diff --git a/.github/workflows/_pr-test-check-changes.yml b/.github/workflows/_pr-test-check-changes.yml
index 870c00118e..b89eb8e167 100644
--- a/.github/workflows/_pr-test-check-changes.yml
+++ b/.github/workflows/_pr-test-check-changes.yml
@@ -29,6 +29,8 @@ on:
value: ${{ jobs.run.outputs.multimodal_gen }}
renderer:
value: ${{ jobs.run.outputs.renderer }}
+ server_args_docs:
+ value: ${{ jobs.run.outputs.server_args_docs }}
rust_workspace:
value: ${{ jobs.run.outputs.rust_workspace }}
partitions:
@@ -53,6 +55,7 @@ jobs:
jit_kernel: ${{ steps.filter.outputs.jit_kernel || steps.run-mode.outputs.run_all_tests }}
multimodal_gen: ${{ steps.filter.outputs.multimodal_gen || steps.run-mode.outputs.run_all_tests }}
renderer: ${{ steps.filter.outputs.renderer || steps.run-mode.outputs.run_all_tests }}
+ server_args_docs: ${{ steps.filter.outputs.server_args_docs || steps.run-mode.outputs.run_all_tests }}
rust_workspace: ${{ steps.filter.outputs.rust_workspace || steps.run-mode.outputs.run_all_tests }}
partitions: ${{ steps.partitions.outputs.partitions }}
partition_model_sha: ${{ steps.partition-model-sha.outputs.sha }}
@@ -115,6 +118,13 @@ jobs:
renderer:
- "rust/sglang-renderer/**"
- "rust/sglang-processor/**"
+ server_args_docs:
+ # test_server_args_docs_coverage.py diffs this doc and its allowlist
+ # against the CLI, so a docs only edit can break that test. Kept out
+ # of main_package which also gates every GPU stage. Only the CPU
+ # stage reads this.
+ - "docs/docs/advanced_features/server_arguments.mdx"
+ - "docs/docs/advanced_features/server_arguments_undocumented.txt"
simulator:
- ".github/workflows/pr-test-extra.yml"
- ".github/workflows/_pr-test-simulator-cpu.yml"
@@ -271,6 +281,7 @@ jobs:
echo "| jit_kernel | ${{ steps.filter.outputs.jit_kernel || steps.run-mode.outputs.run_all_tests }} |"
echo "| multimodal_gen | ${{ steps.filter.outputs.multimodal_gen || steps.run-mode.outputs.run_all_tests }} |"
echo "| rust_workspace | ${{ steps.filter.outputs.rust_workspace || steps.run-mode.outputs.run_all_tests }} |"
+ echo "| server_args_docs | ${{ steps.filter.outputs.server_args_docs || steps.run-mode.outputs.run_all_tests }} |"
echo "| b200_runner | ${{ steps.set-runner.outputs.b200_runner }} |"
echo "| enable_retry | ${{ steps.set-retry.outputs.enable_retry }} |"
echo "| continue_on_error | ${{ steps.set-continue-on-error.outputs.continue_on_error }} |"
diff --git a/.github/workflows/_pr-test-stage-cpu.yml b/.github/workflows/_pr-test-stage-cpu.yml
index 9206c8761d..d98c0feff2 100644
--- a/.github/workflows/_pr-test-stage-cpu.yml
+++ b/.github/workflows/_pr-test-stage-cpu.yml
@@ -13,7 +13,7 @@ on:
type: string
required: true
check_changes:
- description: 'toJson(needs.check-changes.outputs). Read via fromJson(...).main_package / renderer / rust_workspace / continue_on_error / partition_model_sha.'
+ description: 'toJson(needs.check-changes.outputs). Read via fromJson(...).main_package / renderer / server_args_docs / rust_workspace / continue_on_error / partition_model_sha.'
