||
intermediate_size: int,
|
||||
column_major_scales: bool,
|
||||
):
|
||||
if torch.cuda.device_count() < world_size:
|
||||
if torch.accelerator.device_count() < world_size:
|
||||
pytest.skip(f"Need at least {world_size} GPUs to run the test")
|
||||
|
||||
num_local_experts = num_experts // world_size
|
||||
|
||||
@@ -253,7 +253,7 @@ def test_eplb_fml(
|
||||
monkeypatch.setenv("VLLM_USE_FLASHINFER_MOE_FP4", "1")
|
||||
monkeypatch.setenv("VLLM_FLASHINFER_MOE_BACKEND", backend)
|
||||
|
||||
if torch.cuda.device_count() < world_size:
|
||||
if torch.accelerator.device_count() < world_size:
|
||||
pytest.skip(f"Need at least {world_size} GPUs to run the test")
|
||||
|
||||
num_local_experts = num_experts // world_size
|
||||
|
||||
@@ -38,7 +38,7 @@ def nccl_symm_mem_allreduce_worker(local_rank: int, world_size: int):
|
||||
m.delenv("CUDA_VISIBLE_DEVICES", raising=False)
|
||||
dtype = torch.bfloat16
|
||||
device = torch.device(f"cuda:{local_rank}")
|
||||
torch.cuda.set_device(device)
|
||||
torch.accelerator.set_device_index(device)
|
||||
torch.set_default_device(device)
|
||||
torch.set_default_dtype(dtype)
|
||||
update_environment_variables(
|
||||
@@ -84,7 +84,7 @@ def nccl_symm_mem_allreduce_worker(local_rank: int, world_size: int):
|
||||
@pytest.mark.parametrize("world_size", [2])
|
||||
@pytest.mark.skipif(envs.VLLM_TARGET_DEVICE not in ["cuda"], reason="Only test on CUDA")
|
||||
def test_nccl_symm_mem_allreduce(monkeypatch: pytest.MonkeyPatch, world_size):
|
||||
if world_size > torch.cuda.device_count():
|
||||
if world_size > torch.accelerator.device_count():
|
||||
pytest.skip("Not enough GPUs to run the test.")
|
||||
|
||||
# Enable SymmMemCommunicator
|
||||
|
||||
@@ -54,7 +54,7 @@ def worker_fn_wrapper(fn):
|
||||
update_environment_variables(env)
|
||||
local_rank = os.environ["LOCAL_RANK"]
|
||||
device = torch.device(f"cuda:{local_rank}")
|
||||
torch.cuda.set_device(device)
|
||||
torch.accelerator.set_device_index(device)
|
||||
init_distributed_environment()
|
||||
fn()
|
||||
|
||||
@@ -73,7 +73,7 @@ def worker_fn():
|
||||
|
||||
|
||||
@pytest.mark.skipif(
|
||||
torch.cuda.device_count() < 2, reason="Need at least 2 GPUs to run the test."
|
||||
torch.accelerator.device_count() < 2, reason="Need at least 2 GPUs to run the test."
|
||||
)
|
||||
def test_pynccl():
|
||||
distributed_run(worker_fn, 2)
|
||||
@@ -102,7 +102,7 @@ def multiple_allreduce_worker_fn():
|
||||
|
||||
|
||||
@pytest.mark.skipif(
|
||||
torch.cuda.device_count() < 4, reason="Need at least 4 GPUs to run the test."
|
||||
torch.accelerator.device_count() < 4, reason="Need at least 4 GPUs to run the test."
|
||||
)
|
||||
def test_pynccl_multiple_allreduce():
|
||||
# this tests pynccl for multiple tp groups, in a standalone way
|
||||
@@ -130,7 +130,7 @@ def multiple_allreduce_with_vllm_worker_fn():
|
||||
|
||||
|
||||
@pytest.mark.skipif(
|
||||
torch.cuda.device_count() < 4, reason="Need at least 4 GPUs to run the test."
|
||||
torch.accelerator.device_count() < 4, reason="Need at least 4 GPUs to run the test."
|
||||
)
|
||||
def test_pynccl_multiple_allreduce_with_vllm():
|
||||
# this tests pynccl for multiple tp groups, together with vllm
|
||||
@@ -185,7 +185,7 @@ def all_gather_worker_fn():
|
||||
|
||||
|
||||
@pytest.mark.skipif(
|
||||
torch.cuda.device_count() < 2, reason="Need at least 2 GPUs to run the test."
|
||||
torch.accelerator.device_count() < 2, reason="Need at least 2 GPUs to run the test."
|
||||
)
|
||||
def test_pynccl_all_gather():
|
||||
distributed_run(all_gather_worker_fn, 2)
|
||||
@@ -220,7 +220,7 @@ def all_gatherv_worker_fn():
|
||||
|
||||
|
||||
@pytest.mark.skipif(
|
||||
torch.cuda.device_count() < 2, reason="Need at least 2 GPUs to run the test."
|
||||
torch.accelerator.device_count() < 2, reason="Need at least 2 GPUs to run the test."
|
||||
)
|
||||
def test_pynccl_all_gatherv():
|
||||
distributed_run(all_gatherv_worker_fn, 2)
|
||||
@@ -260,7 +260,7 @@ def reduce_scatter_worker_fn():
|
||||
|
||||
|
||||
@pytest.mark.skipif(
|
||||
torch.cuda.device_count() < 2, reason="Need at least 2 GPUs to run the test."
|
||||
torch.accelerator.device_count() < 2, reason="Need at least 2 GPUs to run the test."
|
||||
)
|
||||
def test_pynccl_reduce_scatter():
|
||||
distributed_run(reduce_scatter_worker_fn, 2)
|
||||
@@ -298,14 +298,14 @@ def reduce_scatterv_worker_fn():
|
||||
|
||||
|
||||
@pytest.mark.skipif(
|
||||
torch.cuda.device_count() < 2, reason="Need at least 2 GPUs to run the test."
|
||||
torch.accelerator.device_count() < 2, reason="Need at least 2 GPUs to run the test."
|
||||
)
|
||||
def test_pynccl_reduce_scatterv():
|
||||
distributed_run(reduce_scatterv_worker_fn, 2)
|
||||
|
||||
|
||||
@pytest.mark.skipif(
|
||||
torch.cuda.device_count() < 2, reason="Need at least 2 GPUs to run the test."
|
||||
torch.accelerator.device_count() < 2, reason="Need at least 2 GPUs to run the test."
|
||||
)
|
||||
def test_pynccl_with_cudagraph():
|
||||
distributed_run(worker_fn_with_cudagraph, 2)
|
||||
@@ -330,7 +330,7 @@ def send_recv_worker_fn():
|
||||
|
||||
|
||||
@pytest.mark.skipif(
|
||||
torch.cuda.device_count() < 2, reason="Need at least 2 GPUs to run the test."
|
||||
torch.accelerator.device_count() < 2, reason="Need at least 2 GPUs to run the test."
|
||||
)
|
||||
def test_pynccl_send_recv():
|
||||
distributed_run(send_recv_worker_fn, 2)
|
||||
@@ -363,14 +363,14 @@ def multiple_send_recv_worker_fn():
|
||||
|
||||
|
||||
@pytest.mark.skipif(
|
||||
torch.cuda.device_count() < 4, reason="Need at least 4 GPUs to run the test."
|
||||
torch.accelerator.device_count() < 4, reason="Need at least 4 GPUs to run the test."
|
||||
)
|
||||
def test_pynccl_multiple_send_recv():
|
||||
distributed_run(multiple_send_recv_worker_fn, 4)
|
||||
|
||||
|
||||
@pytest.mark.skipif(
|
||||
torch.cuda.device_count() < 4, reason="Need at least 4 GPUs to run the test."
|
||||
torch.accelerator.device_count() < 4, reason="Need at least 4 GPUs to run the test."
