1823 lines
74 KiB
Python
1823 lines
74 KiB
Python
"""Unit tests for runtime_context: delegation, singletons, and override()."""
|
|
|
|
from sglang.test.ci.ci_register import register_cpu_ci
|
|
|
|
register_cpu_ci(est_time=33, suite="base-a-test-cpu")
|
|
|
|
import dataclasses
|
|
import json
|
|
import os
|
|
import pathlib as _pathlib
|
|
import shutil
|
|
import tempfile
|
|
import unittest
|
|
import warnings
|
|
from unittest.mock import patch
|
|
|
|
import msgspec
|
|
import msgspec.structs
|
|
|
|
import sglang as _sglang
|
|
import sglang.srt.server_args as server_args_module
|
|
from sglang.srt.arg_groups import prefill_buffer_ceiling
|
|
from sglang.srt.arg_groups.arg_utils import NS, A, Arg
|
|
from sglang.srt.arg_groups.model_override_base import resolving_view
|
|
from sglang.srt.arg_groups.overrides import (
|
|
attention_backends_of,
|
|
)
|
|
from sglang.srt.arg_groups.overrides import (
|
|
mamba_cache_chunk_size as mamba_cache_chunk_size_of,
|
|
)
|
|
from sglang.srt.arg_groups.overrides import (
|
|
max_prefill_buffer_tokens as max_prefill_buffer_tokens_of,
|
|
)
|
|
from sglang.srt.arg_groups.overrides import (
|
|
resolution_result,
|
|
resolved_view,
|
|
)
|
|
from sglang.srt.runtime_context import (
|
|
Flags,
|
|
ParallelContext,
|
|
RuntimeContext,
|
|
_FlagGroupBase,
|
|
assert_published,
|
|
derive_parallel_widths,
|
|
get_context,
|
|
get_exec,
|
|
get_flags,
|
|
get_parallel,
|
|
get_schedule,
|
|
get_server_args,
|
|
max_prefill_buffer_tokens,
|
|
max_speculative_num_draft_tokens,
|
|
publish,
|
|
publish_role,
|
|
reset_context,
|
|
)
|
|
from sglang.srt.server_args import ServerArgs
|
|
from sglang.test.test_utils import CustomTestCase
|
|
|
|
_SRT = _pathlib.Path(next(iter(_sglang.__path__))).resolve() / "srt"
|
|
_PS = "sglang.srt.distributed.parallel_state"
|
|
_DP = "sglang.srt.layers.dp_attention"
|
|
|
|
# Ranks and the world size read the live group: they are not implied by
|
|
# anything, so there is nothing to derive them from. The quotients used to be
|
|
# in this table and are not any more -- `attn_tp_size` and its siblings are
|
|
# functions of the configured leaves, and `TestDerivedWidthsComeFromTheLeaves`
|
|
# is what pins them.
|
|
SIZE_RANK_DELEGATIONS = [
|
|
("world_size", f"{_PS}.get_world_size"),
|
|
("world_rank", f"{_PS}.get_world_rank"),
|
|
("tp_rank", f"{_PS}.get_tensor_model_parallel_rank"),
|
|
("dcp_rank", f"{_PS}.get_dcp_rank"),
|
|
("pp_rank", f"{_PS}.get_pipeline_model_parallel_rank"),
|
|
("moe_ep_rank", f"{_PS}.get_moe_expert_parallel_rank"),
|
|
("moe_dp_rank", f"{_PS}.get_moe_data_parallel_rank"),
|
|
("moe_tp_rank", f"{_PS}.get_moe_tensor_parallel_rank"),
|
|
("attn_tp_rank", f"{_PS}.get_attn_tensor_model_parallel_rank"),
|
|
("attn_cp_rank", f"{_PS}.get_attn_context_model_parallel_rank"),
|
|
("attn_dp_rank", f"{_DP}.get_attention_dp_rank"),
|
|
]
|
|
|
|
GROUP_DELEGATIONS = [
|
|
("world_group", f"{_PS}.get_world_group"),
|
|
("tp_group", f"{_PS}.get_tp_group"),
|
|
("dcp_group", f"{_PS}.get_dcp_group"),
|
|
("pp_group", f"{_PS}.get_pp_group"),
|
|
("moe_ep_group", f"{_PS}.get_moe_ep_group"),
|
|
("moe_dp_group", f"{_PS}.get_moe_dp_group"),
|
|
("moe_tp_group", f"{_PS}.get_moe_tp_group"),
|
|
("attn_tp_group", f"{_PS}.get_attn_tp_group"),
|
|
("attn_cp_group", f"{_PS}.get_attn_cp_group"),
|
|
]
|
|
|
|
|
|
class TestRuntimeContextSingletons(CustomTestCase):
|
|
def test_singletons(self):
|
|
self.assertIs(get_parallel(), get_parallel())
|
|
self.assertIsInstance(get_parallel(), ParallelContext)
|
|
self.assertIsInstance(get_context(), RuntimeContext)
|
|
self.assertIs(get_context().parallel, get_parallel())
|
|
|
|
|
|
class _IsolatedOverrides(CustomTestCase):
|
|
"""Give each test a clean override map, restoring afterward only the overrides
|
|
installed outside it (e.g. by another test file sharing the process)."""
|
|
|
|
def setUp(self):
|
|
super().setUp()
|
|
p = get_parallel()
|
|
self._saved_overrides = dict(p._overrides)
|
|
p._overrides.clear()
|
|
|
|
def tearDown(self):
|
|
p = get_parallel()
|
|
p._overrides.clear()
|
|
p._overrides.update(self._saved_overrides)
|
|
super().tearDown()
|
|
|
|
|
|
class TestParallelDelegation(_IsolatedOverrides):
|
|
def test_size_rank_delegate_to_canonical_getters(self):
|
|
# Patch each getter to a distinct sentinel: a miswired attribute would read
|
|
# a different (unpatched) getter and fail.
|
|
for i, (attr, target) in enumerate(SIZE_RANK_DELEGATIONS):
|
|
sentinel = 1000 + i
|
|
with patch(target, return_value=sentinel):
|
|
self.assertEqual(
|
|
getattr(get_parallel(), attr),
|
|
sentinel,
|
|
msg=f"{attr} must delegate to {target}",
|
|
)
|
|
|
|
def test_groups_delegate_to_canonical_getters(self):
|
|
for attr, target in GROUP_DELEGATIONS:
|
|
sentinel = object()
|
|
with patch(target, return_value=sentinel):
|
|
self.assertIs(
|
|
getattr(get_parallel(), attr),
|
|
sentinel,
|
|
msg=f"{attr} must delegate to {target}",
|
|
)
|
|
|
|
def test_wrapper_holds_no_resolved_state(self):
|
|
# __slots__: no __dict__; the only instance state is the override hook.
|
|
self.assertFalse(hasattr(get_parallel(), "__dict__"))
|
|
# tp_group IS exposed: live delegation handles PD-multiplexing / the tp patch.
|
|
self.assertTrue(hasattr(ParallelContext, "tp_group"))
|
|
# local_attn_dp is intentionally not part of the wrapper surface.
|
|
self.assertFalse(hasattr(ParallelContext, "local_attn_dp_size"))
|
|
|
|
|
|
class TestParallelOverride(_IsolatedOverrides):
|
|
def test_override_takes_precedence(self):
|
|
p = get_parallel()
|
|
with p.override(tp_size=99, tp_rank=3, attn_dp_size=8):
|
|
self.assertEqual(p.tp_size, 99)
|
|
self.assertEqual(p.tp_rank, 3)
|
|
self.assertEqual(p.attn_dp_size, 8)
|
|
# same singleton: a fresh get_parallel() sees the override too
|
|
self.assertEqual(get_parallel().tp_size, 99)
|
|
self.assertEqual(p._overrides, {})
|
|
|
|
def test_override_can_force_groups(self):
|
|
sentinel = object()
|
|
with get_parallel().override(tp_group=sentinel):
|
|
self.assertIs(get_parallel().tp_group, sentinel)
|
|
|
|
def test_override_nests_and_restores(self):
|
|
p = get_parallel()
|
|
with p.override(tp_size=2):
|
|
self.assertEqual(p.tp_size, 2)
|
|
with p.override(tp_size=4, pp_size=2):
|
|
self.assertEqual(p.tp_size, 4)
|
|
self.assertEqual(p.pp_size, 2)
|
|
self.assertEqual(p.tp_size, 2)
|
|
self.assertNotIn("pp_size", p._overrides)
|
|
|
|
def test_override_unknown_key_raises_and_does_not_mutate(self):
|
|
p = get_parallel()
|
|
with self.assertRaises(ValueError):
|
|
with p.override(tp_sizee=1): # typo
|
|
pass
|
|
self.assertEqual(p._overrides, {})
|
|
|
|
|
|
class TestParallelDCP(_IsolatedOverrides):
|
|
"""The DCP width is a quotient; the DCP rank is a live reading.
|
|
|
|
They used to be tested the same way, by mocking the group getters, because
|
|
the width read the group too. It does not: `attn_dcp_size` is
|
|
`dcp_size if dcp_enabled else 1`, so the way to state it is to state the
|
|
leaves.
|
|
"""
|
|
|
|
def _published(self, **fields):
|
|
reset_context()
|
|
self.addCleanup(reset_context)
|
|
publish(ServerArgs(model_path="dummy", **fields), role="test")
|
|
return get_parallel()
|
|
|
|
def test_attn_dcp_is_one_when_dcp_is_off(self):
|
|
parallel = self._published(tp_size=8, dcp_size=1)
|
|
self.assertFalse(parallel.dcp_enabled)
|
|
self.assertEqual(parallel.attn_dcp_size, 1)
|
|
|
|
def test_attn_dcp_is_the_configured_width_when_on(self):
|
|
parallel = self._published(tp_size=8, dcp_size=8)
|
|
self.assertTrue(parallel.dcp_enabled)
|
|
self.assertEqual(parallel.attn_dcp_size, 8)
|
|
|
|
def test_the_dcp_rank_still_reads_the_group(self):
|
|
"""A rank is not implied by the configuration, so it reads the group --
|
|
gated on a width that is."""
