Visualize per-token information in dump comparator (#19594)

This commit is contained in:
fzyzcjy
2026-03-01 10:32:59 +08:00
committed by GitHub
parent f5a10e04cd
commit 67810828cf
11 changed files with 575 additions and 2 deletions

View File

@@ -19,6 +19,8 @@ from sglang.srt.debug_utils.comparator.aligner.token_aligner.types import (
TokenAlignerPlan,
)
from sglang.srt.debug_utils.comparator.dims import (
SEQ_DIM_NAME,
TOKEN_DIM_NAME,
apply_dim_names,
resolve_dim_names,
)
@@ -50,6 +52,7 @@ def compare_bundle_pair(
x=None, y=None
),
viz_output_dir: Optional[Path] = None,
compute_per_token: bool = False,
) -> Union[ComparisonRecord, SkipRecord, NonTensorRecord]:
with warning_sink.context() as collected_warnings:
result = _compare_bundle_pair_inner(
@@ -61,6 +64,7 @@ def compare_bundle_pair(
diff_threshold=diff_threshold,
thd_seq_lens_by_step_pair=thd_seq_lens_by_step_pair,
viz_output_dir=viz_output_dir,
compute_per_token=compute_per_token,
)
return result.model_copy(update={"warnings": collected_warnings})
@@ -78,6 +82,7 @@ def _compare_bundle_pair_inner(
x=None, y=None
),
viz_output_dir: Optional[Path] = None,
compute_per_token: bool = False,
) -> Union[ComparisonRecord, SkipRecord, NonTensorRecord]:
# 1. Load all successfully loaded values
all_pair: Pair[list[ValueWithMeta]] = Pair(
@@ -104,6 +109,7 @@ def _compare_bundle_pair_inner(
diff_threshold=diff_threshold,
thd_seq_lens_by_step_pair=thd_seq_lens_by_step_pair,
viz_output_dir=viz_output_dir,
compute_per_token=compute_per_token,
)
@@ -117,6 +123,7 @@ def _compare_bundle_pair_tensor_type(
x=None, y=None
),
viz_output_dir: Optional[Path] = None,
compute_per_token: bool = False,
) -> Union[ComparisonRecord, SkipRecord]:
if not valid_pair.x or not valid_pair.y:
reason = "baseline_load_failed" if not valid_pair.x else "target_load_failed"
@@ -153,6 +160,11 @@ def _compare_bundle_pair_tensor_type(
reason: str = f"{side_name}_load_failed"
return SkipRecord(name=name, reason=reason)
# Resolve seq_dim for per-token computation
seq_dim: Optional[int] = (
_resolve_seq_dim(aligner_result.tensors.y) if compute_per_token else None
)
# Compare
aligned_baseline: torch.Tensor = aligner_result.tensors.x.rename(None)
aligned_target: torch.Tensor = aligner_result.tensors.y.rename(None)
@@ -162,6 +174,7 @@ def _compare_bundle_pair_tensor_type(
x_target=aligned_target,
name=name,
diff_threshold=diff_threshold,
seq_dim=seq_dim,
)
record = ComparisonRecord(**info.model_dump(), aligner_plan=plan)
@@ -209,6 +222,19 @@ def _try_generate_viz(
)
def _resolve_seq_dim(tensor: torch.Tensor) -> Optional[int]:
"""Find the token/seq dimension index from the tensor's named dims."""
if tensor.names[0] is None:
return None
names: tuple[Optional[str], ...] = tensor.names
for target_name in (TOKEN_DIM_NAME, SEQ_DIM_NAME):
if target_name in names:
return list(names).index(target_name)
return None
def _compare_bundle_pair_non_tensor_type(
*,
name: str,

