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

[CI] Refactor CI (#952)

Browse Source 
1. remove some useless test func and file
2. fix format.sh problem
3. enable full test for singlecard and multicard
4. move long term test to long_term folder. For this kind of test, it
only runs by labeled and daily test. Include: spec decode、accuracy test

## After refactor:
There are 4 test modules
- `singlecard`: contains the test running on one NPU. It'll be run for
each PR and daily test.
- `multicard`: contains the test running on multi NPUs. It'll be run for
each PR and daily test.
- `long_term`: contains the test that cost much time(Now include `spec
decode` and `accuracy` test). It'll be run for the PR with
`long-term-test` labeled and daily test.
- `e2e`: contains the test for doc and pd feature. It'll be run for the
PR with `pd-test` labeled and daily test.

## Todo:
1. some test are skipped, they should be fixed and reenabled in the
future.
2. pyhccl test for multicard doesn't work at all. It should be enabled
as well.
3. ensure long-term-test pass by daily test.

### Know issue
Now, `ready` labels is required to start pd test or long term test. And
when `long-term-test` or `pd-test` is labeled after another one, the old
labeled test will be re-run again. So the labeled test should be ran in
the following step:

1. decide which test need run, then label it. `long-term-test` or
`pd-test` or both.
2. add `ready-for-test` label, then the test will be ran.

Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
This commit is contained in:
wangxiyuan
2025-05-28 06:31:35 +08:00
committed by GitHub
parent 9f5ab59e30
commit e2a0c19cea
34 changed files with 171 additions and 1288 deletions
Download Patch File Download Diff File Expand all files Collapse all files

0
tests/ops/__init__.py
View File

100
tests/ops/test_fused_moe.py
View File

@@ -1,100 +0,0 @@
# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
# Copyright 2023 The vLLM team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# SPDX-License-Identifier: Apache-2.0
# This file is a part of the vllm-ascend project.
# Adapted from vllm/tests/kernels/test_moe.py
"""Tests for the MOE layers.
Run `pytest tests/ops/test_fused_moe.py`.
"""
# fused moe ops test will hit the infer_schema error, we need add the patch
# here to make the test pass.
import vllm_ascend.patch.worker.patch_common.patch_utils # type: ignore[import] # isort: skip # noqa
import pytest
import torch
from vllm.model_executor.layers.activation import SiluAndMul
from vllm_ascend.ops.fused_moe import fused_experts
NUM_EXPERTS = [8, 64]
EP_SIZE = [1, 4]
TOP_KS = [2, 6]
DEVICE = ["npu"]
def torch_moe(a, w1, w2, topk_weights, topk_ids, topk, expert_map):
B, D = a.shape
a = a.view(B, -1, D).repeat(1, topk, 1).reshape(-1, D)
out = torch.zeros(B * topk, w2.shape[1], dtype=a.dtype, device=a.device)
topk_weights = topk_weights.view(-1)
topk_ids = topk_ids.view(-1)
if expert_map is not None:
topk_ids = expert_map[topk_ids]
for i in range(w1.shape[0]):
mask = topk_ids == i
if mask.sum():
out[mask] = SiluAndMul()(
a[mask] @ w1[i].transpose(0, 1)) @ w2[i].transpose(0, 1)
return (out.view(B, -1, w2.shape[1]) *
topk_weights.view(B, -1, 1).to(out.dtype)).sum(dim=1)
@pytest.mark.parametrize("m", [1, 33, 64, 222, 1024 * 128])
@pytest.mark.parametrize("n", [128, 1024, 2048])
@pytest.mark.parametrize("k", [128, 511, 1024])
@pytest.mark.parametrize("e", NUM_EXPERTS)
@pytest.mark.parametrize("topk", TOP_KS)
@pytest.mark.parametrize("ep_size", EP_SIZE)
@pytest.mark.parametrize("dtype", [torch.float16, torch.bfloat16])
@pytest.mark.parametrize("device", DEVICE)
def test_fused_experts(
m: int,
n: int,
k: int,
e: int,
topk: int,
ep_size: int,
dtype: torch.dtype,
device: str,
):
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)
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)
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()