type: string
required: true
caller_inputs:
@@ -49,7 +49,7 @@ jobs:
# emits a skipped entry for wait-for-jobs to match. The schedule /
# parallel-dispatch / !failure() half stays on the caller, where the
# `needs` context it guards actually lives.
- if: fromJson(inputs.check_changes).main_package == 'true' || fromJson(inputs.check_changes).renderer == 'true'
+ if: fromJson(inputs.check_changes).main_package == 'true' || fromJson(inputs.check_changes).renderer == 'true' || fromJson(inputs.check_changes).server_args_docs == 'true'
runs-on: ubuntu-latest
timeout-minutes: 240
env:
diff --git a/docs/.mintignore b/docs/.mintignore
index 9922f06dc8..98923e24fe 100644
--- a/docs/.mintignore
+++ b/docs/.mintignore
@@ -5,3 +5,6 @@
# Draft content
drafts/
*.draft.mdx
+
+# Data read by test_server_args_docs_coverage.py, not a page
+advanced_features/server_arguments_undocumented.txt
diff --git a/docs/docs/advanced_features/server_arguments.mdx b/docs/docs/advanced_features/server_arguments.mdx
index cb3fea10a5..9a065c0c29 100644
--- a/docs/docs/advanced_features/server_arguments.mdx
+++ b/docs/docs/advanced_features/server_arguments.mdx
@@ -780,12 +780,6 @@ Please consult the documentation below and [server_args.py](https://github.com/s
`False` |
bool flag (set to enable) |
-
- | `--custom-sigquit-handler` |
- Register a custom sigquit handler so you can do additional cleanup after the server is shutdown. This is only available for Engine, not for CLI. |
- `None` |
- Type: str |
-
| `--batch-notify-size` |
Number of streaming notifications to batch before yielding to the event loop. Reduces asyncio wakeup overhead under high concurrency. |
@@ -1611,6 +1605,12 @@ Combining `--enable-response-store` with `--disaggregation-mode=prefill` or `dec
The accept probability of a draft token is raised from its target probability p to min(1, p / threshold_acc). |
`1.0` |
Type: float |
+
+
+ | `--speculative-use-block-verification` |
+ Use block verification for EAGLE/EAGLE3/NEXTN on CUDA or ROCm (requires topk=1). |
+ `False` |
+ Type: bool |
| `--speculative-token-map` |
@@ -2968,12 +2968,6 @@ This option does not disable required attention or MoE collectives, including ex
The input filename for dumping tensors |
`None` |
Type: str |
-
-
- | `--debug-tensor-dump-inject` |
- Inject the outputs from jax as the input of every layer. |
- `False` |
- Type: str |
@@ -3050,8 +3044,8 @@ This option does not disable required attention or MoE collectives, including ex
bool flag (set to enable) |
- | `--optimistic-prefill-retries` |
- Number of optimistic prefill retries that will skip the bootstrap wait. |
+ `--optimistic-prefill-attempts` |
+ Number of optimistic prefill forward passes that skip the bootstrap wait. |
`0` |
Type: int |
@@ -3434,36 +3428,3 @@ This option does not disable required attention or MoE collectives, including ex
-
-## Deprecated arguments
-
-
-
-
-
-
-
-
-
- | Argument |
- Description |
- Defaults |
- Options |
-
-
-
-
-
-
-
-
-
- | `--hybrid-kvcache-ratio` |
- Mix ratio in [0,1] between uniform and hybrid kv buffers (0.0 = pure uniform: swa_size / full_size = 1)(1.0 = pure hybrid: swa_size / full_size = local_attention_size / context_length) |
- `None` |
- Optional[float] |
-
-
-
-
-
diff --git a/docs/docs/advanced_features/server_arguments_undocumented.txt b/docs/docs/advanced_features/server_arguments_undocumented.txt
new file mode 100644
index 0000000000..c56089778c
--- /dev/null
+++ b/docs/docs/advanced_features/server_arguments_undocumented.txt
@@ -0,0 +1,95 @@
+# Live ServerArgs flags with no row in server_arguments.mdx. One flag per line.
+# test/registered/unit/server_args/test_server_args_docs_coverage.py reads this
+# list. It may only shrink: when you add a row for a flag, delete it here.
