mirror of
https://github.com/headroomlabs-ai/headroom.git
synced 2026-11-22 13:29:19 +08:00
fix(transforms/kompress): preserve line boundaries and tabular output in Kompress (#3119)
## Description KompressCompressor flattened `tool_result` content into a space-joined word list, discarding line boundaries and column padding — so compression could glue words from unrelated shell commands onto one fabricated line, and drop `ls -l` table fields inconsistently row to row. Now tracks each word's source line and never joins across a line boundary, and protects fixed-width tabular rows as atomic units. Closes #3099 ## Type of Change - [x] Bug fix (non-breaking change that fixes an issue) - [ ] New feature (non-breaking change that adds functionality) - [ ] Breaking change (fix or feature that would cause existing functionality to change) - [ ] Documentation update - [ ] Performance improvement - [ ] Code refactoring (no functional changes) ## Changes Made - `KompressCompressor.compress()`/`compress_batch()`: track each word's source line, never join kept words from different lines onto one line. - Add `_tabular_protected_line_ids()`: runs of ≥3 consistent fixed-width-column lines (`ls -l`, `ps`) are kept whole, not partially crushed. Reuses the existing `HEADROOM_KOMPRESS_MUST_KEEP` toggle. - Add `tests/test_kompress_line_integrity.py` reproducing both bugs. ## Testing - [x] Unit tests pass (`pytest`) - [x] Linting passes (`ruff check .`) - [x] Type checking passes (`mypy headroom`) - [x] New tests added for new functionality - [x] Manual testing performed ### Test Output ```text tests/test_kompress_line_integrity.py .... [4 passed] 66 passed, 1 warning in 18.15s ruff check .: All checks passed! ``` ## Real Behavior Proof - Environment: macOS/Apple Silicon (MPS), Python 3.12, real cached `chopratejas/kompress-v2-base` model. - Exact command / steps: ran `compress()`/`compress_batch()` on Bash-transcript content shaped like the issue, before and after the fix, including an aggressive `target_ratio=0.15` run to force lines to fully empty out. - Observed result: unpatched code collapsed a 24-line transcript onto one line and dropped `ls -la` fields — bug reproduced live. Patched code: no merges under any condition, tabular rows stayed intact even at ~22-29% overall keep ratio. - Not tested: full HTTP proxy path; ONNX backend with a real model. ## Runtime Rollout Safety - Rollout-managed feature(s): None new — reuses `HEADROOM_KOMPRESS_MUST_KEEP` - Minimum rollout channel: N/A - Stable/default behavior changed: Yes — compression output is now line/row-safe by default (correctness fix, not opt-in) - Kill switch / disable path: `HEADROOM_KOMPRESS_MUST_KEEP=0` disables tabular-row protection; the line-join fix itself has no toggle (never valid to merge lines) - Unsafe override required: No - Qualification impact: None known - Rollback path: revert this commit/PR ## Review Readiness - [x] I have performed a self-review - [x] This PR is ready for human review ## Checklist - [x] My code follows the project's style guidelines - [x] I have performed a self-review of my code - [x] I have commented my code, particularly in hard-to-understand areas - [ ] I have made corresponding changes to the documentation - [x] My changes generate no new warnings - [x] I have added tests that prove my fix is effective or that my feature works - [x] New and existing unit tests pass locally with my changes - [x] I did **not** edit `CHANGELOG.md` — it is generated by release-please from my Conventional Commit PR title (a CI guard enforces this) --------- Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com> Co-authored-by: JerrettDavis <mxjerrett@gmail.com>
This commit is contained in:
co-authored by
Claude Sonnet 5
JerrettDavis
parent
49f69be283
commit
fe2ed2b545
@@ -17,6 +17,7 @@ from __future__ import annotations
|
||||
import contextlib
|
||||
import gc
|
||||
import hashlib
|
||||
import itertools
|
||||
import logging
|
||||
import os
|
||||
import re
|
||||
@@ -98,6 +99,56 @@ def _add_kompress_must_keep_words(
|
||||
kept_ids.add(word_idx + chunk_start)
|
||||
|
||||
|
||||
_TABULAR_ROW_SPLIT_RE = re.compile(r"\s{2,}")
|
||||
|
||||
|
||||
def _words_by_line(content: str) -> tuple[list[str], list[int]]:
|
||||
"""Split into words like content.split(), plus each word's source line index.