|
|
with (
|
|
get_parallel().override(tp_size=8, dcp_size=8, dcp_enabled=False),
|
|
patch(f"{_PS}.get_dcp_rank", side_effect=AssertionError),
|
|
):
|
|
self.assertEqual(get_parallel().attn_dcp_rank, 0)
|
|
with (
|
|
get_parallel().override(tp_size=8, dcp_size=8, dcp_enabled=True),
|
|
patch(f"{_PS}.get_dcp_rank", return_value=3),
|
|
):
|
|
self.assertEqual(get_parallel().attn_dcp_rank, 3)
|
|
|
|
def test_the_width_does_not_consult_the_platform(self):
|
|
with patch("sglang.srt.utils.is_cuda", return_value=False) as is_cuda:
|
|
parallel = self._published(tp_size=8, dcp_size=8)
|
|
self.assertTrue(parallel.dcp_enabled)
|
|
self.assertEqual(parallel.attn_dcp_size, 8)
|
|
is_cuda.assert_not_called()
|
|
|
|
|
|
class _IsolatedServerArgs(CustomTestCase):
|
|
"""Save/restore the published ServerArgs around each test (the slot is
|
|
process-global; another test file sharing the process may have published)."""
|
|
|
|
def setUp(self):
|
|
super().setUp()
|
|
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)
|
|
super().tearDown()
|
|
|
|
|
|
class TestServerArgsOwnership(_IsolatedServerArgs):
|
|
"""V2b: the context owns the slot; the legacy getters are identity shims."""
|
|
|
|
def test_legacy_setter_publishes_into_context(self):
|
|
# Identity, not equality: the slot holds the very object published.
|
|
sentinel = ServerArgs(model_path="dummy")
|
|
server_args_module.set_global_server_args_for_scheduler(sentinel)
|
|
self.assertIs(get_server_args(), sentinel)
|
|
self.assertIs(get_context().server_args, sentinel)
|
|
|
|
def test_the_retired_accessor_raises_and_names_the_replacement(self):
|
|
"""`get_global_server_args` is retired: it answered with the record,
|
|
so a caller reading a field resolution had decided got a stale value
|
|
and no error at all.
|
|
|
|
`RuntimeError` unconditionally, not a warning first: a
|
|
`DeprecationWarning` is filtered by default outside `__main__`, so no
|
|
production caller would have seen it, and under
|
|
`-W error::DeprecationWarning` it would have changed the exception a
|
|
caller catches. The message has to name where to read instead, since
|
|
the answer differs by what the caller wanted.
|
|
"""
|
|
with self.assertRaises(RuntimeError) as cm:
|
|
server_args_module.get_global_server_args()
|
|
message = str(cm.exception)
|
|
self.assertIn("runtime_context", message)
|
|
self.assertIn("get_server_args()", message)
|
|
|
|
# And the type does not change when warnings are errors.
|
|
with warnings.catch_warnings():
|
|
warnings.simplefilter("error")
|
|
with self.assertRaises(RuntimeError):
|
|
server_args_module.get_global_server_args()
|
|
|
|
def test_tokenizer_alias_is_distinct_role_shim(self):
|
|
# Deliberately NOT an alias: the two legacy setters publish with
|
|
# different process roles (scheduler vs tokenizer).
|
|
self.assertIsNot(
|
|
server_args_module.set_global_server_args_for_tokenizer,
|
|
server_args_module.set_global_server_args_for_scheduler,
|
|
)
|
|
|
|
def test_pre_publish_error_verbatim(self):
|
|
reset_context()
|
|
with self.assertRaises(ValueError) as cm:
|
|
get_server_args()
|
|
self.assertEqual(str(cm.exception), "Global server args is not set yet!")
|
|
|
|
def test_republish_overwrite_allowed(self):
|
|
first = ServerArgs(model_path="dummy")
|
|
second = ServerArgs(model_path="dummy")
|
|
server_args_module.set_global_server_args_for_scheduler(first)
|
|
server_args_module.set_global_server_args_for_scheduler(second)
|
|
self.assertIs(get_server_args(), second)
|
|
|
|
def test_reset_context_clears_owned_store(self):
|
|
server_args_module.set_global_server_args_for_scheduler(
|
|
ServerArgs(model_path="dummy")
|
|
)
|
|
reset_context()
|
|
with self.assertRaises(ValueError):
|
|
get_server_args()
|
|
|
|
|
|
class TestAssertPublished(_IsolatedServerArgs):
|
|
"""Publishing is the process entry's job; the constructors only check.
|
|
|
|
`ModelRunner`, `TokenizerManager` and `MMEncoder` assert. A publish inside
|
|
a process that has already published re-projects the bags, discarding every
|
|
`override()` taken since and the provenance log with it, so a constructor
|
|
that finds nothing published fails loud.
|
|
"""
|
|
|
|
def _record(self, **fields):
|
|
return ServerArgs(model_path="dummy", **fields)
|
|
|
|
def test_the_check_leaves_a_live_process_alone(self):
|
|
record = self._record(grammar_backend="xgrammar")
|
|
publish(record, role="scheduler")
|
|
get_context().override("grammar.import_fallback", grammar_backend="none")
|
|
|
|
assert_published(record, role="scheduler")
|
|
|
|
self.assertEqual(
|
|
get_exec().kernel.grammar_backend,
|
|
"none",
|
|
"the check re-projected the bags, so the import fallback was "
|
|
"discarded and the process reports a backend it is not using",
|
|
)
|
|
self.assertEqual(
|
|
len(get_context().overrides_log()),
|
|
1,
|
|
"the provenance of the override went with it",
|
|
)
|
|
|
|
def test_a_different_record_fails(self):
|
|
first = self._record(grammar_backend="xgrammar")
|
|
publish(first, role="scheduler")
|
|
second = self._record(grammar_backend="llguidance")
|
|
|
|
with self.assertRaisesRegex(RuntimeError, "a different record is published"):
|
|
assert_published(second, role="scheduler")
|
|
|
|
self.assertIs(
|
|
get_server_args(),
|
|
first,
|
|
"the failing check published anyway",
|
|
)
|
|
|
|
def test_an_empty_slot_fails(self):
|
|
"""An empty slot fails."""
|
|
reset_context()
|
|
record = self._record(grammar_backend="xgrammar")
|
|
|
|
with self.assertRaisesRegex(
|
|
RuntimeError, "nothing is published in this process"
|
|
):
|
|
assert_published(record, role="scheduler")
|
|
|
|
def test_the_same_record_under_a_different_role_fails(self):
|
|
"""The role decides which namespaces this process may read."""
|
|
record = self._record()
|
|
publish(record, role="tokenizer")
|
|
|
|
with self.assertRaisesRegex(RuntimeError, "published under role 'tokenizer'"):
|
|
assert_published(record, role="scheduler")
|
|
|
|
self.assertEqual(publish_role(), "tokenizer")
|
|
|
|
|
|
class TestServerArgsScopedOverride(_IsolatedServerArgs):
|
|
"""ctx.override_server_args: the config tier's scoped test override —
|
|
tests force execution paths by overriding the context, not by
|
|
hand-building and publishing config objects."""
|
|
|
|
def test_install_publishes_fresh_config_with_fields(self):
|
|
reset_context()
|
|
override = get_context().override_server_args(
|
|
attention_backend="triton", chunked_prefill_size=-1
|
|
)
|
|
published = override.install()
|
|
self.assertIs(get_server_args(), published)
|
|
# The hook declares; the record keeps the operator's input, so the
|
|
# values are read where resolution puts them.
|
|
self.assertEqual(resolution_result(published, "attention_backend"), "triton")
|
|
self.assertEqual(resolution_result(published, "chunked_prefill_size"), -1)
|
|
# unnamed fields keep their dataclass defaults
|
|
self.assertEqual(resolution_result(published, "tp_size"), 1)
|
|
|
|
def test_unknown_fields_are_rejected(self):
|
|
with self.assertRaises(ValueError):
|
|
get_context().override_server_args(not_a_config_field=1).install()
|
|
|
|
def test_restore_reinstates_previous_publish(self):
|
|
previous = object()
|
|
get_context().set_server_args(previous)
|
|
override = get_context().override_server_args(tp_size=8)
|
|
override.install()
|
|
self.assertEqual(get_parallel().tp_size, 8)
|
|
override.restore()
|
|
self.assertIs(get_server_args(), previous)
|
|
|
|
def test_restore_reinstates_the_empty_slot(self):
|
|
reset_context()
|
|
with get_context().override_server_args():
|
|
get_server_args() # published inside the scope
|
|
with self.assertRaises(ValueError):
|
|
get_server_args()
|
|
|
|
def test_nesting_restores_in_order(self):
|
|
reset_context()
|
|
with get_context().override_server_args(tp_size=2) as outer:
|
|
with get_context().override_server_args(tp_size=4):
|
|
self.assertEqual(get_parallel().tp_size, 4)
|
|
self.assertIs(get_server_args(), outer)
|
|
self.assertEqual(get_parallel().tp_size, 2)
|
|
|
|
def test_private_attribute_seeding(self):
|
|
# Property caches (e.g. _mamba_cache_chunk_size) are seeded through
|
|
# the same call; the strict guard exempts underscore names.
|
|
published = (
|
|
get_context().override_server_args(_mamba_cache_chunk_size=64).install()
|
|
)
|
|
self.assertEqual(mamba_cache_chunk_size_of(published), 64)
|
|
|
|
def test_an_underscore_field_is_declared_like_any_other(self):
|
|
"""The split is fields vs not-fields, not the leading underscore.
|
|
|
|
`_speculative_draft_quantization_explicitly_set` is a real field
|
|
published under `spec`. Seeding it as a raw attribute instead of
|
|
declaring it would leave the earlier resolution authoritative, so both
|
|
the resolution and the bag would keep answering the pre-override value
|
|
while the record said otherwise.
|
|
"""
|
|
from sglang.srt.arg_groups.overrides import resolution_result
|
|
from sglang.srt.runtime_context import get_spec
|
|
|
|
name = "_speculative_draft_quantization_explicitly_set"
|
|
self.assertIn(name, ServerArgs.__struct_fields__)
|
|
|
|
published = get_context().override_server_args(**{name: True}).install()
|
|
# The record keeps the operator's input, as it does for every other
|
|
# field; the override travels as a declaration.
|
|
self.assertIsNone(getattr(published, name))
|
|
self.assertIs(resolution_result(published, name), True)
|
|
self.assertIs(getattr(get_spec(), name), True)
|
|
|
|
def test_installed_config_arms_the_strict_guard(self):
|
|
# The published dummy must behave like a resolved config: bare writes
|
|
# raise.
|
|
published = get_context().override_server_args(tp_size=2).install()
|
|
with self.assertRaises(AttributeError):
|
|
published.tp_size = 4
|
|
self.assertEqual(resolution_result(published, "tp_size"), 2)
|
|
|
|
def test_restore_resets_the_capture_seed(self):
|
|
# install() seeds flags.capture from the published dummy; restore()
|
|
# must put back the pre-install runtime state on both restore paths.
|
|
reset_context()
|
|
self.assertFalse(get_flags().capture.enable_torch_compile)
|
|
override = get_context().override_server_args(enable_torch_compile=True)
|
|
override.install()
|
|
self.assertTrue(get_flags().capture.enable_torch_compile)
|
|
override.restore()
|
|
self.assertFalse(get_flags().capture.enable_torch_compile)
|
|
|
|
def test_double_install_rejected(self):
|
|
override = get_context().override_server_args()
|
|
override.install()
|
|
with self.assertRaises(AssertionError):
|
|
override.install()
|
|
|
|
|
|
class _FakeCaptureGroup(_FlagGroupBase):
|
|
gamma: int = 0
|
|
|
|
|
|
class TestFlagsTier(_IsolatedServerArgs):
|
|
"""Runtime-flags tier: typed groups, typo-safe writes, override primitive.
