chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[CI] Migrate Attention Backend tests to test/registered/attention/ (#15563)

Browse Source
This commit is contained in:
Alison Shao
2025-12-22 22:17:52 -08:00
committed by GitHub
parent 989d4b3012
commit 883747ced1
18 changed files with 94 additions and 22 deletions
Download Patch File Download Diff File Expand all files Collapse all files

38
.github/workflows/pr-test.yml vendored
View File

@@ -522,6 +522,43 @@ jobs:
cd test/
python3 run_suite.py --hw cuda --suite stage-b-test-large-2-gpu
stage-c-test-large-4-gpu:
needs: [check-changes, call-gate, stage-b-test-small-1-gpu, stage-b-test-large-1-gpu, stage-b-test-large-2-gpu, sgl-kernel-build-wheels]
if: |
always() &&
(
(inputs.target_stage == 'stage-c-test-large-4-gpu') ||
(
!inputs.target_stage &&
(github.event_name == 'schedule' || (!failure() && !cancelled())) &&
((needs.check-changes.outputs.main_package == 'true') || (needs.check-changes.outputs.sgl_kernel == 'true'))
)
)
runs-on: 4-gpu-h100
env:
RUNNER_LABELS: 4-gpu-h100
steps:
- name: Checkout code
uses: actions/checkout@v4
- name: Download artifacts
if: needs.check-changes.outputs.sgl_kernel == 'true'
uses: actions/download-artifact@v4
with:
path: sgl-kernel/dist/
merge-multiple: true
pattern: wheel-python3.10-cuda12.9
- name: Install dependencies
run: |
CUSTOM_BUILD_SGL_KERNEL=${{needs.check-changes.outputs.sgl_kernel}} bash scripts/ci/ci_install_dependency.sh
- name: Run test
timeout-minutes: 30
run: |
cd test/
python3 run_suite.py --hw cuda --suite stage-c-test-large-4-gpu
multimodal-gen-test-1-gpu:
needs: [check-changes, call-gate, sgl-kernel-build-wheels]
if: |
@@ -1402,6 +1439,7 @@ jobs:
stage-b-test-small-1-gpu,
stage-b-test-large-1-gpu,
stage-b-test-large-2-gpu,
stage-c-test-large-4-gpu,
quantization-test,
unit-test-backend-1-gpu,
unit-test-backend-2-gpu,

1
scripts/ci/slash_command_handler.py
View File

@@ -147,6 +147,7 @@ def handle_rerun_stage(
"stage-b-test-small-1-gpu",
"stage-b-test-large-1-gpu",
"stage-b-test-large-2-gpu",
"stage-c-test-large-4-gpu",
"multimodal-gen-test-1-gpu",
"multimodal-gen-test-2-gpu",
"quantization-test",

4
test/srt/test_create_kvindices.py → test/registered/attention/test_create_kvindices.py
View File

@@ -4,8 +4,12 @@ import numpy as np
import torch
from sglang.srt.layers.attention.utils import create_flashinfer_kv_indices_triton
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.test_utils import CustomTestCase
# Triton kernel unit test for KV indices creation
register_cuda_ci(est_time=10, suite="stage-b-test-small-1-gpu")
class TestCreateKvIndices(CustomTestCase):
@classmethod

5
test/srt/test_fa3.py → test/registered/attention/test_fa3.py
View File

@@ -5,6 +5,7 @@ import requests
from sglang.srt.environ import envs
from sglang.srt.utils import get_device_sm, kill_process_tree
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
from sglang.test.test_utils import (
DEFAULT_MODEL_NAME_FOR_TEST,
@@ -17,6 +18,10 @@ from sglang.test.test_utils import (
popen_launch_server,
)
# FlashAttention3 integration tests (requires SM 90+ / H100)
# Multiple test classes: FA3, FA3+MLA, FA3+SpecDecode variants
register_cuda_ci(est_time=300, suite="stage-b-test-large-1-gpu")
GSM_DATASET_PATH = None
# In case of some machine lack internet connection, we can set OFFLINE_MODE to True.

