[Refactor] Clean up parallel runtime comments (#40632)

test/registered/dcp/test_reduce_scatter_along_dim.py

@@ -131,8 +131,6 @@ def init_distributed():
         local_rank=local_rank,
         backend="nccl",
     )
-    # The context answers a handle from what was stated on it, not from
-    # this module global, so a rank stood up by hand says so itself.
     get_parallel().override_permanently(world_group=coord)
 
     cpu_group = coord.cpu_group

test/registered/kernels/benchmark/communication/bench_custom_all_reduce.py

@@ -62,8 +62,6 @@ def _init_cpu_group() -> dist.ProcessGroup:
         local_rank=local_rank,
         backend="nccl",
     )
-    # The context answers a handle from what was stated on it, not from
-    # this module global, so a rank stood up by hand says so itself.
     get_parallel().override_permanently(world_group=coord)
     atexit.register(dist.destroy_process_group)
     torch.cuda.set_stream(torch.cuda.Stream())

test/registered/kernels/benchmark/communication/bench_symm_mem_all_gather.py

@@ -78,8 +78,6 @@ def _init_cpu_group() -> dist.ProcessGroup:
         local_rank=local_rank,
         backend="nccl",
     )
-    # The context answers a handle from what was stated on it, not from
-    # this module global, so a rank stood up by hand says so itself.
     get_parallel().override_permanently(world_group=ps._WORLD)
     atexit.register(dist.destroy_process_group)
     logging.disable(logging.INFO)

test/registered/kernels/benchmark/communication/bench_tp_qknorm.py

@@ -109,8 +109,6 @@ def _init_cpu_group() -> dist.ProcessGroup:
         local_rank=local_rank,
         backend="nccl",
     )
-    # The context answers a handle from what was stated on it, not from
-    # this module global, so a rank stood up by hand says so itself.
     get_parallel().override_permanently(world_group=coord)
     atexit.register(dist.destroy_process_group)
     logging.disable(logging.INFO)

test/registered/kernels/ops/communication/test_custom_all_reduce.py

@@ -130,8 +130,6 @@ def _init_cpu_group_once() -> dist.ProcessGroup:
         local_rank=local_rank,
         backend="nccl",
     )
-    # The context answers a handle from what was stated on it, not from
-    # this module global, so a rank stood up by hand says so itself.
     get_parallel().override_permanently(world_group=coord)
     atexit.register(dist.destroy_process_group)
     cpu_group = coord.cpu_group

test/registered/kernels/ops/communication/test_symm_mem_all_gather.py

@@ -71,8 +71,6 @@ def _init_cpu_group_once() -> dist.ProcessGroup:
         local_rank=local_rank,
         backend="nccl",
     )
-    # The context answers a handle from what was stated on it, not from
-    # this module global, so a rank stood up by hand says so itself.
     get_parallel().override_permanently(world_group=ps._WORLD)
     atexit.register(dist.destroy_process_group)
     logging.disable(logging.INFO)

test/registered/kernels/ops/communication/test_tp_qknorm.py

@@ -93,8 +93,6 @@ def _init_cpu_group_once() -> dist.ProcessGroup:
         local_rank=local_rank,
         backend="nccl",
     )
-    # The context answers a handle from what was stated on it, not from
-    # this module global, so a rank stood up by hand says so itself.
     get_parallel().override_permanently(world_group=coord)
     atexit.register(dist.destroy_process_group)
     cpu_group = coord.cpu_group

test/registered/kernels/ops/kimi_k3/test_ar_fusion.py

@@ -66,8 +66,6 @@ def _init_world():
         local_rank=local_rank,
         backend="nccl",
     )
-    # The context answers a handle from what was stated on it, not from
-    # this module global, so a rank stood up by hand says so itself.
     get_parallel().override_permanently(world_group=coord)
     atexit.register(dist.destroy_process_group)
     logging.disable(logging.INFO)

test/registered/kernels/ops/kimi_k3/test_collectives.py

@@ -54,8 +54,6 @@ def _init_world():
         local_rank=local_rank,
         backend="nccl",
     )
-    # The context answers a handle from what was stated on it, not from
-    # this module global, so a rank stood up by hand says so itself.
     get_parallel().override_permanently(world_group=coord)
     atexit.register(dist.destroy_process_group)
     cpu_group = coord.cpu_group

test/registered/kernels/ops/layernorm/test_mhc_kernels.py

@@ -13,12 +13,7 @@ register_cuda_ci(est_time=10, stage="base-b", runner_config="1-gpu-large")
 