|
|
|
|
Resolved configuration lives on server_args fields (materialized at the
|
|
end of __post_init__); the flags tier only carries runtime state
|
|
(today: the capture lifecycle)."""
|
|
|
|
def test_wiring_and_groups(self):
|
|
flags = get_flags()
|
|
self.assertIs(flags, get_context().flags)
|
|
self.assertIsInstance(flags, Flags)
|
|
self.assertTrue(hasattr(flags, "capture"))
|
|
|
|
def test_typo_safety(self):
|
|
group = _FakeCaptureGroup()
|
|
with self.assertRaises(AttributeError):
|
|
group.gamma_misspelled = 2 # undeclared leaf
|
|
with self.assertRaises(AttributeError):
|
|
get_flags().not_a_flag = 1
|
|
|
|
def test_override_is_transactional(self):
|
|
group = _FakeCaptureGroup()
|
|
with group.override(gamma=99):
|
|
self.assertEqual(group.gamma, 99)
|
|
self.assertEqual(group.gamma, 0)
|
|
with self.assertRaises(ValueError):
|
|
with group.override(gamma=2, delta=3): # delta undeclared
|
|
pass
|
|
self.assertEqual(group.gamma, 0) # validated before any write
|
|
|
|
def test_reset_context_installs_fresh_flags(self):
|
|
old = get_flags()
|
|
old.capture.enable_torch_compile = True
|
|
reset_context()
|
|
self.assertIsNot(get_flags(), old)
|
|
self.assertFalse(get_flags().capture.enable_torch_compile)
|
|
|
|
|
|
@dataclasses.dataclass
|
|
class _FakeResolvedArgs:
|
|
"""Publishable fixture with a resolvable whitelist (real flat leaves)."""
|
|
|
|
page_size: A[int | None, Arg(help="p", resolvable=True), NS("schedule")] = None
|
|
sampling_backend: A[
|
|
str | None, Arg(help="s", resolvable=True), NS("exec.kernel")
|
|
] = None
|
|
attention_backend: A[str | None, Arg(help="ab"), NS("exec.kernel")] = None
|
|
prefill_attention_backend: A[str | None, Arg(help="pab"), NS("exec.kernel")] = None
|
|
decode_attention_backend: A[str | None, Arg(help="dab"), NS("exec.kernel")] = None
|
|
disable_radix_cache: A[bool, Arg(help="drc"), NS("memory")] = False
|
|
mamba_radix_cache_strategy: A[str, Arg(help="mrcs"), NS("exec.mamba")] = "auto"
|
|
speculative_algorithm: A[str | None, Arg(help="sa"), NS("spec")] = None
|
|
speculative_num_draft_tokens: A[int | None, Arg(help="d"), NS("spec")] = None
|
|
speculative_adaptive: A[bool, Arg(help="a"), NS("spec")] = False
|
|
speculative_adaptive_config: A[str | None, Arg(help="c"), NS("spec")] = None
|
|
load_format: A[str, Arg(help="lf"), NS("model")] = "auto"
|
|
remote_instance_weight_loader_backend: A[str, Arg(help="rb"), NS("model")] = "nccl"
|
|
remote_instance_weight_loader_start_seed_via_transfer_engine: A[
|
|
bool, Arg(help="rs"), NS("model")
|
|
] = False
|
|
modelexpress_config: A[str | None, Arg(help="mx"), NS("model")] = None
|
|
disaggregation_mode: A[str, Arg(help="dm"), NS("disagg")] = "null"
|
|
max_running_requests: A[int | None, Arg(help="mrr"), NS("schedule")] = None
|
|
chunked_prefill_size: A[int, Arg(help="cps"), NS("schedule")] = -1
|
|
max_prefill_tokens: A[int, Arg(help="mpt"), NS("schedule")] = 16384
|
|
enable_dynamic_chunking: A[bool, Arg(help="edc"), NS("schedule")] = False
|
|
cuda_graph_config: A[object | None, Arg(help="cgc"), NS("exec.graph")] = None
|
|
tp_size: A[int, Arg(help="tp"), NS("parallel")] = 1
|
|
pp_size: A[int, Arg(help="pp"), NS("parallel")] = 1
|
|
_resolved_overrides: list = dataclasses.field(default_factory=list)
|
|
|
|
|
|
class TestMoeFlagsGroup(_IsolatedServerArgs):
|
|
"""flags.moe: materialized by initialize_moe_config; the ACTIVE backends
|
|
swap under the speculative contexts and restore on exit."""
|
|
|
|
def _init(self, **kw):
|
|
from sglang.srt.layers.moe.utils import initialize_moe_config
|
|
|
|
defaults = dict(
|
|
moe_a2a_backend="none",
|
|
moe_runner_backend="auto",
|
|
speculative_moe_runner_backend=None,
|
|
speculative_moe_a2a_backend=None,
|
|
deepep_mode="auto",
|
|
deepep_config=None,
|
|
enable_two_batch_overlap=False,
|
|
enable_single_batch_overlap=False,
|
|
tbo_token_distribution_threshold=0.48,
|
|
disable_flashinfer_cutlass_moe_fp4_allgather=False,
|
|
quantization=None,
|
|
disable_shared_experts_fusion=False,
|
|
)
|
|
defaults.update(kw)
|
|
# The flags are seeded from the bags, so the test publishes a config
|
|
# carrying these values.
|
|
override = get_context().override_server_args(**defaults)
|
|
override.install()
|
|
self.addCleanup(override.restore)
|
|
initialize_moe_config()
|
|
|
|
def test_lazy_defaults_before_initialize(self):
|
|
from sglang.srt.layers.moe.utils import (
|
|
get_moe_a2a_backend,
|
|
get_moe_runner_backend,
|
|
is_tbo_enabled,
|
|
)
|
|
|
|
reset_context()
|
|
self.assertTrue(get_moe_a2a_backend().is_none())
|
|
self.assertEqual(get_moe_runner_backend().name, "AUTO")
|
|
self.assertFalse(is_tbo_enabled())
|
|
|
|
def test_initialize_materializes_group(self):
|
|
from sglang.srt.layers.moe.utils import get_moe_a2a_backend, is_tbo_enabled
|
|
|
|
self._init(moe_a2a_backend="deepep", enable_two_batch_overlap=True)
|
|
self.assertTrue(get_moe_a2a_backend().is_deepep())
|
|
self.assertTrue(is_tbo_enabled())
|
|
self.assertEqual(get_flags().moe.deepep_config, "")
|
|
|
|
def test_speculative_swap_and_restore(self):
|
|
from sglang.srt.layers.moe.utils import (
|
|
get_moe_a2a_backend,
|
|
get_moe_runner_backend,
|
|
speculative_moe_a2a_backend_context,
|
|
speculative_moe_backend_context,
|
|
)
|
|
|
|
self._init(
|
|
moe_a2a_backend="deepep",
|
|
moe_runner_backend="triton",
|
|
speculative_moe_runner_backend="auto",
|
|
speculative_moe_a2a_backend="none",
|
|
)
|
|
with speculative_moe_backend_context(), speculative_moe_a2a_backend_context():
|
|
self.assertEqual(get_moe_runner_backend().name, "AUTO")
|
|
self.assertTrue(get_moe_a2a_backend().is_none())
|
|
# MTP layers are unquantized: fp4 allgather is forced off
|
|
self.assertTrue(get_flags().moe.disable_fp4_allgather)
|
|
self.assertTrue(get_flags().moe.speculative_context)
|
|
self.assertEqual(get_moe_runner_backend().name, "TRITON")
|
|
self.assertTrue(get_moe_a2a_backend().is_deepep())
|
|
self.assertFalse(get_flags().moe.disable_fp4_allgather)
|
|
self.assertFalse(get_flags().moe.speculative_context)
|
|
|
|
def test_swap_restores_on_exception(self):
|
|
from sglang.srt.layers.moe.utils import (
|
|
get_moe_runner_backend,
|
|
speculative_moe_backend_context,
|
|
)
|
|
|
|
self._init(moe_runner_backend="triton", speculative_moe_runner_backend="auto")
|
|
with self.assertRaises(RuntimeError):
|
|
with speculative_moe_backend_context():
|
|
raise RuntimeError("boom")
|
|
self.assertEqual(get_moe_runner_backend().name, "TRITON")
|
|
|
|
|
|
class TestDpFlagsGroup(_IsolatedServerArgs):
|
|
"""flags.dp: the DP-attention runtime flags; is_dp_attention_enabled is a
|
|
thin shim over the group leaf."""
|
|
|
|
def test_shim_reads_the_leaf(self):
|
|
from sglang.srt.layers.dp_attention import is_dp_attention_enabled
|
|
|
|
reset_context()
|
|
self.assertFalse(is_dp_attention_enabled())
|
|
get_flags().dp.enabled = True
|
|
self.assertTrue(is_dp_attention_enabled())
|
|
|
|
def test_scoped_override_forces_the_predicate(self):
|
|
from sglang.srt.layers.dp_attention import is_dp_attention_enabled
|
|
|
|
reset_context()
|
|
with get_flags().dp.override(enabled=True):
|
|
self.assertTrue(is_dp_attention_enabled())
|
|
self.assertFalse(is_dp_attention_enabled())
|
|
|
|
|
|
class TestResources(_IsolatedServerArgs):
|
|
"""ctx.resources: named slots for process-level resource handles with one
|
|
reset lifecycle; owning accessors keep their creation/publish semantics."""