4
test/srt/test_flash_attention_4.py → test/registered/attention/test_flash_attention_4.py
View File

@@ -3,6 +3,7 @@ from types import SimpleNamespace
from urllib.parse import urlparse
from sglang.srt.utils import get_device_sm, kill_process_tree
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
from sglang.test.test_utils import (
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
@@ -10,6 +11,9 @@ from sglang.test.test_utils import (
popen_launch_server,
)
# FlashAttention4 integration test (requires SM 100+ / Blackwell B200)
register_cuda_ci(est_time=200, suite="stage-b-test-large-1-gpu")
@unittest.skipIf(get_device_sm() < 100, "Test requires CUDA SM 100 or higher")
class TestFlashAttention4(unittest.TestCase):

5
test/srt/test_hybrid_attn_backend.py → test/registered/attention/test_hybrid_attn_backend.py
View File

@@ -5,6 +5,7 @@ import requests
from sglang.srt.environ import envs
from sglang.srt.utils import get_device_sm, kill_process_tree
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
from sglang.test.test_utils import (
DEFAULT_EAGLE_DRAFT_MODEL_FOR_TEST,
@@ -17,6 +18,10 @@ from sglang.test.test_utils import (
popen_launch_server,
)
# Hybrid attention backend tests (FA3 prefill + FlashInfer decode, requires SM 90+ / H100)
# Multiple test classes: base, MLA, TorchCompile, SpecDecode variants
register_cuda_ci(est_time=200, suite="stage-b-test-large-1-gpu")
GSM_DATASET_PATH = None
# Default server arguments shared across all tests

4
test/srt/test_local_attn.py → test/registered/attention/test_local_attn.py
View File

@@ -5,6 +5,7 @@ from types import SimpleNamespace
import requests
from sglang.srt.utils import get_device_sm, kill_process_tree
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
from sglang.test.test_utils import (
DEFAULT_MODEL_NAME_FOR_TEST_LOCAL_ATTENTION,
@@ -14,6 +15,9 @@ from sglang.test.test_utils import (
popen_launch_server,
)
# Local attention with FA3 (requires SM 90+ / H100, tp=4)
register_cuda_ci(est_time=200, suite="stage-c-test-large-4-gpu")
@unittest.skipIf(get_device_sm() < 90, "Test requires CUDA SM 90 or higher")
class TestFlashAttention3LocalAttn(CustomTestCase):

4
test/srt/test_radix_attention.py → test/registered/attention/test_radix_attention.py
View File

@@ -1,6 +1,7 @@
import unittest
from sglang.srt.environ import envs
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.kits.radix_cache_server_kit import run_radix_attention_test
from sglang.test.test_utils import (
DEFAULT_SMALL_MODEL_NAME_FOR_TEST,
@@ -12,6 +13,9 @@ from sglang.test.test_utils import (
popen_launch_server,
)
# RadixAttention server integration tests
register_cuda_ci(est_time=100, suite="stage-b-test-small-1-gpu")
class TestRadixCacheFCFS(CustomTestCase):
@classmethod

5
test/srt/test_radix_cache_unit.py → test/registered/attention/test_radix_cache_unit.py
View File

@@ -17,6 +17,11 @@ Usage:
python -m pytest test_radix_cache_unit.py::TestRadixCache::test_insert_basic
"""
from sglang.test.ci.ci_register import register_cuda_ci
# CPU-based unit test, runs quickly on any GPU runner
register_cuda_ci(est_time=5, suite="stage-b-test-small-1-gpu")
import time
import unittest
import unittest.mock

6
test/srt/test_torch_native_attention_backend.py → test/registered/attention/test_torch_native_attention_backend.py
View File

@@ -1,12 +1,13 @@
"""
Usage:
python3 -m unittest test_triton_attention_backend.TestTritonAttnBackend.test_mmlu
python3 -m unittest test_torch_native_attention_backend.TestTorchNativeAttnBackend.test_mmlu
"""
import unittest
from types import SimpleNamespace
from sglang.srt.utils import kill_process_tree
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.run_eval import run_eval
from sglang.test.test_utils import (
DEFAULT_MODEL_NAME_FOR_TEST,
@@ -16,6 +17,9 @@ from sglang.test.test_utils import (
popen_launch_server,
)
# Torch native attention backend integration test with MMLU eval
register_cuda_ci(est_time=150, suite="stage-b-test-small-1-gpu")
class TestTorchNativeAttnBackend(CustomTestCase):
def test_mmlu(self):

4
test/srt/test_triton_attention_backend.py → test/registered/attention/test_triton_attention_backend.py
View File

