test: recover the config-namespace-migration deferrals (#33171)
The module-skipped tests injected config by faking get_server_args (a SimpleNamespace stand-in patched onto the module) or by writing fields onto a ServerArgs instance post-publish — both invisible to the namespace accessors the production code now reads. Re-enable them by publishing the config they need (get_context().override_server_args seeding, scoped per test), asserting bag state where the old assertions checked instance write-through (declare_load_time_override is bag-only), and extending the per-runner stubs the code genuinely reads (kv_cache_dtype_str, max_total_tokens, context_len). The unified-radix-cache file (which grew a large hicache/insert-walk suite while skipped) is recovered in the same change: - test_cache_finished_req_strips_thinking (19 parametrized classes) wrote strip_thinking_cache onto the ServerArgs instance; the cache reads get_serving().strip_thinking_cache — use the serving bag's scoped override. - test_shallower_crossing_backs_up_above_backuped_middle staged its broken-backup-continuity setup through insert_host, which now deliberately drops refills below an un-backed-up node under write-through (host_insert_dropped). Build the same tree state through an explicit backup + device eviction. Every config-namespace-migration deferral is recovered, so the deferral ratchet (test_migration_deferral_ratchet.py) has done its job and is retired.
This commit is contained in:
Cheng Wan
@@ -7,11 +7,7 @@ import torch.nn as nn
|
||||
import torch.nn.functional as F
|
||||
|
||||
from sglang.srt.layers.logits_processor import LogitsProcessor
|
||||
from sglang.srt.runtime_context import get_server_args
|
||||
from sglang.srt.server_args import (
|
||||
ServerArgs,
|
||||
set_global_server_args_for_scheduler,
|
||||
)
|
||||
from sglang.srt.runtime_context import get_context
|
||||
from sglang.srt.utils import get_device
|
||||
from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
|
||||
|
||||
@@ -19,21 +15,6 @@ register_cuda_ci(est_time=9, stage="base-b", runner_config="1-gpu-small")
|
||||
register_amd_ci(est_time=15, suite="stage-b-test-1-gpu-small-amd")
|
||||
|
||||
|
||||
import pytest as _pytest_defer
|
||||
|
||||
_DEFER_REASON = (
|
||||
"Temporarily skipped during the ServerArgs config-namespace migration; "
|
||||
"re-enabled once the runtime-config accessor API stabilizes."
|
||||
)
|
||||
pytestmark = _pytest_defer.mark.skip(reason=_DEFER_REASON)
|
||||
|
||||
|
||||
def setUpModule():
|
||||
import unittest
|
||||
|
||||
raise unittest.SkipTest(_DEFER_REASON)
|
||||
|
||||
|
||||
class LMHeadStub(nn.Module):
|
||||
def __init__(self, vocab, hidden, dtype, device=get_device()):
|
||||
super().__init__()
|
||||
@@ -58,8 +39,11 @@ class TestLMHeadFP32(unittest.TestCase):
|
||||
raise unittest.SkipTest("needs CUDA GPU or XPU")
|
||||
|
||||
def _make_logprocessor(self, vocab_size, enable_fp32):
|
||||
set_global_server_args_for_scheduler(ServerArgs(model_path="dummy"))
|
||||
get_server_args().enable_fp32_lm_head = enable_fp32
|
||||
# LogitsProcessor reads get_exec().features.enable_fp32_lm_head
|
||||
# from the published config.
|
||||
override = get_context().override_server_args(enable_fp32_lm_head=enable_fp32)
|
||||
override.install()
|
||||
self.addCleanup(override.restore)
|
||||
cfg = SimpleNamespace(vocab_size=vocab_size, final_logit_softcapping=None)
|
||||
return LogitsProcessor(cfg, skip_all_gather=True, logit_scale=None)
|
||||
|
||||
|
||||
@@ -4,7 +4,7 @@
|
||||
``ForwardBatch.num_token_non_padded`` is a scalar tensor on the model device
|
||||
(see ``ForwardBatch.compute``, which does ``.to(device, ...)``). The eager TBO
|
||||
split path already honors this -- ``compute_tbo_children_num_token_non_padded_raw``
|
||||
moves the tensor to ``get_server_args().device`` -- but
|
||||
moves the tensor to ``get_device().device`` -- but
|
||||
``TboCudaGraphRunnerPlugin`` preallocated its persistent buffer with a bare
|
||||
``torch.zeros((2,), dtype=torch.int32)``, leaving it on CPU.
|
||||
|
||||
@@ -18,44 +18,27 @@ CPU-only: a ``meta`` device makes the buffer placement observable without a GPU.
|
||||
"""
|
||||
|
||||
import unittest
|
||||
from types import SimpleNamespace
|
||||
from unittest.mock import patch
|
||||
|
||||
import torch
|
||||
|
||||
import sglang.srt.batch_overlap.two_batch_overlap as tbo
|
||||
from sglang.srt.batch_overlap.two_batch_overlap import (
|
||||
TboCudaGraphRunnerPlugin,
|
||||
TboForwardBatchPreparer,
|
||||
)
|
||||
from sglang.srt.runtime_context import get_context, get_device
|
||||
from sglang.test.ci.ci_register import register_cpu_ci
|
||||
from sglang.test.test_utils import CustomTestCase
|
||||
|
||||
register_cpu_ci(est_time=5, suite="base-a-test-cpu")
|
||||
|
||||
|
||||
import pytest as _pytest_defer
|
||||
|
||||
_DEFER_REASON = (
|
||||
"Temporarily skipped during the ServerArgs config-namespace migration; "
|
||||
"re-enabled once the runtime-config accessor API stabilizes."