|
|
|
|
def test_graph_pool_lazy_create_and_reuse(self):
|
|
from types import SimpleNamespace
|
|
|
|
from sglang.srt.model_executor.runner_utils.pool import (
|
|
get_global_graph_memory_pool,
|
|
get_or_create_global_graph_memory_pool,
|
|
)
|
|
|
|
reset_context()
|
|
self.assertIsNone(get_global_graph_memory_pool())
|
|
dev = SimpleNamespace(graph_pool_handle=lambda: object())
|
|
handle = get_or_create_global_graph_memory_pool(dev)
|
|
self.assertIs(get_or_create_global_graph_memory_pool(dev), handle)
|
|
|
|
def test_expert_recorder_noop_default_and_injection(self):
|
|
from sglang.srt.eplb.expert_distribution import (
|
|
get_global_expert_distribution_recorder,
|
|
)
|
|
from sglang.srt.runtime_context import get_resources
|
|
|
|
reset_context()
|
|
self.assertEqual(
|
|
type(get_global_expert_distribution_recorder()).__name__,
|
|
"_ExpertDistributionRecorderNoop",
|
|
)
|
|
with get_resources().override(expert_distribution_recorder="mock"):
|
|
self.assertEqual(get_global_expert_distribution_recorder(), "mock")
|
|
|
|
def test_expert_location_metadata_publish_once_until_reset(self):
|
|
from sglang.srt.eplb.expert_location import (
|
|
get_global_expert_location_metadata,
|
|
set_global_expert_location_metadata,
|
|
)
|
|
|
|
reset_context()
|
|
self.assertIsNone(get_global_expert_location_metadata())
|
|
set_global_expert_location_metadata("meta")
|
|
with self.assertRaises(AssertionError):
|
|
set_global_expert_location_metadata("again")
|
|
reset_context()
|
|
self.assertIsNone(get_global_expert_location_metadata())
|
|
|
|
|
|
class TestNamedStreams(_IsolatedServerArgs):
|
|
"""ctx.get_stream(name): keyed get-or-create (the persistent-buffer
|
|
pattern); set_stream installs explicitly."""
|
|
|
|
def test_get_or_create_shares_by_name(self):
|
|
from unittest.mock import patch
|
|
|
|
reset_context()
|
|
created = []
|
|
|
|
class _FakeStream:
|
|
def __init__(self):
|
|
created.append(self)
|
|
|
|
with patch("torch.cuda.Stream", _FakeStream):
|
|
a = get_context().get_stream("alt")
|
|
b = get_context().get_stream("alt")
|
|
c = get_context().get_stream("other")
|
|
self.assertIs(a, b)
|
|
self.assertIsNot(a, c)
|
|
self.assertEqual(len(created), 2)
|
|
|
|
def test_get_buffer_keyed_lazy(self):
|
|
reset_context()
|
|
created = []
|
|
|
|
def factory():
|
|
created.append(object())
|
|
return created[-1]
|
|
|
|
a = get_context().get_buffer("ws", factory)
|
|
b = get_context().get_buffer("ws", factory)
|
|
self.assertIs(a, b)
|
|
self.assertEqual(len(created), 1)
|
|
self.assertIsNot(get_context().get_buffer("other", factory), a)
|
|
|
|
def test_set_stream_installs_explicitly(self):
|
|
reset_context()
|
|
sentinel = object()
|
|
get_context().set_stream("alt", sentinel)
|
|
self.assertIs(get_context().get_stream("alt"), sentinel)
|
|
|
|
def test_reset_clears_the_registry(self):
|
|
reset_context()
|
|
get_context().set_stream("alt", object())
|
|
reset_context()
|
|
self.assertEqual(get_context().resources.streams, {})
|
|
|
|
def test_capturer_slots_roundtrip_and_reset(self):
|
|
from sglang.srt.state_capturer.indexer_topk import (
|
|
get_global_indexer_capturer,
|
|
set_global_indexer_capturer,
|
|
)
|
|
from sglang.srt.state_capturer.routed_experts import (
|
|
get_global_experts_capturer,
|
|
set_global_experts_capturer,
|
|
)
|
|
|
|
reset_context()
|
|
self.assertIsNone(get_global_indexer_capturer())
|
|
self.assertIsNone(get_global_experts_capturer())
|
|
indexer, experts = object(), object()
|
|
set_global_indexer_capturer(indexer)
|
|
set_global_experts_capturer(experts)
|
|
self.assertIs(get_global_indexer_capturer(), indexer)
|
|
self.assertIs(get_global_experts_capturer(), experts)
|
|
reset_context()
|
|
self.assertIsNone(get_global_indexer_capturer())
|
|
self.assertIsNone(get_global_experts_capturer())
|
|
|
|
def test_tcp_store_slot_roundtrip_and_reset(self):
|
|
from sglang.srt.distributed.utils import (
|
|
get_global_tcp_store,
|
|
set_global_tcp_store,
|
|
)
|
|
|
|
reset_context()
|
|
self.assertIsNone(get_global_tcp_store())
|
|
store = object()
|
|
set_global_tcp_store(store)
|
|
self.assertIs(get_global_tcp_store(), store)
|
|
reset_context()
|
|
self.assertIsNone(get_global_tcp_store())
|
|
|
|
def test_trace_level_env_seeded_lazy_default(self):
|
|
from sglang.srt.observability.trace import (
|
|
get_global_trace_level,
|
|
set_global_trace_level,
|
|
)
|
|
|
|
reset_context()
|
|
with patch.dict(os.environ, {}, clear=False):
|
|
os.environ.pop("SGLANG_TRACE_LEVEL", None)
|
|
self.assertEqual(get_global_trace_level(), 3)
|
|
set_global_trace_level(5)
|
|
self.assertEqual(get_global_trace_level(), 5)
|
|
reset_context()
|
|
with patch.dict(os.environ, {"SGLANG_TRACE_LEVEL": "1"}):
|
|
self.assertEqual(get_global_trace_level(), 1)
|
|
|
|
|
|
class TestEpBufferState(_IsolatedServerArgs):
|
|
"""EP dispatcher buffer managers: state lives on ctx.resources; the
|
|
facade keeps the mode-transition and clean semantics."""
|
|
|
|
def test_deepep_dispatch_mode_transitions_and_reset(self):
|
|
try:
|
|
from sglang.srt.layers.moe.token_dispatcher.deepep import DeepEPBuffer
|
|
except ImportError:
|
|
self.skipTest("deep_ep not installed")
|
|
|
|
reset_context()
|
|
cleans = []
|
|
|
|
class _FakeBuffer:
|
|
low_latency_mode = True
|
|
|
|
def clean_low_latency_buffer(self, *args):
|
|
cleans.append(args)
|
|
|
|
state = DeepEPBuffer._state()
|
|
state.buffer = _FakeBuffer()
|
|
state.hidden_size = 7168
|
|
state.num_max_dispatch_tokens_per_rank = 128
|
|
state.num_experts = 256
|
|
|
|
DeepEPBuffer.set_dispatch_mode_as_normal()
|
|
# NORMAL -> LOW_LATENCY must clean the low-latency buffer once.
|
|
DeepEPBuffer.set_dispatch_mode_as_low_latency()
|
|
self.assertEqual(cleans, [(128, 7168, 256)])
|
|
# LOW_LATENCY -> LOW_LATENCY must not clean again.
|
|
DeepEPBuffer.set_dispatch_mode_as_low_latency()
|
|
self.assertEqual(len(cleans), 1)
|
|
|
|
reset_context()
|
|
self.assertIsNone(DeepEPBuffer._state().buffer)
|
|
|
|
|
|
class TestForwardFlags(_IsolatedServerArgs):
|
|
"""ctx.forward: contextvar-backed per-forward flags; scoped() restores,
|
|
threads see defaults."""
|
|
|
|
def test_scoped_set_restore_and_nesting(self):
|
|
from sglang.srt.runtime_context import get_forward
|
|
|
|
reset_context()
|
|
fwd = get_forward()
|
|
self.assertFalse(fwd.multi_stream)
|
|
with fwd.scoped(multi_stream=True):
|
|
self.assertTrue(fwd.multi_stream)
|
|
with fwd.scoped(multi_stream=False):
|
|
self.assertFalse(fwd.multi_stream)
|
|
self.assertTrue(fwd.multi_stream)
|
|
self.assertFalse(fwd.multi_stream)
|
|
|
|
def test_scoped_restores_on_exception_and_validates_keys(self):
|
|
from sglang.srt.runtime_context import get_forward
|
|
|
|
reset_context()
|
|
fwd = get_forward()
|
|
with self.assertRaises(RuntimeError):
|
|
with fwd.scoped(moe_output_buffer="buf"):
|
|
raise RuntimeError("boom")
|
|
self.assertIsNone(fwd.moe_output_buffer)
|
|
with self.assertRaises(ValueError):
|
|
with fwd.scoped(nope=1):
|
|
pass
|
|
with self.assertRaises(AttributeError):
|
|
fwd.multi_stream = True # attribute writes are rejected
|
|
|
|
def test_threads_see_defaults(self):
|
|
import threading
|
|
|
|
from sglang.srt.runtime_context import get_forward
|
|
|
|
reset_context()
|
|
fwd = get_forward()
|
|
seen = {}
|
|
with fwd.scoped(multi_stream=True):
|
|
|
|
def probe():
|
|
seen["value"] = get_forward().multi_stream
|
|
|
|
worker = threading.Thread(target=probe)
|
|
worker.start()
|
|
worker.join()
|
|
self.assertFalse(seen["value"]) # a new thread sees the default
|
|
|
|
def test_graph_visible_flags_trace_under_torch_compile(self):