 @pytest.fixture
 def stated_tp_group():
-    """A TP group for a test that runs in a process without one.
-
-    The production call passes the group *into* `use_symmetric_memory`, so
-    stubbing that context manager does not stop the read -- the argument is
-    evaluated first. Stating it on the context answers every spelling.
-    """
+    """Provide a TP-group placeholder for kernels with mocked symmetric memory."""
     from sglang.srt.runtime_context import get_parallel
 
     with get_parallel().override(tp_group=None):
@@ -35,10 +30,7 @@ def test_mhc_fused_post_pre_matches_unfused(
         pytest.skip("CUDA is required for TileLang mHC kernels")
 
     monkeypatch.setattr(mhc, "is_dsa_prefill_cp_interleave", lambda: False)
-    # This is a single-process kernel unit test with no TP group initialized.
-    # mhc_pre / mhc_fused_post_pre allocate the MoE input in the symmetric-memory
-    # pool, which asks for the TP group; bypassing the allocation is enough, and
-    # then nothing asks. Mirrors the workaround in test_mxfp4_sm90_cutlass.py.
+    # Disable symmetric-memory allocation for this single-process kernel test.
     monkeypatch.setattr(mhc, "use_symmetric_memory", lambda *a, **kw: nullcontext())
     monkeypatch.setattr(mhc, "is_allocation_symmetric", lambda: False)
     torch.manual_seed(0)

test/registered/kernels/ops/moe/test_minimax_quant_scatter.py

@@ -37,12 +37,7 @@ dev = "cuda"
 
 @pytest.fixture
 def stated_tp_group():
-    """A TP group for a test that runs in a process without one.
-
-    The production call passes the group *into* `use_symmetric_memory`, so
-    stubbing that context manager does not stop the read -- the argument is
-    evaluated first. Stating it on the context answers every spelling.
-    """
+    """Provide a TP-group placeholder for kernels with mocked symmetric memory."""
     from sglang.srt.runtime_context import get_parallel
 
     with get_parallel().override(tp_group=None):

test/registered/layers/mamba/test_mamba2_mixer.py

@@ -129,9 +129,6 @@ def mixer2_gated_norm_tensor_parallel(
         )
         mixer.weight.weight_loader(mixer.weight, weight)
 
-    # m2 reads tp via get_parallel().tp_size/rank — state a single-rank topology
-    # through the context. Every width that follows from `tp_size` is named:
-    # narrowing one leaf and leaving the quotients behind describes no layout.
     with get_parallel().override(
         tp_size=1,
         tp_rank=0,

test/registered/ops/test_aiter_greedy_sample_amd.py

@@ -37,10 +37,7 @@ def _mock_global_server_args(backend="pytorch"):
     class _DummyTPGroup:
         device_group = None
 
-    # `Sampler.__init__` asks the context for the group; state one for the rest
-    # of the process, since this process has no distributed init. Not the scoped
-    # `override()`: its context manager would be collected here and take the
-    # value back down with it.
+    # Provide a TP group for sampler initialization without distributed setup.
     get_parallel().override_permanently(tp_group=_DummyTPGroup())
     from sglang.srt.runtime_context import get_flags
 

test/registered/unit/constrained/test_grammar_manager.py

@@ -47,12 +47,9 @@ register_cpu_ci(est_time=5, suite="stage-b-test-cpu-intel")
 
 
 def _make_scheduler(grammar_backend_name="none", skip_tokenizer=False):
-    """Create a mock scheduler with necessary attributes.
+    """Create a mock scheduler and publish its configuration and placement.
 
-    The grammar manager reads its config and its place in the pipeline from
-    the context, so the settings that used to be hung off the mock are
-    published instead. The caller resets the context; every test here goes
-    through `_GrammarFixture`.
+    The caller must reset the context during teardown.
     """
     reset_context()
     server_args = ServerArgs(
@@ -778,9 +775,7 @@ class TestGrammarManagerPPSync(unittest.TestCase):
         enter_override(
             self, get_context().override_server_args(skip_tokenizer_init=True)
         )
-        # After that override, not before: installing a server-args override
-        # re-resolves the parallel bag from defaults, which puts `pp_size`
-        # back to 1 whatever was published.
+        # Override ranks after the server-args override rebuilds the config bags.
         enter_scope(self, get_parallel().override(pp_size=pp_size, pp_rank=pp_rank))
         scheduler.pp_group = pp_group
         mgr = GrammarManager(scheduler)