190
tests/ops/test_multi_step.py
View File

@@ -1,190 +0,0 @@
# Copyright (c) China Merchants Bank Co., Ltd. 2025. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#/
# to run this test, you need to cd to the upper package which is 'tests',
# and run with command 'pytest -s ops/test_multi_step.py'
import torch
import torch_npu # noqa: F401
DTYPES = [torch.int32, torch.int64]
DEVICES = [f"npu:{0}"]
# Set tolerance to 0 for equals
DEFAULT_ATOL = 0
DEFAULT_RTOL = 0
# test custom ops of https://github.com/vllm-project/vllm-ascend/tree/main/csrc/kernels/advance_step.cpp
@torch.inference_mode()
def test_single_generation_multi_step() -> None:
input_tokens_data = [2926]
input_tokens_ascendc = torch.tensor(input_tokens_data, device='npu:0')
input_tokens_python = torch.tensor(input_tokens_data, device='npu:0')
sampled_token_ids_data = [[13]]
sampled_token_ids = torch.tensor(sampled_token_ids_data, device='npu:0')
input_positions_data = [5]
input_positions_ascendc = torch.tensor(input_positions_data,
device='npu:0')
input_positions_python = torch.tensor(input_positions_data, device='npu:0')
seq_lens_data = [6]
seq_lens_ascendc = torch.tensor(seq_lens_data,
device='npu:0',
dtype=torch.int32)
seq_lens_python = torch.tensor(seq_lens_data,
device='npu:0',
dtype=torch.int32)
slot_mapping_data = [5]
slot_mapping_ascendc = torch.tensor(slot_mapping_data,
device='npu:0',
dtype=torch.int32)
slot_mapping_python = torch.tensor(slot_mapping_data,
device='npu:0',
dtype=torch.int32)
block_tables_data = [[0]]
block_tables = torch.tensor(block_tables_data,
device='npu:0',
dtype=torch.int32)
torch.ops._C.advance_step_flashattn_ascendc(
1, 1, 128, input_tokens_ascendc, sampled_token_ids,
input_positions_ascendc, seq_lens_ascendc, slot_mapping_ascendc,
block_tables)
normal(1, 1, 128, input_tokens_python, sampled_token_ids,
input_positions_python, seq_lens_python, slot_mapping_python,
block_tables)
# Compare the results.
torch.testing.assert_close(input_tokens_ascendc,
input_tokens_python,
atol=DEFAULT_ATOL,
rtol=DEFAULT_RTOL)
torch.testing.assert_close(input_positions_ascendc,
input_positions_python,
atol=DEFAULT_ATOL,
rtol=DEFAULT_RTOL)
torch.testing.assert_close(seq_lens_ascendc,
seq_lens_python,
atol=DEFAULT_ATOL,
rtol=DEFAULT_RTOL)
torch.testing.assert_close(slot_mapping_ascendc,
slot_mapping_python,
atol=DEFAULT_ATOL,
rtol=DEFAULT_RTOL)
@torch.inference_mode()
def test_multi_result_generation_multi_step() -> None:
input_tokens_data = [2926, 279, 12095, 1588]
input_tokens_ascendc = torch.tensor(input_tokens_data, device='npu:0')
input_tokens_python = torch.tensor(input_tokens_data, device='npu:0')
sampled_token_ids_data = [[13], [1968], [13], [13]]
sampled_token_ids = torch.tensor(sampled_token_ids_data, device='npu:0')
input_positions_data = [5, 7, 5, 5]
input_positions_ascendc = torch.tensor(input_positions_data,
device='npu:0')
input_positions_python = torch.tensor(input_positions_data, device='npu:0')
seq_lens_data = [6, 8, 6, 6]
seq_lens_ascendc = torch.tensor(seq_lens_data,
device='npu:0',
dtype=torch.int32)
seq_lens_python = torch.tensor(seq_lens_data,
device='npu:0',
dtype=torch.int32)
slot_mapping_data = [5, 135, 261, 389]
slot_mapping_ascendc = torch.tensor(slot_mapping_data,
device='npu:0',
dtype=torch.int32)
slot_mapping_python = torch.tensor(slot_mapping_data,
device='npu:0',
dtype=torch.int32)
block_tables_data = [[0], [1], [2], [3]]
block_tables = torch.tensor(block_tables_data,
device='npu:0',
dtype=torch.int32)
torch.ops._C.advance_step_flashattn_ascendc(
4, 4, 128, input_tokens_ascendc, sampled_token_ids,
input_positions_ascendc, seq_lens_ascendc, slot_mapping_ascendc,
block_tables)
normal(4, 4, 128, input_tokens_python, sampled_token_ids,
input_positions_python, seq_lens_python, slot_mapping_python,
block_tables)
# Compare the results.
torch.testing.assert_close(input_tokens_ascendc,
input_tokens_python,
atol=DEFAULT_ATOL,
rtol=DEFAULT_RTOL)
torch.testing.assert_close(input_positions_ascendc,
input_positions_python,
atol=DEFAULT_ATOL,
rtol=DEFAULT_RTOL)
torch.testing.assert_close(seq_lens_ascendc,
seq_lens_python,
atol=DEFAULT_ATOL,
rtol=DEFAULT_RTOL)
torch.testing.assert_close(slot_mapping_ascendc,
slot_mapping_python,
atol=DEFAULT_ATOL,
rtol=DEFAULT_RTOL)
def normal(num_seqs: int, num_queries: int, block_size: int,
input_tokens: torch.Tensor, sampled_token_ids: torch.Tensor,
input_positions: torch.Tensor, seq_lens_tensor: torch.Tensor,
slot_mapping: torch.Tensor, block_tables: torch.Tensor) -> None:
sampled_token_ids_list = sampled_token_ids[:num_queries].squeeze(-1)
input_tokens[:num_queries] = sampled_token_ids_list
# get seq_lens and input_positions
seq_lens = seq_lens_tensor[:num_queries]
next_seq_lens = seq_lens + 1
next_input_pos = next_seq_lens - 1
# update seq_lens and input_positions
seq_lens_tensor[:num_queries] = next_seq_lens
input_positions[:num_queries] = next_input_pos # type: ignore
# get block index and offset
block_idx = next_input_pos // block_size
block_offset = next_input_pos % block_size
current_block_table = block_tables.gather(
1, block_idx.unsqueeze(-1)).squeeze(-1)
slot_num = current_block_table * block_size + block_offset
# update slot_mapping
slot_mapping[:num_queries] = slot_num