|
||||
|
||||
Reconstruction must never join words that came from different source lines
|
||||
onto one line -- doing so can glue unrelated facts together (#3099).
|
||||
"""
|
||||
words: list[str] = []
|
||||
line_ids: list[int] = []
|
||||
for line_idx, line in enumerate(content.split("\n")):
|
||||
for word in line.split():
|
||||
words.append(word)
|
||||
line_ids.append(line_idx)
|
||||
return words, line_ids
|
||||
|
||||
|
||||
def _join_kept_words_by_line(words: list[str], line_ids: list[int], kept_ids: set[int]) -> str:
|
||||
"""Rejoin kept words, starting a new line wherever they crossed a line boundary."""
|
||||
parts: list[str] = []
|
||||
prev_line: int | None = None
|
||||
for i in sorted(kept_ids):
|
||||
if prev_line is not None:
|
||||
parts.append("\n" if line_ids[i] != prev_line else " ")
|
||||
parts.append(words[i])
|
||||
prev_line = line_ids[i]
|
||||
return "".join(parts)
|
||||
|
||||
|
||||
def _tabular_protected_line_ids(content: str) -> set[int]:
|
||||
"""Flag lines in a run of >=3 consecutive lines with a consistent fixed-width
|
||||
column count (ls -l, ps, etc.), so compression keeps each such row whole
|
||||
instead of dropping individual fields inconsistently row to row (#3099).
|
||||
Shares the must-keep toggle: HEADROOM_KOMPRESS_MUST_KEEP=0 disables this too.
|
||||
"""
|
||||
if os.environ.get(_KOMPRESS_MUST_KEEP_ENV, "1") == "0":
|
||||
return set()
|
||||
cols_per_line = [
|
||||
len(_TABULAR_ROW_SPLIT_RE.split(line.strip())) if line.strip() else 0
|
||||
for line in content.split("\n")
|
||||
]
|
||||
protected: set[int] = set()
|
||||
for cols, group in itertools.groupby(enumerate(cols_per_line), key=lambda p: p[1]):
|
||||
indices = [idx for idx, _ in group]
|
||||
if cols >= 3 and len(indices) >= 3:
|
||||
protected.update(indices)
|
||||
return protected
|
||||
|
||||
|
||||
# ONNX artifacts are resolved against the model repo in this order, falling
|
||||
# through on download miss OR session-load failure:
|
||||
#
|
||||
@@ -1602,12 +1653,14 @@ class KompressCompressor(Transform):
|
||||
KompressResult with compressed text.
|
||||
"""
|
||||
t_deadline = time.perf_counter() if _deadline_started_at is None else _deadline_started_at
|
||||
words = content.split()
|
||||
words, line_ids = _words_by_line(content)
|
||||
n_words = len(words)
|
||||
|
||||
if n_words < max(10, self.config.min_input_words) or self._degraded_reason is not None:
|
||||
return self._passthrough(content, n_words)
|
||||
|
||||
protected_line_ids = _tabular_protected_line_ids(content)
|
||||
|
||||
# Cooperative wall-clock budget (#1171): kompress ONNX inference is
|
||||
# O(tokens) and non-preemptible once the request's asyncio timeout fires,
|
||||
# so one large block can run for minutes holding a worker (the leak ->
|
||||
@@ -1788,6 +1841,10 @@ class KompressCompressor(Transform):