|
||||
)
|
||||
def test_pynccl_broadcast():
|
||||
distributed_run(broadcast_worker_fn, 4)
|
||||
|
||||
@@ -39,7 +39,7 @@ def graph_quickreduce(
|
||||
with monkeypatch.context() as m:
|
||||
m.delenv("CUDA_VISIBLE_DEVICES", raising=False)
|
||||
device = torch.device(f"cuda:{rank}")
|
||||
torch.cuda.set_device(device)
|
||||
torch.accelerator.set_device_index(device)
|
||||
init_test_distributed_environment(tp_size, pp_size, rank, distributed_init_port)
|
||||
ensure_model_parallel_initialized(tp_size, pp_size)
|
||||
group = get_tp_group().device_group
|
||||
@@ -65,12 +65,10 @@ def graph_quickreduce(
|
||||
for sz in test_sizes:
|
||||
for dtype in [torch.float16, torch.bfloat16]:
|
||||
with graph_capture(device=device) as graph_capture_context:
|
||||
inp1 = torch.randint(
|
||||
1, 23, (sz,), dtype=dtype, device=torch.cuda.current_device()
|
||||
)
|
||||
inp2 = torch.randint(
|
||||
-23, 1, (sz,), dtype=dtype, device=torch.cuda.current_device()
|
||||
)
|
||||
device_idx = torch.accelerator.current_device_index()
|
||||
inp1 = torch.randint(1, 23, (sz,), dtype=dtype, device=device_idx)
|
||||
inp2 = torch.randint(-23, 1, (sz,), dtype=dtype, device=device_idx)
|
||||
|
||||
torch.accelerator.synchronize()
|
||||
graph = torch.cuda.CUDAGraph()
|
||||
with torch.cuda.graph(graph, stream=graph_capture_context.stream):
|
||||
@@ -95,7 +93,7 @@ def eager_quickreduce(
|
||||
with monkeypatch.context() as m:
|
||||
m.delenv("CUDA_VISIBLE_DEVICES", raising=False)
|
||||
device = torch.device(f"cuda:{rank}")
|
||||
torch.cuda.set_device(device)
|
||||
torch.accelerator.set_device_index(device)
|
||||
|
||||
init_test_distributed_environment(tp_size, pp_size, rank, distributed_init_port)
|
||||
|
||||
@@ -130,7 +128,7 @@ def test_custom_quick_allreduce(
|
||||
quant_mode,
|
||||
):
|
||||
world_size = tp_size * pipeline_parallel_size
|
||||
if world_size > torch.cuda.device_count():
|
||||
if world_size > torch.accelerator.device_count():
|
||||
pytest.skip("Not enough GPUs to run the test.")
|
||||
|
||||
monkeypatch.setenv("VLLM_ROCM_QUICK_REDUCE_QUANTIZATION", quant_mode)
|
||||
@@ -145,7 +143,7 @@ def qr_variable_input(rank, world_size):
|
||||
has been observed with the gpt_oss model).
|
||||
"""
|
||||
device = torch.device(f"cuda:{rank}")
|
||||
torch.cuda.set_device(device)
|
||||
torch.accelerator.set_device_index(device)
|
||||
qr_max_size = None # MB
|
||||
_ptr = ops.init_custom_qr(rank, world_size, qr_max_size)
|
||||
ranks = []
|
||||
@@ -169,14 +167,13 @@ def qr_variable_input(rank, world_size):
|
||||
s1 = 1024
|
||||
while num < 50000: # 50000 is sufficient to identify issues.
|
||||
dtype = torch.float16
|
||||
device_idx = torch.accelerator.current_device_index()
|
||||
if num % 2 == 0:
|
||||
s2 = 1024
|
||||
inp1 = torch.zeros(
|
||||
(s1, s2), dtype=dtype, device=torch.cuda.current_device()
|
||||
)
|
||||
inp1 = torch.zeros((s1, s2), dtype=dtype, device=device_idx)
|
||||
else:
|
||||
s2 = 2048
|
||||
inp1 = torch.ones((s1, s2), dtype=dtype, device=torch.cuda.current_device())
|
||||
inp1 = torch.ones((s1, s2), dtype=dtype, device=device_idx)
|
||||
result = torch.empty_like(inp1)
|
||||
# FP = 0 INT8 = 1 INT6 = 2 INT4 = 3 NONE = 4
|
||||
ops.qr_all_reduce(_ptr, inp1, result, 3, cast_bf2half=True)
|
||||
@@ -198,7 +195,7 @@ def qr_variable_input(rank, world_size):
|
||||
@pytest.mark.parametrize("pipeline_parallel_size", [1])
|
||||
def test_custom_quick_allreduce_variable_input(tp_size, pipeline_parallel_size):
|
||||
world_size = tp_size * pipeline_parallel_size
|
||||
if world_size > torch.cuda.device_count():
|
||||
if world_size > torch.accelerator.device_count():
|
||||
pytest.skip("Not enough GPUs to run the test.")
|
||||
|
||||
multiprocessing.set_start_method("spawn", force=True)
|
||||
|
||||
@@ -39,7 +39,7 @@ def symm_mem_allreduce_worker(local_rank: int, world_size: int, q: mp.Queue):
|
||||
m.delenv("CUDA_VISIBLE_DEVICES", raising=False)
|
||||
dtype = torch.bfloat16
|
||||
device = torch.device(f"cuda:{local_rank}")
|
||||
torch.cuda.set_device(device)
|
||||
torch.accelerator.set_device_index(device)
|
||||
torch.set_default_device(device)
|
||||
torch.set_default_dtype(dtype)
|
||||
update_environment_variables(
|
||||
@@ -105,7 +105,7 @@ def test_symm_mem_allreduce(
|
||||
monkeypatch: pytest.MonkeyPatch, tp_size, pipeline_parallel_size
|
||||
):
|
||||
world_size = tp_size * pipeline_parallel_size
|
||||
if world_size > torch.cuda.device_count():
|
||||
if world_size > torch.accelerator.device_count():
|
||||
pytest.skip("Not enough GPUs to run the test.")
|
||||
q = mp.get_context("spawn").Queue()
|
||||
mp.spawn(symm_mem_allreduce_worker, args=(world_size, q), nprocs=world_size)
|
||||
@@ -126,7 +126,7 @@ def test_symm_mem_allreduce(
|
||||
@pytest.mark.skipif(envs.VLLM_TARGET_DEVICE not in ["cuda"], reason="Only test on CUDA")
|
||||
def test_dp_with_symm_mem_allreduce(monkeypatch: pytest.MonkeyPatch):
|
||||
world_size = 4
|
||||
if world_size > torch.cuda.device_count():
|
||||
if world_size > torch.accelerator.device_count():
|
||||
pytest.skip("Not enough GPUs to run the test.")
|
||||
# Verify that the DataParallel runs without error
|
||||
engine_args = EngineArgs(
|
||||
|
||||
@@ -66,7 +66,7 @@ def cpu_worker(rank, WORLD_SIZE, port1, port2):
|
||||
|
||||
|
||||
def gpu_worker(rank, WORLD_SIZE, port1, port2):
|
||||
torch.cuda.set_device(rank)
|
||||
torch.accelerator.set_device_index(rank)
|
||||
pg1 = StatelessProcessGroup.create(
|
||||
host="127.0.0.1", port=port1, rank=rank, world_size=WORLD_SIZE
|
||||
)
|
||||
|
||||
@@ -203,7 +203,7 @@ class TestEngineRegistry:
|
||||
|
||||
def test_nccl_receive_weights_without_init_raises():
|
||||
"""Test that receive_weights raises if init_transfer_engine wasn't called."""