|
||||
)
|
||||
pytestmark = _pytest_defer.mark.skip(reason=_DEFER_REASON)
|
||||
|
||||
|
||||
def setUpModule():
|
||||
import unittest
|
||||
|
||||
raise unittest.SkipTest(_DEFER_REASON)
|
||||
|
||||
|
||||
class TestTboCudaGraphNumTokenDevice(CustomTestCase):
|
||||
def test_plugin_buffer_on_model_device(self):
|
||||
# Use 'meta' so the configured device differs from the implicit CPU
|
||||
# default; a bare torch.zeros() would leave the buffer on CPU and fail.
|
||||
fake_args = SimpleNamespace(device="meta")
|
||||
with patch.object(tbo, "get_server_args", lambda: fake_args):
|
||||
plugin = TboCudaGraphRunnerPlugin()
|
||||
with get_context().override_server_args():
|
||||
with get_device().override(device="meta"):
|
||||
plugin = TboCudaGraphRunnerPlugin()
|
||||
|
||||
buf = plugin._tbo_children_num_token_non_padded
|
||||
self.assertEqual(tuple(buf.shape), (2,))
|
||||
@@ -65,14 +48,12 @@ class TestTboCudaGraphNumTokenDevice(CustomTestCase):
|
||||
def test_graph_and_eager_paths_agree_on_device(self):
|
||||
# Both the preallocated cuda-graph buffer and the eager split tensor must
|
||||
# land on the same (model) device, matching ForwardBatch's contract.
|
||||
fake_args = SimpleNamespace(device="meta")
|
||||
with patch.object(tbo, "get_server_args", lambda: fake_args):
|
||||
eager = (
|
||||
TboForwardBatchPreparer.compute_tbo_children_num_token_non_padded_raw(
|
||||
with get_context().override_server_args():
|
||||
with get_device().override(device="meta"):
|
||||
eager = TboForwardBatchPreparer.compute_tbo_children_num_token_non_padded_raw(
|
||||
tbo_split_token_index=3, num_token_non_padded=8
|
||||
)
|
||||
)
|
||||
plugin = TboCudaGraphRunnerPlugin()
|
||||
plugin = TboCudaGraphRunnerPlugin()
|
||||
|
||||
self.assertEqual(
|
||||
eager.device.type,
|
||||
@@ -82,13 +63,11 @@ class TestTboCudaGraphNumTokenDevice(CustomTestCase):
|
||||
def test_eager_split_values(self):
|
||||
# value_a = min(split, n); value_b = max(0, n - split). Computed on CPU
|
||||
# so the values are materializable.
|
||||
fake_args = SimpleNamespace(device="cpu")
|
||||
with patch.object(tbo, "get_server_args", lambda: fake_args):
|
||||
eager = (
|
||||
TboForwardBatchPreparer.compute_tbo_children_num_token_non_padded_raw(
|
||||
with get_context().override_server_args():
|
||||
with get_device().override(device="cpu"):
|
||||
eager = TboForwardBatchPreparer.compute_tbo_children_num_token_non_padded_raw(
|
||||
tbo_split_token_index=3, num_token_non_padded=8
|
||||
)
|
||||
)
|
||||
self.assertEqual(eager.dtype, torch.int32)
|
||||
self.assertEqual(eager.tolist(), [3, 5])
|
||||
|
||||
|
||||
@@ -6,36 +6,18 @@ crashed TBO cuda-graph capture until reset. CPU-only.
|
||||
"""
|
||||
|
||||
import unittest
|
||||
from types import SimpleNamespace
|
||||
from unittest.mock import patch
|
||||
|
||||
import torch
|
||||
|
||||
import sglang.srt.batch_overlap.two_batch_overlap as tbo
|
||||
from sglang.srt.batch_overlap.two_batch_overlap import TboForwardBatchPreparer
|
||||
from sglang.srt.model_executor.forward_batch_info import ForwardBatch, ForwardMode
|
||||
from sglang.srt.runtime_context import get_parallel
|
||||
from sglang.srt.runtime_context import get_context, get_parallel
|
||||
from sglang.test.ci.ci_register import register_cpu_ci
|
||||
from sglang.test.test_utils import CustomTestCase
|
||||
|
||||
register_cpu_ci(est_time=5, suite="base-a-test-cpu")
|
||||
|
||||
|
||||
import pytest as _pytest_defer
|
||||
|
||||
_DEFER_REASON = (
|
||||
"Temporarily skipped during the ServerArgs config-namespace migration; "
|
||||
"re-enabled once the runtime-config accessor API stabilizes."
|
||||
)
|
||||
pytestmark = _pytest_defer.mark.skip(reason=_DEFER_REASON)
|
||||
|
||||
|
||||
def setUpModule():
|
||||
import unittest
|
||||
|
||||
raise unittest.SkipTest(_DEFER_REASON)
|
||||
|
||||
|
||||
def _make_target_verify_batch(bs: int) -> ForwardBatch:
|
||||
return ForwardBatch(
|
||||
forward_mode=ForwardMode.TARGET_VERIFY,
|
||||
@@ -52,10 +34,11 @@ def _make_target_verify_batch(bs: int) -> ForwardBatch:
|
||||
|
||||
|
||||
def _filter(batch: ForwardBatch, *, lo: int, hi: int) -> ForwardBatch:
|
||||
fake_args = SimpleNamespace(moe_dense_tp_size=None, attention_backend="fa3")
|
||||
with get_parallel().override(attn_tp_size=1), patch.object(
|
||||
tbo, "get_server_args", lambda: fake_args
|
||||
):
|
||||
# filter_batch reads attention_backend (get_server_args) and
|
||||
# moe_dense_tp_size (get_parallel) from the published config.
|
||||
with get_context().override_server_args(
|
||||
attention_backend="fa3", moe_dense_tp_size=None
|
||||
), get_parallel().override(attn_tp_size=1):
|
||||
return TboForwardBatchPreparer.filter_batch(
|
||||
batch,
|
||||
start_token_index=lo,
|
||||
|
||||
@@ -30,20 +30,6 @@ from sglang.test.ci.ci_register import register_cpu_ci
|
||||
|
||||
register_cpu_ci(2.0, "base-a-test-cpu")
|
||||
|
||||
import pytest as _pytest_defer
|
||||
|
||||
_DEFER_REASON = (
|
||||
"Temporarily skipped during the ServerArgs config-namespace migration; "
|
||||
"re-enabled once the runtime-config accessor API stabilizes."
