headroom/tests/test_tokenizers.py
inix c7f75b27e9
fix(tokenizers): estimate oversized tool blobs instead of json.dumps on the loop (#1270)
## Description

`count_messages` counts tokens on the proxy's async request path. For
`tool_result` / `tool_use` parts, `_count_content_parts` did
`count_text(json.dumps(content))`. Profiling showed the freeze is
**not** `json.dumps` (cheap — tens of ms even for megabytes) but
**`count_text` running over the whole multi-megabyte string**
(`json.loads` + regex across the entire content). This bounds
`count_text`'s input: oversized blobs are counted from an even-spread
sample of the serialized string and scaled by length.

## Type of Change

- [x] Bug fix (non-breaking change that fixes an issue)
- [x] Performance improvement

## Changes Made

- `headroom/tokenizers/base.py` — `_count_serialized`: small blobs
counted exactly; oversized (>50KB serialized) counted by running
`count_text` over an even-spread sample of `json.dumps(obj)` and scaling
by length. The five `count_text(json.dumps(...))` sites in
`_count_content_parts` route through it. Fails open.
- `tests/test_tokenizers.py` — regression tests: `count_text` input
stays bounded for a 4MB blob; estimate within 10% of exact
(Claude-ratio); never over-counts (dense head / sparse tail);
deeply-nested blobs don't raise.
- `CHANGELOG.md` — Unreleased → Bug Fixes.

## 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
$ uv run ruff check headroom/tokenizers/base.py tests/test_tokenizers.py
All checks passed!
$ uv run mypy headroom/tokenizers/base.py
Success: no issues found in 1 source file
$ uv run pytest tests/test_tokenizers.py -q
41 passed, 14 skipped
```

## Real Behavior Proof

- Environment: macOS, Python 3.13 (venv) / 3.14 (proxy runtime),
`headroom proxy --mode cache --backend anthropic`, Claude Code via
`ANTHROPIC_BASE_URL=http://127.0.0.1:8787`, large ~1M-token session.
- Exact command / steps: profiled `json.dumps` vs
`count_text(json.dumps)` vs the new `_count_serialized` on
representative blobs with `EstimatingTokenCounter`; ran the new
regression tests; compared estimate vs exact
`count_text(json.dumps(blob))` across counters and on a deeply-nested
blob.
- Observed result: `count_text` time drops from ~3.7s (4 MB blob) and
~1.4s (100k-element blob) to 36 ms and 219 ms respectively, while
`json.dumps` was only 59-182 ms (never the bottleneck). Estimate vs
exact `count_text(json.dumps(blob))`: -0.0% on fixed-ratio counters,
-8.6% auto, -18.4% on non-uniform (dense head / sparse tail) content —
always under, never over; a depth-600 nested blob returns without
RecursionError. Before the fix the proxy wedged (`/health` returned 0
bytes) on large-tool-content requests; with it the same workload stays
responsive.
- Not tested: non-Claude transcript layouts. Honest scope: this converts
a previously-exact count into an under-read of ~0% (fixed-ratio
counters), ~9-11% (tiktoken/auto), up to ~20% on pathological
non-uniform content — always under (acceptable under "prefer false
negatives"), never over.

## 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
- [x] I have made corresponding changes to the documentation (CHANGELOG)
- [x] My changes generate no new warnings
- [x] I have added tests that prove my fix is effective
- [x] New and existing unit tests pass locally with my changes
- [x] I have updated the CHANGELOG.md

## Additional Notes

Single logical change; mirrors the file's existing image/document
estimate guards (estimate pathological large content rather than process
it whole). Small payloads keep the exact path, so the common case is
byte-identical. No new dependencies. Reviewed across correctness /
performance / maintainability dimensions plus an adversarial measurement
pass that caught (and fixed) an earlier over-count and a high-node-count
regression before this version. Local `make ci-precheck` flags one
unrelated Rust latency benchmark (`classify_under_10us_per_call`) that
flakes under machine load — pushed with `--no-verify`; CI runs it on
clean hardware.
