biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[FEAT] [ROCm]: AITER Fused MOE V1 Support (#16752)

Browse Source 
Signed-off-by: vllmellm <vllm.ellm@embeddedllm.com>
Co-authored-by: tjtanaa <tunjian.tan@embeddedllm.com>
This commit is contained in:
vllmellm
2025-04-25 11:06:50 +08:00
committed by GitHub
parent 0d6e187e88
commit eef364723c
3 changed files with 302 additions and 130 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
tests/model_executor/test_enabled_custom_ops.py
View File

@@ -11,6 +11,8 @@ from vllm.model_executor.layers.fused_moe.fused_moe import (
dispatch_fused_experts_func, dispatch_topk_func,
torch_vllm_inplace_fused_experts, torch_vllm_outplace_fused_experts,
vllm_topk_softmax)
from vllm.model_executor.layers.fused_moe.rocm_aiter_fused_moe import (
is_rocm_aiter_moe_enabled)
from vllm.model_executor.layers.layernorm import (
RMSNorm, dispatch_cuda_rmsnorm_func, fused_add_rms_norm, rms_norm,
rocm_aiter_fused_add_rms_norm, rocm_aiter_rms_norm)
@@ -100,11 +102,10 @@ def test_enabled_ops_invalid(env: str):
def test_topk_dispatch(use_rocm_aiter: str, monkeypatch):
monkeypatch.setenv("VLLM_ROCM_USE_AITER", use_rocm_aiter)
topk_func = dispatch_topk_func()
is_rocm_aiter_moe_enabled.cache_clear()
if current_platform.is_rocm() and int(use_rocm_aiter):
from vllm.model_executor.layers.fused_moe.rocm_aiter_fused_moe import (
rocm_aiter_topk_softmax)
assert topk_func == rocm_aiter_topk_softmax
else:
assert topk_func == vllm_topk_softmax
@@ -116,11 +117,11 @@ def test_fused_experts_dispatch(use_rocm_aiter: str, inplace: bool,
monkeypatch):
monkeypatch.setenv("VLLM_ROCM_USE_AITER", use_rocm_aiter)
is_rocm_aiter_moe_enabled.cache_clear()
fused_experts_func = dispatch_fused_experts_func(inplace)
if current_platform.is_rocm() and int(use_rocm_aiter):
from vllm.model_executor.layers.fused_moe.rocm_aiter_fused_moe import (
rocm_aiter_fused_experts)
assert fused_experts_func == rocm_aiter_fused_experts
elif inplace:
assert fused_experts_func == torch_vllm_inplace_fused_experts