|
||||
)
|
||||
pytestmark = _pytest_defer.mark.skip(reason=_DEFER_REASON)
|
||||
|
||||
|
||||
def setUpModule():
|
||||
import unittest
|
||||
|
||||
raise unittest.SkipTest(_DEFER_REASON)
|
||||
|
||||
|
||||
register_cpu_ci(est_time=7, suite="base-c-test-cpu")
|
||||
|
||||
|
||||
@@ -5,30 +5,26 @@ from sglang.test.ci.ci_register import register_cpu_ci
|
||||
register_cpu_ci(est_time=5, suite="base-a-test-cpu")
|
||||
|
||||
|
||||
import pytest as _pytest_defer
|
||||
|
||||
_DEFER_REASON = (
|
||||
"Temporarily skipped during the ServerArgs config-namespace migration; "
|
||||
"re-enabled once the runtime-config accessor API stabilizes."
|
||||
)
|
||||
pytestmark = _pytest_defer.mark.skip(reason=_DEFER_REASON)
|
||||
|
||||
|
||||
def setUpModule():
|
||||
import unittest
|
||||
|
||||
raise unittest.SkipTest(_DEFER_REASON)
|
||||
|
||||
|
||||
import unittest
|
||||
from unittest.mock import MagicMock, call, patch
|
||||
|
||||
from sglang.srt.runtime_context import get_context
|
||||
from sglang.test.test_utils import CustomTestCase
|
||||
|
||||
|
||||
class TestRegisterToBootstrap(CustomTestCase):
|
||||
"""Tests for CommonKVManager.register_to_bootstrap retry/backoff behavior."""
|
||||
|
||||
def setUp(self):
|
||||
# register_to_bootstrap reads get_parallel().load_balance_method /
|
||||
# .enable_dsa_cache_layer_split and get_serving().port from the
|
||||
# published config.
|
||||
override = get_context().override_server_args(
|
||||
load_balance_method="follow_bootstrap_room", port=30000
|
||||
)
|
||||
override.install()
|
||||
self.addCleanup(override.restore)
|
||||
|
||||
@patch("sglang.srt.disaggregation.common.conn.time")
|
||||
@patch("sglang.srt.disaggregation.common.conn.requests.put")
|
||||
def test_succeeds_on_first_attempt(self, mock_put, mock_time):
|
||||
@@ -282,11 +278,8 @@ class TestRegisterToBootstrap(CustomTestCase):
|
||||
|
||||
mgr.kv_args = MagicMock()
|
||||
mgr.kv_args.page_size = 16
|
||||
|
||||
mgr.server_args = MagicMock()
|
||||
mgr.server_args.kv_cache_dtype = "auto"
|
||||
mgr.server_args.load_balance_method = "follow_bootstrap_room"
|
||||
mgr.server_args.port = 30000
|
||||
# Resolved per-runner value threaded through KVArgs (the payload field).
|
||||
mgr.kv_cache_dtype_str = "auto"
|
||||
|
||||
return mgr
|
||||
|
||||
|
||||
@@ -21,25 +21,12 @@ import unittest
|
||||
from types import SimpleNamespace
|
||||
from unittest import mock
|
||||
|
||||
from sglang.srt.runtime_context import get_context
|
||||
from sglang.test.ci.ci_register import register_cpu_ci, register_mlx_ci
|
||||
from sglang.test.test_utils import CustomTestCase
|
||||
|
||||
register_cpu_ci(est_time=1, suite="base-a-test-cpu")
|
||||
|
||||
import pytest as _pytest_defer
|
||||
|
||||
_DEFER_REASON = (
|
||||
"Temporarily skipped during the ServerArgs config-namespace migration; "
|
||||
"re-enabled once the runtime-config accessor API stabilizes."
|
||||
)
|
||||
pytestmark = _pytest_defer.mark.skip(reason=_DEFER_REASON)
|
||||
|
||||
|
||||
def setUpModule():
|
||||
import unittest
|
||||
|
||||
raise unittest.SkipTest(_DEFER_REASON)
|
||||
|
||||
|
||||
register_mlx_ci(est_time=1, suite="stage-a-unit-test-mlx")
|
||||
|
||||
@@ -69,9 +56,19 @@ def _arch(hybrid):
|
||||
)
|
||||
|
||||
|
||||
def _published(stub):
|
||||
"""Publish the config the resolver reads (get_schedule() /
|
||||
get_memory())."""
|
||||
return get_context().override_server_args(
|
||||
max_running_requests=stub._max_running_requests,
|
||||
max_mamba_cache_size=stub._max_mamba_cache_size,
|
||||
disable_radix_cache=stub._disable_radix_cache,
|
||||
)
|
||||
|
||||
|
||||
def _resolve(stub, hybrid=False):
|
||||
"""Run the resolver with the model architecture patched (see _arch)."""
|
||||
with _arch(hybrid):
|
||||
with _arch(hybrid), _published(stub):
|
||||
return stub._resolve_max_running_requests()
|
||||
|
||||
|
||||
@@ -85,13 +82,11 @@ def _stub(
|
||||
"""A stub carrying only what _resolve_max_running_requests reads."""
