Support overriding and post-hoc providing metadata in dump comparator (#19598)

This commit is contained in:
fzyzcjy
2026-03-01 10:35:06 +08:00
committed by GitHub
parent e41164af1c
commit e78f1283f7
5 changed files with 817 additions and 101 deletions

View File

@@ -25,6 +25,7 @@ from sglang.srt.debug_utils.comparator.dims import (
resolve_dim_names,
)
from sglang.srt.debug_utils.comparator.dp_utils import filter_to_non_empty_dp_rank
from sglang.srt.debug_utils.comparator.meta_overrider import MetaOverrider
from sglang.srt.debug_utils.comparator.output_types import (
ComparisonRecord,
GeneralWarning,
@@ -54,6 +55,7 @@ def compare_bundle_pair(
),
viz_output_dir: Optional[Path] = None,
compute_per_token: bool = False,
meta_overrider: Optional[MetaOverrider] = None,
) -> Union[ComparisonRecord, SkipRecord, NonTensorRecord]:
with warning_sink.context() as collected_warnings:
result = _compare_bundle_pair_inner(
@@ -66,6 +68,7 @@ def compare_bundle_pair(
thd_seq_lens_by_step_pair=thd_seq_lens_by_step_pair,
viz_output_dir=viz_output_dir,
compute_per_token=compute_per_token,
meta_overrider=meta_overrider,
)
return result.model_copy(update={"warnings": collected_warnings})
@@ -84,6 +87,7 @@ def _compare_bundle_pair_inner(
),
viz_output_dir: Optional[Path] = None,
compute_per_token: bool = False,
meta_overrider: Optional[MetaOverrider] = None,
) -> Union[ComparisonRecord, SkipRecord, NonTensorRecord]:
# 1. Load all successfully loaded values
all_pair: Pair[list[ValueWithMeta]] = Pair(
@@ -101,6 +105,24 @@ def _compare_bundle_pair_inner(
y=filter_to_non_empty_dp_rank(all_pair.y),
)
# 1c. Dims override: patch meta["dims"] before downstream reads it
if meta_overrider is not None and not meta_overrider.is_empty:
_apply = meta_overrider.apply_to_meta
all_pair = Pair(
x=[
ValueWithMeta(
value=v.value, meta=_apply(name=name, meta=v.meta, side="baseline")
)
for v in all_pair.x
],
y=[
ValueWithMeta(
value=v.value, meta=_apply(name=name, meta=v.meta, side="target")
)
for v in all_pair.y
],
)
# 2. Check if any side has non-tensor values → non-tensor display path
has_non_tensor: bool = any(
not isinstance(it.value, torch.Tensor) for it in [*all_pair.x, *all_pair.y]

View File

@@ -22,6 +22,7 @@ from sglang.srt.debug_utils.comparator.bundle_matcher import (
match_bundles,
)
from sglang.srt.debug_utils.comparator.display import emit_display_records
from sglang.srt.debug_utils.comparator.meta_overrider import MetaOverrider
from sglang.srt.debug_utils.comparator.output_types import (
ComparisonRecord,
ConfigRecord,
@@ -85,6 +86,13 @@ def run(args: argparse.Namespace) -> None:
Path(args.visualize_per_token) if args.visualize_per_token else None
)
meta_overrider: MetaOverrider = MetaOverrider.from_args_and_config(
override_dims=args.override_dims,
override_baseline_dims=args.override_baseline_dims,
override_target_dims=args.override_target_dims,
override_config=Path(args.override_config) if args.override_config else None,
)
comparison_records = _compare_bundle_pairs(
bundle_info_pairs=bundle_info_pairs,
baseline_path=Path(args.baseline_path),
@@ -94,6 +102,7 @@ def run(args: argparse.Namespace) -> None:
thd_seq_lens_by_step_pair=ta_result.thd_seq_lens_by_step_pair,
viz_output_dir=viz_output_dir,
compute_per_token=visualize_per_token is not None,
meta_overrider=meta_overrider,
)
_consume_comparison_records(
comparison_records=comparison_records,
@@ -155,6 +164,7 @@ def _compare_bundle_pairs(
thd_seq_lens_by_step_pair: Pair[Optional[dict[int, list[int]]]],
viz_output_dir: Optional[Path] = None,
compute_per_token: bool = False,
meta_overrider: Optional[MetaOverrider] = None,
) -> Iterator[Union[ComparisonRecord, SkipRecord, NonTensorRecord]]:
for bundle_info_pair in bundle_info_pairs:
if not bundle_info_pair.y:
@@ -174,6 +184,7 @@ def _compare_bundle_pairs(
thd_seq_lens_by_step_pair=thd_seq_lens_by_step_pair,
viz_output_dir=viz_output_dir,
compute_per_token=compute_per_token,
meta_overrider=meta_overrider,
)
@@ -252,4 +263,31 @@ def _parse_args() -> argparse.Namespace:
default=None,
help="Output path for per-token relative difference heatmap PNG",
)
# Dims override
parser.add_argument(
"--override-dims",
action="append",
default=[],
help="Override dims for both sides: 'name:dims_string' (repeatable)",
)
parser.add_argument(
"--override-baseline-dims",
action="append",
default=[],
help="Override dims for baseline only: 'name:dims_string' (repeatable)",
)
parser.add_argument(
"--override-target-dims",
action="append",
default=[],
help="Override dims for target only: 'name:dims_string' (repeatable)",
)
parser.add_argument(
"--override-config",
type=str,
default=None,
help="Path to YAML override config file (dims overrides, etc.)",
)
return parser.parse_args()