test/registered/unit/disaggregation/test_register_to_bootstrap.py

@@ -204,8 +204,6 @@ class TestRegisterToBootstrap(CustomTestCase):
     def test_rust_attention_dp_replicates_complete_topology_across_hosts(
         self, mock_put
     ):
-        # The consumer reads the group through `get_parallel()`, so the
-        # stub is stated there rather than in the module the build writes.
         mock_world_group = MagicMock()
         success_resp = MagicMock()
         success_resp.status_code = 200

test/registered/unit/entrypoints/test_rust_server_dp_ports.py

@@ -45,9 +45,6 @@ def test_dp_leaders_reuse_node_local_ports(
                 server_args=SimpleNamespace(),
                 model_config=SimpleNamespace(is_multimodal=False),
             )
-            # Where this rank sits, stated whole: the attention rank follows
-            # from the TP rank and the attention-TP width, and the identities
-            # refuse the combination if it describes no real layout.
             with parallel.override(
                 tp_rank=tp_rank,
                 attn_dp_rank=dp_rank,

test/registered/unit/layers/attention/test_flashattention_pa_swa_prefill_lens_size.py

@@ -93,8 +93,6 @@ def _make_prefill_aware_swa_runner(
         page_size=1,
         attn_cp_size=1,
         tp_size=1,
-        # The backend still reads the runner's frozen record for these two;
-        # same single-rank placement, stated where it looks for it.
         ps=SimpleNamespace(attn_cp_size=1, tp_size=1),
         is_draft_worker=False,
         server_args=server_args,

test/registered/unit/layers/quantization/test_mxfp4_sm100_trtllm_gen.py

@@ -33,12 +33,7 @@ GROUP_SIZE = 32  # MXFP4 block size
 
 @pytest.fixture
 def stated_tp_group():
-    """A TP group for a test that runs in a process without one.
-
-    The production call passes the group *into* `use_symmetric_memory`, so
-    stubbing that context manager does not stop the read -- the argument is
-    evaluated first. Stating it on the context answers every spelling.
-    """
+    """Provide a TP-group placeholder for kernels with mocked symmetric memory."""
     from sglang.srt.runtime_context import get_parallel
 
     with get_parallel().override(tp_group=None):

test/registered/unit/layers/quantization/test_mxfp4_sm120_cutlass.py

@@ -19,12 +19,7 @@ register_cuda_ci(est_time=14, stage="base-b", runner_config="1-gpu-small")
 
 @pytest.fixture
 def stated_tp_group():
-    """A TP group for a test that runs in a process without one.
-
-    The production call passes the group *into* `use_symmetric_memory`, so
-    stubbing that context manager does not stop the read -- the argument is
-    evaluated first. Stating it on the context answers every spelling.
-    """
+    """Provide a TP-group placeholder for kernels with mocked symmetric memory."""
     from sglang.srt.runtime_context import get_parallel
 
     with get_parallel().override(tp_group=None):

test/registered/unit/layers/quantization/test_mxfp4_sm90_cutlass.py

@@ -65,12 +65,7 @@ GROUP_SIZE = 32  # MXFP4 block size
 
 @pytest.fixture
 def stated_tp_group():
-    """A TP group for a test that runs in a process without one.
-
-    The production call passes the group *into* `use_symmetric_memory`, so
-    stubbing that context manager does not stop the read -- the argument is
-    evaluated first. Stating it on the context answers every spelling.
-    """
+    """Provide a TP-group placeholder for kernels with mocked symmetric memory."""
     from sglang.srt.runtime_context import get_parallel
 
     with get_parallel().override(tp_group=None):

test/registered/unit/managers/test_disagg_idle_step_counters.py

@@ -64,8 +64,6 @@ def load_mlx_scheduler_module():
 class TestSchedulerIdleStepCounters(CustomTestCase):
     def setUp(self):
         super().setUp()
-        # The loop asks the context where this process sits; nothing here
-        # builds a process group, so the placement arrives by publishing one.
         enter_scope(self, published_topology(role="scheduler"))
 
     @parameterized.expand(
@@ -184,11 +182,7 @@ class TestSchedulerIdleStepCounters(CustomTestCase):
                         f"{PDMUX_MODULE}.torch.cuda.stream",
                         side_effect=lambda stream: nullcontext(),
                     ),
-                    # The prefill section runs under the duplicate communicator
-                    # `--enable-pdmux` builds, in place of the module flag this
-                    # replaces. The loop has no process groups at all, so stand
-                    # one in: the scope refuses to open without it rather than
-                    # letting prefill quietly share the decode communicator.
+                    # PD multiplexing requires a separate prefill communicator.
                     patch.object(
                         parallel_state,
                         "_PDMUX_PREFILL_TP_GROUP",