View File

@@ -30,6 +30,9 @@ from sglang.srt.debug_utils.comparator.output_types import (
SummaryRecord,
print_record,
)
from sglang.srt.debug_utils.comparator.per_token_visualizer import (
generate_per_token_heatmap,
)
from sglang.srt.debug_utils.comparator.utils import Pair
from sglang.srt.debug_utils.comparator.warning_sink import warning_sink
from sglang.srt.debug_utils.dump_loader import read_meta, read_tokenizer_path
@@ -78,6 +81,10 @@ def run(args: argparse.Namespace) -> None:
Path(args.viz_output_dir) if args.viz_bundle_details else None
)
visualize_per_token: Optional[Path] = (
Path(args.visualize_per_token) if args.visualize_per_token else None
)
comparison_records = _compare_bundle_pairs(
bundle_info_pairs=bundle_info_pairs,
baseline_path=Path(args.baseline_path),
@@ -86,9 +93,12 @@ def run(args: argparse.Namespace) -> None:
diff_threshold=args.diff_threshold,
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,
)
_consume_comparison_records(
comparison_records=comparison_records, output_format=args.output_format
comparison_records=comparison_records,
output_format=args.output_format,
visualize_per_token=visualize_per_token,
)
@@ -144,6 +154,7 @@ def _compare_bundle_pairs(
diff_threshold: float,
thd_seq_lens_by_step_pair: Pair[Optional[dict[int, list[int]]]],
viz_output_dir: Optional[Path] = None,
compute_per_token: bool = False,
) -> Iterator[Union[ComparisonRecord, SkipRecord, NonTensorRecord]]:
for bundle_info_pair in bundle_info_pairs:
if not bundle_info_pair.y:
@@ -162,6 +173,7 @@ def _compare_bundle_pairs(
diff_threshold=diff_threshold,
thd_seq_lens_by_step_pair=thd_seq_lens_by_step_pair,
viz_output_dir=viz_output_dir,
compute_per_token=compute_per_token,
)
@@ -169,18 +181,28 @@ def _consume_comparison_records(
*,
comparison_records: Iterator[Union[ComparisonRecord, SkipRecord, NonTensorRecord]],
output_format: str,
visualize_per_token: Optional[Path] = None,
) -> None:
counts: dict[str, int] = {"passed": 0, "failed": 0, "skipped": 0}
collected_comparisons: list[ComparisonRecord] = []
for record in comparison_records:
counts[record.category] += 1
print_record(record, output_format=output_format)
if visualize_per_token is not None and isinstance(record, ComparisonRecord):
collected_comparisons.append(record)
print_record(
SummaryRecord(total=sum(counts.values()), **counts),
output_format=output_format,
)
if visualize_per_token is not None and collected_comparisons:
generate_per_token_heatmap(
records=collected_comparisons,
output_path=visualize_per_token,
)
def _parse_args() -> argparse.Namespace:
parser = argparse.ArgumentParser()
@@ -224,4 +246,10 @@ def _parse_args() -> argparse.Namespace:
default="/tmp/comparator_viz/",
help="Output directory for visualization PNGs (default: /tmp/comparator_viz/)",
)
parser.add_argument(
"--visualize-per-token",
type=str,
default=None,
help="Output path for per-token relative difference heatmap PNG",
)
return parser.parse_args()

View File

@@ -0,0 +1,83 @@
"""Per-token relative difference heatmap generator.
Produces a single PNG with rows = tensor names, columns = token positions,
color = log10(rel_diff).
"""
from __future__ import annotations
from pathlib import Path
from typing import Optional
from sglang.srt.debug_utils.comparator.output_types import ComparisonRecord
def generate_per_token_heatmap(
*,
records: list[ComparisonRecord],
output_path: Path,
) -> Optional[Path]:
"""Generate a per-token relative difference heatmap PNG.
Returns the output path if a file was written, or None if no data was available.
"""
rows_data: list[tuple[str, list[float]]] = _collect_per_token_data(records=records)
if not rows_data:
return None
_render_heatmap(rows_data=rows_data, output_path=output_path)
return output_path
def _collect_per_token_data(
*,
records: list[ComparisonRecord],
) -> list[tuple[str, list[float]]]:
rows: list[tuple[str, list[float]]] = []
for record in records:
if record.diff is None or record.diff.per_token_rel_diff is None:
continue
rows.append((record.name, record.diff.per_token_rel_diff))
return rows
def _render_heatmap(
*,
rows_data: list[tuple[str, list[float]]],
output_path: Path,
) -> None:
import matplotlib
import numpy as np
matplotlib.use("Agg")
import matplotlib.pyplot as plt
max_len: int = max(len(vals) for _, vals in rows_data)
labels: list[str] = [label for label, _ in rows_data]
matrix: np.ndarray = np.full((len(rows_data), max_len), np.nan, dtype=np.float64)
for i, (_, vals) in enumerate(rows_data):
matrix[i, : len(vals)] = vals
fig_width: float = max(12.0, max_len * 0.15)
fig_height: float = max(6.0, len(rows_data) * 0.3)
fig, ax = plt.subplots(figsize=(fig_width, fig_height))
im = ax.imshow(
np.log10(matrix + 1e-10), aspect="auto", cmap="hot", interpolation="nearest"
)
ax.set_xlabel("Token Position")
ax.set_ylabel("Tensor")
ax.set_yticks(range(len(labels)))
ax.set_yticklabels(labels, fontsize=8)
colorbar = fig.colorbar(im, ax=ax)
colorbar.set_label("log10(rel_diff)")
ax.set_title("Per-Token Relative Difference Heatmap")
fig.tight_layout()
output_path.parent.mkdir(parents=True, exist_ok=True)
fig.savefig(str(output_path), dpi=150)
plt.close(fig)