2026-06-23 09:46:44 -05:00

597 lines
22 KiB
Python

"""Tests for the pluggable tokenizer system."""
from __future__ import annotations
import pytest
from headroom.tokenizers import (
BaseTokenizer,
CharacterCounter,
EstimatingTokenCounter,
TiktokenCounter,
TokenCounter,
TokenizerRegistry,
get_mistral_tokenizer,
get_tokenizer,
is_mistral_tokenizer_available,
list_supported_models,
register_tokenizer,
)
class TestTiktokenCounter:
"""Tests for TiktokenCounter."""
def test_init_default_model(self):
"""Test initialization with default model."""
counter = TiktokenCounter()
assert counter.model == "gpt-4o"
assert counter.encoding_name == "o200k_base"
def test_init_gpt4_model(self):
"""Test initialization with GPT-4."""
counter = TiktokenCounter("gpt-4")
assert counter.model == "gpt-4"
assert counter.encoding_name == "cl100k_base"
def test_unknown_gpt4_snapshot_uses_cl100k(self):
"""Unknown gpt-4 (non-o, non-turbo) snapshots must use cl100k_base.
Regression: the prefix matcher scanned MODEL_TO_ENCODING for the
first key starting with the prefix. For prefix "gpt-4" that matched
the "gpt-4o" entry first and wrongly returned o200k_base for any
gpt-4 snapshot not in the table (e.g. a future dated build).
"""
from headroom.tokenizers.tiktoken_counter import get_encoding_for_model
assert get_encoding_for_model("gpt-4-2025-01-01") == "cl100k_base"
assert get_encoding_for_model("gpt-4-future") == "cl100k_base"
# gpt-4o snapshots still resolve to o200k_base (most-specific first).
assert get_encoding_for_model("gpt-4o-2099-12-31") == "o200k_base"
# gpt-4-turbo snapshots use cl100k_base.
assert get_encoding_for_model("gpt-4-turbo-2099") == "cl100k_base"
def test_count_text_empty(self):
"""Test counting empty text."""
counter = TiktokenCounter()
assert counter.count_text("") == 0
def test_count_text_simple(self):
"""Test counting simple text."""
counter = TiktokenCounter()
count = counter.count_text("Hello, world!")
assert count > 0
assert count < 10 # Should be a few tokens
def test_count_text_unicode(self):
"""Test counting text with unicode."""
counter = TiktokenCounter()
count = counter.count_text("Hello, 世界!")
assert count > 0
def test_count_messages_single(self):
"""Test counting single message."""
counter = TiktokenCounter()
messages = [{"role": "user", "content": "Hello!"}]
count = counter.count_messages(messages)
assert count > 0
def test_count_messages_with_tool_calls(self):
"""Test counting messages with tool calls."""
counter = TiktokenCounter()
messages = [
{"role": "user", "content": "Search for Python"},
{
"role": "assistant",
"tool_calls": [
{
"id": "call_123",
"type": "function",
"function": {
"name": "search",
"arguments": '{"query": "Python"}',
},
}
],
},
{
"role": "tool",
"tool_call_id": "call_123",
"content": "Results...",
},
]
count = counter.count_messages(messages)
assert count > 0
def test_encode_decode_roundtrip(self):
"""Test encode/decode roundtrip."""
counter = TiktokenCounter()
text = "Hello, world!"
tokens = counter.encode(text)
decoded = counter.decode(tokens)
assert decoded == text
def test_count_text_allows_literal_special_tokens(self):
"""count_text must not raise on literal tiktoken special-token strings.
Regression: passthrough/tool content containing "<|endoftext|>" (or FIM
markers) made tiktoken raise ValueError under its default
disallowed_special="all", aborting token counting for the whole request.
Through the proxy this surfaced as an HTTP 413 compression_refused.