|
||||
stub = MlxModelRunnerStub.__new__(MlxModelRunnerStub)
|
||||
stub.model_config = SimpleNamespace() # only handed to the patched _arch fn
|
||||
stub.server_args = SimpleNamespace(
|
||||
max_running_requests=max_running_requests,
|
||||
max_mamba_cache_size=max_mamba_cache_size,
|
||||
disable_radix_cache=disable_radix_cache,
|
||||
)
|
||||
stub._max_running_requests = max_running_requests
|
||||
stub._max_mamba_cache_size = max_mamba_cache_size
|
||||
stub._disable_radix_cache = disable_radix_cache
|
||||
stub.max_total_num_tokens = max_total_num_tokens
|
||||
stub.dp_size = dp_size
|
||||
stub.ps = SimpleNamespace(attn_dp_size=dp_size)
|
||||
return stub
|
||||
|
||||
|
||||
@@ -101,14 +96,14 @@ def _hybrid_stub_for_initialize(
|
||||
"""A stub carrying what the real initialize() reads (hybrid path)."""
|
||||
stub = MlxModelRunnerStub.__new__(MlxModelRunnerStub)
|
||||
stub._mlx_pool_size = pool
|
||||
stub.dp_size = 1
|
||||
stub.ps = SimpleNamespace(attn_dp_size=1)
|
||||
stub.device = "cpu" # read by init_ngram_embedding_manager
|
||||
stub.server_args = SimpleNamespace(
|
||||
enable_memory_saver=False,
|
||||
max_running_requests=max_running_requests,
|
||||
max_mamba_cache_size=max_mamba_cache_size,
|
||||
disable_radix_cache=disable_radix_cache,
|
||||
)
|
||||
# Evaluated as a call argument in init_ngram_embedding_manager before
|
||||
# the use_ngram_embedding short-circuit; never read.
|
||||
stub.server_args = SimpleNamespace()
|
||||
stub._max_running_requests = max_running_requests
|
||||
stub._max_mamba_cache_size = max_mamba_cache_size
|
||||
stub._disable_radix_cache = disable_radix_cache
|
||||
stub.model_config = SimpleNamespace(
|
||||
is_hybrid_swa=False,
|
||||
sliding_window_size=None,
|
||||
@@ -220,7 +215,7 @@ class TestMlxHybridInitializeAllocation(CustomTestCase):
|
||||
stub = _hybrid_stub_for_initialize(
|
||||
max_running_requests=8, max_mamba_cache_size=2 * RATIO
|
||||
)
|
||||
with _arch(hybrid=True):
|
||||
with _arch(hybrid=True), _published(stub):
|
||||
stub.initialize()
|
||||
self.assertEqual(stub.max_running_requests, 2)
|
||||
pool = stub.req_to_token_pool
|
||||
@@ -233,7 +228,7 @@ class TestMlxHybridInitializeAllocation(CustomTestCase):
|
||||
stub = _hybrid_stub_for_initialize(
|
||||
max_running_requests=4, max_mamba_cache_size=2
|
||||
)
|
||||
with _arch(hybrid=True), self.assertRaisesRegex(
|
||||
with _arch(hybrid=True), _published(stub), self.assertRaisesRegex(
|
||||
RuntimeError, "max_mamba_cache_size"
|
||||
):
|
||||
stub.initialize()
|
||||
@@ -247,7 +242,7 @@ class TestMlxHybridInitializeAllocation(CustomTestCase):
|
||||
max_mamba_cache_size=8,
|
||||
disable_radix_cache=True,
|
||||
)
|
||||
with _arch(hybrid=True):
|
||||
with _arch(hybrid=True), _published(stub):
|
||||
stub.initialize()
|
||||
self.assertEqual(stub.max_running_requests, 8)
|
||||
pool = stub.req_to_token_pool
|
||||
@@ -262,7 +257,7 @@ class TestMlxHybridInitializeAllocation(CustomTestCase):
|
||||
max_mamba_cache_size=None,
|
||||
disable_radix_cache=True,
|
||||
)
|
||||
with _arch(hybrid=True):
|
||||
with _arch(hybrid=True), _published(stub):
|
||||
stub.initialize()
|
||||
self.assertEqual(stub.max_running_requests, 3)
|
||||
self.assertEqual(stub.req_to_token_pool.auxiliary_state_pool.size, 3)
|
||||
@@ -282,7 +277,7 @@ class TestMlxHybridInitializeAllocation(CustomTestCase):
|
||||
max_mamba_cache_size=2,
|
||||
disable_radix_cache=True,
|
||||
)
|
||||
with _arch(hybrid=True):
|
||||
with _arch(hybrid=True), _published(stub):
|
||||
stub.initialize()
|
||||
pool = stub.req_to_token_pool
|
||||
aux_capacity = pool.auxiliary_state_pool.available_size()
|
||||
@@ -305,7 +300,7 @@ class TestMlxHybridInitializeAllocation(CustomTestCase):
|
||||
max_mamba_cache_size=2 * RATIO,
|
||||
disable_radix_cache=False,
|
||||
)
|
||||
with _arch(hybrid=True):
|
||||
with _arch(hybrid=True), _published(stub):
|
||||
stub.initialize()
|
||||
pool = stub.req_to_token_pool
|
||||
free_before = pool.auxiliary_state_pool.available_size()
|
||||
@@ -322,7 +317,7 @@ class TestMlxHybridInitializeAllocation(CustomTestCase):
|
||||
stub = _hybrid_stub_for_initialize(
|
||||
max_running_requests=2, max_mamba_cache_size=None
|
||||
)
|
||||
with _arch(hybrid=True):
|
||||
with _arch(hybrid=True), _published(stub):
|
||||
stub.initialize()
|
||||
self.assertEqual(stub.max_running_requests, 2)
|
||||
self.assertEqual(stub.req_to_token_pool.auxiliary_state_pool.size, 2 * RATIO)
|
||||
|
||||
@@ -1,28 +1,13 @@
|
||||
import json
|
||||
from pathlib import Path
|
||||
from types import SimpleNamespace
|
||||
|
||||
from sglang.test.ci.ci_register import register_cpu_ci
|
||||
|
||||
register_cpu_ci(est_time=3, suite="base-a-test-cpu")
|
||||
|
||||
|
||||
import pytest as _pytest_defer
|
||||
|
||||
_DEFER_REASON = (
|
||||
"Temporarily skipped during the ServerArgs config-namespace migration; "
|
||||
"re-enabled once the runtime-config accessor API stabilizes."