|
||||
if torch.cuda.device_count() < 1:
|
||||
if torch.accelerator.device_count() < 1:
|
||||
pytest.skip("Need at least 1 GPU for this test")
|
||||
|
||||
config = WeightTransferConfig(backend="nccl")
|
||||
@@ -336,7 +336,7 @@ def inference_receive_tensor(
|
||||
|
||||
|
||||
@pytest.mark.skipif(
|
||||
torch.cuda.device_count() < 2,
|
||||
torch.accelerator.device_count() < 2,
|
||||
reason="Need at least 2 GPUs to run NCCL weight transfer test.",
|
||||
)
|
||||
def test_nccl_weight_transfer_between_processes():
|
||||
@@ -382,7 +382,7 @@ class TestIPCWeightTransferUpdateInfoValidation:
|
||||
|
||||
def test_valid_update_info(self):
|
||||
"""Test creating valid IPCWeightTransferUpdateInfo."""
|
||||
if torch.cuda.device_count() < 1:
|
||||
if torch.accelerator.device_count() < 1:
|
||||
pytest.skip("Need at least 1 GPU for this test")
|
||||
|
||||
# Create a dummy tensor and IPC handle
|
||||
@@ -404,7 +404,7 @@ class TestIPCWeightTransferUpdateInfoValidation:
|
||||
|
||||
def test_mismatched_dtype_names_raises(self):
|
||||
"""Test that mismatched dtype_names length raises ValueError."""
|
||||
if torch.cuda.device_count() < 1:
|
||||
if torch.accelerator.device_count() < 1:
|
||||
pytest.skip("Need at least 1 GPU for this test")
|
||||
|
||||
dummy_tensor = torch.ones(10, 10, device="cuda:0")
|
||||
@@ -422,7 +422,7 @@ class TestIPCWeightTransferUpdateInfoValidation:
|
||||
|
||||
def test_mismatched_shapes_raises(self):
|
||||
"""Test that mismatched shapes length raises ValueError."""
|
||||
if torch.cuda.device_count() < 1:
|
||||
if torch.accelerator.device_count() < 1:
|
||||
pytest.skip("Need at least 1 GPU for this test")
|
||||
|
||||
dummy_tensor = torch.ones(10, 10, device="cuda:0")
|
||||
@@ -440,7 +440,7 @@ class TestIPCWeightTransferUpdateInfoValidation:
|
||||
|
||||
def test_mismatched_ipc_handles_raises(self):
|
||||
"""Test that mismatched ipc_handles length raises ValueError."""
|
||||
if torch.cuda.device_count() < 1:
|
||||
if torch.accelerator.device_count() < 1:
|
||||
pytest.skip("Need at least 1 GPU for this test")
|
||||
|
||||
dummy_tensor = torch.ones(10, 10, device="cuda:0")
|
||||
@@ -458,7 +458,7 @@ class TestIPCWeightTransferUpdateInfoValidation:
|
||||
|
||||
def test_valid_update_info_from_pickled(self, monkeypatch):
|
||||
"""Test creating IPCWeightTransferUpdateInfo from pickled handles."""
|
||||
if torch.cuda.device_count() < 1:
|
||||
if torch.accelerator.device_count() < 1:
|
||||
pytest.skip("Need at least 1 GPU for this test")
|
||||
|
||||
monkeypatch.setenv("VLLM_ALLOW_INSECURE_SERIALIZATION", "1")
|
||||
@@ -493,7 +493,7 @@ class TestIPCWeightTransferUpdateInfoValidation:
|
||||
|
||||
def test_both_handles_and_pickled_raises(self):
|
||||
"""Test that providing both ipc_handles and ipc_handles_pickled raises."""
|
||||
if torch.cuda.device_count() < 1:
|
||||
if torch.accelerator.device_count() < 1:
|
||||
pytest.skip("Need at least 1 GPU for this test")
|
||||
|
||||
dummy_tensor = torch.ones(10, 10, device="cuda:0")
|
||||
@@ -540,7 +540,7 @@ class TestIPCEngineParsing:
|
||||
|
||||
def test_parse_update_info_valid(self):
|
||||
"""Test parsing valid update info dict."""
|
||||
if torch.cuda.device_count() < 1:
|
||||
if torch.accelerator.device_count() < 1:
|
||||
pytest.skip("Need at least 1 GPU for this test")
|
||||
|
||||
config = WeightTransferConfig(backend="ipc")
|
||||
@@ -572,7 +572,7 @@ class TestIPCEngineParsing:
|
||||
|
||||
def test_parse_update_info_pickled(self, monkeypatch):
|
||||
"""Test parsing update info with pickled IPC handles (HTTP path)."""
|
||||
if torch.cuda.device_count() < 1:
|
||||
if torch.accelerator.device_count() < 1:
|
||||
pytest.skip("Need at least 1 GPU for this test")
|
||||
|
||||
monkeypatch.setenv("VLLM_ALLOW_INSECURE_SERIALIZATION", "1")
|
||||
@@ -731,7 +731,7 @@ def inference_receive_ipc_tensor(
|
||||
|
||||
|
||||
@pytest.mark.skipif(
|
||||
torch.cuda.device_count() < 1,
|
||||
torch.accelerator.device_count() < 1,
|
||||
reason="Need at least 1 GPU to run IPC weight transfer test.",
|
||||
)
|
||||
@pytest.mark.parametrize("mode", ["ray", "http"])
|
||||
@@ -789,7 +789,7 @@ def test_ipc_weight_transfer_between_processes(mode: str):
|
||||
|
||||
def test_ipc_receive_weights_missing_gpu_uuid_raises():
|
||||
"""Test that receive_weights raises if GPU UUID not found in IPC handles."""
|
||||
if torch.cuda.device_count() < 1:
|
||||
if torch.accelerator.device_count() < 1:
|
||||
pytest.skip("Need at least 1 GPU for this test")
|
||||
|
||||
config = WeightTransferConfig(backend="ipc")
|
||||
|
||||
@@ -13,7 +13,7 @@ from ...utils import create_new_process_for_each_test
|
||||
@pytest.mark.parametrize("backend", ["mp", "ray"])
|
||||
@create_new_process_for_each_test()
|
||||
def test_collective_rpc(tp_size, backend, monkeypatch):
|
||||
if torch.cuda.device_count() < tp_size:
|
||||
if torch.accelerator.device_count() < tp_size:
|
||||
pytest.skip(f"Not enough GPUs for tensor parallelism {tp_size}")
|
||||
if tp_size == 1 and backend == "ray":
|
||||
pytest.skip("Skip duplicate test case")
|
||||
|
||||
@@ -106,7 +106,7 @@ def mock_create_engine(config, parallel_config):
|
||||
@create_new_process_for_each_test()
|
||||
def test_get_world_size_tp1():
|
||||
"""Test world_size is correctly configured for TP=1."""
|
||||
if torch.cuda.device_count() < 1:
|
||||
if torch.accelerator.device_count() < 1:
|
||||
pytest.skip("Need at least 1 GPU for this test")
|
||||
|
||||
llm = LLM(
|
||||
@@ -125,7 +125,7 @@ def test_get_world_size_tp1():
|
||||
def test_init_weight_transfer_engine_calls_engine():
|
||||
"""Test that init_weight_transfer_engine calls the engine's
|
||||
init_transfer_engine method."""
|
||||
if torch.cuda.device_count() < 1:
|
||||
if torch.accelerator.device_count() < 1:
|
||||
pytest.skip("Need at least 1 GPU for this test")
|
||||
|
||||
# Run in-process so mock.patch works (spawn won't inherit the mock)
|
||||
@@ -174,7 +174,7 @@ def test_init_weight_transfer_engine_calls_engine():
|
||||
@create_new_process_for_each_test()
|
||||
def test_update_weights_calls_engine():
|
||||
"""Test that update_weights calls the engine's receive_weights method."""
|
||||
if torch.cuda.device_count() < 1:
|
||||
if torch.accelerator.device_count() < 1:
|
||||
pytest.skip("Need at least 1 GPU for this test")
|
||||
|
||||
# Run in-process so mock.patch works (spawn won't inherit the mock)
|
||||
@@ -233,7 +233,7 @@ def test_update_weights_calls_engine():
|
||||
@create_new_process_for_each_test()
|
||||
def test_full_weight_transfer_flow():
|
||||
"""Test the complete weight transfer flow: init -> update."""
