[ROCm][FEAT] AITER Fused Allreduce + RMSNorm (#37646)

Signed-off-by: vllmellm <vllm.ellm@embeddedllm.com>
Signed-off-by: Rita Brugarolas Brufau <rita.brugarolasbrufau@amd.com>
Signed-off-by: junkang1991 <junkangchow@gmail.com>
Co-authored-by: Rita Brugarolas <Rita.BrugarolasBrufau@amd.com>
Co-authored-by: junkang1991 <junkangchow@gmail.com>
Co-authored-by: Luka Govedič <ProExpertProg@users.noreply.github.com>
Co-authored-by: TJian <tunjian.tan@embeddedllm.com>
This commit is contained in:
vllmellm
2026-05-01 23:07:18 +08:00
committed by GitHub
co-authored by Rita Brugarolas junkang1991 Luka Govedič TJian
parent bc635fad23
commit 529c671e80
9 changed files with 454 additions and 20 deletions
+19 -3
View File
@@ -19,6 +19,8 @@ from .models import (
FLASHINFER_ATTN,
FLASHINFER_MLA_ATTN,
FLASHMLA_SPARSE_ATTN,
ROCM_AITER_UNIFIED_ATTN,
ROCM_ATTN,
TRITON_ATTN,
deepseek_coder_v2_lite_fp8,
deepseek_r1_fp4,
@@ -34,7 +36,9 @@ from .models import (
qwen3_a3b_fp8,
)
pytestmark = pytest.mark.skipif(not current_platform.is_cuda(), reason="Only test CUDA")
pytestmark = pytest.mark.skipif(
not current_platform.is_cuda_alike(), reason="Only test CUDA/ROCm"
)
@multi_gpu_test(num_gpus=2)
@@ -55,6 +59,7 @@ pytestmark = pytest.mark.skipif(not current_platform.is_cuda(), reason="Only tes
@pytest.mark.parametrize("n_layers", [4])
@pytest.mark.parametrize("custom_ops", custom_ops_combos("quant_fp8", "rms_norm"))
@pytest.mark.parametrize("inductor_graph_partition", INDUCTOR_GRAPH_PARTITION)
@pytest.mark.skipif(not current_platform.is_cuda(), reason="Only test CUDA")
def test_tp2_ar_rms_fp8_fusions(
model_name: str,
matches_fn: Callable[[int], Matches],
@@ -124,6 +129,7 @@ def test_tp2_ar_rms_fp8_fusions(
@pytest.mark.parametrize("custom_ops", custom_ops_combos("rms_norm"))
@pytest.mark.parametrize("inductor_graph_partition", INDUCTOR_GRAPH_PARTITION)
@pytest.mark.skipif(not is_blackwell(), reason="Blackwell required for fp4")
@pytest.mark.skipif(not current_platform.is_cuda(), reason="Only test CUDA")
def test_tp2_ar_rms_fp4_fusions(
model_name: str,
matches_fn: Callable[[int], Matches],
@@ -176,10 +182,19 @@ def test_tp2_ar_rms_fp4_fusions(
"model_name, matches_fn, model_kwargs, hf_overrides",
[llama3_8b, qwen3_a3b, gpt_oss_20b],
)
@pytest.mark.parametrize("attn_backend", [TRITON_ATTN])
@pytest.mark.parametrize(
"attn_backend",
[
TRITON_ATTN,
FLASHINFER_ATTN,
ROCM_ATTN,
ROCM_AITER_UNIFIED_ATTN,
],
)
@pytest.mark.parametrize("n_layers", [4])
@pytest.mark.parametrize("custom_ops", custom_ops_combos("rms_norm"))
@pytest.mark.parametrize("custom_ops", tuple(custom_ops_combos("rms_norm")))
@pytest.mark.parametrize("inductor_graph_partition", INDUCTOR_GRAPH_PARTITION)
@pytest.mark.skipif(not current_platform.is_cuda_alike(), reason="Only test CUDA/ROCm")
def test_tp2_ar_rms_fusions(
model_name: str,
matches_fn: Callable[[int], Matches],
@@ -221,4 +236,5 @@ def test_tp2_ar_rms_fusions(
compilation_config,
matches_check,
tp_size=2,
use_aiter=current_platform.is_rocm(),
)