|
|
# Regression: dynamo cannot trace ContextVar.get, and these flags are
|
|
# read inside compiled model code (vocab embedding, communicator, DP
|
|
# gather) — they must stay plain-slot backed. fullgraph=True turns
|
|
# any graph break back into a failure.
|
|
import torch
|
|
|
|
from sglang.srt.runtime_context import get_forward
|
|
|
|
reset_context()
|
|
|
|
@torch.compile(fullgraph=True, backend="eager", dynamic=False)
|
|
def probe(x):
|
|
fwd = get_forward()
|
|
if fwd.attn_input_scattered:
|
|
x = x + 1
|
|
if fwd.is_extend_in_batch:
|
|
x = x + 2
|
|
if fwd.fuse_mlp_allreduce:
|
|
x = x + 4
|
|
if fwd.mlp_reduce_scatter:
|
|
x = x + 8
|
|
if fwd.flashinfer_trtllm_bypass:
|
|
x = x + 16
|
|
return x
|
|
|
|
self.assertEqual(probe(torch.zeros(())).item(), 0)
|
|
with get_forward().scoped(attn_input_scattered=True):
|
|
self.assertEqual(probe(torch.zeros(())).item(), 1)
|
|
get_forward().set("is_extend_in_batch", True)
|
|
self.assertEqual(probe(torch.zeros(())).item(), 2)
|
|
get_forward().set("is_extend_in_batch", False)
|
|
with get_forward().scoped(
|
|
fuse_mlp_allreduce=True,
|
|
mlp_reduce_scatter=True,
|
|
flashinfer_trtllm_bypass=True,
|
|
):
|
|
self.assertEqual(probe(torch.zeros(())).item(), 28)
|
|
self.assertEqual(probe(torch.zeros(())).item(), 0)
|
|
|
|
def test_parallel_config_leaves_trace_under_torch_compile(self):
|
|
# Regression: gate helpers such as ``enable_moe_dense_fully_dp()`` read
|
|
# parallel config leaves inside compiled model forwards, which must
|
|
# stay dynamo-traceable (``object.__getattribute__`` graph-breaks).
|
|
# fullgraph=True turns any graph break back into a failure.
|
|
import torch
|
|
|
|
from sglang.srt.runtime_context import get_parallel
|
|
|
|
reset_context()
|
|
with get_context().override_server_args(moe_dense_tp_size=1, dwdp_size=4):
|
|
|
|
@torch.compile(fullgraph=True, backend="eager", dynamic=False)
|
|
def probe(x):
|
|
par = get_parallel()
|
|
if par.enable_prefill_cp:
|
|
x = x + 1
|
|
if par.moe_dense_tp_size == 1:
|
|
x = x + 2
|
|
if par.dwdp_size > 1:
|
|
x = x + 4
|
|
return x
|
|
|
|
self.assertEqual(probe(torch.zeros(())).item(), 6)
|
|
|
|
def test_graph_visible_flags_are_process_visible_across_threads(self):
|
|
# Documented divergence from the contextvar-backed flags: plain slots
|
|
# are process-global (the storage form these flags had before the
|
|
# tier), so another thread sees the current value, not the default.
|
|
import threading
|
|
|
|
from sglang.srt.runtime_context import get_forward
|
|
|
|
reset_context()
|
|
seen = {}
|
|
with get_forward().scoped(attn_input_scattered=True):
|
|
|
|
def probe():
|
|
seen["value"] = get_forward().attn_input_scattered
|
|
|
|
worker = threading.Thread(target=probe)
|
|
worker.start()
|
|
worker.join()
|
|
self.assertTrue(seen["value"])
|
|
self.assertFalse(get_forward().attn_input_scattered)
|
|
|
|
def test_multi_stream_shims(self):
|
|
from sglang.srt.utils.multi_stream_utils import (
|
|
do_multi_stream,
|
|
with_multi_stream,
|
|
)
|
|
|
|
reset_context()
|
|
self.assertFalse(do_multi_stream())
|
|
with with_multi_stream(True):
|
|
self.assertTrue(do_multi_stream())
|
|
self.assertFalse(do_multi_stream())
|
|
|
|
def test_attn_tp_context_per_forward_slots(self):
|
|
from types import SimpleNamespace
|
|
|
|
from sglang.srt.layers.communicator import get_attn_tp_context
|
|
from sglang.srt.runtime_context import get_forward
|
|
|
|
reset_context()
|
|
ctx = get_attn_tp_context()
|
|
self.assertFalse(ctx.input_scattered)
|
|
fb = SimpleNamespace(
|
|
forward_mode=SimpleNamespace(
|
|
is_extend=lambda: False, is_target_verify=lambda: False
|
|
),
|
|
input_ids=None,
|
|
can_run_tbo=False,
|
|
)
|
|
sentinel = SimpleNamespace(fetch_qkv_latent=lambda: "qkv")
|
|
with ctx.maybe_input_scattered(fb):
|
|
ctx.set_attn_inputs(sentinel)
|
|
self.assertEqual(ctx.fetch_qkv_latent(), "qkv")
|
|
# attn inputs are cleared at scope exit, flag restored
|
|
self.assertIsNone(get_forward().attn_inputs)
|
|
self.assertFalse(ctx.input_scattered)
|
|
|
|
def test_dp_buffer_state_split(self):
|
|
import torch
|
|
|
|
from sglang.srt.layers.dp_attention import _DpGatheredBufferWrapper as wrapper
|
|
from sglang.srt.layers.dp_attention import (
|
|
get_dp_dtype,
|
|
get_dp_global_num_tokens,
|
|
get_global_dp_buffer_len,
|
|
is_dp_max_padding,
|
|
set_dp_buffer_len,
|
|
)
|
|
|
|
reset_context()
|
|
# metadata is init-static (flags.dp); sizing is per-forward sticky
|
|
wrapper.set_metadata(64, torch.float16, torch.device("cpu"))
|
|
self.assertEqual(get_dp_dtype(), torch.float16)
|
|
set_dp_buffer_len(128, 32, True, [64, 64])
|
|
self.assertEqual(get_global_dp_buffer_len(), 128)
|
|
self.assertTrue(is_dp_max_padding())
|
|
self.assertEqual(get_dp_global_num_tokens(), [64, 64])
|
|
set_dp_buffer_len(256, 64, False) # sticky until the next write
|
|
self.assertEqual(get_global_dp_buffer_len(), 256)
|
|
self.assertFalse(is_dp_max_padding())
|
|
self.assertIsNone(get_dp_global_num_tokens())
|
|
reset_context()
|
|
self.assertIsNone(get_dp_dtype())
|
|
|
|
def test_is_extend_in_batch_sticky_within_thread(self):
|
|
from sglang.srt.layers.dp_attention import (
|
|
get_is_extend_in_batch,
|
|
set_is_extend_in_batch,
|
|
)
|
|
|
|
reset_context()
|
|
self.assertFalse(get_is_extend_in_batch())
|
|
set_is_extend_in_batch(True)
|
|
self.assertTrue(get_is_extend_in_batch()) # sticky until next write
|
|
set_is_extend_in_batch(False)
|
|
self.assertFalse(get_is_extend_in_batch())
|
|
|
|
def test_moe_output_buffer_ctx(self):
|
|
from sglang.srt.layers.moe.moe_runner.base import moe_output_buffer_ctx
|
|
from sglang.srt.runtime_context import get_forward
|
|
|
|
reset_context()
|
|
sentinel = object()
|
|
with moe_output_buffer_ctx(sentinel):
|
|
self.assertIs(get_forward().moe_output_buffer, sentinel)
|
|
self.assertIsNone(get_forward().moe_output_buffer)
|
|
|
|
def test_mlp_comm_forward_flags(self):
|
|
"""Decoder-published MLP collective flags: scoped restore + skip helpers."""
|
|
from sglang.srt.layers.moe.utils import (
|
|
should_skip_mlp_all_reduce,
|
|
should_skip_post_experts_all_reduce,
|
|
)
|
|
from sglang.srt.runtime_context import get_forward
|
|
|
|
reset_context()
|
|
fwd = get_forward()
|
|
self.assertFalse(fwd.fuse_mlp_allreduce)
|
|
self.assertFalse(fwd.mlp_reduce_scatter)
|
|
self.assertFalse(fwd.flashinfer_trtllm_bypass)
|
|
self.assertFalse(should_skip_mlp_all_reduce())
|
|
|
|
with fwd.scoped(fuse_mlp_allreduce=True):
|
|
self.assertTrue(fwd.fuse_mlp_allreduce)
|
|
self.assertTrue(should_skip_mlp_all_reduce())
|
|
# Fusion alone is enough to skip post-experts AR.
|
|
self.assertTrue(should_skip_post_experts_all_reduce(is_tp_path=True))
|
|
self.assertFalse(fwd.fuse_mlp_allreduce)
|
|
self.assertFalse(should_skip_mlp_all_reduce())
|
|
|
|
with fwd.scoped(mlp_reduce_scatter=True):
|
|
self.assertTrue(fwd.mlp_reduce_scatter)
|
|
self.assertTrue(should_skip_mlp_all_reduce())
|
|
self.assertFalse(fwd.mlp_reduce_scatter)
|
|
|
|
with fwd.scoped(flashinfer_trtllm_bypass=True):
|
|
self.assertTrue(fwd.flashinfer_trtllm_bypass)
|
|
self.assertFalse(fwd.flashinfer_trtllm_bypass)
|
|
|
|
def test_dp_reduce_scatterv_requires_single_rank_attention_dp_shards(self):
|
|
from sglang.srt.layers.moe.utils import should_use_dp_reduce_scatterv
|
|
|
|
reset_context()
|
|
with patch(
|
|
"sglang.srt.layers.moe.utils.is_dp_attention_enabled",
|
|
return_value=True,
|
|
):
|
|
# The optimized path is valid when the collective group and the
|
|
# variable-split list have the same number of entries.
|
|
with get_parallel().override(tp_size=8, attn_dp_size=8, moe_ep_size=8):
|
|
self.assertTrue(should_use_dp_reduce_scatterv())
|
|
|
|
# Otherwise the standard all-reduce plus scatter path must be used.
|
|
with get_parallel().override(tp_size=8, attn_dp_size=2, moe_ep_size=2):
|
|
self.assertFalse(should_use_dp_reduce_scatterv())
|
|
|
|
|
|
class TestPublishLifecycle(_IsolatedServerArgs):
|
|
"""Publish installs the resolved server_args and seeds the capture tier."""
|
|
|
|
def _publish(self, **kw):
|
|
args = _FakeResolvedArgs(**kw)
|
|
get_context().set_server_args(args)
|
|
return args
|
|
|
|
def test_capture_tier_seeded_at_publish(self):
|
|
args = self._publish(page_size=1)
|
|
args.enable_torch_compile = True
|
|
get_context().set_server_args(args) # re-publish picks up the value
|
|
self.assertTrue(get_flags().capture.enable_torch_compile)
|
|
# capture-time write (B4) targets the capture leaf
|
|
get_flags().capture.enable_torch_compile = False
|
|
self.assertFalse(get_flags().capture.enable_torch_compile)
|
|
|
|
def test_capture_tier_defaults_for_sentinel_publish(self):
|
|
get_context().set_server_args(object())
|
|
self.assertFalse(get_flags().capture.enable_torch_compile)
|
|
|
|
|
|
class TestDerivedPredicatesAgreeAcrossTiers(_IsolatedServerArgs):
|
|
"""One definition per predicate, checked rather than asserted in prose.