|
||||
if torch.cuda.device_count() < 1:
|
||||
if torch.accelerator.device_count() < 1:
|
||||
pytest.skip("Need at least 1 GPU for this test")
|
||||
|
||||
# Run in-process so mock.patch works (spawn won't inherit the mock)
|
||||
@@ -294,7 +294,7 @@ def test_full_weight_transfer_flow():
|
||||
@create_new_process_for_each_test()
|
||||
def test_weight_transfer_config_backend():
|
||||
"""Test that WeightTransferConfig backend is properly configured."""
|
||||
if torch.cuda.device_count() < 1:
|
||||
if torch.accelerator.device_count() < 1:
|
||||
pytest.skip("Need at least 1 GPU for this test")
|
||||
|
||||
# Test with nccl backend
|
||||
|
||||
@@ -36,7 +36,9 @@ BLOCK_SIZES = [16, 32]
|
||||
USE_ALIBI = [False, True]
|
||||
KV_CACHE_DTYPE = ["auto", "fp8"]
|
||||
SEEDS = [0]
|
||||
CUDA_DEVICES = [f"cuda:{i}" for i in range(1 if torch.cuda.device_count() == 1 else 2)]
|
||||
CUDA_DEVICES = [
|
||||
f"cuda:{i}" for i in range(1 if torch.accelerator.device_count() == 1 else 2)
|
||||
]
|
||||
|
||||
|
||||
def ref_masked_attention(
|
||||
|
||||
@@ -35,7 +35,9 @@ NUM_BLOCKS = [1024, 10000]
|
||||
|
||||
NUM_MAPPINGS = [256] # Arbitrary values for testing
|
||||
SEEDS = [0]
|
||||
CUDA_DEVICES = [f"cuda:{i}" for i in range(1 if torch.cuda.device_count() == 1 else 2)]
|
||||
CUDA_DEVICES = [
|
||||
f"cuda:{i}" for i in range(1 if torch.accelerator.device_count() == 1 else 2)
|
||||
]
|
||||
|
||||
# We assume fp8 is always enabled for testing.
|
||||
KV_CACHE_DTYPE = ["auto", "fp8"]
|
||||
@@ -69,7 +71,7 @@ def test_reshape_and_cache(
|
||||
pytest.skip()
|
||||
set_random_seed(seed)
|
||||
torch.set_default_device(device)
|
||||
torch.cuda.set_device(device)
|
||||
torch.accelerator.set_device_index(device)
|
||||
# Create a random slot mapping.
|
||||
num_slots = block_size * num_blocks
|
||||
slot_mapping_lst = random.sample(range(num_slots), num_tokens)
|
||||
@@ -192,7 +194,7 @@ def test_reshape_and_cache_flash(
|
||||
) -> None:
|
||||
set_random_seed(seed)
|
||||
torch.set_default_device(device)
|
||||
torch.cuda.set_device(device)
|
||||
torch.accelerator.set_device_index(device)
|
||||
assert implementation in ["cuda", "triton"]
|
||||
if implementation == "triton" and kv_cache_layout == "HND":
|
||||
pytest.skip("Triton implementation only supports NHD layout.")
|
||||
@@ -553,7 +555,7 @@ def test_concat_and_cache_mla(
|
||||
) -> None:
|
||||
set_random_seed(seed)
|
||||
torch.set_default_device(device)
|
||||
torch.cuda.set_device(device)
|
||||
torch.accelerator.set_device_index(device)
|
||||
|
||||
total_slots = num_blocks * block_size
|
||||
slot_mapping_lst = random.sample(range(total_slots), num_tokens)
|
||||
@@ -632,7 +634,7 @@ def test_concat_and_cache_ds_mla(
|
||||
kv_cache_dtype = "fp8_ds_mla"
|
||||
set_random_seed(seed)
|
||||
torch.set_default_device(device)
|
||||
torch.cuda.set_device(device)
|
||||
torch.accelerator.set_device_index(device)
|
||||
|
||||
total_slots = num_blocks * block_size
|
||||
slot_mapping_lst = random.sample(range(total_slots), num_tokens)
|
||||
@@ -744,7 +746,7 @@ def test_swap_blocks_mla(
|
||||
) -> None:
|
||||
set_random_seed(seed)
|
||||
torch.set_default_device(device)
|
||||
torch.cuda.set_device(device)
|
||||
torch.accelerator.set_device_index(device)
|
||||
|
||||
entry_size = kv_lora_rank + qk_rope_head_dim
|
||||
|
||||
|
||||
@@ -69,7 +69,7 @@ def test_cutlass_mla_decode(
|
||||
init_dtype = torch.bfloat16 if torch_dtype == torch.float8_e4m3fn else torch_dtype
|
||||
torch.set_default_dtype(init_dtype)
|
||||
torch.set_default_device(device)
|
||||
torch.cuda.set_device(device)
|
||||
torch.accelerator.set_device_index(device)
|
||||
torch.manual_seed(42)
|
||||
random.seed(42)
|
||||
|
||||
|
||||
@@ -57,7 +57,7 @@ def test_flash_mla(
|
||||
init_dtype = torch.bfloat16 if torch_dtype == torch.float8_e4m3fn else torch_dtype
|
||||
torch.set_default_dtype(init_dtype)
|
||||
torch.set_default_device(device)
|
||||
torch.cuda.set_device(device)
|
||||
torch.accelerator.set_device_index(device)
|
||||
torch.manual_seed(0)
|
||||
random.seed(0)
|
||||
|
||||
|
||||
@@ -21,7 +21,9 @@ NUM_HEADS = [64]
|
||||
NUM_QUERIES_PER_KV = [1, 64]
|
||||
HEAD_SIZES = [24, 128]
|
||||
DTYPES = [torch.float16]
|
||||
CUDA_DEVICES = [f"cuda:{i}" for i in range(1 if torch.cuda.device_count() == 1 else 2)]
|
||||
CUDA_DEVICES = [
|
||||
f"cuda:{i}" for i in range(1 if torch.accelerator.device_count() == 1 else 2)
|
||||
]
|
||||
SLIDING_WINDOW = [0, 16, 2048]
|
||||
KV_CACHE_DTYPES = ["auto", "fp8", "fp8_e5m2"]
|
||||
|
||||
@@ -135,7 +137,7 @@ def test_contexted_kv_attention(
|
||||
# for GPU 1 would run on both GPU0 and GPU1 and things would hang
|
||||
#
|
||||
# see also similar issue: https://github.com/Dao-AILab/flash-attention/issues/523
|
||||
torch.cuda.set_device(device)
|
||||
torch.accelerator.set_device_index(device)
|
||||
|
||||
MAX_SEQ_LEN = 1024
|
||||
MAX_CTX_LEN = 1024
|
||||
@@ -356,7 +358,7 @@ def test_contexted_kv_attention_alibi(
|
||||
# for GPU 1 would run on both GPU0 and GPU1 and things would hang
|
||||
#
|
||||
# see also similar issue: https://github.com/Dao-AILab/flash-attention/issues/523
|
||||
torch.cuda.set_device(device)
|
||||
torch.accelerator.set_device_index(device)
|
||||
|
||||
def _get_alibi_slopes(total_num_heads: int) -> torch.Tensor:
|
||||
# Fork from: vllm/vllm/model_executor/models/bloom.py#L44
|
||||
|
||||
@@ -26,7 +26,9 @@ DTYPES = [torch.half, torch.bfloat16, torch.float]
|
||||
NUM_TOKENS = [7, 83, 2048] # Arbitrary values for testing
|
||||
D = [512, 13824] # Arbitrary values for testing
|
||||
SEEDS = [0]
|
||||
CUDA_DEVICES = [f"cuda:{i}" for i in range(1 if torch.cuda.device_count() == 1 else 2)]
|
||||
CUDA_DEVICES = [
|
||||
f"cuda:{i}" for i in range(1 if torch.accelerator.device_count() == 1 else 2)
|
||||
]
|
||||
|
||||
|
||||
@pytest.mark.parametrize(
|
||||
|
||||
@@ -33,7 +33,9 @@ SCALE_UBS = [True, False]
|
||||
GROUP_SIZES = [None, [1, 64], [1, 128]]
|
||||
TMA_ALIGNMENTS = [0, 4]
|
||||
SEEDS = [0]
|
||||
CUDA_DEVICES = [f"cuda:{i}" for i in range(1 if torch.cuda.device_count() == 1 else 2)]
|
||||
CUDA_DEVICES = [
|
||||