198
tests/ops/test_rotary_embedding.py
View File

@@ -1,198 +0,0 @@
# Copyright 2023 The vLLM team.
# Copyright (c) Huawei Technologies Co., Ltd. 2024-2025. All rights reserved.
# Adapted from
# https://github.com/vllm-project/vllm/blob/main/vllm/tests/kernels/test_rotary_embedding.py
from typing import Optional, Tuple, Union
import pytest
import torch
import torch.nn as nn
import vllm_ascend.platform # noqa: F401
# Only Neox style true scenario is supported for now
IS_NEOX_STYLE = [True]
DTYPES = [torch.half]
HEAD_SIZES = [64, 96, 128, 256]
ROTARY_DIMS = [None, 32] # None means rotary dim == head size
NUM_HEADS = [17] # Arbitrary values for testing
BATCH_SIZES = [5] # Arbitrary values for testing
SEQ_LENS = [11, 4096] # Arbitrary values for testing
SEEDS = [0]
DEVICES = [f"npu:{0}"]
# Set tolerance to 1 for quant ops
DEFAULT_ATOL = 1e-3
DEFAULT_RTOL = 1e-3
def _apply_rotary_emb(
x: torch.Tensor,
cos: torch.Tensor,
sin: torch.Tensor,
is_neox_style: bool,
) -> torch.Tensor:
"""
Args:
x: [num_tokens, num_heads, head_size]
cos: [num_tokens, head_size // 2]
sin: [num_tokens, head_size // 2]
is_neox_style: Whether to use the Neox-style or GPT-J-style rotary
positional embeddings.
"""
cos = cos.unsqueeze(-2).to(x.dtype)
sin = sin.unsqueeze(-2).to(x.dtype)
if is_neox_style:
x1, x2 = torch.chunk(x, 2, dim=-1)
else:
x1 = x[..., ::2]
x2 = x[..., 1::2]
o1 = x1 * cos - x2 * sin
o2 = x2 * cos + x1 * sin
if is_neox_style:
return torch.cat((o1, o2), dim=-1)
else:
return torch.stack((o1, o2), dim=-1).flatten(-2)
# adapted from https://github.com/vllm-project/vllm/vllm/model_executor/layers/rotary_embedding.py
class RotaryEmbedding(nn.Module):
"""Original rotary positional embedding."""
def __init__(
self,
head_size: int,
rotary_dim: int,
max_position_embeddings: int,
base: int,
is_neox_style: bool,
dtype: torch.dtype,
) -> None:
super().__init__()
self.head_size = head_size
self.rotary_dim = rotary_dim
self.max_position_embeddings = max_position_embeddings
self.base = base
self.is_neox_style = is_neox_style
self.dtype = dtype
cache = self._compute_cos_sin_cache()
cache = cache.to(dtype)
self.cos_sin_cache: torch.Tensor
self.register_buffer("cos_sin_cache", cache, persistent=False)
def _compute_inv_freq(self, base: Union[int, float]) -> torch.Tensor:
"""Compute the inverse frequency."""
# NOTE(woosuk): To exactly match the HF implementation, we need to
# use CPU to compute the cache and then move it to GPU. However, we
# create the cache on GPU for faster initialization. This may cause
# a slight numerical difference between the HF implementation and ours.
inv_freq = 1.0 / (base**(torch.arange(
0, self.rotary_dim, 2, dtype=torch.float) / self.rotary_dim))
return inv_freq
def _compute_cos_sin_cache(self) -> torch.Tensor:
"""Compute the cos and sin cache."""
inv_freq = self._compute_inv_freq(self.base)
t = torch.arange(self.max_position_embeddings, dtype=torch.float)