+--c128-page-size
+--constrained-json-max-whitespace-cnt
+--cuda-graph-prefill-max-context
+--decoupled-spec-bind-endpoint
+--decoupled-spec-connect-endpoints
+--decoupled-spec-rank
+--decoupled-spec-role
+--deepep-v2-mode
+--disable-aiter-allreduce-fusion-in-decode
+--disable-aiter-allreduce-fusion-in-prefill
+--disaggregation-decode-extra-slots
+--disaggregation-decode-host-receive-threshold
+--disaggregation-decode-retraction-backup
+--disaggregation-enable-kv-checksum
+--dsa-paged-mqa-logits-backend
+--dsv4-attn-backend
+--dsv4-prefill-backend
+--dwdp-size
+--elastic-ep-initial-size
+--elastic-ep-join-mode
+--elastic-ep-join-rank-offset
+--elastic-ep-scale-timeout
+--enable-cp-decode-attn-tp
+--enable-decoder-swa-bounded-replay
+--enable-dense-mlp-attn-tp
+--enable-dsa-cache-layer-split
+--enable-dsa-fused-indexer
+--enable-encoder-swa-bounded-replay
+--enable-flexkv
+--enable-layernorm-sp
+--enable-lean-attention
+--enable-linear-replayssm
+--enable-linear-replayssm-spec
+--enable-linker-mla-dedup
+--enable-pd-role-switch
+--enable-response-store
+--enable-scattered-sconv
+--enable-session-radix-cache
+--enable-shared-experts-attn-tp
+--enable-tp-lm-head-all-to-all
+--enable-unified-cache-external-linker
+--enable-w4a4-mxfp4-megamoe
+--flexkv-config-file
+--fuseep-mode
+--gated-launch-port
+--grpc-port
+--grpc-response-timeout-secs
+--hicache-host-memory-fraction
+--hicache-host-memory-mode
+--hicache-storage-prefetch-retry-max-attempts
+--hicache-storage-prefetch-retry-poll-interval
+--http2-initial-connection-window-size
+--linear-attn-verify-backend
+--linear-replayssm-cache-len
+--mamba-max-states-per-path
+--max-ep-size
+--min-free-slots-delay
+--mm-feature-transport
+--mm-global-cache-backend
+--mm-io-worker-num
+--mm-preprocess-cache-size-mb
+--mm-processor-worker-num
+--otlp-service-name
+--prefill-decode-interval
+--prefill-kv-cache-dequant-dtype
+--qsa-indexer-dtype
+--radix-eviction-policy-config
+--return-input-ids
+--return-output-ids
+--sampling-mask-max-tokens
+--shared-experts-tp-size
+--sidecar
+--sidecar-args
+--smg-grpc-mode
+--spec-trace-dir
+--speculative-domino-candidate-pool-size
+--speculative-draft-kv-cache-dtype
+--speculative-draft-sink-size
+--speculative-dspark-align-verify-tokens-to-graph-tier
+--speculative-dspark-block-size
+--speculative-dspark-confidence-sts-path
+--speculative-dspark-sps-table-path
+--speculative-use-rejection-sampling
+--startup-weight-load-mode
+--swa-prefix-tails
+--trust-mm-content-hashes
+--trust-request-chat-template
+--unified-cache-external-linker-backend
+--uno-lora-path
+--weight-cache-mode
+--weight-cache-socket
+--weight-cache-timeout
diff --git a/test/registered/unit/server_args/test_server_args_docs_coverage.py b/test/registered/unit/server_args/test_server_args_docs_coverage.py
new file mode 100644
index 0000000000..fa127755be
--- /dev/null
+++ b/test/registered/unit/server_args/test_server_args_docs_coverage.py
@@ -0,0 +1,145 @@
+"""Every live ServerArgs flag needs a row in server_arguments.mdx and every flag
+the doc names must still register. The allowlist may only shrink.
+"""
+
+import argparse
+import re
+import unittest
+from collections.abc import Iterable
+from pathlib import Path
+
+from sglang.srt.arg_groups.argparse_actions import (
+ DeprecatedAction,
+ DeprecatedAliasStoreAction,
+ DeprecatedStoreConstAction,
+ DeprecatedStoreTrueAction,
+)
+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")
+
+_DOC_DIR = Path(__file__).resolve().parents[4] / "docs" / "docs" / "advanced_features"
+_DOC_PATH = _DOC_DIR / "server_arguments.mdx"
+# Live flags with no row. May only shrink. Kept beside the doc, outside the
+# main_package CI paths, so a PR that backfills rows runs only the CPU stage.
+_UNDOCUMENTED_PATH = _DOC_DIR / "server_arguments_undocumented.txt"
+
+_DEPRECATED_ACTION_TYPES = (
+ DeprecatedAction,
+ DeprecatedStoreTrueAction,
+ DeprecatedStoreConstAction,
+ DeprecatedAliasStoreAction,
+)
+
+# Argument, Description, Defaults, Options.
+_DOC_COLUMNS = 4
+
+# Only delimited mentions count, so a prose wildcard like `--cuda-graph-*`
+# never becomes a fabricated flag name.