f"cuda:{i}" for i in range(1 if torch.accelerator.device_count() == 1 else 2)
|
||||
]
|
||||
|
||||
EPS = 1e-6
|
||||
|
||||
@@ -182,7 +184,7 @@ def test_rms_norm(
|
||||
if torch.cuda.is_available():
|
||||
torch.cuda.manual_seed(seed)
|
||||
torch.set_default_device(device)
|
||||
torch.cuda.set_device(device)
|
||||
torch.accelerator.set_device_index(device)
|
||||
|
||||
if group_size is not None and hidden_size % group_size[1] != 0:
|
||||
# skip
|
||||
|
||||
@@ -14,7 +14,9 @@ NUM_TOKENS = [7, 83, 4096] # Arbitrary values for testing
|
||||
HIDDEN_SIZES = [8, 768, 769, 5120, 5125, 8192] # Arbitrary values for testing
|
||||
ADD_RESIDUAL = [False, True]
|
||||
SEEDS = [0]
|
||||
CUDA_DEVICES = [f"cuda:{i}" for i in range(1 if torch.cuda.device_count() == 1 else 2)]
|
||||
CUDA_DEVICES = [
|
||||
f"cuda:{i}" for i in range(1 if torch.accelerator.device_count() == 1 else 2)
|
||||
]
|
||||
|
||||
|
||||
@pytest.mark.parametrize("num_tokens", NUM_TOKENS)
|
||||
|
||||
@@ -19,7 +19,9 @@ NUM_HEADS = [17] # Arbitrary values for testing
|
||||
BATCH_SIZES = [5] # Arbitrary values for testing
|
||||
SEQ_LENS = [11, 8192] # Arbitrary values for testing
|
||||
SEEDS = [0]
|
||||
CUDA_DEVICES = [f"cuda:{i}" for i in range(1 if torch.cuda.device_count() == 1 else 2)]
|
||||
CUDA_DEVICES = [
|
||||
f"cuda:{i}" for i in range(1 if torch.accelerator.device_count() == 1 else 2)
|
||||
]
|
||||
USE_KEY = [True, False]
|
||||
|
||||
|
||||
|
||||
@@ -28,7 +28,8 @@ from vllm.utils.torch_utils import set_random_seed
|
||||
@pytest.mark.parametrize("block_size", [16, 64, 256])
|
||||
@pytest.mark.parametrize("seed", [0])
|
||||
@pytest.mark.parametrize(
|
||||
"device", [f"cuda:{i}" for i in range(1 if torch.cuda.device_count() == 1 else 2)]
|
||||
"device",
|
||||
[f"cuda:{i}" for i in range(1 if torch.accelerator.device_count() == 1 else 2)],
|
||||
)
|
||||
@torch.inference_mode()
|
||||
def test_concat_and_cache_mla_rope_fused(
|
||||
|
||||
@@ -6,7 +6,9 @@ import torch
|
||||
from vllm.utils.platform_utils import is_uva_available
|
||||
from vllm.utils.torch_utils import get_accelerator_view_from_cpu_tensor
|
||||
|
||||
CUDA_DEVICES = [f"cuda:{i}" for i in range(1 if torch.cuda.device_count() == 1 else 2)]
|
||||
CUDA_DEVICES = [
|
||||
f"cuda:{i}" for i in range(1 if torch.accelerator.device_count() == 1 else 2)
|
||||
]
|
||||
|
||||
|
||||
@pytest.mark.skipif(not is_uva_available(), reason="UVA is not available.")
|
||||
|
||||
@@ -71,7 +71,7 @@ def mixer2_gated_norm_tensor_parallel(
|
||||
set_random_seed(0)
|
||||
|
||||
device = torch.device(f"cuda:{local_rank}")
|
||||
torch.cuda.set_device(device)
|
||||
torch.accelerator.set_device_index(device)
|
||||
torch.set_default_device(device)
|
||||
torch.set_default_dtype(dtype)
|
||||
|
||||
|
||||
@@ -322,7 +322,7 @@ class WeightTensors:
|
||||
)
|
||||
|
||||
def to_current_device(self):
|
||||
device = torch.cuda.current_device()
|
||||
device = torch.accelerator.current_device_index()
|
||||
self.w1 = self.w1.to(device=device)
|
||||
self.w2 = self.w2.to(device=device)
|
||||
|
||||
@@ -392,7 +392,8 @@ class RankTensors:
|
||||
Return hidden_states
|
||||
"""
|
||||
m, k, dtype = (config.M, config.K, config.dtype)
|
||||
a = torch.randn((m, k), device=torch.cuda.current_device(), dtype=dtype) / 15.0
|
||||
device = torch.accelerator.current_device_index()
|
||||
a = torch.randn((m, k), device=device, dtype=dtype) / 15.0
|
||||
|
||||
if config.quant_dtype is None:
|
||||
return a, None
|
||||
@@ -428,9 +429,10 @@ class RankTensors:
|
||||
topk_weights, topk_ids, _ = fused_topk(hidden_states, score, topk, False)
|
||||
|
||||
# distribute topk_ids evenly
|
||||
device = torch.accelerator.current_device_index()
|
||||
for mi in range(m):
|
||||
topk_ids[mi] = torch.randperm(config.E)[:topk]
|
||||
topk_ids = topk_ids.to(device=torch.cuda.current_device())
|
||||
topk_ids = topk_ids.to(device=device)
|
||||
|
||||
expert_map = None
|
||||
if config.world_size > 1 and config.supports_expert_map():
|
||||
@@ -440,9 +442,7 @@ class RankTensors:
|
||||
s = pgi.rank * num_local_experts
|
||||
e = s + num_local_experts
|
||||
expert_map[s:e] = torch.tensor(list(range(num_local_experts)))
|
||||
expert_map = expert_map.to(
|
||||
device=torch.cuda.current_device(), dtype=torch.int32
|
||||
)
|
||||
expert_map = expert_map.to(device=device, dtype=torch.int32)
|
||||
|
||||
return RankTensors(
|
||||
hidden_states=hidden_states,
|
||||
@@ -558,7 +558,9 @@ def reference_moe_impl(
|
||||
|
||||
def _make_gscale(num_experts: int) -> torch.Tensor:
|
||||
return torch.ones(
|
||||
(num_experts,), device=torch.cuda.current_device(), dtype=torch.float32
|
||||
(num_experts,),
|
||||
device=torch.accelerator.current_device_index(),
|
||||
dtype=torch.float32,
|
||||
)
|
||||
|
||||
|
||||
|
||||
@@ -66,7 +66,7 @@ def _worker_parallel_launch(
|
||||
**kwargs: P.kwargs,
|
||||
) -> None:
|
||||
rank = node_rank * world_local_size + local_rank
|
||||
torch.cuda.set_device(local_rank)
|
||||
torch.accelerator.set_device_index(local_rank)
|
||||
device = torch.device("cuda", local_rank)
|
||||
torch.distributed.init_process_group(
|
||||
backend="cpu:gloo,cuda:nccl",
|
||||
|
||||
@@ -34,7 +34,8 @@ def do_profile(
|
||||
record_shapes=True,
|
||||
) as tprof:
|
||||
fn(**fn_kwargs)
|
||||
torch.accelerator.synchronize(torch.cuda.current_device())
|
||||
device = torch.accelerator.current_device_index()
|
||||
torch.accelerator.synchronize(device=device)
|
||||
|
||||
# TODO (varun): Add a descriptive trace file name
|
||||
tprof.export_chrome_trace(
|
||||
|
||||
@@ -52,7 +52,7 @@ def _worker_parallel_launch(
|
||||
**kwargs: P.kwargs,
|
||||
) -> None:
|
||||
rank = node_rank * world_local_size + local_rank
|
||||
torch.cuda.set_device(local_rank)
|
||||
torch.accelerator.set_device_index(local_rank)
|
||||
device = torch.device("cuda", local_rank)
|
||||
torch.distributed.init_process_group(
|
||||
backend="cpu:gloo,cuda:nccl",
|
||||
|
||||
@@ -134,10 +134,8 @@ class TestTensors:
|
||||
|
||||
fp8_info = torch.finfo(torch.float8_e4m3fn)
|
||||
fp8_max, fp8_min = fp8_info.max, fp8_info.min
|
||||
|
||||
rank_tokens = (
|
||||