|
||||
# reconstruct from context. Disable via HEADROOM_KOMPRESS_MUST_KEEP=0.
|
||||
_add_kompress_must_keep_words(kept_ids, chunk_words, chunk_start)
|
||||
|
||||
for i, line_id in enumerate(line_ids):
|
||||
if line_id in protected_line_ids:
|
||||
kept_ids.add(i)
|
||||
|
||||
if not kept_ids:
|
||||
if inference_ms >= 1000.0:
|
||||
logger.info(
|
||||
@@ -1802,7 +1859,7 @@ class KompressCompressor(Transform):
|
||||
return self._passthrough(content, n_words)
|
||||
|
||||
compressed_words = [words[w] for w in sorted(kept_ids) if w < n_words]
|
||||
compressed = " ".join(compressed_words)
|
||||
compressed = _join_kept_words_by_line(words, line_ids, kept_ids)
|
||||
compressed_count = len(compressed_words)
|
||||
cache_key: str | None = None
|
||||
original_tokens = n_words
|
||||
@@ -2021,7 +2078,14 @@ class KompressCompressor(Transform):
|
||||
]
|
||||
|
||||
results: list[KompressResult | None] = [None] * n
|
||||
word_lists: list[list[str]] = [c.split() for c in contents]
|
||||
word_lists: list[list[str]] = []
|
||||
line_id_lists: list[list[int]] = []
|
||||
protected_line_ids_list: list[set[int]] = []
|
||||
for c in contents:
|
||||
words, line_ids = _words_by_line(c)
|
||||
word_lists.append(words)
|
||||
line_id_lists.append(line_ids)
|
||||
protected_line_ids_list.append(_tabular_protected_line_ids(c))
|
||||
|
||||
# Short texts short-circuit to passthrough — no model call needed.
|
||||
max_chunk_words = self.config.chunk_words
|
||||
@@ -2203,14 +2267,19 @@ class KompressCompressor(Transform):
|
||||
continue
|
||||
content = contents[text_idx]
|
||||
words = word_lists[text_idx]
|
||||
line_ids = line_id_lists[text_idx]
|
||||
n_words = len(words)
|
||||
|
||||
for i, line_id in enumerate(line_ids):
|
||||
if line_id in protected_line_ids_list[text_idx]:
|
||||
kept_ids.add(i)
|
||||
|
||||
if not kept_ids:
|
||||
results[text_idx] = self._passthrough(content, n_words)
|
||||
continue
|
||||
|
||||
compressed_words = [words[w] for w in sorted(kept_ids) if w < n_words]
|
||||
compressed = " ".join(compressed_words)
|
||||
compressed = _join_kept_words_by_line(words, line_ids, kept_ids)
|
||||
compressed_count = len(compressed_words)
|
||||
cache_key: str | None = None
|
||||
original_tokens = n_words
|
||||
|
||||
@@ -0,0 +1,108 @@
|
||||
"""Regression tests for issue #3099: compression merging words across line
|
||||
boundaries, and dropping fields from tabular rows inconsistently row to row.
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import pytest
|
||||
|
||||
from headroom.transforms import kompress_compressor as kc
|
||||
from headroom.transforms.kompress_compressor import KompressCompressor, KompressConfig
|
||||
|
||||
|
||||
class _Enc(dict):
|
||||
def word_ids(self, batch_index=0):
|
||||
return self["_word_ids"][batch_index]
|
||||
|
||||
|
||||
class _Tok:
|
||||
def __call__(self, chunk_words, **kw):
|
||||
if chunk_words and isinstance(chunk_words[0], list):
|
||||
batch_words = chunk_words
|
||||
else:
|
||||
batch_words = [chunk_words]
|
||||
return _Enc(
|
||||
input_ids=[[0] * len(words) for words in batch_words],
|
||||
attention_mask=[[1] * len(words) for words in batch_words],
|
||||
_word_ids=[list(range(len(words))) for words in batch_words],
|
||||
)
|
||||
|
||||
|
||||
class _KeepIndicesModel:
|
||||
"""Fake model that keeps exactly the given (global, single-chunk) word indices."""