|
|
|
|
Each of these exists twice by construction -- once over a config-shaped
|
|
object (the resolution pipeline's `*_of` helper, which `ServerArgs`
|
|
delegates to) and once over the published bags. The pair must agree on
|
|
every input, or a decision made before publish differs from the same
|
|
decision made after it.
|
|
"""
|
|
|
|
_STRATEGIES = ("auto", "no_buffer", "extra_buffer", "extra_buffer_lazy")
|
|
|
|
def test_the_mamba_extra_buffer_predicate_has_one_answer(self):
|
|
"""It used to be asserted that two spellings agreed. There is one now:
|
|
the declaration computes it at publish, and the bag carries it."""
|
|
for disable_radix_cache in (False, True):
|
|
for strategy in self._STRATEGIES:
|
|
with self.subTest(radix=disable_radix_cache, strategy=strategy):
|
|
reset_context()
|
|
publish(
|
|
ServerArgs(
|
|
model_path="dummy",
|
|
disable_radix_cache=disable_radix_cache,
|
|
mamba_radix_cache_strategy=strategy,
|
|
),
|
|
role="test",
|
|
)
|
|
expected = disable_radix_cache is False and strategy in (
|
|
"extra_buffer",
|
|
"extra_buffer_lazy",
|
|
)
|
|
self.assertEqual(
|
|
get_exec().mamba.enable_mamba_extra_buffer, expected
|
|
)
|
|
self.assertEqual(
|
|
get_exec().mamba.enable_mamba_extra_buffer_lazy,
|
|
disable_radix_cache is False
|
|
and strategy == "extra_buffer_lazy",
|
|
)
|
|
|
|
def test_prefill_buffer_ceiling_matches_the_member(self):
|
|
from sglang.srt.runtime_context import max_prefill_buffer_tokens
|
|
|
|
for chunked in (-1, 0, 1024, 8192):
|
|
for dynamic in (False, True):
|
|
for pp in (1, 4):
|
|
for max_prefill in (0, 2048, 16384):
|
|
with self.subTest(
|
|
chunked=chunked,
|
|
dynamic=dynamic,
|
|
pp=pp,
|
|
max_prefill=max_prefill,
|
|
):
|
|
args = _FakeResolvedArgs(
|
|
chunked_prefill_size=chunked,
|
|
enable_dynamic_chunking=dynamic,
|
|
pp_size=pp,
|
|
max_prefill_tokens=max_prefill,
|
|
)
|
|
get_context().set_server_args(args)
|
|
self.assertEqual(
|
|
max_prefill_buffer_tokens_of(args),
|
|
max_prefill_buffer_tokens(),
|
|
)
|
|
|
|
def test_prefill_buffer_ceiling_hook_honored_across_tiers(self):
|
|
args = _FakeResolvedArgs(
|
|
chunked_prefill_size=8192,
|
|
enable_dynamic_chunking=True,
|
|
pp_size=4,
|
|
max_prefill_tokens=16384,
|
|
)
|
|
|
|
def provider(record, default_ceiling):
|
|
self.assertIs(record, args)
|
|
return default_ceiling + 5
|
|
|
|
with patch.object(prefill_buffer_ceiling, "_prefill_buffer_ceiling_fn", None):
|
|
register = prefill_buffer_ceiling.register_prefill_buffer_ceiling
|
|
self.assertEqual(max_prefill_buffer_tokens_of(args), 16384)
|
|
self.assertIs(register(provider), provider)
|
|
register(provider)
|
|
with self.assertRaisesRegex(RuntimeError, "already registered"):
|
|
register(lambda record, default_ceiling: default_ceiling)
|
|
for record_or_view in (args, resolving_view(args), resolved_view(args)):
|
|
self.assertEqual(max_prefill_buffer_tokens_of(record_or_view), 16389)
|
|
get_context().set_server_args(args)
|
|
self.assertEqual(max_prefill_buffer_tokens(), 16389)
|
|
with get_schedule().override(max_prefill_tokens=32768):
|
|
self.assertEqual(max_prefill_buffer_tokens(), 32773)
|
|
self.assertEqual(args.max_prefill_tokens, 16384)
|
|
|
|
def test_prefill_buffer_ceiling_provider_can_preserve_defaults(self):
|
|
args = _FakeResolvedArgs(chunked_prefill_size=4096)
|
|
|
|
def provider(record, default_ceiling):
|
|
return default_ceiling
|
|
|
|
with patch.object(prefill_buffer_ceiling, "_prefill_buffer_ceiling_fn", None):
|
|
prefill_buffer_ceiling.register_prefill_buffer_ceiling(provider)
|
|
for record_or_view in (args, resolving_view(args), resolved_view(args)):
|
|
self.assertEqual(max_prefill_buffer_tokens_of(record_or_view), 4096)
|
|
get_context().set_server_args(args)
|
|
self.assertEqual(max_prefill_buffer_tokens(), 4096)
|
|
|
|
def test_activation_reserve_matches_the_member(self):
|
|
from types import SimpleNamespace
|
|
|
|
from sglang.srt.arg_groups.overrides import (
|
|
pre_capture_activation_reserve_mb_of,
|
|
)
|
|
from sglang.srt.runtime_context import pre_capture_activation_reserve_mb
|
|
|
|
graph = SimpleNamespace(decode=SimpleNamespace(max_bs=64))
|
|
cases = (
|
|
dict(disaggregation_mode="null", chunked_prefill_size=8192),
|
|
dict(disaggregation_mode="null", chunked_prefill_size=-1),
|
|
dict(
|
|
disaggregation_mode="null",
|
|
chunked_prefill_size=-1,
|
|
max_prefill_tokens=1024,
|
|
),
|
|
dict(disaggregation_mode="decode", max_running_requests=32),
|
|
dict(disaggregation_mode="decode", max_running_requests=None),
|
|
dict(
|
|
disaggregation_mode="decode",
|
|
max_running_requests=None,
|
|
speculative_num_draft_tokens=4,
|
|
),
|
|
dict(
|
|
disaggregation_mode="null",
|
|
chunked_prefill_size=8192,
|
|
tp_size=8,
|
|
pp_size=2,
|
|
),
|
|
)
|
|
for case in cases:
|
|
for gpu_mem in (None, 20 * 1024, 80 * 1024):
|
|
with self.subTest(gpu_mem=gpu_mem, **case):
|
|
args = _FakeResolvedArgs(cuda_graph_config=graph, **case)
|
|
get_context().set_server_args(args)
|
|
self.assertEqual(
|
|
pre_capture_activation_reserve_mb_of(args, gpu_mem),
|
|
pre_capture_activation_reserve_mb(gpu_mem),
|
|
)
|
|
|
|
def test_remote_instance_transfer_engine_matches_the_member(self):
|
|
from sglang.srt.runtime_context import remote_instance_transfer_engine_enabled
|
|
|
|
backends = ("nccl", "transfer_engine", "modelexpress")
|
|
transports = (None, '{"transport": "transfer_engine"}', '{"transport": "nixl"}')
|
|
for seed_via_te in (False, True):
|
|
for load_format in ("auto", "remote_instance"):
|
|
for backend in backends:
|
|
for mx in transports:
|
|
with self.subTest(
|
|
seed=seed_via_te,
|
|
load_format=load_format,
|
|
backend=backend,
|
|
modelexpress=mx,
|
|
):
|
|
args = _FakeResolvedArgs(
|
|
load_format=load_format,
|
|
remote_instance_weight_loader_backend=backend,
|
|
remote_instance_weight_loader_start_seed_via_transfer_engine=seed_via_te,
|
|
modelexpress_config=mx,
|
|
)
|
|
get_context().set_server_args(args)
|
|
for override in (None, "remote_instance", "auto"):
|
|
self.assertEqual(
|
|
ServerArgs.remote_instance_weight_loader_use_transfer_engine(
|
|
args, override
|
|
),
|
|
remote_instance_transfer_engine_enabled(override),
|
|
)
|
|
|
|
def test_attention_backends_match_the_member(self):
|
|
from sglang.srt.runtime_context import attention_backends
|
|
|
|
backends = (None, "fa3", "triton")
|
|
for base in backends:
|
|
for prefill in backends:
|
|
for decode in backends:
|
|
with self.subTest(base=base, prefill=prefill, decode=decode):
|
|
args = _FakeResolvedArgs(
|
|
attention_backend=base,
|
|
prefill_attention_backend=prefill,
|
|
decode_attention_backend=decode,
|
|
)
|
|
get_context().set_server_args(args)
|
|
self.assertEqual(
|
|
attention_backends_of(resolved_view(args)),
|
|
attention_backends(),
|
|
)
|
|
|
|
|
|
class TestAdaptiveDraftBoundLifecycle(_IsolatedServerArgs):
|
|
"""The adaptive draft-token bound is snapshotted at each publication."""
|
|
|
|
def _write_config(self, steps):
|
|
path = os.path.join(tempfile.mkdtemp(prefix="adaptive_cfg_"), "adaptive.json")
|
|
self.addCleanup(shutil.rmtree, os.path.dirname(path), ignore_errors=True)
|
|
with open(path, "w") as handle:
|
|
json.dump({"1": {"candidate_steps": steps}}, handle)
|
|
return path
|
|
|
|
def test_republishing_recomputes_the_bound(self):
|
|
path = self._write_config([2])
|
|
get_context().set_server_args(
|
|
_FakeResolvedArgs(
|
|
speculative_num_draft_tokens=3,
|
|
speculative_adaptive=True,
|
|
speculative_adaptive_config=path,
|
|
)
|
|
)
|
|
self.assertEqual(max_speculative_num_draft_tokens(), 3)
|
|
|
|
with open(path, "w") as handle:
|
|
json.dump({"1": {"candidate_steps": [4]}}, handle)
|
|
# The new publication must not retain the previous capacity.
|
|
get_context().set_server_args(
|
|
_FakeResolvedArgs(
|
|
speculative_num_draft_tokens=3,
|
|
speculative_adaptive=True,
|
|
speculative_adaptive_config=path,
|
|
)
|
|
)
|
|
self.assertEqual(max_speculative_num_draft_tokens(), 5)
|
|
|
|
def test_reset_clears_the_bound(self):
|
|
path = self._write_config([2])
|
|
get_context().set_server_args(
|
|
_FakeResolvedArgs(
|
|
speculative_num_draft_tokens=3,
|
|
speculative_adaptive=True,
|
|
speculative_adaptive_config=path,
|
|
)
|
|
)
|
|
self.assertEqual(max_speculative_num_draft_tokens(), 3)
|
|
reset_context()
|
|
with open(path, "w") as handle:
|
|
json.dump({"1": {"candidate_steps": [6]}}, handle)
|
|
get_context().set_server_args(
|
|
_FakeResolvedArgs(
|
|
speculative_num_draft_tokens=3,
|
|
speculative_adaptive=True,
|
|
speculative_adaptive_config=path,
|
|
)
|
|
)
|
|
self.assertEqual(max_speculative_num_draft_tokens(), 7)
|
|
|
|
|
|
class TestParallelLeafReads(_IsolatedServerArgs):
|
|
"""The contract ``ParallelContext.__getattr__`` answers a parallel leaf on."""