@@ -7,6 +7,7 @@ import unittest
from types import SimpleNamespace
from sglang.srt.utils import kill_process_tree
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.run_eval import run_eval
from sglang.test.test_utils import (
DEFAULT_MODEL_NAME_FOR_TEST,
@@ -18,6 +19,9 @@ from sglang.test.test_utils import (
run_bench_offline_throughput,
)
# Triton attention backend integration test with latency benchmark and MMLU eval
register_cuda_ci(est_time=200, suite="stage-b-test-small-1-gpu")
class TestTritonAttnBackend(CustomTestCase):
def test_latency(self):

4
test/srt/test_triton_attention_kernels.py → test/registered/attention/test_triton_attention_kernels.py
View File

@@ -19,8 +19,12 @@ from sglang.srt.layers.attention.triton_ops.prefill_attention import (
context_attention_fwd,
)
from sglang.srt.utils import get_device
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.test_utils import CustomTestCase
# Triton attention kernel unit tests (decode, extend, prefill)
register_cuda_ci(est_time=30, suite="stage-b-test-small-1-gpu")
def extend_attention_fwd_torch(
q: torch.Tensor, # [extend_tokens, H_Q, D]

4
test/srt/test_triton_sliding_window.py → test/registered/attention/test_triton_sliding_window.py
View File

@@ -4,6 +4,7 @@ from types import SimpleNamespace
import requests
from sglang.srt.utils import kill_process_tree
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.run_eval import run_eval
from sglang.test.test_utils import (
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
@@ -13,6 +14,9 @@ from sglang.test.test_utils import (
popen_launch_server,
)
# Sliding window attention with Triton backend (Gemma-3 model)
register_cuda_ci(est_time=100, suite="stage-b-test-small-1-gpu")
class TestSlidingWindowAttentionTriton(CustomTestCase):
"""Test sliding window attention functionality with triton backend."""

4
test/srt/test_wave_attention_kernels.py → test/registered/attention/test_wave_attention_kernels.py
View File

@@ -21,6 +21,10 @@ from sglang.srt.layers.attention.wave_ops.extend_attention import extend_attenti
from sglang.srt.layers.attention.wave_ops.prefill_attention import (
prefill_attention_wave,
)
from sglang.test.ci.ci_register import register_amd_ci
# Wave attention kernel unit tests (AMD only - requires wave_lang)
register_amd_ci(est_time=60, suite="stage-a-test-1")
class TestWaveAttention(unittest.TestCase):

2
test/registered/spec/eagle/test_eagle_infer_a.py
View File

@@ -22,7 +22,7 @@ from sglang.test.test_utils import (
popen_launch_server,
)
register_cuda_ci(est_time=470, suite="stage-b-test-small-1-gpu")
register_cuda_ci(est_time=500, suite="stage-b-test-small-1-gpu")
torch_dtype = torch.float16
prefill_tolerance = 5e-2

2
test/registered/spec/eagle/test_eagle_infer_b.py
View File

@@ -20,7 +20,7 @@ from sglang.test.test_utils import (
run_logprob_check,
)
register_cuda_ci(est_time=473, suite="stage-b-test-small-1-gpu")
register_cuda_ci(est_time=1100, suite="stage-b-test-small-1-gpu")
class TestEAGLEServerBasic(EagleServerBase):

1
test/run_suite.py
View File

@@ -24,6 +24,7 @@ PER_COMMIT_SUITES = {
"stage-b-test-small-1-gpu",
"stage-b-test-large-1-gpu",
"stage-b-test-large-2-gpu",
"stage-c-test-large-4-gpu",
"stage-b-test-4-gpu-b200",
],
HWBackend.NPU: [],