View File

@@ -0,0 +1,107 @@
"""Meta overrider: replace metadata fields without re-running dumps.
Currently only overrides 'dims', but the design supports overriding
additional meta fields (e.g. parallel_info) in the future.
"""
from __future__ import annotations
import re
from pathlib import Path
from typing import Any, Literal, Optional
import yaml
from sglang.srt.debug_utils.comparator.utils import _StrictBase
class MetaOverrideRule(_StrictBase):
"""Single override rule: regex match on tensor name → replacement meta field(s).
Currently only 'dims' is supported; more fields may be added in the future.
"""
match: str
dims: str
side: Literal["both", "baseline", "target"] = "both"
class MetaOverrideConfig(_StrictBase):
"""YAML top-level config for overriding comparator behavior."""
overrides: list[MetaOverrideRule] = []
class MetaOverrider:
"""Holds override rules and applies first-match-wins replacement."""
def __init__(self, rules: list[MetaOverrideRule]) -> None:
self._rules: list[MetaOverrideRule] = rules
@property
def is_empty(self) -> bool:
return len(self._rules) == 0
@classmethod
def from_args_and_config(
cls,
*,
override_dims: list[str],
override_baseline_dims: list[str],
override_target_dims: list[str],
override_config: Optional[Path],
) -> "MetaOverrider":
per_side_args: list[tuple[list[str], Literal["both", "baseline", "target"]]] = [
(override_dims, "both"),
(override_baseline_dims, "baseline"),
(override_target_dims, "target"),
]
cli_rules: list[MetaOverrideRule] = [
MetaOverrideRule(match=name, dims=dims_str, side=side)
for raw_args, side in per_side_args
for name, dims_str in [_parse_cli_override_arg(raw) for raw in raw_args]
]
yaml_rules: list[MetaOverrideRule] = (
_load_yaml_rules(override_config) if override_config is not None else []
)
return cls(rules=cli_rules + yaml_rules)
def apply_to_meta(
self,
*,
name: str,
meta: dict[str, Any],
side: Literal["baseline", "target"],
) -> dict[str, Any]:
"""First-match-wins: return meta with dims replaced by the first matching rule for this side."""
for rule in self._rules:
if rule.side not in ("both", side):
continue
if re.search(rule.match, name):
return {**meta, "dims": rule.dims}
return meta
def _parse_cli_override_arg(raw: str) -> tuple[str, str]:
"""Parse 'name:dims_string' from a CLI --override-* argument."""
parts: list[str] = raw.split(":", maxsplit=1)
if len(parts) != 2 or not parts[0].strip() or not parts[1].strip():
raise ValueError(
f"Invalid override format: {raw!r}; expected 'name:dims_string'"
)
return parts[0].strip(), parts[1].strip()
def _load_yaml_rules(path: Path) -> list[MetaOverrideRule]:
"""Load override rules from a YAML config file."""
with open(path) as f:
raw_data: Any = yaml.safe_load(f)
if raw_data is None:
return []
config: MetaOverrideConfig = MetaOverrideConfig.model_validate(raw_data)
return config.overrides