|
|
|
|
def test_a_leaf_answers_what_resolution_decided(self):
|
|
from sglang.srt.arg_groups.overrides import resolution_result
|
|
|
|
with get_context().override_server_args() as server_args:
|
|
self.assertEqual(
|
|
resolution_result(server_args, "nccl_port"),
|
|
get_parallel().nccl_port,
|
|
"a parallel leaf read off the context disagreed with what "
|
|
"resolution decided",
|
|
)
|
|
|
|
def test_before_publish_the_error_names_the_namespace(self):
|
|
with self.assertRaisesRegex(ValueError, r"'parallel' not published"):
|
|
getattr(ParallelContext(), "nccl_port")
|
|
|
|
def test_an_unknown_name_is_still_an_attribute_error(self):
|
|
with self.assertRaisesRegex(AttributeError, r"has no 'not_a_leaf'"):
|
|
getattr(ParallelContext(), "not_a_leaf")
|
|
|
|
|
|
class TestDerivedWidths(_IsolatedOverrides):
|
|
"""The widths no flag sets are computed from the leaves and permanently
|
|
overridable.
|
|
|
|
`attn_tp_size` and its siblings used to be read back off the group
|
|
coordinator that was built from them, which made the answer depend on
|
|
distributed init and, after an elastic scale, disagree with the leaves.
|
|
"""
|
|
|
|
def setUp(self):
|
|
super().setUp()
|
|
parallel = get_parallel()
|
|
self._saved_derived = dict(parallel._derived)
|
|
parallel.clear_derived_widths()
|
|
self.addCleanup(
|
|
lambda: (
|
|
parallel.clear_derived_widths(),
|
|
parallel.override_permanently(**self._saved_derived),
|
|
)
|
|
)
|
|
|
|
def test_the_published_configuration_decides_the_widths(self):
|
|
"""The quotients are computed once, at publish, from the leaves.
|
|
|
|
Every input is a record field, so there is nothing to recompute on a
|
|
read: `publish` fills the bag and the bag is the answer.
|
|
"""
|
|
reset_context()
|
|
self.addCleanup(reset_context)
|
|
publish(
|
|
ServerArgs(
|
|
model_path="dummy", tp_size=8, dp_size=2, enable_dp_attention=True
|
|
),
|
|
role="test",
|
|
)
|
|
self.assertEqual(get_parallel().attn_tp_size, 4)
|
|
self.assertEqual(get_parallel().attn_dp_size, 2)
|
|
self.assertEqual(get_parallel().moe_tp_size, 8)
|
|
|
|
reset_context()
|
|
publish(
|
|
ServerArgs(model_path="dummy", tp_size=8, ep_size=4, moe_dp_size=2),
|
|
role="test",
|
|
)
|
|
self.assertEqual(get_parallel().moe_tp_size, 1)
|
|
|
|
def test_a_topology_is_stated_by_naming_the_width(self):
|
|
"""Overriding a leaf does not move the quotient -- the quotient is not
|
|
recomputed on read. Naming it is how a test states one."""
|
|
reset_context()
|
|
self.addCleanup(reset_context)
|
|
publish(ServerArgs(model_path="dummy", tp_size=8), role="test")
|
|
self.assertEqual(get_parallel().attn_tp_size, 8)
|
|
with get_parallel().override(tp_size=2):
|
|
self.assertEqual(get_parallel().attn_tp_size, 8)
|
|
with get_parallel().override(attn_tp_size=4):
|
|
self.assertEqual(get_parallel().attn_tp_size, 4)
|
|
|
|
def test_an_unstated_topology_still_fails(self):
|
|
"""Neutral leaves are for the dimensions a caller is not using, not for
|
|
a caller that stated nothing: every width would come back 1, which is a
|
|
plausible-looking number invented out of nothing."""
|
|
with self.assertRaises(RuntimeError) as caught:
|
|
get_parallel().attn_tp_size
|
|
self.assertIn("not available", str(caught.exception))
|
|
|
|
def test_a_permanent_override_and_a_live_group_both_win_over_the_leaves(self):
|
|
"""Order is permanent override, then live group, then the leaves.
|
|
Where a group exists it is the truth -- elastic scale-up moves the
|
|
group without a fresh override -- so the leaf derivation only
|
|
answers where there is none.
|
|
"""
|
|
parallel = get_parallel()
|
|
parallel.override_permanently(attn_tp_size=7)
|
|
self.addCleanup(parallel.clear_derived_widths)
|
|
with parallel.override(tp_size=8, attn_dp_size=2):
|
|
self.assertEqual(parallel.attn_tp_size, 7)
|
|
|
|
def test_the_quotients_come_from_the_leaves(self):
|
|
widths = derive_parallel_widths(
|
|
tp_size=8,
|
|
attn_cp_size=1,
|
|
attn_dp_size=2,
|
|
moe_ep_size=4,
|
|
moe_dp_size=2,
|
|
dcp_size=1,
|
|
dcp_enabled=False,
|
|
)
|
|
self.assertEqual(widths["attn_tp_size"], 8 // 2 // 1)
|
|
self.assertEqual(widths["moe_tp_size"], 8 // 4 // 2)
|
|
self.assertEqual(widths["attn_dcp_size"], 1)
|
|
|
|
def test_the_world_size_is_not_permanently_overridden(self):
|
|
"""It is not a quotient, and the live getter is right at every moment.
|
|
A value fixed when the groups are built would answer with the launch
|
|
count after `try_admit_scale_ranks` expands WORLD, and with the joining
|
|
cohort's own width on a scale-joiner, which lays its groups out at
|
|
`tp * pp` while WORLD spans `ep_join_rank_offset + tp * pp`."""
|
|
widths = derive_parallel_widths(
|
|
tp_size=4,
|
|
attn_cp_size=1,
|
|
attn_dp_size=1,
|
|
moe_ep_size=1,
|
|
moe_dp_size=1,
|
|
dcp_size=1,
|
|
dcp_enabled=False,
|
|
)
|
|
self.assertNotIn("world_size", widths)
|
|
parallel = get_parallel()
|
|
parallel.override_permanently(attn_tp_size=4)
|
|
with patch(f"{_PS}.get_world_size", return_value=9):
|
|
self.assertEqual(parallel.world_size, 9)
|
|
|
|
def test_a_permanently_overridden_width_is_what_the_reader_answers_with(self):
|
|
parallel = get_parallel()
|
|
parallel.override_permanently(attn_tp_size=4, moe_tp_size=1)
|
|
with patch(
|
|
f"{_PS}.get_attn_tensor_model_parallel_world_size",
|
|
side_effect=AssertionError("the group must not be asked"),
|
|
):
|
|
self.assertEqual(parallel.attn_tp_size, 4)
|
|
|
|
def test_a_scoped_override_still_wins_over_the_permanent_one(self):
|
|
parallel = get_parallel()
|
|
parallel.override_permanently(attn_tp_size=4)
|
|
with parallel.override(attn_tp_size=1):
|
|
self.assertEqual(parallel.attn_tp_size, 1)
|
|
self.assertEqual(parallel.attn_tp_size, 4)
|
|
|
|
def test_the_group_is_never_consulted(self):
|
|
"""There is no third source. A quotient comes from a scoped override, a
|
|
permanent override, or the published leaf -- never from a group
|
|
coordinator."""
|
|
reset_context()
|
|
self.addCleanup(reset_context)
|
|
with patch(
|
|
f"{_PS}.get_attn_tensor_model_parallel_world_size",
|
|
side_effect=AssertionError("the group must not be consulted"),
|
|
):
|
|
publish(
|
|
ServerArgs(
|
|
model_path="dummy", tp_size=8, dp_size=2, enable_dp_attention=True
|
|
),
|
|
role="test",
|
|
)
|
|
self.assertEqual(get_parallel().attn_tp_size, 4)
|
|
|
|
def test_with_neither_the_failure_names_the_cause(self):
|
|
with patch(
|
|
f"{_PS}.get_attn_tensor_model_parallel_world_size",
|
|
side_effect=AssertionError("attention tp group is not initialized"),
|
|
):
|
|
with self.assertRaisesRegex(RuntimeError, r"derived parallel width"):
|
|
get_parallel().attn_tp_size
|
|
|
|
def test_a_temporary_disable_beats_the_permanent_override(self):
|
|
"""`disable_dp_size()` runs a draft scope without DP attention. It moves
|
|
the module global the legacy getter reads, so it has to move the derived
|
|
width too -- the scoped override wins over the permanent one, and a
|
|
scope that left it alone would answer with the target model's width
|
|
for its duration."""
|
|
from sglang.srt.layers import dp_attention
|
|
|
|
parallel = get_parallel()
|
|
parallel.override_permanently(attn_dp_size=4)
|
|
with patch.object(dp_attention, "_ATTN_DP_SIZE", 4):
|
|
with dp_attention.disable_dp_size():
|
|
self.assertEqual(dp_attention.get_attention_dp_size(), 1)
|
|
self.assertEqual(parallel.attn_dp_size, 1)
|
|
self.assertEqual(parallel.attn_dp_size, 4)
|
|
|
|
def test_the_permanent_override_is_cleared_and_reset(self):
|
|
parallel = get_parallel()
|
|
parallel.override_permanently(attn_dp_size=2)
|
|
self.assertEqual(parallel.attn_dp_size, 2)
|
|
# Elastic scaling overrides again where it updates the live width.
|
|
parallel.override_permanently(attn_dp_size=4)
|
|
self.assertEqual(parallel.attn_dp_size, 4)
|
|
parallel.clear_derived_widths()
|
|
with parallel.override(tp_size=8, attn_dp_size=1):
|
|
self.assertEqual(parallel.attn_dp_size, 1)
|
|
|
|
def test_reset_context_drops_the_permanent_override(self):
|
|
"""The permanent override belongs to the lifecycle that made it.