freqs = torch.einsum("i,j -> ij", t, inv_freq)
cos = freqs.cos()
sin = freqs.sin()
cache = torch.cat((cos, sin), dim=-1)
return cache
def forward_native(
self,
positions: torch.Tensor,
query: torch.Tensor,
key: torch.Tensor,
offsets: Optional[torch.Tensor] = None,
) -> Tuple[torch.Tensor, torch.Tensor]:
"""A PyTorch-native implementation of forward()."""
if offsets is not None:
positions = positions + offsets
positions = positions.flatten()
num_tokens = positions.shape[0]
cos_sin = self.cos_sin_cache.index_select(0, positions)
cos, sin = cos_sin.chunk(2, dim=-1)
query_shape = query.shape
query = query.view(num_tokens, -1, self.head_size)
query_rot = query[..., :self.rotary_dim]
query_pass = query[..., self.rotary_dim:]
query_rot = _apply_rotary_emb(query_rot, cos, sin, self.is_neox_style)
query = torch.cat((query_rot, query_pass), dim=-1).reshape(query_shape)
key_shape = key.shape
key = key.view(num_tokens, -1, self.head_size)
key_rot = key[..., :self.rotary_dim]
key_pass = key[..., self.rotary_dim:]
key_rot = _apply_rotary_emb(key_rot, cos, sin, self.is_neox_style)
key = torch.cat((key_rot, key_pass), dim=-1).reshape(key_shape)
return query, key
# test with leading dimension and merge seqlen and batch_size as num_tokens
@pytest.mark.parametrize("is_neox_style", IS_NEOX_STYLE)
@pytest.mark.parametrize("batch_size", BATCH_SIZES)
@pytest.mark.parametrize("seq_len", SEQ_LENS)
@pytest.mark.parametrize("num_heads", NUM_HEADS)
@pytest.mark.parametrize("head_size", HEAD_SIZES)
@pytest.mark.parametrize("rotary_dim", ROTARY_DIMS)
@pytest.mark.parametrize("dtype", DTYPES)
@pytest.mark.parametrize("seed", SEEDS)
@pytest.mark.parametrize("device", DEVICES)
@torch.inference_mode()
def test_rotary_embedding_quant_with_leading_dim(
is_neox_style: bool,
batch_size: int,
seq_len: int,
num_heads: int,
head_size: int,
rotary_dim: Optional[int],
dtype: torch.dtype,
seed: int,
device: str,
max_position: int = 8192,
base: int = 10000,
) -> None:
if rotary_dim is None:
rotary_dim = head_size
torch.set_default_device(device)
if rotary_dim is None:
rotary_dim = head_size
rope = RotaryEmbedding(head_size, rotary_dim, max_position, base,
is_neox_style, dtype)
rope = rope.to(dtype=dtype)
num_tokens = batch_size * seq_len
positions = torch.randint(0, max_position, (batch_size * seq_len, ))
qkv_tensor = torch.randn(num_tokens,
num_heads * head_size * 3,
dtype=dtype)
query, key, _ = qkv_tensor.split(
[num_heads * head_size, num_heads * head_size, num_heads * head_size],
dim=-1,
)
ref_query, ref_key = rope.forward_native(positions, query, key)
query, key = torch.ops._C.rotary_embedding(
positions,
query,
key,
rope.head_size,
rope.cos_sin_cache,
rope.is_neox_style,
)
# Compare the results.
torch.testing.assert_close(query.view(ref_query.size()),
ref_query,
atol=DEFAULT_ATOL,
rtol=DEFAULT_RTOL)
torch.testing.assert_close(key.view(ref_key.size()),
ref_key,
atol=DEFAULT_ATOL,
rtol=DEFAULT_RTOL)