config: retire the last process-global config field reads (#33338)
`get_server_args().<field>` reads one process's startup record. Nine sites still did that for a value that has a namespace: the attention backend (5), `skip_tokenizer_init` (2), the draft-aware `load_format`, and a chunked-prefill size in `sglang.kernels`. They now read `get_exec().kernel` / `get_serving()` / `get_model()` / `get_schedule()`, so they see the resolved value including post-publish overrides. The multimodal processor's device selection moves to the instance it was constructed with rather than to a namespace: `base_gpu_id` differs per worker (the encode-server DP workers each specialise their own copy), so no process-global value can stand in for it, and engines sharing a tokenizer process each need their own. Branch order, the NPU preprocess patches, and the case that leaves "device" unset are unchanged. What stays on `get_server_args()` is the derived API — `@property` and method members computed from several fields plus the HF config (`mamba_cache_chunk_size`, `get_model_config()`, `enable_mamba_extra_buffer*`) — plus three config-intent reads of live-shadowed sizes, each of which needs an answer the live topology property cannot give (the DSA indexer's PP gate must short-circuit before touching the PP group, `allocation`'s DCP gate asks whether DCP was configured at all, and the CUDA-IPC recycler runs where no group exists). A new AST ratchet pins both shapes it can see — the direct call and an alias bound from it in the same function — at 0 and 12 respectively, exempting the derived APIs and those three sites by name. The alias-form baseline is not zero: those reads are mostly per-runner fields in model code, and lowering them is the next slice. Two fixtures stopped faking config: `test_dllm_fdfo_kv_reuse` rebound `allocation.get_server_args` to a SimpleNamespace, which silently stops intercepting the moment a reader migrates; it publishes a real config instead.
This commit is contained in:
Cheng Wan
@@ -7,9 +7,9 @@ from types import SimpleNamespace
|
||||
import torch
|
||||
|
||||
from sglang.srt.dllm.mixin.scheduler import DllmManager
|
||||
from sglang.srt.mem_cache import allocation
|
||||
from sglang.srt.mem_cache.allocation import alloc_for_extend
|
||||
from sglang.srt.mem_cache.memory_pool import ReqToTokenPool
|
||||
from sglang.srt.runtime_context import get_context
|
||||
from sglang.test.ci.ci_register import register_cpu_ci
|
||||
|
||||
register_cpu_ci(est_time=5, suite="base-a-test-cpu")
|
||||
@@ -117,16 +117,11 @@ class TestDllmFdfoKvReuse(unittest.TestCase):
|
||||
self.pool = ReqToTokenPool(
|
||||
size=8, max_context_len=64, device="cpu", enable_memory_saver=False
|
||||
)
|
||||
self._old_support_triton = allocation.support_triton
|
||||
self._old_get_server_args = allocation.get_server_args
|
||||
allocation.support_triton = lambda _: False
|
||||
allocation.get_server_args = lambda: SimpleNamespace(
|
||||
override = get_context().override_server_args(
|
||||
attention_backend="torch_native", dcp_size=1
|
||||
)
|
||||
|
||||
def tearDown(self):
|
||||
allocation.support_triton = self._old_support_triton
|
||||
allocation.get_server_args = self._old_get_server_args
|
||||
override.install()
|
||||
self.addCleanup(override.restore)
|
||||
|
||||
def test_alloc_for_extend_mixed_reuse_allocates_only_fresh_and_writes_rows(self):
|
||||
allocator = _FakeAllocator(base=200)
|
||||
|
||||
@@ -0,0 +1,81 @@
|
||||
"""The fast-image-processor device comes from the processor's own ServerArgs.
|
||||
|
||||
Regression: the device decision read the published global ServerArgs, which is
|
||||
last-publish-wins. Two engines in one tokenizer process then shared whichever
|
||||
config published last, so one engine's images were preprocessed on the other
|
||||
engine's GPU.