torch.randn((m, k), device=torch.cuda.current_device(), dtype=dtype) / 10.0
|
||||
)
|
||||
device = torch.accelerator.current_device_index()
|
||||
rank_tokens = torch.randn((m, k), device=device, dtype=dtype) / 10.0
|
||||
rank_tokens = rank_tokens.clamp(min=fp8_min, max=fp8_max)
|
||||
rank_token_scales = None
|
||||
|
||||
@@ -145,11 +143,13 @@ class TestTensors:
|
||||
low=0,
|
||||
high=config.num_experts,
|
||||
size=(m, topk),
|
||||
device=torch.cuda.current_device(),
|
||||
device=device,
|
||||
).to(dtype=torch.int64)
|
||||
|
||||
topk_weights = torch.randn(
|
||||
topk_ids.shape, dtype=torch.float32, device=torch.cuda.current_device()
|
||||
topk_ids.shape,
|
||||
dtype=torch.float32,
|
||||
device=device,
|
||||
)
|
||||
|
||||
return TestTensors(
|
||||
@@ -296,7 +296,8 @@ def deepep_deepgemm_moe_impl(
|
||||
s = pgi.rank * num_local_experts
|
||||
e = s + num_local_experts
|
||||
expert_map[s:e] = torch.tensor(list(range(num_local_experts)))
|
||||
return expert_map.to(device=torch.cuda.current_device(), dtype=torch.int32)
|
||||
device = torch.accelerator.current_device_index()
|
||||
return expert_map.to(device=device, dtype=torch.int32)
|
||||
|
||||
quant_config = fp8_w8a8_moe_quant_config(
|
||||
w1_scale=w1_scale,
|
||||
@@ -376,10 +377,11 @@ def _test_deepep_deepgemm_moe(
|
||||
|
||||
set_random_seed(pgi.rank)
|
||||
|
||||
w1 = w1.to(device=torch.cuda.current_device())
|
||||
w2 = w2.to(device=torch.cuda.current_device())
|
||||
w1_scale = w1_scale.to(device=torch.cuda.current_device())
|
||||
w2_scale = w2_scale.to(device=torch.cuda.current_device())
|
||||
device = torch.accelerator.current_device_index()
|
||||
w1 = w1.to(device=device)
|
||||
w2 = w2.to(device=device)
|
||||
w1_scale = w1_scale.to(device=device)
|
||||
w2_scale = w2_scale.to(device=device)
|
||||
|
||||
pg = torch.distributed.new_group(list(range(pgi.world_size)))
|
||||
test_tensors = TestTensors.make(config, pgi.rank)
|
||||
|
||||
@@ -210,7 +210,8 @@ def deep_ep_moe_impl(
|
||||
s = pgi.rank * num_local_experts
|
||||
e = s + num_local_experts
|
||||
expert_map[s:e] = torch.tensor(list(range(num_local_experts)))
|
||||
return expert_map.to(device=torch.cuda.current_device(), dtype=torch.int32)
|
||||
device = torch.accelerator.current_device_index()
|
||||
return expert_map.to(device=device, dtype=torch.int32)
|
||||
|
||||
hidden_size = test_tensors.rank_tokens.size(1)
|
||||
is_quantized = w1.dtype == torch.float8_e4m3fn
|
||||
@@ -365,15 +366,13 @@ def _deep_ep_moe(
|
||||
)
|
||||
|
||||
is_quantized = w1.dtype == torch.float8_e4m3fn
|
||||
w1 = w1.to(device=torch.cuda.current_device())
|
||||
w2 = w2.to(device=torch.cuda.current_device())
|
||||
device_idx = torch.accelerator.current_device_index()
|
||||
w1 = w1.to(device=device_idx)
|
||||
w2 = w2.to(device=device_idx)
|
||||
if is_quantized:
|
||||
w1_scale = w1_scale.to( # type: ignore
|
||||
device=torch.cuda.current_device()
|
||||
)
|
||||
w2_scale = w2_scale.to( # type: ignore
|
||||
device=torch.cuda.current_device()
|
||||
)
|
||||
assert w1_scale is not None and w2_scale is not None
|
||||
w1_scale = w1_scale.to(device=device_idx)
|
||||
w2_scale = w2_scale.to(device=device_idx)
|
||||
|
||||
pg = torch.distributed.new_group(list(range(pgi.world_size)))
|
||||
test_tensors = TestTensors.make(config, low_latency_mode)
|
||||
|
||||
@@ -716,7 +716,7 @@ def test_mixtral_moe(
|
||||
monkeypatch.setenv("MASTER_ADDR", "localhost")
|
||||
monkeypatch.setenv("MASTER_PORT", "12345")
|
||||
init_distributed_environment()
|
||||
init_workspace_manager(torch.cuda.current_device())
|
||||
init_workspace_manager(torch.accelerator.current_device_index())
|
||||
|
||||
# Instantiate our and huggingface's MoE blocks
|
||||
vllm_config.compilation_config.static_forward_context = dict()
|
||||
|
||||
@@ -71,10 +71,10 @@ def enable_pickle(monkeypatch):
|
||||
)
|
||||
@pytest.mark.skipif(not QUARK_MXFP4_AVAILABLE, reason="amd-quark>=0.9 is not available")
|
||||
def test_mxfp4_loading_and_execution_moe(vllm_runner, model_case: ModelCase):
|
||||
if torch.cuda.device_count() < model_case.tp:
|
||||
if torch.accelerator.device_count() < model_case.tp:
|
||||
pytest.skip(
|
||||
f"This test requires >={model_case.tp} gpus, got only "
|
||||
f"{torch.cuda.device_count()}"
|
||||
f"{torch.accelerator.device_count()}"
|
||||
)
|
||||
|
||||
# `cudagraph_capture_sizes=[16]` to reduce load time.
|
||||
|
||||
@@ -15,7 +15,9 @@ from vllm.model_executor.layers.quantization.utils.w8a8_utils import (
|
||||
)
|
||||
from vllm.platforms import current_platform
|
||||
|
||||
CUDA_DEVICES = [f"cuda:{i}" for i in range(1 if torch.cuda.device_count() == 1 else 2)]
|
||||
CUDA_DEVICES = [
|
||||
f"cuda:{i}" for i in range(1 if torch.accelerator.device_count() == 1 else 2)
|
||||
]
|
||||
|
||||
capability = current_platform.get_device_capability()
|
||||
capability = capability[0] * 10 + capability[1]
|
||||
|
||||
@@ -40,7 +40,9 @@ MNK_FACTORS = [
|
||||
(512, 24576, 128),
|
||||
]
|
||||
|
||||
CUDA_DEVICES = [f"cuda:{i}" for i in range(1 if torch.cuda.device_count() == 1 else 2)]
|
||||
CUDA_DEVICES = [
|
||||
f"cuda:{i}" for i in range(1 if torch.accelerator.device_count() == 1 else 2)
|
||||
]
|
||||
|
||||
# -1 means full extent in that dimension
|
||||
TENSORWISE_GROUP_SHAPE = (-1, -1)
|
||||
|
||||
@@ -29,7 +29,9 @@ if current_platform.is_rocm():
|
||||
allow_module_level=True,
|
||||
)
|
||||
|
||||
CUDA_DEVICES = [f"cuda:{i}" for i in range(1 if torch.cuda.device_count() == 1 else 2)]
|
||||
CUDA_DEVICES = [
|
||||
f"cuda:{i}" for i in range(1 if torch.accelerator.device_count() == 1 else 2)
|
||||
]
|
||||
|
||||
# TODO: in future PR refactor this and `is_quant_method_supported` in the kernel
|
||||
# unit tests to a common utility function. Currently the use of
|
||||
|
||||
@@ -13,7 +13,7 @@ except ImportError:
|
||||
)
|
||||
|
||||
|
||||
@pytest.mark.skipif(torch.cuda.device_count() < 1, reason="Need CUDA device")
|
||||
@pytest.mark.skipif(torch.accelerator.device_count() < 1, reason="Need CUDA device")
|
||||
def test_gather_cache_oob():
|
||||
"""
|
||||
Tests for OOB read in gather_and_maybe_dequant_cache (Issue #27909).