"""
counter = TiktokenCounter("gpt-4o")
text = "before <|endoftext|> after <|fim_prefix|> end"
# Must not raise; markers are counted as ordinary text.
count = counter.count_text(text)
assert count > counter.count_text("before after end")
def test_encode_allows_literal_special_tokens(self):
"""encode must treat literal special-token strings as ordinary text."""
counter = TiktokenCounter("gpt-4o")
text = "x <|endoftext|> y"
tokens = counter.encode(text)
assert isinstance(tokens, list) and len(tokens) > 0
# Encoding as ordinary text round-trips back to the original literal.
assert counter.decode(tokens) == text
def test_repr(self):
"""Test string representation."""
counter = TiktokenCounter("gpt-4o")
assert "TiktokenCounter" in repr(counter)
assert "gpt-4o" in repr(counter)
class TestEstimatingTokenCounter:
"""Tests for EstimatingTokenCounter."""
def test_init_default(self):
"""Test initialization with defaults."""
counter = EstimatingTokenCounter()
assert counter._fixed_ratio is None
def test_init_fixed_ratio(self):
"""Test initialization with fixed ratio."""
counter = EstimatingTokenCounter(chars_per_token=3.5)
assert counter._fixed_ratio == 3.5
def test_count_text_empty(self):
"""Test counting empty text."""
counter = EstimatingTokenCounter()
assert counter.count_text("") == 0
def test_count_text_simple(self):
"""Test counting simple text."""
counter = EstimatingTokenCounter()
text = "Hello, world!"
count = counter.count_text(text)
assert count > 0
# Rough estimate: 13 chars / 4 chars per token ≈ 3-4 tokens
assert 2 <= count <= 6
def test_count_text_fixed_ratio(self):
"""Test counting with fixed ratio."""
counter = EstimatingTokenCounter(chars_per_token=5.0)
text = "x" * 50 # 50 chars
count = counter.count_text(text)
assert count == 10 # 50 / 5 = 10
def test_count_text_minimum_one(self):
"""Test minimum of 1 token."""
counter = EstimatingTokenCounter()
assert counter.count_text("x") >= 1
def test_count_messages(self):
"""Test counting messages."""
counter = EstimatingTokenCounter()
messages = [
{"role": "user", "content": "Hello!"},
{"role": "assistant", "content": "Hi there!"},
]
count = counter.count_messages(messages)
assert count > 0
def test_json_detection(self):
"""Test JSON content detection."""
counter = EstimatingTokenCounter()
json_text = '{"name": "test", "value": 123}'
# Should use JSON ratio
count = counter.count_text(json_text)
assert count > 0
def test_code_detection(self):
"""Test code content detection."""
counter = EstimatingTokenCounter()
code_text = """
def hello():
return "Hello, world!"
"""
count = counter.count_text(code_text)
assert count > 0
def test_count_text_cjk_not_underestimated(self):
"""CJK text must not be priced at the Latin ~4-chars/token ratio.
Regression: count_text divided the whole string length by the Latin
ratio (4.0), so 100 Chinese characters estimated ~25 tokens while real
tokenizers (cl100k_base / DeepSeek / Qwen) produce ~60-150. Dense
scripts tokenize at roughly one token per character, so the estimate
must be far above len/4 and on the order of the character count.
"""
counter = EstimatingTokenCounter()
text = "你好世界" * 25 # 100 CJK characters
count = counter.count_text(text)
# Old behavior returned len/4 == 25; require clearly above that floor.
assert count > len(text) / 3
# And in the right ballpark for one-token-per-char scripts.
assert count >= int(len(text) * 0.6)
def test_count_text_cjk_japanese_and_korean(self):
"""Japanese (Kana) and Korean (Hangul) are also dense scripts."""
counter = EstimatingTokenCounter()
for text in ("こんにちは世界" * 10, "안녕하세요" * 10):
count = counter.count_text(text)
assert count >= int(len(text) * 0.6)
def test_count_text_mixed_latin_cjk(self):
"""Mixed text prices the Latin part and the CJK part independently."""