|
||||
"""
|
||||
|
||||
import unittest
|
||||
from unittest.mock import patch
|
||||
|
||||
from sglang.srt.multimodal.processors.base_processor import BaseMultimodalProcessor
|
||||
from sglang.srt.server_args import ServerArgs
|
||||
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")
|
||||
|
||||
BASE = "sglang.srt.multimodal.processors.base_processor"
|
||||
|
||||
|
||||
class _Processor:
|
||||
pass
|
||||
|
||||
|
||||
class _StubProcessor(BaseMultimodalProcessor):
|
||||
async def process_mm_data_async(self, *args, **kwargs):
|
||||
raise NotImplementedError
|
||||
|
||||
|
||||
def _make(**fields):
|
||||
processor = _StubProcessor.__new__(_StubProcessor)
|
||||
processor.server_args = ServerArgs(model_path="dummy", **fields)
|
||||
return processor
|
||||
|
||||
|
||||
class TestFastImageProcessorDevice(CustomTestCase):
|
||||
def _device(self, processor, **platform):
|
||||
flags = {"_is_cpu": False, "_is_xpu": False, "_is_npu": False}
|
||||
flags.update(platform)
|
||||
with patch.multiple(BASE, **flags):
|
||||
return processor._fast_image_processor_device(_Processor())
|
||||
|
||||
def test_device_follows_the_instance_base_gpu_id(self):
|
||||
self.assertEqual(self._device(_make(base_gpu_id=3)), "cuda:3")
|
||||
|
||||
def test_engines_in_one_process_keep_their_own_device(self):
|
||||
first, second = _make(base_gpu_id=0), _make(base_gpu_id=5)
|
||||
self.assertEqual(self._device(first), "cuda:0")
|
||||
self.assertEqual(self._device(second), "cuda:5")
|
||||
|
||||
def test_publishing_another_config_does_not_move_the_device(self):
|
||||
from sglang.srt.runtime_context import get_context
|
||||
|
||||
processor = _make(base_gpu_id=2)
|
||||
override = get_context().override_server_args(base_gpu_id=7)
|
||||
override.install()
|
||||
self.addCleanup(override.restore)
|
||||
self.assertEqual(self._device(processor), "cuda:2")
|
||||
|
||||
def test_rl_on_policy_target_forces_cpu(self):
|
||||
processor = _make(base_gpu_id=3, rl_on_policy_target="fsdp")
|
||||
self.assertEqual(self._device(processor), "cpu")
|
||||
|
||||
def test_cpu_and_xpu_platforms_win_over_base_gpu_id(self):
|
||||
processor = _make(base_gpu_id=3)
|
||||
self.assertEqual(self._device(processor, _is_cpu=True), "cpu")
|
||||
self.assertEqual(self._device(processor, _is_xpu=True), "xpu")
|
||||
|
||||
def test_npu_glm4v_leaves_the_device_unset(self):
|
||||
class Glm4vProcessor:
|
||||
pass
|
||||
|
||||
processor = _make(base_gpu_id=3)
|
||||
with patch.multiple(BASE, _is_cpu=False, _is_xpu=False, _is_npu=True):
|
||||
device = processor._fast_image_processor_device(Glm4vProcessor())
|
||||
self.assertIsNone(device)
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
unittest.main()
|
||||
@@ -0,0 +1,165 @@
|
||||
"""Ratchet guard: process-global config reads may only decrease.
|
||||
|
||||
``get_server_args()`` returns the published ``ServerArgs`` — one process's
|
||||
startup record. Config decisions read the namespace accessors instead
|
||||
(``get_exec()`` / ``get_memory()`` / …), which carry the resolved value
|
||||
including post-publish overrides, and per-runner values come from the runner
|
||||
that owns them.
|
||||
|
||||
Two shapes count as a read: the direct ``get_server_args().field``, and the
|
||||
alias ``sa = get_server_args()`` followed by ``sa.field`` in the same function.
|
||||
A whole-object pass (``def f(server_args)``) is not a global read and is not
|
||||
counted — there the caller decided which instance to hand over.
|
||||
|
||||
What legitimately remains:
|
||||
|
||||
- **Derived APIs.** ``@property`` and method members of ``ServerArgs``
|
||||
(``mamba_cache_chunk_size``, ``get_model_config()``,
|
||||
``enable_mamba_extra_buffer()``, …) are computed from several fields plus the
|
||||
HF config, so they are not namespace leaves and ``ServerArgs`` is their only
|
||||
home. Exempt by name below.
|
||||
- **Config-intent reads of live-shadowed sizes.** ``get_parallel()`` shadows
|
||||
``tp/pp/dcp/attn_cp/moe_dp_size`` with the live topology, so a config-intent
|
||||
read of one has nowhere else to go. Each exempt site needs an answer the live
|
||||
property cannot give:
|
||||
|
||||
- ``dsa_indexer.pp_size`` gates ``pp_size > 1 and not get_pp_group()...``, and
|
||||
the short circuit is the point: with PP off the group is never touched, which
|
||||
is what lets the ``Indexer`` be constructed before distributed init. The live
|
||||
property would demand the group either way.
|
||||
- ``allocation.dcp_size`` asks whether DCP was *configured*; the live property
|
||||
reads ``get_dcp_group()``, and that group is only installed when DCP is on.
|
||||
- ``cuda_ipc_transport_utils.tp_size`` runs in the tokenizer process, which has
|
||||
no groups at all (the call site already guards for "not published yet").
|
||||
- The alias-form baseline is not zero yet. Lowering it is the next slice; the
|
||||
failure message lists the sites whenever the count moves.