|
||||
)
|
||||
pytestmark = _pytest_defer.mark.skip(reason=_DEFER_REASON)
|
||||
|
||||
|
||||
def setUpModule():
|
||||
import unittest
|
||||
|
||||
raise unittest.SkipTest(_DEFER_REASON)
|
||||
|
||||
|
||||
from sglang.srt.layers.moe.moe_runner.triton_utils import fused_moe_triton_config
|
||||
from sglang.srt.runtime_context import get_context
|
||||
|
||||
|
||||
def test_h200_bf16_config_is_available_for_current_triton_runtime():
|
||||
@@ -47,11 +32,6 @@ def test_down_moe_reuses_tuned_up_config_when_separate_config_is_absent(
|
||||
|
||||
monkeypatch.setenv("SGLANG_MOE_CONFIG_DIR", str(tmp_path))
|
||||
monkeypatch.setattr(fused_moe_triton_config.triton, "__version__", "3.6.0")
|
||||
monkeypatch.setattr(
|
||||
fused_moe_triton_config,
|
||||
"get_server_args",
|
||||
lambda: SimpleNamespace(enable_deterministic_inference=False),
|
||||
)
|
||||
monkeypatch.setattr(
|
||||
fused_moe_triton_config,
|
||||
"get_config_file_name",
|
||||
@@ -60,9 +40,11 @@ def test_down_moe_reuses_tuned_up_config_when_separate_config_is_absent(
|
||||
fused_moe_triton_config.get_moe_configs.cache_clear()
|
||||
|
||||
try:
|
||||
assert fused_moe_triton_config.get_moe_configs(
|
||||
32, 768, None, down_moe=True
|
||||
) == {128: {"BLOCK_SIZE_M": 64}}
|
||||
# get_moe_configs reads get_exec().deterministic.
|
||||
with get_context().override_server_args(enable_deterministic_inference=False):
|
||||
assert fused_moe_triton_config.get_moe_configs(
|
||||
32, 768, None, down_moe=True
|
||||
) == {128: {"BLOCK_SIZE_M": 64}}
|
||||
finally:
|
||||
fused_moe_triton_config.get_moe_configs.cache_clear()
|
||||
|
||||
|
||||
@@ -11,21 +11,6 @@ register_cuda_ci(est_time=9, stage="base-b", runner_config="1-gpu-small")
|
||||
register_amd_ci(est_time=1, suite="stage-b-test-1-gpu-small-amd")
|
||||
|
||||
|
||||
import pytest as _pytest_defer
|
||||
|
||||
_DEFER_REASON = (
|
||||
"Temporarily skipped during the ServerArgs config-namespace migration; "
|
||||
"re-enabled once the runtime-config accessor API stabilizes."
|
||||
)
|
||||
pytestmark = _pytest_defer.mark.skip(reason=_DEFER_REASON)
|
||||
|
||||
|
||||
def setUpModule():
|
||||
import unittest
|
||||
|
||||
raise unittest.SkipTest(_DEFER_REASON)
|
||||
|
||||
|
||||
class TestMmProcessConfigValidation(unittest.TestCase):
|
||||
"""Server-args validation for mm_process_config."""
|
||||
|
||||
|
||||
@@ -15,27 +15,13 @@ from sglang.srt.disaggregation.decode import ( # noqa: E402
|
||||
from sglang.srt.disaggregation.utils import DisaggregationMode # noqa: E402
|
||||
from sglang.srt.managers.schedule_batch import FINISH_ABORT, Req # noqa: E402
|
||||
from sglang.srt.managers.scheduler import Scheduler # noqa: E402
|
||||
from sglang.srt.runtime_context import get_context # noqa: E402
|
||||
from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
|
||||
|
||||
register_cuda_ci(est_time=5, stage="base-b", runner_config="1-gpu-small")
|
||||
register_amd_ci(est_time=5, suite="stage-b-test-1-gpu-small-amd")
|
||||
|
||||
|
||||
import pytest as _pytest_defer
|
||||
|
||||
_DEFER_REASON = (
|
||||
"Temporarily skipped during the ServerArgs config-namespace migration; "
|
||||
"re-enabled once the runtime-config accessor API stabilizes."
|
||||
)
|
||||
pytestmark = _pytest_defer.mark.skip(reason=_DEFER_REASON)
|
||||
|
||||
|
||||
def setUpModule():
|
||||
import unittest
|
||||
|
||||
raise unittest.SkipTest(_DEFER_REASON)
|
||||
|
||||
|
||||
class TestDisaggregationPriorityQueueing(unittest.TestCase):
|
||||
def _new_scheduler(self, disaggregation_mode: DisaggregationMode) -> Scheduler:
|
||||
scheduler = Scheduler.__new__(Scheduler)
|
||||
@@ -442,9 +428,8 @@ class TestDecodePrebuiltPriority(unittest.TestCase):
|
||||
scheduler.enable_overlap = False
|
||||
scheduler.spec_algorithm = MagicMock()
|
||||
scheduler.max_running_requests = 1
|
||||
scheduler.server_args = SimpleNamespace(
|
||||
disaggregation_decode_enable_radix_cache=False
|
||||
)
|
||||
# Passed whole into the (mocked) batch's process_prebuilt; never read.
|
||||
scheduler.server_args = SimpleNamespace()
|
||||
scheduler.future_map = MagicMock()
|
||||
scheduler.policy = MagicMock()
|
||||
scheduler.policy.calc_priority.side_effect = lambda waiting_queue, _: (
|
||||
@@ -452,10 +437,14 @@ class TestDecodePrebuiltPriority(unittest.TestCase):
|
||||
)
|
||||
|
||||
new_batch = MagicMock()