19
test/srt/run_suite.py
View File

@@ -54,12 +54,10 @@ suites = {
TestFile("dllm/test_llada2_mini.py", 520),
TestFile("test_abort.py", 131),
TestFile("test_chunked_prefill.py", 312),
TestFile("test_create_kvindices.py", 7),
TestFile("test_deterministic.py", 228),
TestFile("test_constrained_decoding.py", 111),
TestFile("test_eval_fp8_accuracy.py", 250),
TestFile("test_external_models.py", 30),
TestFile("test_fa3.py", 420),
TestFile("test_flashmla.py", 230),
TestFile("test_fp8_utils.py", 9),
TestFile("rotary_embedding/test_mrope.py", 10),
@@ -67,7 +65,6 @@ suites = {
TestFile("test_gpt_oss_1gpu.py", 402),
TestFile("test_harmony_parser.py", 6),
TestFile("test_hidden_states.py", 55),
TestFile("test_hybrid_attn_backend.py", 379),
TestFile("test_input_embeddings.py", 38),
TestFile("test_io_struct.py", 8),
TestFile("test_jinja_template_utils.py", 7),
@@ -91,8 +88,6 @@ suites = {
TestFile("test_prefill_adder.py", 1),
TestFile("test_priority_scheduling.py", 130),
TestFile("test_pytorch_sampling_backend.py", 66),
TestFile("test_radix_attention.py", 105),
TestFile("test_radix_cache_unit.py", 8),
TestFile("test_reasoning_parser.py", 5),
TestFile("test_request_queue_validation.py", 47),
TestFile("test_retract_decode.py", 259),
@@ -110,13 +105,8 @@ suites = {
TestFile("test_torch_compile.py", 190),
TestFile("test_torch_compile_moe.py", 210),
TestFile("test_triton_fused_moe.py", 12),
TestFile("test_torch_native_attention_backend.py", 221),
TestFile("test_torchao.py", 103),
TestFile("test_triton_attention_kernels.py", 4),
TestFile("test_triton_attention_backend.py", 203),
TestFile("test_triton_attention_kernels.py", 4),
TestFile("test_triton_moe_channel_fp8_kernel.py", 16),
TestFile("test_triton_sliding_window.py", 84),
TestFile("test_utils_update_weights.py", 29),
TestFile("test_video_utils.py", 5),
TestFile("test_vision_chunked_prefill.py", 150),
@@ -144,7 +134,6 @@ suites = {
"per-commit-4-gpu": [
TestFile("models/test_qwen3_next_models.py", 650),
TestFile("test_gpt_oss_4gpu.py", 300),
TestFile("test_local_attn.py", 411),
TestFile("test_multi_instance_release_memory_occupation.py", 64),
TestFile("test_pp_single_node.py", 500),
TestFile("test_epd_disaggregation.py", 150),
@@ -166,7 +155,6 @@ suites = {
],
"per-commit-4-gpu-b200": [
TestFile("test_deepseek_v3_fp4_4gpu.py", 1500),
TestFile("test_flash_attention_4.py", 90),
TestFile("test_fp8_blockwise_gemm.py", 280),
TestFile("test_gpt_oss_4gpu.py", 700),
TestFile("test_llama31_fp4.py", 90),
@@ -253,7 +241,6 @@ suite_amd = {
TestFile("test_abort.py", 51),
TestFile("test_bench_typebaseddispatcher.py", 10),
TestFile("test_chunked_prefill.py", 312),
TestFile("test_create_kvindices.py", 2),
TestFile("test_eval_fp8_accuracy.py", 303),
TestFile("test_fused_moe.py", 30),
TestFile("test_harmony_parser.py", 20),
@@ -268,7 +255,6 @@ suite_amd = {
TestFile("test_page_size.py", 60),
TestFile("test_penalty.py", 180),
TestFile("test_pytorch_sampling_backend.py", 66),
TestFile("test_radix_attention.py", 105),
TestFile("test_reasoning_parser.py", 5),
TestFile("test_constrained_decoding.py", 120),
TestFile("test_retract_decode.py", 450),
@@ -279,12 +265,7 @@ suite_amd = {
TestFile("test_srt_engine.py", 261),
TestFile("test_torch_compile.py", 169),
# TestFile("test_torch_compile_moe.py", 210), # Disabled temporarily, see https://github.com/sgl-project/sglang/issues/13107
TestFile("test_torch_native_attention_backend.py", 123),
# TestFile("test_triton_attention_kernels.py", 4),
TestFile("test_triton_attention_backend.py", 150),
TestFile("test_triton_sliding_window.py", 250),
TestFile("test_type_based_dispatcher.py", 10),
TestFile("test_wave_attention_kernels.py", 2),
# Disabled temporarily
# TestFile("test_vlm_input_format.py", 300),
# TestFile("models/test_embedding_models.py", 73), # Disabled temporarily, see https://github.com/sgl-project/sglang/issues/11127