View File

@@ -12,6 +12,7 @@ from sglang.srt.debug_utils.comparator.tensor_comparator.types import (
from sglang.srt.debug_utils.comparator.utils import (
Pair,
argmax_coord,
calc_per_token_rel_diff,
calc_rel_diff,
compute_smaller_dtype,
try_unify_shape,
@@ -27,6 +28,7 @@ def compare_tensor_pair(
x_target: torch.Tensor,
name: str = "",
diff_threshold: float = 1e-3,
seq_dim: Optional[int] = None,
) -> TensorComparisonInfo:
baseline_info = TensorInfo(
shape=list(x_baseline.shape),
@@ -59,6 +61,7 @@ def compare_tensor_pair(
x_baseline=x_baseline_f,
x_target=x_target_f,
diff_threshold=diff_threshold,
seq_dim=seq_dim,
)
needs_sample = diff.max_abs_diff > SAMPLE_DIFF_THRESHOLD
@@ -122,6 +125,7 @@ def _compute_diff(
x_baseline: torch.Tensor,
x_target: torch.Tensor,
diff_threshold: float = 1e-3,
seq_dim: Optional[int] = None,
) -> DiffInfo:
if x_baseline.numel() == 0:
return DiffInfo(
@@ -145,6 +149,12 @@ def _compute_diff(
include_quantiles: bool = raw_abs_diff.numel() < QUANTILE_NUMEL_THRESHOLD
per_token_rel_diff: Optional[list[float]] = None
if seq_dim is not None and x_baseline.dim() > seq_dim:
per_token_rel_diff = calc_per_token_rel_diff(
x_baseline, x_target, seq_dim=seq_dim
).tolist()
return DiffInfo(
rel_diff=rel_diff,
max_abs_diff=max_abs_diff,
@@ -157,4 +167,5 @@ def _compute_diff(
target_at_max=x_target[max_diff_coord].item(),
diff_threshold=diff_threshold,
passed=rel_diff <= diff_threshold,
per_token_rel_diff=per_token_rel_diff,
)

View File

@@ -31,6 +31,7 @@ class DiffInfo(_StrictBase):
target_at_max: float
diff_threshold: float
passed: bool
per_token_rel_diff: Optional[list[float]] = None
class TensorComparisonInfo(_StrictBase):

View File

@@ -66,3 +66,23 @@ def calc_rel_diff(x: torch.Tensor, y: torch.Tensor) -> torch.Tensor:
denominator = (x * x + y * y).sum()
sim = 2 * (x * y).sum() / denominator
return 1 - sim
def calc_per_token_rel_diff(
x: torch.Tensor, y: torch.Tensor, *, seq_dim: int
) -> torch.Tensor:
"""Cosine-distance-like metric per token position.
Sums over all dims except seq_dim.
"""
x, y = x.double(), y.double()
other_dims: list[int] = [d for d in range(x.dim()) if d != seq_dim]
if other_dims:
denominator: torch.Tensor = (x * x + y * y).sum(dim=other_dims)
sim: torch.Tensor = 2 * (x * y).sum(dim=other_dims) / (denominator + 1e-10)
else:
denominator = x * x + y * y
sim = 2 * (x * y) / (denominator + 1e-10)
return (1 - sim).float()