|
||||
# get_new_prebuilt_batch reads the published disagg config
|
||||
# (disaggregation_decode_enable_radix_cache).
|
||||
with patch(
|
||||
"sglang.srt.disaggregation.decode.ScheduleBatch.init_new",
|
||||
return_value=new_batch,
|
||||
) as init_new:
|
||||
) as init_new, get_context().override_server_args(
|
||||
disaggregation_decode_enable_radix_cache=False
|
||||
):
|
||||
ret = SchedulerDisaggregationDecodeMixin.get_new_prebuilt_batch(
|
||||
scheduler, scheduler.running_batch
|
||||
)
|
||||
|
||||
@@ -30,21 +30,6 @@ from sglang.srt.observability.req_time_stats import APIServerReqTimeStats
|
||||
register_cpu_ci(est_time=15, suite="base-a-test-cpu")
|
||||
|
||||
|
||||
import pytest as _pytest_defer
|
||||
|
||||
_DEFER_REASON = (
|
||||
"Temporarily skipped during the ServerArgs config-namespace migration; "
|
||||
"re-enabled once the runtime-config accessor API stabilizes."
|
||||
)
|
||||
pytestmark = _pytest_defer.mark.skip(reason=_DEFER_REASON)
|
||||
|
||||
|
||||
def setUpModule():
|
||||
import unittest
|
||||
|
||||
raise unittest.SkipTest(_DEFER_REASON)
|
||||
|
||||
|
||||
_NOT_FINISHED = object() # Sentinel: request has not finished yet
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
@@ -15,21 +15,6 @@ from sglang.test.ci.ci_register import register_cpu_ci
|
||||
register_cpu_ci(est_time=5, suite="base-a-test-cpu")
|
||||
|
||||
|
||||
import pytest as _pytest_defer
|
||||
|
||||
_DEFER_REASON = (
|
||||
"Temporarily skipped during the ServerArgs config-namespace migration; "
|
||||
"re-enabled once the runtime-config accessor API stabilizes."
|
||||
)
|
||||
pytestmark = _pytest_defer.mark.skip(reason=_DEFER_REASON)
|
||||
|
||||
|
||||
def setUpModule():
|
||||
import unittest
|
||||
|
||||
raise unittest.SkipTest(_DEFER_REASON)
|
||||
|
||||
|
||||
class _FakeAllocator:
|
||||
def __init__(self, base=1000, page_size=1):
|
||||
self.base = base
|
||||
|
||||
@@ -81,7 +81,7 @@ from sglang.srt.mem_cache.unified_radix_cache import (
|
||||
_OngoingPrefetch,
|
||||
_OngoingWriteThrough,
|
||||
)
|
||||
from sglang.srt.runtime_context import get_server_args
|
||||
from sglang.srt.runtime_context import get_server_args, get_serving
|
||||
from sglang.srt.sampling.sampling_params import SamplingParams
|
||||
from sglang.srt.server_args import (
|
||||
ServerArgs,
|
||||
@@ -95,21 +95,6 @@ register_cuda_ci(est_time=16, stage="base-b", runner_config="1-gpu-small")
|
||||
register_amd_ci(est_time=16, suite="stage-b-test-1-gpu-small-amd")
|
||||
|
||||
|
||||
import pytest as _pytest_defer
|
||||
|
||||
_DEFER_REASON = (
|
||||
"Temporarily skipped during the ServerArgs config-namespace migration; "
|
||||
"re-enabled once the runtime-config accessor API stabilizes."
|
||||
)
|
||||
pytestmark = _pytest_defer.mark.skip(reason=_DEFER_REASON)
|
||||
|
||||
|
||||
def setUpModule():
|
||||
import unittest
|
||||
|
||||
raise unittest.SkipTest(_DEFER_REASON)
|
||||
|
||||
|
||||
@dataclass(frozen=True)
|
||||
class CacheConfig:
|
||||
# Tree
|
||||
@@ -947,15 +932,13 @@ class UnifiedRadixCacheSuite:
|
||||
req.mamba_last_track_seqlen = kv_len
|
||||
req.reasoning_tokens = 1
|
||||
|
||||
get_server_args().strip_thinking_cache = True
|
||||
try:
|
||||
# cache_finished_req reads get_serving().strip_thinking_cache
|
||||
with get_serving().override(strip_thinking_cache=True):
|
||||
avail_before = allocator.available_size()
|
||||
cache.cache_finished_req(
|
||||
req, is_insert=True, kv_len_to_handle=req.effective_kv_committed_len()
|
||||
)
|
||||
start_p, end_p = req.effective_kv_committed_len(), req.kv.kv_allocated_len
|
||||
finally:
|
||||
get_server_args().strip_thinking_cache = False
|
||||
if ps > 1:
|
||||
start_p = ((start_p + ps - 1) // ps) * ps
|
||||
if start_p < end_p:
|
||||
@@ -5740,19 +5723,17 @@ class TestResumableInsertWalk(_InsertWalkSuite):
|
||||
self._insert(cache, allocator, req_to_token_pool, [1, 2, 3, 4])
|
||||
top = next(iter(cache.root_node.children.values()))
|
||||
|
||||
# A storage-prefetch completion host-inserts a backuped node below the
|
||||
# still-unbacked top, legitimately breaking backup continuity.
|
||||
host_indices = cache.cache_controller.mem_pool_host.alloc(8)
|
||||
host_result = cache.tree_core.insert_host(
|
||||
cache.root_node.id,
|
||||
RadixKey(array("q", list(range(1, 9)))),
|
||||
host_indices,
|
||||
[f"h{i}" for i in range(8)],
|
||||
)
|
||||
cache.cache_controller.mem_pool_host.free(
|
||||
host_indices[: host_result.prefix_len]
|
||||
)
