EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity

[V1] Make V1 engine backward compatible (#637)

Browse Source 
### What this PR does / why we need it?
Enforce eager mode in the V1 engine ahead of the upcoming CANN and
torch_npu releases.

### Does this PR introduce _any_ user-facing change?
After this change, users will no longer need to manually set
enforce_eager=True.

### How was this patch tested?
Test it with regular offline inference examples.

Signed-off-by: Yizhou Liu <liu_yizhou@outlook.com>
This commit is contained in:
yiz-liu
2025-04-24 17:20:11 +08:00
committed by GitHub
parent bd70ce828c
commit d785e78563
4 changed files with 43 additions and 46 deletions
Download Patch File Download Diff File Expand all files Collapse all files

52
tests/ops/test_fused_moe.py
View File

@@ -22,7 +22,6 @@ Run `pytest tests/ops/test_fused_moe.py`.
import pytest
import torch
from vllm.config import VllmConfig, set_current_vllm_config
from vllm.model_executor.layers.activation import SiluAndMul
from vllm_ascend.ops.fused_moe import fused_experts
@@ -68,36 +67,31 @@ def test_fused_experts(
dtype: torch.dtype,
device: str,
):
vllm_config = VllmConfig()
with set_current_vllm_config(vllm_config):
a = torch.randn((m, k), device=device, dtype=dtype) / 10
w1 = torch.randn((e, 2 * n, k), device=device, dtype=dtype) / 10
w2 = torch.randn((e, k, n), device=device, dtype=dtype) / 10
a = torch.randn((m, k), device=device, dtype=dtype) / 10
w1 = torch.randn((e, 2 * n, k), device=device, dtype=dtype) / 10
w2 = torch.randn((e, k, n), device=device, dtype=dtype) / 10
score = torch.randn((m, e), device=device, dtype=dtype)
score = torch.randn((m, e), device=device, dtype=dtype)
if ep_size > 1:
local_e = e // ep_size
e_ids = torch.randint(0,
e, (local_e, ),
device=device,
dtype=torch.int32)
e_map = torch.full((e, ), -1, device=device, dtype=torch.int32)
e_map[e_ids] = torch.arange(local_e,
device=device,
dtype=torch.int32)
w1 = w1[e_ids]
w2 = w2[e_ids]
else:
e_map = None
if ep_size > 1:
local_e = e // ep_size
e_ids = torch.randint(0,
e, (local_e, ),
device=device,
dtype=torch.int32)
e_map = torch.full((e, ), -1, device=device, dtype=torch.int32)
e_map[e_ids] = torch.arange(local_e, device=device, dtype=torch.int32)
w1 = w1[e_ids]
w2 = w2[e_ids]
else:
e_map = None
score = torch.softmax(score, dim=-1, dtype=dtype)
topk_weights, topk_ids = torch.topk(score, topk)
topk_ids = topk_ids.to(torch.int32)
score = torch.softmax(score, dim=-1, dtype=dtype)
topk_weights, topk_ids = torch.topk(score, topk)
topk_ids = topk_ids.to(torch.int32)
output = fused_experts(a, w1, w2, topk_weights, topk_ids, topk, e_map)
torch_output = torch_moe(a, w1, w2, topk_weights, topk_ids, topk,
e_map)
# TODO: The native params are: atol=2e-2, rtol=0, maybe related to the nan problem
torch.testing.assert_close(output, torch_output, atol=4e-2, rtol=1)
output = fused_experts(a, w1, w2, topk_weights, topk_ids, topk, e_map)
torch_output = torch_moe(a, w1, w2, topk_weights, topk_ids, topk, e_map)
# TODO: The native params are: atol=2e-2, rtol=0, maybe related to the nan problem
torch.testing.assert_close(output, torch_output, atol=4e-2, rtol=1)
torch.npu.empty_cache()