|
||||
|
||||
@@ -13,7 +13,9 @@ QUANT_DTYPES = [current_platform.fp8_dtype()]
|
||||
NUM_TOKENS = [1, 17, 86, 1234, 3045] # Arbitrary values for testing
|
||||
HIDDEN_SIZES = [16, 48, 128, 1562, 4096] # Arbitrary values for testing
|
||||
SEEDS = [0]
|
||||
CUDA_DEVICES = [f"cuda:{i}" for i in range(1 if torch.cuda.device_count() == 1 else 2)]
|
||||
CUDA_DEVICES = [
|
||||
f"cuda:{i}" for i in range(1 if torch.accelerator.device_count() == 1 else 2)
|
||||
]
|
||||
|
||||
|
||||
def ref_impl(
|
||||
|
||||
@@ -638,7 +638,7 @@ def use_fused_moe_lora_kernel_tensor_parallel(
|
||||
set_random_seed(seed)
|
||||
|
||||
device = torch.device(f"cuda:{local_rank}")
|
||||
torch.cuda.set_device(device)
|
||||
torch.accelerator.set_device_index(device)
|
||||
torch.set_default_device(device)
|
||||
torch.set_default_dtype(dtype)
|
||||
|
||||
|
||||
@@ -61,7 +61,7 @@ pytestmark = pytest.mark.skipif(
|
||||
)
|
||||
|
||||
DEVICES = (
|
||||
[f"cuda:{i}" for i in range(1 if torch.cuda.device_count() == 1 else 2)]
|
||||
[f"cuda:{i}" for i in range(1 if torch.accelerator.device_count() == 1 else 2)]
|
||||
if current_platform.is_cuda_alike()
|
||||
else ["cpu"]
|
||||
)
|
||||
@@ -260,7 +260,7 @@ def test_embeddings(
|
||||
# device, see: https://github.com/triton-lang/triton/issues/2925
|
||||
# Same below.
|
||||
if current_platform.is_cuda_alike():
|
||||
torch.cuda.set_device(device)
|
||||
torch.accelerator.set_device_index(device)
|
||||
|
||||
torch.set_default_device(device)
|
||||
max_loras = 8
|
||||
@@ -359,7 +359,7 @@ def test_lm_head_logits_processor(
|
||||
default_vllm_config, dist_init, num_loras, device, vocab_size, stage
|
||||
) -> None:
|
||||
if current_platform.is_cuda_alike():
|
||||
torch.cuda.set_device(device)
|
||||
torch.accelerator.set_device_index(device)
|
||||
|
||||
torch.set_default_device(device)
|
||||
max_loras = 8
|
||||
@@ -476,7 +476,7 @@ def test_lm_head_logits_processor_invalid_vocab_size(
|
||||
) -> None:
|
||||
"""Test that LogitsProcessorWithLoRA raises ValueError for invalid vocab sizes."""
|
||||
if current_platform.is_cuda_alike():
|
||||
torch.cuda.set_device(device)
|
||||
torch.accelerator.set_device_index(device)
|
||||
|
||||
torch.set_default_device(device)
|
||||
max_loras = 8
|
||||
@@ -505,7 +505,7 @@ def test_linear_replicated(
|
||||
stage,
|
||||
) -> None:
|
||||
if current_platform.is_cuda_alike():
|
||||
torch.cuda.set_device(device)
|
||||
torch.accelerator.set_device_index(device)
|
||||
|
||||
max_loras = 8
|
||||
torch.set_default_device(device)
|
||||
@@ -612,7 +612,7 @@ def test_linear_parallel(
|
||||
default_vllm_config, dist_init, num_loras, orientation, fully_shard, device, stage
|
||||
) -> None:
|
||||
if current_platform.is_cuda_alike():
|
||||
torch.cuda.set_device(device)
|
||||
torch.accelerator.set_device_index(device)
|
||||
|
||||
max_loras = 8
|
||||
torch.set_default_device(device)
|
||||
@@ -737,7 +737,7 @@ def test_column_parallel_packed(
|
||||
default_vllm_config, dist_init, num_loras, repeats, fully_shard, device, stage
|
||||
) -> None:
|
||||
if current_platform.is_cuda_alike():
|
||||
torch.cuda.set_device(device)
|
||||
torch.accelerator.set_device_index(device)
|
||||
|
||||
max_loras = 8
|
||||
torch.set_default_device(device)
|
||||
@@ -885,7 +885,7 @@ def test_merged_column_parallel_variable_slice(
|
||||
default_vllm_config, dist_init, num_loras, num_slices, device, stage
|
||||
) -> None:
|
||||
if current_platform.is_cuda_alike():
|
||||
torch.cuda.set_device(device)
|
||||
torch.accelerator.set_device_index(device)
|
||||
|
||||
max_loras = 8
|
||||
torch.set_default_device(device)
|
||||
|
||||
@@ -37,7 +37,7 @@ EMBEDDING_MODULES = {
|
||||
|
||||
|
||||
DEVICES = (
|
||||
[f"cuda:{i}" for i in range(1 if torch.cuda.device_count() == 1 else 2)]
|
||||
[f"cuda:{i}" for i in range(1 if torch.accelerator.device_count() == 1 else 2)]
|
||||
if current_platform.is_cuda_alike()
|
||||
else ["cpu"]
|
||||
)
|
||||
|
||||
@@ -34,7 +34,7 @@ def do_sample(
|
||||
def test_mixtral_lora(mixtral_lora_files, tp_size):
|
||||
"""Original test, the LoRA model has the common target modules, not all"""
|
||||
if (
|
||||
torch.cuda.device_count() < tp_size
|
||||
torch.accelerator.device_count() < tp_size
|
||||
and tp_size > 1
|
||||
and current_platform.is_cuda_alike()
|
||||
):
|
||||
|
||||
@@ -395,7 +395,7 @@ def test_kernels(
|
||||
Tests LoRA kernels.
|
||||
"""
|
||||
torch.set_default_device(device)
|
||||
torch.cuda.set_device(device)
|
||||
torch.accelerator.set_device_index(device)
|
||||
set_random_seed(seed)
|
||||
|
||||
if op_type == "shrink":
|
||||
@@ -448,7 +448,7 @@ def test_kernels_hidden_size(
|
||||
Tests SGMV and LoRA kernels.