|
||||
# Break backup continuity: a backuped (then device-evicted) middle
|
||||
# below the still-unbacked top. insert_host refills below an
|
||||
# un-backed-up node are dropped under write-through
|
||||
# (host_insert_dropped), so the state is built through an explicit
|
||||
# backup + device eviction.
|
||||
self._insert(cache, allocator, req_to_token_pool, list(range(1, 9)))
|
||||
middle = next(iter(top.children.values()))
|
||||
self.assertGreater(_write_backup(cache, middle, write_back=True), 0)
|
||||
cache.writing_check(write_back=True)
|
||||
cache.evict(EvictParams(num_tokens=4))
|
||||
self.assertTrue(middle.evicted)
|
||||
self.assertTrue(middle.backuped)
|
||||
self.assertFalse(top.backuped)
|
||||
|
||||
|
||||
@@ -17,21 +17,6 @@ from sglang.test.ci.ci_register import register_cpu_ci
|
||||
register_cpu_ci(est_time=10, suite="base-a-test-cpu")
|
||||
|
||||
|
||||
import pytest as _pytest_defer
|
||||
|
||||
_DEFER_REASON = (
|
||||
"Temporarily skipped during the ServerArgs config-namespace migration; "
|
||||
"re-enabled once the runtime-config accessor API stabilizes."
|
||||
)
|
||||
pytestmark = _pytest_defer.mark.skip(reason=_DEFER_REASON)
|
||||
|
||||
|
||||
def setUpModule():
|
||||
import unittest
|
||||
|
||||
raise unittest.SkipTest(_DEFER_REASON)
|
||||
|
||||
|
||||
@contextlib.contextmanager
|
||||
def mock_cpu_env(kv_size=2, tp_size=1, swa_eviction_interval=4):
|
||||
"""Mock GPU-dependent functions for CPU-only testing.
|
||||
@@ -116,10 +101,12 @@ def _make_model_runner(
|
||||
mc.hf_config = SimpleNamespace(architectures=["LlamaForCausalLM"])
|
||||
mc.hf_config.get_text_config = lambda: mc.hf_config
|
||||
mc.linear_attn_registry_result = None
|
||||
mc.context_len = 8192
|
||||
mr.model_config = mc
|
||||
mr.kv_cache_dtype = "fake_bf16"
|
||||
|
||||
sa = SimpleNamespace()
|
||||
sa.max_total_tokens = None
|
||||
sa.swa_full_tokens_ratio = swa_full_tokens_ratio
|
||||
sa.page_size = page_size
|
||||
sa.disable_radix_cache = disable_radix_cache
|
||||
|
||||
@@ -2,40 +2,16 @@ import unittest
|
||||
from types import SimpleNamespace
|
||||
|
||||
from sglang.srt.models.deepseek_v4 import DeepseekV4ForCausalLM
|
||||
from sglang.srt.runtime_context import get_context, reset_context
|
||||
from sglang.srt.server_args import ServerArgs
|
||||
from sglang.srt.runtime_context import get_context, get_exec
|
||||
from sglang.test.ci.ci_register import register_cpu_ci
|
||||
|
||||
register_cpu_ci(est_time=4, suite="base-a-test-cpu")
|
||||
|
||||
|
||||
import pytest as _pytest_defer
|
||||
|
||||
_DEFER_REASON = (
|
||||
"Temporarily skipped during the ServerArgs config-namespace migration; "
|
||||
"re-enabled once the runtime-config accessor API stabilizes."
|
||||
)
|
||||
pytestmark = _pytest_defer.mark.skip(reason=_DEFER_REASON)
|
||||
|
||||
|
||||
def setUpModule():
|
||||
import unittest
|
||||
|
||||
raise unittest.SkipTest(_DEFER_REASON)
|
||||
|
||||
|
||||
class TestDeepseekV4SharedExpertFusionPolicy(unittest.TestCase):
|
||||
"""The disable decision is a load-time resolution: it writes through to
|
||||
the published config via declare_load_time_override."""
|
||||
|
||||
def setUp(self):
|
||||
self._saved_server_args = get_context()._server_args
|
||||
|
||||
def tearDown(self):
|
||||
if self._saved_server_args is None:
|
||||
reset_context()
|
||||
else:
|
||||
get_context().set_server_args(self._saved_server_args)
|
||||
"""The disable decision is a load-time resolution: it lands on the
|
||||
published config bag via declare_load_time_override (bag-only; the
|
||||
ServerArgs instance stays pristine)."""
|
||||
|
||||
def _make_model(self, n_shared_experts=1):
|
||||
return SimpleNamespace(
|
||||
@@ -43,29 +19,30 @@ class TestDeepseekV4SharedExpertFusionPolicy(unittest.TestCase):
|
||||
)
|
||||
|
||||
def _publish(self, enforce):
|
||||
server_args = ServerArgs(model_path="dummy")
|
||||
server_args.enforce_shared_experts_fusion = enforce
|
||||
get_context().set_server_args(server_args)
|
||||
return server_args
|
||||
override = get_context().override_server_args(
|
||||
enforce_shared_experts_fusion=enforce
|
||||
)
|
||||
override.install()
|
||||
self.addCleanup(override.restore)
|
||||
|
||||
def test_disables_shared_fusion_without_enforce(self):
|
||||
server_args = self._publish(enforce=False)
|
||||
self._publish(enforce=False)
|
||||
model = self._make_model()
|
||||
|
||||
DeepseekV4ForCausalLM.determine_num_fused_shared_experts(model)
|
||||
|
||||
self.assertEqual(model.num_fused_shared_experts, 0)
|
||||
# post-init declaration writes through to the published config
|
||||
self.assertTrue(server_args.disable_shared_experts_fusion)
|
||||
# post-init declaration lands on the published config bag
|
||||
self.assertTrue(get_exec().moe.disable_shared_experts_fusion)
|
||||
|
||||
def test_enables_shared_fusion_when_enforced(self):
|
||||
server_args = self._publish(enforce=True)
|
||||
self._publish(enforce=True)
|
||||
model = self._make_model()
|
||||
|
||||
DeepseekV4ForCausalLM.determine_num_fused_shared_experts(model)
|
||||
|
||||
self.assertEqual(model.num_fused_shared_experts, 1)
|
||||
self.assertFalse(server_args.disable_shared_experts_fusion)
|
||||
self.assertFalse(get_exec().moe.disable_shared_experts_fusion)
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
|
||||
@@ -13,6 +13,7 @@ from unittest.mock import MagicMock, patch
|
||||
import torch
|
||||
|
||||
from sglang.srt.model_executor.forward_batch_info import ForwardMode
|
||||
from sglang.srt.runtime_context import get_context
|
||||
from sglang.srt.speculative.adaptive_runtime_state import SpecRuntimeState
|
||||
from sglang.srt.speculative.eagle_utils import organize_draft_results
|
||||
from sglang.srt.speculative.eagle_worker_v2 import EagleDraftWorker, EAGLEWorkerV2
|
||||
@@ -26,20 +27,6 @@ from sglang.test.test_utils import CustomTestCase
|
||||
register_cuda_ci(est_time=20, stage="base-b", runner_config="1-gpu-small")
|
||||
register_amd_ci(est_time=20, stage="stage-b", runner_config="1-gpu-small-amd")
|
||||
|
||||
import pytest as _pytest_defer
|
||||
|
||||
_DEFER_REASON = (
|
||||
"Temporarily skipped during the ServerArgs config-namespace migration; "
|
||||
"re-enabled once the runtime-config accessor API stabilizes."