+_ROW_FLAG_RE = re.compile(r"`(--[a-zA-Z0-9][\w-]*)`|(--[a-zA-Z0-9][\w-]*)")
+
+
+def _undocumented() -> frozenset[str]:
+ lines = _UNDOCUMENTED_PATH.read_text(encoding="utf-8").splitlines()
+ return frozenset(
+ line.strip() for line in lines if line.strip() and not line.startswith("#")
+ )
+
+
+_UNDOCUMENTED = _undocumented()
+
+
+def _parser_actions() -> tuple[list[argparse.Action], list[argparse.Action]]:
+ parser = argparse.ArgumentParser(add_help=False)
+ ServerArgs.add_cli_args(parser)
+ live, deprecated = [], []
+ for action in parser._actions:
+ (deprecated if isinstance(action, _DEPRECATED_ACTION_TYPES) else live).append(
+ action
+ )
+ return live, deprecated
+
+
+def _doc_rows() -> list[list[str]]:
+ text = _DOC_PATH.read_text(encoding="utf-8-sig")
+ rows = [
+ re.findall(r"]*>(.*?) | ", row, re.S)
+ for row in re.findall(r"]*>(.*?)
", text, re.S)
+ ]
+ # Header rows hold only cells.
+ return [cells for cells in rows if cells]
+
+
+def _flags_in(cells: Iterable[str]) -> set[str]:
+ return {
+ match.group(1) or match.group(2)
+ for cell in cells
+ for match in _ROW_FLAG_RE.finditer(cell)
+ }
+
+
+class TestServerArgsDocsCoverage(CustomTestCase):
+ @classmethod
+ def setUpClass(cls):
+ cls.live_actions, deprecated_actions = _parser_actions()
+ cls.rows = _doc_rows()
+ argument_column_flags = _flags_in(cells[0] for cells in cls.rows)
+ cls.uncovered = {
+ action.option_strings[0]
+ for action in cls.live_actions
+ # Any alias counts.
+ if argument_column_flags.isdisjoint(action.option_strings)
+ }
+ registered = {
+ o for a in cls.live_actions + deprecated_actions for o in a.option_strings
+ }
+ cls.stale = _flags_in(cell for cells in cls.rows for cell in cells) - registered
+
+ def test_every_row_has_four_cells(self):
+ """Coverage reads only the first cell. A row the cell regex splits wrongly
+ can pass description text off as the Argument column and hide a gap."""
+ misshapen = [cells[0][:80] for cells in self.rows if len(cells) != _DOC_COLUMNS]
+ self.assertFalse(
+ misshapen,
+ f"these rows do not have {_DOC_COLUMNS} cells, so their first cell may "
+ "not be the Argument column:\n " + "\n ".join(misshapen),
+ )
+
+ def test_no_new_undocumented_flags(self):
+ added = sorted(self.uncovered - _UNDOCUMENTED)
+ self.assertFalse(
+ added,
+ "these live flags have no row in server_arguments.mdx and are not "
+ "on the allow list. Add a row (lift the help text from the "
+ "field's Arg(help=...)) or if it is a deliberate omission, add "
+ f"it to {_UNDOCUMENTED_PATH.name}:\n " + "\n ".join(added),
+ )
+
+ def test_undocumented_allowlist_is_current(self):
+ live_options = {a.option_strings[0] for a in self.live_actions}
+ outdated = []
+ for flag in sorted(_UNDOCUMENTED - self.uncovered):
+ reason = "now documented" if flag in live_options else "flag is gone"
+ outdated.append(f"{flag} ({reason})")
+ self.assertFalse(
+ outdated,
+ f"these entries in {_UNDOCUMENTED_PATH.name} are out of date. "
+ "Remove them from the file:\n " + "\n ".join(outdated),
+ )
+
+ def test_no_stale_rows(self):
+ stale = sorted(self.stale)
+ self.assertFalse(
+ stale,
+ "server_arguments.mdx names these flags but they no longer register "
+ "as a current or deprecated flag. The flag was renamed (possibly via "
+ "cli_name=) or removed, or its field has no CLI surface (no A[] "
+ "annotation, or Arg(no_cli=True)). Fix or delete the mention:\n "
+ + "\n ".join(stale),
+ )
+
+
+if __name__ == "__main__":
+ unittest.main()
|