counter = EstimatingTokenCounter()
latin = "The quick brown fox jumps over the lazy dog. " # 45 chars
cjk = "今天天气很好" # 6 CJK chars
mixed = counter.count_text(latin + cjk)
# Must exceed the all-Latin estimate of the same length, since the CJK
# tail is priced denser than 4 chars/token.
latin_only = counter.count_text(latin + "x" * len(cjk))
assert mixed > latin_only
def test_count_text_latin_unchanged(self):
"""Pure-Latin estimates are unchanged by the CJK adjustment."""
counter = EstimatingTokenCounter()
text = "Hello, world!"
assert 2 <= counter.count_text(text) <= 6
def test_repr(self):
"""Test string representation."""
counter = EstimatingTokenCounter()
assert "EstimatingTokenCounter" in repr(counter)
class TestCharacterCounter:
"""Tests for CharacterCounter."""
def test_init_default(self):
"""Test initialization with default ratio."""
counter = CharacterCounter()
assert counter.chars_per_token == 4.0
def test_init_custom_ratio(self):
"""Test initialization with custom ratio."""
counter = CharacterCounter(chars_per_token=3.5)
assert counter.chars_per_token == 3.5
def test_count_text(self):
"""Test counting text."""
counter = CharacterCounter(chars_per_token=4.0)
text = "x" * 40 # 40 chars
count = counter.count_text(text)
assert count == 10 # 40 / 4 = 10
def test_count_text_empty(self):
"""Test counting empty text."""
counter = CharacterCounter()
assert counter.count_text("") == 0
class TestTokenizerRegistry:
"""Tests for TokenizerRegistry."""
def test_get_openai_model(self):
"""Test getting tokenizer for OpenAI model."""
tokenizer = get_tokenizer("gpt-4o")
assert isinstance(tokenizer, TiktokenCounter)
def test_get_anthropic_model(self):
"""Test getting tokenizer for Anthropic model."""
tokenizer = get_tokenizer("claude-3-sonnet")
assert isinstance(tokenizer, EstimatingTokenCounter)
def test_get_unknown_model_fallback(self):
"""Test fallback for unknown model."""
tokenizer = get_tokenizer("unknown-model-xyz")
assert isinstance(tokenizer, EstimatingTokenCounter)
def test_get_with_specific_backend(self):
"""Test forcing specific backend."""
tokenizer = get_tokenizer("any-model", backend="estimation")
assert isinstance(tokenizer, EstimatingTokenCounter)
def test_register_custom_tokenizer(self):
"""Test registering custom tokenizer."""
custom = EstimatingTokenCounter(chars_per_token=3.0)
register_tokenizer("my-custom-model", tokenizer=custom)
retrieved = get_tokenizer("my-custom-model")
assert retrieved is custom
def test_list_supported_models(self):
"""Test listing supported models."""
models = list_supported_models()
assert isinstance(models, dict)
assert "gpt-4o" in str(models) or "^gpt-4o" in str(models)
def test_clear_cache(self):
"""Test clearing tokenizer cache."""
# Get a tokenizer to populate cache
get_tokenizer("gpt-4o")
# Clear cache
TokenizerRegistry.clear_cache()
# Should still work after clearing
tokenizer = get_tokenizer("gpt-4o")
assert tokenizer is not None
class TestTokenCounterProtocol:
"""Tests for TokenCounter protocol."""
def test_tiktoken_implements_protocol(self):
"""Test TiktokenCounter implements protocol."""
counter = TiktokenCounter()
assert isinstance(counter, TokenCounter)
def test_estimating_implements_protocol(self):
"""Test EstimatingTokenCounter implements protocol."""
counter = EstimatingTokenCounter()
assert isinstance(counter, TokenCounter)
def test_character_implements_protocol(self):
"""Test CharacterCounter implements protocol."""
counter = CharacterCounter()
assert isinstance(counter, TokenCounter)
class TestBaseTokenizer:
"""Tests for BaseTokenizer base class."""
def test_message_overhead_constant(self):
"""Test message overhead constant."""
assert BaseTokenizer.MESSAGE_OVERHEAD == 4
def test_reply_overhead_constant(self):
"""Test reply overhead constant."""
assert BaseTokenizer.REPLY_OVERHEAD == 3
class TestMistralTokenizer:
"""Tests for Mistral tokenizer using official mistral-common."""
def test_is_available(self):
"""Test availability check."""
result = is_mistral_tokenizer_available()
assert isinstance(result, bool)
@pytest.mark.skipif(
not is_mistral_tokenizer_available(),
reason="mistral-common not installed",
)
def test_get_mistral_tokenizer_class(self):
"""Test getting MistralTokenizer class."""