|
|
|
|
`_derived_width` prefers it over the published leaf, so one that
|
|
outlived `reset_context()` would let the next test read the previous
|
|
topology.
|
|
"""
|
|
parallel = get_parallel()
|
|
parallel.override_permanently(attn_tp_size=4)
|
|
self.assertEqual(parallel.attn_tp_size, 4)
|
|
reset_context()
|
|
self.addCleanup(reset_context)
|
|
publish(ServerArgs(model_path="dummy", tp_size=1), role="test")
|
|
self.assertEqual(get_parallel().attn_tp_size, 1)
|
|
|
|
def test_the_rank_helper_agrees_with_the_override(self):
|
|
"""`compute_dp_attention_world_info` keeps the ranks and takes the
|
|
widths from the same derivation `override_permanently`'s callers use."""
|
|
from sglang.srt.layers.dp_attention import compute_dp_attention_world_info
|
|
|
|
for tp_size, dp_size, attn_cp_size in ((8, 2, 1), (8, 2, 2), (16, 4, 2)):
|
|
_, attn_tp_size, _, attn_dp_size = compute_dp_attention_world_info(
|
|
True, 0, tp_size, dp_size, attn_cp_size
|
|
)
|
|
widths = derive_parallel_widths(
|
|
tp_size=tp_size,
|
|
attn_cp_size=attn_cp_size,
|
|
attn_dp_size=attn_dp_size,
|
|
moe_ep_size=1,
|
|
moe_dp_size=1,
|
|
dcp_size=1,
|
|
dcp_enabled=False,
|
|
)
|
|
self.assertEqual(attn_tp_size, widths["attn_tp_size"])
|
|
self.assertEqual(attn_dp_size, widths["attn_dp_size"])
|
|
|
|
def test_recomputing_from_published_leaves_matches_the_publish_bag(self):
|
|
"""`initialize_model_parallel` no longer overrides anything -- see
|
|
`test_initialize_model_parallel_no_longer_touches_the_bag` below --
|
|
which makes this the load-bearing half of 16-field-registry-design.md
|
|
§6e: every real caller must forward leaves that already match its own
|
|
published config, because nothing corrects a mismatch anymore.
|
|
`scheduler.py`'s `ps.attn_dp_size`/`ps.moe_ep_size`/etc, and the
|
|
weight-cache daemon's own already-published config, both do -- this
|
|
pins that the formula they'd recompute from those leaves
|
|
(`derive_attention_widths`, `derive_parallel_widths`, the same ones
|
|
`publish` itself used) agrees with what's already in the bag, across
|
|
the widths `test_the_rank_helper_agrees_with_the_override` does not
|
|
vary -- moe_ep_size, moe_dp_size, and dcp_size -- using real
|
|
`publish()`.
|
|
|
|
A caller that does NOT keep the two in sync is a bug in that caller,
|
|
not something this framework silently corrects: two real ones existed
|
|
(`test/registered/eplb/test_lplb_distributed.py` and
|
|
`test/manual/ep/test_flashinfer_dispatcher.py`, both publishing a
|
|
placeholder config and then building real groups at a width it never
|
|
reflected) and were fixed by publishing the actual width instead of
|
|
relying on a correction to paper over the mismatch.
|
|
"""
|
|
shapes = (
|
|
dict(tp_size=8),
|
|
dict(tp_size=8, dp_size=2, enable_dp_attention=True),
|
|
dict(tp_size=8, ep_size=4, moe_dp_size=2),
|
|
dict(tp_size=8, dcp_size=8),
|
|
)
|
|
for shape in shapes:
|
|
with self.subTest(shape=shape):
|
|
reset_context()
|
|
self.addCleanup(reset_context)
|
|
publish(ServerArgs(model_path="dummy", **shape), role="test")
|
|
parallel = get_parallel()
|
|
published = {
|
|
"attn_tp_size": parallel.attn_tp_size,
|
|
"attn_dp_size": parallel.attn_dp_size,
|
|
"moe_ep_size": parallel.moe_ep_size,
|
|
"moe_tp_size": parallel.moe_tp_size,
|
|
"dcp_enabled": parallel.dcp_enabled,
|
|
"attn_dcp_size": parallel.attn_dcp_size,
|
|
}
|
|
# What every real `initialize_model_parallel` caller forwards:
|
|
# its own already-published leaves, through the same two
|
|
# functions the bag was projected with.
|
|
recomputed = derive_parallel_widths(
|
|
tp_size=parallel.tp_size,
|
|
attn_cp_size=parallel.attn_cp_size,
|
|
attn_dp_size=(
|
|
parallel.dp_size if parallel.enable_dp_attention else 1
|
|
),
|
|
moe_ep_size=parallel.ep_size,
|
|
moe_dp_size=parallel.moe_dp_size,
|
|
dcp_size=parallel.dcp_size,
|
|
dcp_enabled=parallel.dcp_size > 1,
|
|
)
|
|
self.assertEqual(published, recomputed)
|
|
|
|
def test_initialize_model_parallel_no_longer_touches_the_bag(self):
|
|
"""§6e, landed: `initialize_model_parallel` used to recompute and
|
|
permanently override the six derived widths on `get_parallel()`
|
|
after building its groups; that call is gone. Publish a placeholder
|
|
config (tp_size defaults to 1), then build real groups at a
|
|
different width -- the published leaf must now stay exactly what it
|
|
was, because nothing corrects it. This is the behavior a caller
|
|
relies on being told about, loudly, the first time it publishes and
|
|
builds inconsistently -- see
|
|
`test_recomputing_from_published_leaves_matches_the_publish_bag`
|
|
for why every real caller must not do that.
|
|
"""
|
|
from unittest.mock import Mock
|
|
|
|
from sglang.srt.distributed import parallel_state
|
|
|
|
reset_context()
|
|
self.addCleanup(reset_context)
|
|
publish(ServerArgs(model_path="dummy"), role="test")
|
|
self.assertEqual(get_parallel().attn_tp_size, 1)
|
|
self.assertEqual(get_parallel().moe_ep_size, 1)
|
|
|
|
world_size = 8
|
|
with (
|
|
patch.object(parallel_state, "_WORLD", None),
|
|
patch.object(parallel_state, "_TP", None),
|
|
patch.object(parallel_state, "_DCP", None),
|
|
patch.object(parallel_state, "_ATTN_CP", None),
|
|
patch.object(parallel_state, "_ATTN_TP", None),
|
|
patch.object(parallel_state, "_MOE_DP", None),
|
|
patch.object(parallel_state, "_MOE_EP", None),
|
|
patch.object(parallel_state, "_MOE_TP", None),
|
|
patch.object(parallel_state, "_PP", None),
|
|
patch.object(parallel_state, "_SELF_PP", None),
|
|
patch("torch.distributed.is_initialized", return_value=True),
|
|
patch("torch.distributed.get_world_size", return_value=world_size),
|
|
patch("torch.distributed.get_rank", return_value=0),
|
|
patch("torch.distributed.get_backend", return_value="nccl"),
|
|
patch.object(
|
|
parallel_state,
|
|
"init_model_parallel_group",
|
|
return_value=Mock(device_group=Mock()),
|
|
),
|
|
patch.object(parallel_state, "get_world_group") as mock_world_group,
|
|
):
|
|
mock_world_group.return_value = Mock(device_group=Mock(), local_rank=0)
|
|
parallel_state.initialize_model_parallel(
|
|
tensor_model_parallel_size=world_size,
|
|
expert_model_parallel_size=world_size,
|
|
)
|
|
self.addCleanup(parallel_state.destroy_model_parallel)
|
|
|
|
self.assertEqual(
|
|
get_parallel().attn_tp_size,
|
|
1,
|
|
"initialize_model_parallel must not touch the published leaf -- "
|
|
"a caller that needs it corrected must publish a config that "
|
|
"already matches the width it is about to build",
|
|
)
|
|
self.assertEqual(get_parallel().moe_ep_size, 1)
|
|
|
|
|
|
class TestTheDerivedHalfIsDeclared(CustomTestCase):
|
|
"""The quotients are declared beside the leaves, in the same class.
|
|
|
|
A namespace is one file and one class. `Parallel` says both what an
|
|
operator can set and what that decides; the quotients are unannotated, so
|
|
they are not dataclass fields and never reach the record.
|
|
`ParallelContext` installs a property per declaration rather than carrying
|
|
its own list, so the two cannot drift.
|
|
"""
|
|
|
|
def test_every_declared_quotient_has_a_property(self):
|
|
from sglang.srt.arg_groups.arg_utils import Derived
|
|
from sglang.srt.arg_groups.fields.parallel import Parallel
|
|
|
|
declared = {
|
|
name for name, value in vars(Parallel).items() if isinstance(value, Derived)
|
|
}
|
|
self.assertTrue(declared, "the derived half is empty")
|
|
for name in declared:
|
|
self.assertIsInstance(
|
|
getattr(type(get_context().parallel), name, None),
|
|
property,
|
|
f"{name} is declared but no property was installed",
|
|
)
|
|
|
|
def test_the_declared_set_is_what_derive_parallel_widths_produces(self):
|
|
"""The declaration is not a second list to keep in step: it names
|
|
exactly the quotients the derivation returns."""
|
|
from sglang.srt.arg_groups.arg_utils import Derived
|
|
from sglang.srt.arg_groups.fields.parallel import Parallel
|
|
|
|
declared = {
|
|
name for name, value in vars(Parallel).items() if isinstance(value, Derived)
|
|
}
|
|
produced = set(
|
|
derive_parallel_widths(
|
|
tp_size=8,
|
|
attn_cp_size=1,
|
|
attn_dp_size=2,
|
|
moe_ep_size=1,
|
|
moe_dp_size=1,
|
|
dcp_size=1,
|
|
dcp_enabled=False,
|
|
)
|
|
)
|
|
self.assertEqual(declared, produced)
|
|
|
|
def test_a_declared_quotient_is_not_a_record_field(self):
|
|
"""It has no operator input to preserve, and the record is what crosses
|
|
a process boundary."""
|
|
|
|
from sglang.srt.arg_groups.arg_utils import Derived
|
|
from sglang.srt.arg_groups.fields.parallel import Parallel
|
|
from sglang.srt.server_args import ServerArgs
|
|
|
|
fields = {f.name for f in msgspec.structs.fields(ServerArgs)}
|
|
for name, value in vars(Parallel).items():
|
|
if isinstance(value, Derived):
|
|
self.assertNotIn(name, fields)
|
|
|
|
|
|
if __name__ == "__main__":
|
|
unittest.main()
|