|
||||
|
||||
def __init__(self, keep_indices: set[int]):
|
||||
self._keep = keep_indices
|
||||
|
||||
def get_keep_mask(self, input_ids, attention_mask):
|
||||
return [[idx in self._keep for idx in range(len(row))] for row in input_ids]
|
||||
|
||||
def get_scores(self, input_ids, attention_mask):
|
||||
return [[1.0 if idx in self._keep else 0.0 for idx in range(len(row))] for row in input_ids]
|
||||
|
||||
|
||||
def _install_fake_kompress(monkeypatch, keep_indices: set[int], *, must_keep: bool = False) -> None:
|
||||
model = _KeepIndicesModel(keep_indices)
|
||||
monkeypatch.setattr(kc, "_load_kompress", lambda *a, **k: (model, _Tok(), "onnx"))
|
||||
monkeypatch.setattr(kc, "_model_device_type", lambda *a, **k: "cpu")
|
||||
if must_keep:
|
||||
monkeypatch.delenv(kc._KOMPRESS_MUST_KEEP_ENV, raising=False)
|
||||
else:
|
||||
monkeypatch.setenv(kc._KOMPRESS_MUST_KEEP_ENV, "0")
|
||||
|
||||
|
||||
def _compress(monkeypatch, content: str, path: str):
|
||||
compressor = KompressCompressor(KompressConfig(enable_ccr=False, min_input_words=10))
|
||||
if path == "compress":
|
||||
monkeypatch.setattr(compressor, "_should_batch_single_content", lambda *a, **k: False)
|
||||
return compressor.compress(content)
|
||||
monkeypatch.setattr(compressor, "_should_use_sequential_fallback", lambda: False)
|
||||
[result] = compressor.compress_batch([content], batch_size=8)
|
||||
return result
|
||||
|
||||
|
||||
@pytest.mark.parametrize("path", ["compress", "compress_batch"])
|
||||
def test_compress_does_not_merge_words_across_lines(monkeypatch, path):
|
||||
"""Reproduces #3099 bug 1: two commands' output on separate lines.
|
||||
|
||||
The model keeps only the last word of line 1 ("rg") and the first word of
|
||||
line 2 ("/opt/homebrew/bin/ast-grep"). Before the fix, these landed adjacent
|
||||
in the flat kept-word list and were joined with a space onto one line,
|
||||
fabricating "rg /opt/homebrew/bin/ast-grep" -- read literally, that implies
|
||||
rg resolves to that path, which is false.
|
||||
"""
|
||||
line1 = "alpha beta gamma delta epsilon zeta eta rg"
|
||||
line2 = "/opt/homebrew/bin/ast-grep theta iota kappa lambda mu nu xi"
|
||||
content = f"{line1}\n{line2}"
|
||||
|
||||
_install_fake_kompress(monkeypatch, keep_indices={7, 8})
|
||||
|
||||
result = _compress(monkeypatch, content, path)
|
||||
|
||||
assert "rg /opt/homebrew/bin/ast-grep" not in result.compressed
|
||||
assert result.compressed.split("\n") == ["rg", "/opt/homebrew/bin/ast-grep"]
|
||||
|
||||
|
||||
@pytest.mark.parametrize("path", ["compress", "compress_batch"])
|
||||
def test_compress_keeps_tabular_row_intact(monkeypatch, path):
|
||||
"""Reproduces #3099 bug 2: an `ls -la`-shaped listing.
|
||||
|
||||
The model keeps nothing on its own, so absent row protection only the
|
||||
must-keep regex would save permission bits/numbers/filenames -- exactly
|
||||
the reported symptom of some fields (the group column, the month)
|
||||
disappearing inconsistently. Once a row is part of a >=3-row run of
|
||||
consistent fixed-width columns, the whole row must survive intact.
|
||||
"""
|
||||
rows = [
|
||||
"-rw-r--r--@ 1 user staff 4921 Aug 12 20:14 daemon.log",
|
||||
"-rw-r--r--@ 1 user staff 10485 Aug 12 20:14 other.log",
|
||||
"-rw-r--r--@ 1 user staff 512 Aug 12 20:15 third.log",
|
||||
]
|
||||
content = "\n".join(rows)
|
||||
|
||||
_install_fake_kompress(monkeypatch, keep_indices=set(), must_keep=True)
|
||||
|
||||
result = _compress(monkeypatch, content, path)
|
||||
|
||||
expected = "\n".join(" ".join(row.split()) for row in rows)
|
||||
assert result.compressed == expected
|
||||
Reference in New Issue
Block a user