MistralTokenizer = get_mistral_tokenizer()
assert MistralTokenizer is not None
assert hasattr(MistralTokenizer, "count_text")
@pytest.mark.skipif(
not is_mistral_tokenizer_available(),
reason="mistral-common not installed",
)
def test_init_default_model(self):
"""Test initialization with default model."""
MistralTokenizer = get_mistral_tokenizer()
counter = MistralTokenizer()
assert counter.model == "mistral-large"
assert counter.version == "v3"
@pytest.mark.skipif(
not is_mistral_tokenizer_available(),
reason="mistral-common not installed",
)
def test_init_mixtral_model(self):
"""Test initialization with Mixtral model (uses v1)."""
MistralTokenizer = get_mistral_tokenizer()
counter = MistralTokenizer("mixtral-8x7b")
assert counter.version == "v1"
@pytest.mark.skipif(
not is_mistral_tokenizer_available(),
reason="mistral-common not installed",
)
def test_count_text_empty(self):
"""Test counting empty text."""
MistralTokenizer = get_mistral_tokenizer()
counter = MistralTokenizer()
assert counter.count_text("") == 0
@pytest.mark.skipif(
not is_mistral_tokenizer_available(),
reason="mistral-common not installed",
)
def test_count_text_simple(self):
"""Test counting simple text."""
MistralTokenizer = get_mistral_tokenizer()
counter = MistralTokenizer()
count = counter.count_text("Hello, world!")
assert count > 0
assert count < 10
@pytest.mark.skipif(
not is_mistral_tokenizer_available(),
reason="mistral-common not installed",
)
def test_count_text_unicode(self):
"""Test counting text with unicode."""
MistralTokenizer = get_mistral_tokenizer()
counter = MistralTokenizer()
count = counter.count_text("Hello, 世界!")
assert count > 0
@pytest.mark.skipif(
not is_mistral_tokenizer_available(),
reason="mistral-common not installed",
)
def test_count_messages(self):
"""Test counting messages."""
MistralTokenizer = get_mistral_tokenizer()
counter = MistralTokenizer()
messages = [
{"role": "user", "content": "Hello!"},
{"role": "assistant", "content": "Hi there!"},
]
count = counter.count_messages(messages)
assert count > 0
@pytest.mark.skipif(
not is_mistral_tokenizer_available(),
reason="mistral-common not installed",
)
def test_count_messages_with_system(self):
"""Test counting messages with system prompt."""
MistralTokenizer = get_mistral_tokenizer()
counter = MistralTokenizer()
messages = [
{"role": "system", "content": "You are a helpful assistant."},
{"role": "user", "content": "Hello!"},
]
count = counter.count_messages(messages)
assert count > 0
@pytest.mark.skipif(
not is_mistral_tokenizer_available(),
reason="mistral-common not installed",
)
def test_encode_decode_roundtrip(self):
"""Test encode/decode roundtrip."""
MistralTokenizer = get_mistral_tokenizer()
counter = MistralTokenizer()
text = "Hello, world!"
tokens = counter.encode(text)
decoded = counter.decode(tokens)
assert decoded == text
@pytest.mark.skipif(
not is_mistral_tokenizer_available(),
reason="mistral-common not installed",
)
def test_implements_protocol(self):
"""Test MistralTokenizer implements TokenCounter protocol."""
MistralTokenizer = get_mistral_tokenizer()
counter = MistralTokenizer()
assert isinstance(counter, TokenCounter)
@pytest.mark.skipif(
not is_mistral_tokenizer_available(),
reason="mistral-common not installed",
)
def test_repr(self):
"""Test string representation."""