test/registered/unit/managers/test_loadstat_wire.py

@@ -97,12 +97,10 @@ class TestLoadPublisherGating(CustomTestCase):
     """
 
     def _build(self, *, config=ZMQ_ENDPOINT, explicit="auto", ranks=None, **topology):
-        """Construct a publisher with the socket bind stubbed out, returning
-        (publisher, captured _open_pub_socket mock). Opts in via explicit="auto"
-        by default (the feature is off without it). The topology is published
-        rather than overridden, so the ranks the publisher reads are the ones a
-        layout of that shape actually produces; every read happens in the
-        constructor."""
+        """Return a publisher and its mocked socket factory under a published topology.
+
+        ``explicit="auto"`` enables load publication by default.
+        """
         with (
             published_topology(ranks=ranks, **topology),
             patch(

test/registered/unit/managers/test_output_streamer_customized_info.py

@@ -103,7 +103,6 @@ class TestOutputStreamerCustomizedInfo(unittest.TestCase):
         )
         serving_patch.start()
         observability_patch.start()
-        # The streamer asks the context which rank it is streaming from.
         enter_scope(self, published_topology(ranks={"dp_rank": 0}))
         self.addCleanup(serving_patch.stop)
         self.addCleanup(observability_patch.stop)

test/registered/unit/managers/test_pp_cp_rank_offsets.py

@@ -23,11 +23,9 @@ register_cpu_ci(est_time=11, suite="base-a-test-cpu")
 
 
 def _published_topology():
-    """The topology these tests run in.
+    """Publish WORLD rank 12 with TP=8 and PP=2.
 
-    World rank 12 of a `tp=8, pp=2` world is `tp_rank=4` on the second stage,
-    which puts this process at `attn_dp_rank=1` with `attn_tp_rank=0`: the
-    context derives all of them from that one number and the widths.
+    This gives TP rank 4, PP rank 1, attention-DP rank 1, and attention-TP rank 0.
     """
     return published_topology(
         role="scheduler",

test/registered/unit/managers/test_scheduler_chunked_abort_race.py

@@ -50,7 +50,6 @@ def _make_scheduler(pending_req, *, chunked_req, running_reqs) -> Scheduler:
 
 class TestPendingChunkedAbortRace(CustomTestCase):
     def setUp(self):
-        # The abort path asks the context for the pipeline width.
         enter_scope(self, published_topology())
 
     def test_req_left_chunked_slot_is_aborted(self):

test/registered/unit/managers/test_scheduler_timeouts.py

@@ -131,7 +131,6 @@ class TestWaitingTimeout(CustomTestCase):
 
 class TestRunningTimeout(CustomTestCase):
     def setUp(self):
-        # The poll asks the context for the pipeline width.
         enter_scope(self, published_topology())
 
     def test_emits_only_stale_unfinished_reqs_without_marking(self):

test/registered/unit/mem_cache/test_mamba_donated_alloc_ratio.py

@@ -297,7 +297,6 @@ class TestPPMambaPoolSizing(unittest.TestCase):
             server_args=SimpleNamespace(),
             spec_algorithm=SimpleNamespace(is_none=lambda: True),
             layer_info=SimpleNamespace(start_layer=start, end_layer=end),
-            # The runner carries its placement as plain attributes.
             attn_dp_size=1,
             pp_size=pp_size,
             hybrid_gdn_config=None,

test/registered/unit/mem_cache/test_mem_pool_host.py

@@ -302,10 +302,6 @@ class TestHostMemoryBudget(CustomTestCase):
         )
 
     def test_ranks_per_host_divides_world_size_by_nodes(self):
-        # The launcher slices ranks uniformly across nodes, so the co-located
-        # rank count is world_size // nnodes — no hostname collective.
-        # tp_size=16 states the launch width the count divides -- the
-        # published configuration is where ranks_per_host reads it from.
         with (
             get_context().override_server_args(nnodes=2, tp_size=16),
             unittest.mock.patch.object(

test/registered/unit/model_executor/test_pool_configurator.py

@@ -35,8 +35,6 @@ def mock_cpu_env(kv_size=2, tp_size=1, swa_eviction_interval=4):
 
     with (
         patch("torch._utils._element_size", return_value=kv_size),
-        # The whole attention triple, not just one leaf: a width that does not
-        # factor describes no layout.
         get_parallel().override(
             tp_size=tp_size,
             attn_tp_size=tp_size,

test/registered/unit/model_loader/test_transformers_fallback.py

@@ -55,8 +55,6 @@ class TestTransformersFallbackSkipSubstrs(CustomTestCase):
             pass
 
