[ROCm][CI] Extend ROCm quick reduce coverage (#40990)

Signed-off-by: Andreas Karatzas <akaratza@amd.com>
This commit is contained in:
Andreas Karatzas
2026-05-26 21:57:13 +08:00
committed by GitHub
parent d565357a90
commit 445ded18c1
4 changed files with 869 additions and 38 deletions
+73 -3
View File
@@ -25,16 +25,34 @@ from ..utils import (
ensure_model_parallel_initialized,
init_test_distributed_environment,
multi_process_parallel,
set_random_seed,
)
torch.manual_seed(42)
random.seed(44)
def on_gfx942() -> bool:
if current_platform.is_rocm():
from vllm.platforms.rocm import on_gfx942 as rocm_on_gfx942
return rocm_on_gfx942()
return False
set_random_seed(42)
_test_size_rng = random.Random(44)
# Size over 8MB is sufficient for custom quick allreduce.
test_sizes = [random.randint(8 * 1024 * 1024, 10 * 1024 * 1024) for _ in range(8)]
test_sizes = [
_test_size_rng.randint(8 * 1024 * 1024, 10 * 1024 * 1024) for _ in range(8)
]
for i, v in enumerate(test_sizes):
test_sizes[i] -= v % 8
def _assert_quickreduce(fa, inp):
assert fa is not None
assert not fa.disabled
assert fa.should_quick_allreduce(inp)
@pytest.fixture
def envs_cache_disabled():
disable_envs_cache()
@@ -216,11 +234,14 @@ def graph_quickreduce(
):
with monkeypatch.context() as m:
m.delenv("CUDA_VISIBLE_DEVICES", raising=False)
m.delenv("HIP_VISIBLE_DEVICES", raising=False)
m.delenv("ROCR_VISIBLE_DEVICES", raising=False)
device = torch.device(f"cuda:{rank}")
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
fa = get_tp_group().device_communicator.qr_comm
# A small all_reduce for warmup.
# this is needed because device communicators might be created lazily
@@ -246,6 +267,8 @@ def graph_quickreduce(
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)
_assert_quickreduce(fa, inp1)
_assert_quickreduce(fa, inp2)
torch.accelerator.synchronize()
graph = torch.cuda.CUDAGraph()
@@ -270,6 +293,8 @@ def eager_quickreduce(
):
with monkeypatch.context() as m:
m.delenv("CUDA_VISIBLE_DEVICES", raising=False)
m.delenv("HIP_VISIBLE_DEVICES", raising=False)
m.delenv("ROCR_VISIBLE_DEVICES", raising=False)
device = torch.device(f"cuda:{rank}")
torch.accelerator.set_device_index(device)
@@ -281,12 +306,42 @@ def eager_quickreduce(
inp = torch.tensor(
[1.0 * ((i) % 23) for i in range(sz)], dtype=torch.float16, device=device
)
_assert_quickreduce(fa, inp)
out = fa.quick_all_reduce(inp)
torch.testing.assert_close(out, inp * tp_size, atol=2.5, rtol=0.1)
inp = torch.tensor(
[1.0 * ((i) % 23) for i in range(sz)], dtype=torch.bfloat16, device=device
)
_assert_quickreduce(fa, inp)
out = fa.quick_all_reduce(inp)
torch.testing.assert_close(out, inp * tp_size, atol=2.5, rtol=0.1)
@ray.remote(num_gpus=1, max_calls=1)
def bf16_cast_quickreduce(
monkeypatch: pytest.MonkeyPatch,
tp_size,
pp_size,
rank,
distributed_init_port,
):
with monkeypatch.context() as m:
m.delenv("CUDA_VISIBLE_DEVICES", raising=False)
m.delenv("HIP_VISIBLE_DEVICES", raising=False)
m.delenv("ROCR_VISIBLE_DEVICES", raising=False)
m.setenv("VLLM_ROCM_QUICK_REDUCE_CAST_BF16_TO_FP16", "1")
device = torch.device(f"cuda:{rank}")
torch.accelerator.set_device_index(device)
init_test_distributed_environment(tp_size, pp_size, rank, distributed_init_port)
sz = 16 * 1024 * 1024
fa = get_tp_group().device_communicator.qr_comm
inp = torch.tensor(
[1.0 * (i % 23) for i in range(sz)], dtype=torch.bfloat16, device=device
)
_assert_quickreduce(fa, inp)
assert fa.use_fp16_kernels
out = fa.quick_all_reduce(inp)
torch.testing.assert_close(out, inp * tp_size, atol=2.5, rtol=0.1)
@@ -308,12 +363,27 @@ def test_custom_quick_allreduce(
world_size = tp_size * pipeline_parallel_size
if world_size > torch.accelerator.device_count():
pytest.skip("Not enough GPUs to run the test.")
if test_target is graph_quickreduce and on_gfx942():
pytest.xfail(
"CUDA graph capture with quick reduce hits "
"hipErrorStreamCaptureInvalidated on gfx942"
)
monkeypatch.setenv("VLLM_ROCM_QUICK_REDUCE_QUANTIZATION", quant_mode)
multi_process_parallel(monkeypatch, tp_size, pipeline_parallel_size, test_target)
@pytest.mark.skipif(
not current_platform.is_rocm(), reason="only test quick allreduce for rocm"
)
def test_custom_quick_allreduce_bf16_cast(monkeypatch: pytest.MonkeyPatch):
if torch.accelerator.device_count() < 2:
pytest.skip("Not enough GPUs to run the test.")
monkeypatch.setenv("VLLM_ROCM_QUICK_REDUCE_QUANTIZATION", "FP")
multi_process_parallel(monkeypatch, 2, 1, bf16_cast_quickreduce)
def qr_variable_input(rank, world_size):
"""
When the tensor parallelism is set to 4 or 8, frequent changes