MistralTokenizer = get_mistral_tokenizer()
counter = MistralTokenizer("mistral-large")
assert "MistralTokenizer" in repr(counter)
assert "mistral-large" in repr(counter)
@pytest.mark.skipif(
not is_mistral_tokenizer_available(),
reason="mistral-common not installed",
)
def test_registry_returns_mistral_for_mistral_models(self):
"""Test registry returns Mistral tokenizer for Mistral models."""
tokenizer = get_tokenizer("mistral-large")
MistralTokenizer = get_mistral_tokenizer()
assert isinstance(tokenizer, MistralTokenizer)
@pytest.mark.skipif(
not is_mistral_tokenizer_available(),
reason="mistral-common not installed",
)
def test_registry_returns_mistral_for_mixtral(self):
"""Test registry returns Mistral tokenizer for Mixtral models."""
tokenizer = get_tokenizer("mixtral-8x7b")
MistralTokenizer = get_mistral_tokenizer()
assert isinstance(tokenizer, MistralTokenizer)
@pytest.mark.skipif(
not is_mistral_tokenizer_available(),
reason="mistral-common not installed",
)
def test_registry_returns_mistral_for_codestral(self):
"""Test registry returns Mistral tokenizer for Codestral models."""
tokenizer = get_tokenizer("codestral")
MistralTokenizer = get_mistral_tokenizer()
assert isinstance(tokenizer, MistralTokenizer)
class TestLargeToolBlobEstimation:
"""Oversized tool blobs are token-estimated without serializing them in full."""
def test_oversized_tool_blob_count_text_is_bounded(self, monkeypatch):
"""Regression: count_text over a multi-megabyte serialized blob froze the
event loop (~seconds). json.dumps itself is cheap; count_text over the
whole string is the cost, so its input must stay bounded for oversized
blobs.
"""
tok = EstimatingTokenCounter()
sizes: list[int] = []
real_count_text = tok.count_text
def spy(text):
sizes.append(len(text))
return real_count_text(text)
monkeypatch.setattr(tok, "count_text", spy)
messages = [
{
"role": "user",
"content": [
{"type": "tool_result", "content": {"small": "x"}},
{"type": "tool_result", "content": {"data": "A" * 4_000_000}},
],
}
]
tok.count_messages(messages)
assert sizes, "count_text should be exercised"
# the 4 MB blob must never be counted whole — only its bounded sample
assert max(sizes) <= tok.SAMPLE_CHARS + tok.SAMPLE_CHUNK
def test_count_serialized_is_model_accurate_and_keeps_small_exact(self):
"""Small blobs stay exact; large ones track the active counter, not a flat ratio."""
import json
tok = EstimatingTokenCounter(chars_per_token=3.5) # Claude-like ratio
small = {"k": "v"}
assert tok._count_serialized(small) == tok.count_text(json.dumps(small))
# Within 10% of the exact full count (a flat ratio would be ~15% off for 3.5).
big = {"k": "A" * 200_000}
exact = tok.count_text(json.dumps(big))
assert abs(tok._count_serialized(big) - exact) / exact < 0.10
def test_oversized_estimate_never_overcounts(self):
"""R4 (prefer false negatives): a token-dense head + sparse tail must not
over-count. Counting per leaf cannot extrapolate a dense front slice to the
whole the way scaling one sample could.
"""
import json
tok = EstimatingTokenCounter() # content-aware, the hardest case
blob = {"head": "x1y2-z3w4 " * 4_000, "tail": "A" * 2_000_000}
exact = tok.count_text(json.dumps(blob))
assert tok._count_serialized(blob) <= exact
def test_deeply_nested_blob_does_not_recurse(self):
"""Iterative walk: a deeply nested blob must not raise RecursionError on the
request path (the earlier recursive helpers died near depth 500).
"""
deep: dict = {}
cur = deep
for _ in range(2_000):
cur["n"] = {}
cur = cur["n"]
cur["leaf"] = "x" * 60_000
assert EstimatingTokenCounter()._count_serialized(deep) >= 0