         with (
-            # `__init__` only stashes the pipeline group, so an empty
-            # stand-in carries it past the read.
             get_parallel().override(pp_group=SimpleNamespace()),
             patch(
                 "sglang.srt.models.transformers.get_hf_text_config",

test/registered/unit/models/test_glm5_next_bfg_fusion.py

@@ -83,9 +83,6 @@ class TestGlm5NextBfgFusion(unittest.TestCase):
             for attn_tp, rank in ((1, 0), (2, 0), (2, 1)):
                 with (
                     self.subTest(route=expected_route, attn_tp=attn_tp, rank=rank),
-                    # A width is a whole topology: the attention triple has
-                    # to factor `tp_size`, and this process has to sit where
-                    # the triple puts it.
                     get_parallel().override(
                         tp_size=4,
                         tp_rank=rank,

test/registered/unit/observability/test_forward_pass_metrics.py

@@ -64,12 +64,7 @@ class _DummyPublisherThread:
 
 
 def _publish_server_args(test, **fields):
-    """Publish a config for the reporter under test and return the instance.
-
-    The collector asks the context where this process sits, so the ranks are
-    stated too: without them a rank read falls through to a process group that
-    a unit test has not built.
-    """
+    """Install reporter configuration and rank overrides, with test cleanup."""
     fields.setdefault("decode_log_interval", 40)
     override = get_context().override_server_args(**fields)
     server_args = override.install()
@@ -297,8 +292,6 @@ class TestForwardPassMetrics(unittest.TestCase):
             forward_pass_metrics_ipc_name=None,
             kv_events_config=None,
         )
-        # The reporter asks the context whether this is the last stage, and
-        # which replica it is reporting for.
         enter_scope(self, get_parallel().override(pp_rank=0, pp_size=1, dp_rank=2))
         scheduler.enable_kv_cache_events = False
 
@@ -336,7 +329,6 @@ class TestForwardPassMetrics(unittest.TestCase):
             forward_pass_metrics_ipc_name=None,
             kv_events_config=None,
         )
-        # The reporter asks the context whether this is the last stage.
         enter_scope(self, get_parallel().override(pp_rank=0, pp_size=2))
         scheduler.enable_kv_cache_events = False
 

test/registered/unit/server_args/test_resolution_declarations.py

@@ -132,15 +132,7 @@ def _stash_overlay(server_args):
 
 
 def _live_topology_leaves():
-    """Names `ParallelContext` answers from a runtime write, not the config.
-
-    Read off the declarations that carry no `fn`, which is what those are.
-    Inferring them from "did the read raise" is wrong -- it raises only while
-    nothing has written the name, so in a process where an earlier test stated
-    one the property answers that value and a leaf check reads it as a config
-    mismatch (`parallel.tp_size: bag=1 resolution=2`). Whether a name is
-    shadowed is a property of the declaration, not of the process.
-    """
+    """Return runtime-only parallel fields, identified by declarations without ``fn``."""
     from sglang.srt.runtime_context import _derived_widths
 
     return frozenset(n for n, d in _derived_widths().items() if not d.fn)

test/registered/unit/spec/test_dflash_logits.py

@@ -249,7 +249,6 @@ def test_worker_folds_a_gate_admitted_quantized_selector_head(monkeypatch):
         block_size=8,
         selector=object(),
         model_runner=SimpleNamespace(tp_rank=0),
-        # The worker rank-gates its logging on its own frozen record.
         ps=SimpleNamespace(tp_rank=0),
         draft_model=SimpleNamespace(lm_head=None),
         device="cpu",
@@ -285,7 +284,6 @@ def test_worker_warns_once_when_selector_sampling_is_disabled(monkeypatch):
         _selector_sampling_enabled=False,
         _warned_sampling_fallback=False,
         model_runner=SimpleNamespace(tp_rank=0),
-        # The worker rank-gates its logging on its own frozen record.
         ps=SimpleNamespace(tp_rank=0),
     )
     batch = SimpleNamespace(sampling_info=SimpleNamespace(is_all_greedy=False))

test/registered/unit/utils/test_weight_versions.py

@@ -316,7 +316,6 @@ class _SchedulerStub:
 
 class TestSchedulerRecordWeightVersionChange(CustomTestCase):
     def _scheduler(self, *args, pp_size=1, **kwargs):
-        # The recording path asks the context for the pipeline width.
         enter_scope(self, published_topology(pp_size=pp_size))
         scheduler = _SchedulerStub(*args, **kwargs)
         for name, value in (