|
||||
)
|
||||
pytestmark = _pytest_defer.mark.skip(reason=_DEFER_REASON)
|
||||
|
||||
|
||||
def setUpModule():
|
||||
import unittest
|
||||
|
||||
raise unittest.SkipTest(_DEFER_REASON)
|
||||
|
||||
|
||||
register_cpu_ci(est_time=20, suite="base-a-test-cpu")
|
||||
|
||||
@@ -109,6 +96,15 @@ def _make_backend_factory(decode_backend, draft_extend_backend, captured_kwargs=
|
||||
|
||||
|
||||
class TestEagleWorkerV2Topk1FastPath(CustomTestCase):
|
||||
def setUp(self):
|
||||
# _rebuild_topk1_chain_buffers sizes its preallocation from the
|
||||
# published config: get_exec().graph.cuda_graph_config stays None on
|
||||
# the dummy-boundary publish (no resolution), so
|
||||
# get_schedule().max_running_requests alone sizes the buffers.
|
||||
override = get_context().override_server_args(max_running_requests=8)
|
||||
override.install()
|
||||
self.addCleanup(override.restore)
|
||||
|
||||
def test_fast_path_matches_slow_path(self):
|
||||
bs = 3
|
||||
for num_steps in (1, 2, 3, 4):
|
||||
@@ -145,6 +141,13 @@ class TestEagleWorkerV2Topk1FastPath(CustomTestCase):
|
||||
|
||||
|
||||
class TestEagleWorkerV2BackendFallback(CustomTestCase):
|
||||
def setUp(self):
|
||||
# The adaptive state-machine paths write live spec switches through
|
||||
# get_context().override, which needs a published config.
|
||||
override = get_context().override_server_args()
|
||||
override.install()
|
||||
self.addCleanup(override.restore)
|
||||
|
||||
def test_missing_seed_cuda_graph_fallback(self):
|
||||
graph_result = (
|
||||
[],
|
||||
|
||||
@@ -1,53 +0,0 @@
|
||||
"""Ratchet guard: config-namespace-migration test deferrals may only decrease.
|
||||
|
||||
A set of unit tests was module-skipped ("Temporarily skipped during the
|
||||
ServerArgs config-namespace migration") because their fixtures inject config
|
||||
in ways the namespace accessors cannot see; they are recovered together with
|
||||
the reader migration. Nothing else enforces that list: without this pin a new
|
||||
skip can ride in unnoticed, and the already-skipped files keep growing test
|
||||
code nobody has ever run. The count is exact: un-skipping a file must lower
|
||||
the baseline to lock in the recovery, and no new deferral may appear.
|
||||
"""
|
||||
|
||||
from sglang.test.ci.ci_register import register_cpu_ci
|
||||
|
||||
register_cpu_ci(est_time=5, suite="base-a-test-cpu")
|
||||
|
||||
import unittest
|
||||
from pathlib import Path
|
||||
|
||||
from sglang.test.test_utils import CustomTestCase
|
||||
|
||||
# test/registered/unit/<this file> -> test/
|
||||
_TEST_ROOT = Path(__file__).resolve().parents[2]
|
||||
|
||||
_MARKER = "config-namespace migration"
|
||||
|
||||
_BASELINE = 15
|
||||
|
||||
|
||||
class TestMigrationDeferralRatchet(CustomTestCase):
|
||||
def test_deferred_test_files_match_the_baseline(self):
|
||||
deferred = sorted(
|
||||
p.relative_to(_TEST_ROOT).as_posix()
|
||||
for p in _TEST_ROOT.rglob("*.py")
|
||||
if p.name != Path(__file__).name and _MARKER in p.read_text(errors="ignore")
|
||||
)
|
||||
count = len(deferred)
|
||||
if count > _BASELINE:
|
||||
self.fail(
|
||||
f"deferred (module-skipped) migration tests grew: {count} > "
|
||||
f"baseline {_BASELINE}: {deferred}. Do not add new deferrals — "
|
||||
"seed the context (override_server_args / publish a real "
|
||||
"ServerArgs) instead of skipping the module."
|
||||
)
|
||||
if count < _BASELINE:
|
||||
self.fail(
|
||||
f"deferred migration tests shrank: {count} < baseline "
|
||||
f"{_BASELINE}. Lower the baseline in this file to lock in the "
|
||||
"recovery."
|
||||
)
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
unittest.main()
|
||||
Reference in New Issue
Block a user