|
||||
"""
|
||||
|
||||
from sglang.test.ci.ci_register import register_cpu_ci
|
||||
|
||||
register_cpu_ci(est_time=5, suite="base-a-test-cpu")
|
||||
|
||||
import ast
|
||||
import unittest
|
||||
from pathlib import Path
|
||||
|
||||
import sglang
|
||||
from sglang.test.test_utils import CustomTestCase
|
||||
|
||||
# srt is the migrated surface; the rest of the package has no reads today and is
|
||||
# scanned so a new one cannot appear there unnoticed.
|
||||
_PACKAGE_ROOT = Path(next(iter(sglang.__path__)))
|
||||
|
||||
_DERIVED_MEMBERS = frozenset(
|
||||
{
|
||||
"cutedsl_moe_max_num_tokens",
|
||||
"enable_mamba_extra_buffer",
|
||||
"enable_mamba_extra_buffer_lazy",
|
||||
"get_attention_backends",
|
||||
"get_model_config",
|
||||
"mamba_cache_chunk_size",
|
||||
"max_speculative_num_draft_tokens",
|
||||
"model_config",
|
||||
"use_mla_backend",
|
||||
}
|
||||
)
|
||||
|
||||
_CONFIG_INTENT_SIZES = frozenset(
|
||||
{
|
||||
("srt/layers/attention/dsa/dsa_indexer.py", "pp_size"),
|
||||
("srt/mem_cache/allocation.py", "dcp_size"),
|
||||
("srt/utils/cuda_ipc_transport_utils.py", "tp_size"),
|
||||
}
|
||||
)
|
||||
|
||||
_DIRECT_BASELINE = 0
|
||||
_ALIAS_BASELINE = 12
|
||||
|
||||
|
||||
def _is_global_call(node) -> bool:
|
||||
return (
|
||||
isinstance(node, ast.Call)
|
||||
and isinstance(node.func, ast.Name)
|
||||
and node.func.id == "get_server_args"
|
||||
)
|
||||
|
||||
|
||||
def _collect(rel: str, tree: ast.AST):
|
||||
"""The (direct, alias) field reads in one module."""
|
||||
direct, alias = [], []
|
||||
|
||||
def counted(attr: str) -> bool:
|
||||
return attr not in _DERIVED_MEMBERS and (rel, attr) not in _CONFIG_INTENT_SIZES
|
||||
|
||||
for node in ast.walk(tree):
|
||||
if (
|
||||
isinstance(node, ast.Attribute)
|
||||
and _is_global_call(node.value)
|
||||
and counted(node.attr)
|
||||
):
|
||||
direct.append(f"{rel}:{node.lineno}: get_server_args().{node.attr}")
|
||||
|
||||
if not isinstance(node, (ast.FunctionDef, ast.AsyncFunctionDef)):
|
||||
continue
|
||||
params = {a.arg for a in list(node.args.args) + list(node.args.kwonlyargs)}
|
||||
bound = {}
|
||||
for inner in ast.walk(node):
|
||||
if isinstance(inner, ast.Assign) and _is_global_call(inner.value):
|
||||
for target in inner.targets:
|
||||
if isinstance(target, ast.Name) and target.id not in params:
|
||||
bound.setdefault(target.id, inner.lineno)
|
||||
if not bound:
|
||||
continue
|
||||
for inner in ast.walk(node):
|
||||
if (
|
||||
isinstance(inner, ast.Attribute)
|
||||
and isinstance(inner.value, ast.Name)
|
||||
and inner.value.id in bound
|
||||
and inner.lineno >= bound[inner.value.id]
|
||||
and counted(inner.attr)
|
||||
):
|
||||
alias.append(
|
||||
f"{rel}:{inner.lineno}: {inner.value.id}.{inner.attr} "
|
||||
f"(bound from get_server_args() at line {bound[inner.value.id]})"
|
||||
)
|
||||
return direct, alias
|
||||
|
||||
|
||||
def _field_reads():
|
||||
direct, alias = [], []
|
||||
for path in sorted(_PACKAGE_ROOT.rglob("*.py")):
|
||||
rel = path.relative_to(_PACKAGE_ROOT).as_posix()
|
||||
try:
|
||||
tree = ast.parse(path.read_text())
|
||||
except SyntaxError:
|
||||
continue
|
||||
module_direct, module_alias = _collect(rel, tree)
|
||||
direct += module_direct
|
||||
alias += module_alias
|
||||
return direct, alias
|
||||
|
||||
|
||||
class TestGlobalConfigReadRatchet(CustomTestCase):
|
||||
def _check(self, kind, reads, baseline):
|
||||
if len(reads) > baseline:
|
||||
self.fail(
|
||||
f"{kind} process-global config field reads grew: {len(reads)} > "
|
||||
f"baseline {baseline}. Read the namespace accessor for the "
|
||||
"field's namespace, or the owning runner for a per-runner "
|
||||
"field:\n" + "\n".join(reads)
|
||||
)
|
||||
if len(reads) < baseline:
|
||||
self.fail(
|
||||
f"{kind} process-global config field reads shrank: {len(reads)} < "
|
||||
f"baseline {baseline}. Lower the baseline in this file to lock "
|
||||
"in the progress."
|
||||
)
|
||||
|
||||
def test_global_field_reads_match_the_baseline(self):
|
||||
direct, alias = _field_reads()
|
||||
self._check("direct", direct, _DIRECT_BASELINE)
|
||||
self._check("alias-form", alias, _ALIAS_BASELINE)
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
unittest.main()
|
||||
Reference in New Issue
Block a user