|
||||
"""
|
||||
torch.set_default_device(device)
|
||||
torch.cuda.set_device(device)
|
||||
torch.accelerator.set_device_index(device)
|
||||
set_random_seed(seed)
|
||||
|
||||
if op_type == "shrink":
|
||||
|
||||
@@ -203,7 +203,7 @@ def test_raise_value_error_on_invalid_load_format(vllm_runner, capfd, model_ref)
|
||||
torch.accelerator.empty_cache()
|
||||
|
||||
|
||||
@pytest.mark.skipif(torch.cuda.device_count() < 2, reason="Requires 2 GPUs")
|
||||
@pytest.mark.skipif(torch.accelerator.device_count() < 2, reason="Requires 2 GPUs")
|
||||
def test_tensorizer_with_tp_path_without_template(vllm_runner, capfd):
|
||||
try:
|
||||
model_ref = "EleutherAI/pythia-1.4b"
|
||||
@@ -231,7 +231,7 @@ def test_tensorizer_with_tp_path_without_template(vllm_runner, capfd):
|
||||
) in combined_output
|
||||
|
||||
|
||||
@pytest.mark.skipif(torch.cuda.device_count() < 2, reason="Requires 2 GPUs")
|
||||
@pytest.mark.skipif(torch.accelerator.device_count() < 2, reason="Requires 2 GPUs")
|
||||
def test_deserialized_encrypted_vllm_model_with_tp_has_same_outputs(
|
||||
vllm_runner, tmp_path
|
||||
):
|
||||
|
||||
@@ -11,7 +11,7 @@ from vllm.model_executor.models.utils import get_draft_quant_config
|
||||
from vllm.platforms import current_platform
|
||||
|
||||
DEVICES = (
|
||||
[f"cuda:{i}" for i in range(1 if torch.cuda.device_count() == 1 else 2)]
|
||||
[f"cuda:{i}" for i in range(1 if torch.accelerator.device_count() == 1 else 2)]
|
||||
if current_platform.is_cuda_alike()
|
||||
else ["cpu"]
|
||||
)
|
||||
@@ -61,7 +61,7 @@ def test_fc_layer_quant_config_usage(default_vllm_config, dist_init, device) ->
|
||||
from vllm.model_executor.layers.linear import ReplicatedLinear
|
||||
|
||||
if current_platform.is_cuda_alike():
|
||||
torch.cuda.set_device(device)
|
||||
torch.accelerator.set_device_index(device)
|
||||
|
||||
torch.set_default_device(device)
|
||||
|
||||
|
||||
@@ -102,7 +102,7 @@ def run_dp_sharded_vision_model_vs_direct(
|
||||
set_random_seed(0)
|
||||
|
||||
device = f"{current_platform.device_name}:{local_rank}"
|
||||
current_platform.set_device(device)
|
||||
torch.accelerator.set_device_index(device)
|
||||
torch.set_default_device(device)
|
||||
|
||||
update_environment_variables(
|
||||
@@ -288,7 +288,7 @@ def run_dp_sharded_mrope_vision_model_vs_direct(
|
||||
# Set random seed for reproducibility
|
||||
set_random_seed(0)
|
||||
device = f"{current_platform.device_name}:{local_rank}"
|
||||
current_platform.set_device(device)
|
||||
torch.accelerator.set_device_index(device)
|
||||
torch.set_default_device(device)
|
||||
|
||||
update_environment_variables(
|
||||
@@ -365,7 +365,7 @@ def run_dp_sharded_mrope_vision_model_empty_input_worker(
|
||||
"""Test run_dp_sharded_mrope_vision_model with empty input."""
|
||||
# Set up distributed environment
|
||||
device = f"{current_platform.device_name}:{local_rank}"
|
||||
current_platform.set_device(device)
|
||||
torch.accelerator.set_device_index(device)
|
||||
torch.set_default_device(device)
|
||||
|
||||
update_environment_variables(
|
||||
@@ -414,7 +414,7 @@ def run_dp_sharded_mrope_vision_model_uneven_load_worker(
|
||||
# Set up distributed environment
|
||||
set_random_seed(123)
|
||||
device = f"{current_platform.device_name}:{local_rank}"
|
||||
current_platform.set_device(device)
|
||||
torch.accelerator.set_device_index(device)
|
||||
torch.set_default_device(device)
|
||||
|
||||
update_environment_variables(
|
||||
|
||||
@@ -210,10 +210,9 @@ WIKITEXT_ACCURACY_CONFIGS = [
|
||||
@pytest.mark.parametrize("config", WIKITEXT_ACCURACY_CONFIGS)
|
||||
@pytest.mark.parametrize("tp_size", [1, 2])
|
||||
def test_ocp_mx_wikitext_correctness(config: AccuracyTestConfig, tp_size: int):
|
||||
if torch.cuda.device_count() < tp_size:
|
||||
pytest.skip(
|
||||
f"This test requires >={tp_size} gpus, got only {torch.cuda.device_count()}"
|
||||
)
|
||||
device_count = torch.accelerator.device_count()
|
||||
if device_count < tp_size:
|
||||
pytest.skip(f"This test requires >={tp_size} gpus, got only {device_count}")
|
||||
|
||||
task = "wikitext"
|
||||
rtol = 0.1
|
||||
@@ -246,10 +245,9 @@ def test_ocp_mx_wikitext_correctness(config: AccuracyTestConfig, tp_size: int):
|
||||
reason="Read access to huggingface.co/amd is required for this test.",
|
||||
)
|
||||
def test_mxfp4_gsm8k_correctness(config: AccuracyTestConfig):
|
||||
if torch.cuda.device_count() < 8:
|
||||
pytest.skip(
|
||||
f"This test requires >=8 gpus, got only {torch.cuda.device_count()}"
|
||||
)
|
||||
device_count = torch.accelerator.device_count()
|
||||
if device_count < 8:
|
||||
pytest.skip(f"This test requires >=8 gpus, got only {device_count}")
|
||||
|
||||
task = "gsm8k"
|
||||
rtol = 0.03
|
||||
|
||||
@@ -32,7 +32,7 @@ MTP_SIMILARITY_RATE = 0.8
|
||||
|
||||
def _skip_if_insufficient_gpus_for_tp(tp_size: int):
|
||||
"""Skip test if available GPUs < tp_size on ROCm."""
|
||||
available_gpus = torch.cuda.device_count()
|
||||
available_gpus = torch.accelerator.device_count()
|
||||
if available_gpus < tp_size:
|
||||
pytest.skip(
|
||||
f"Test requires {tp_size} GPUs, but only {available_gpus} available"
|
||||
|
||||
@@ -148,7 +148,7 @@ def test_shared_storage_connector_hashes(tmp_path, attn_backend):
|
||||
)
|
||||
|
||||
# don't put this import at the top level
|
||||
# it will call torch.cuda.device_count()
|
||||
# it will call torch.accelerator.device_count()
|
||||
from transformers import AutoProcessor
|
||||
|
||||
# Create processor to handle the chat prompt
|
||||
|
||||
@@ -1570,7 +1570,7 @@ def test_register_kv_caches(
|
||||
]
|
||||
],
|
||||
cache_dtype=torch.bfloat16,
|
||||
device=torch.cuda.current_device(),
|
||||
device=torch.accelerator.current_device_index(),
|
||||
kernel_block_sizes=[block_size],
|
||||
)
|
||||
)
|
||||
|
||||
@@ -141,7 +141,7 @@ def get_attention_backend_params() -> list[str]:
|
||||
|
||||
|
||||
def get_tp_size_params() -> list[pytest.param]:
|
||||
num_gpus = torch.cuda.device_count() if torch.cuda.is_available() else 1
|
||||
num_gpus = torch.accelerator.device_count() if torch.cuda.is_available() else 1
|
||||
return [pytest.param(tp, id=f"tp{tp}") for tp in TP_SIZES if tp <= num_gpus]
|
||||
|
||||
|
||||
|
||||
@@ -117,7 +117,8 @@ def worker_process(
|
||||
|
||||
|
||||
@pytest.mark.skipif(
|
||||
torch.cuda.device_count() < 2, reason="Need at least 2 GPUs for tensor parallelism"
|
||||
torch.accelerator.device_count() < 2,
|
||||
reason="Need at least 2 GPUs for tensor parallelism",
|
||||
)
|
||||
def test_init_distributed_is_called_before_memory_snapshot():
|
||||
"""Test that distributed env is setup before memory snapshot.
|
||||
|
||||
Reference in New Issue
Block a user