litellm/.github/scripts/verify_linux_native_wheel.py
devin-ai-integration[bot] 0abd9267c1
feat(tokenizer): preserve Python defaults with opt-in Rust dispatch (#42174)
* ci: benchmark and gate an installed release wheel

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

* ci: simplify installed-wheel benchmark check

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

* feat(rust): add native tokenizer codec

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

* refactor(tokenizer): route Python tokenization through the Rust extension

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

* style(lint): format tokenizer call

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

* fix(packaging): restore runtime dependencies and native images

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

* fix(tokenizer): preserve Python SDK behavior with Rust tokenizers

* fix(tokenizer): restore compatibility paths

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

* refactor(tokenizer): count custom tokenizers directly

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

* fix(tokenizer): preserve caller-supplied Python tokenizer counts

* fix(tokenizer): reuse packaged vocabularies in the native wheel

* refactor(rust_bridge): route token counting through the catalog as RUST_OPT_IN

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

* fix(spend_tracking): compare tokenizer groups by value

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

* chore(deps): re-resolve filelock under the <4.0 pin

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

* fix(llms): align transformation override signatures with base configs

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

* build(rust): use fat LTO to keep the native wheel under the 35 MB limit

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

* feat(tokenizer): preserve Python defaults with opt-in Rust dispatch

* test(proxy): tolerate missing litellm.utils.Tokenizer when patching it

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

* test(proxy): patch the tokenizer dispatch function instead of the removed alias

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

* feat(tokenizer): give the Rust wrappers the tiktoken and tokenizers surface

Callers of litellm.encoding and litellm.create_tokenizer must see the same
read-only API whichever backend the catalog selects.

- OpenAIEncoding mirrors tiktoken.Encoding: n_vocab, max_token_value,
  token_byte_values, encode_single_token, encode_with_unstable,
  encode_to_numpy, decode_with_offsets, is_special_token, repr; the Rust
  tiktoken crate keeps a Vocabulary beside each CoreBPE and reports the
  requested encoding name (gpt2 stays gpt2).
- HuggingFaceTokenizer mirrors the read-only tokenizers.Tokenizer surface
  (token_to_id, id_to_token, get_vocab, get_vocab_size,
  get_added_tokens_decoder, num_special_tokens_to_add, padding, truncation,
  encode_special_tokens, from_buffer); HuggingFaceEncoding gains the
  char/word/token lookups, pad, truncate, set_sequence_id and merge.
  Mutators stay on the Python tokenizer.
- from_json/from_pretrained claim the fork gate only when the huggingface
  feature is compiled in; the surrogate fallback matches on the Codec.
- Tokenizer caching is keyed on the same catalog Context the dispatch runs
  on; rust_tokenizer reads the encoding name without loading an encoding;
  LITELLM_RUST parsing is cached.
- Drop the unused tiktoken_encoding_for_model export and Error::Download.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

* fix(tokenizer): close the exhaustive matches with assert_never

CodeQL reads a `match` over a Literal with no default arm as an implicit
`None` return. `assert_never` makes the exhaustiveness explicit for both the
HuggingFace tokenizer loader and the Rust token-counter factory.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

* feat(tokenizer): derive the fast counter from the shared tokenizer

The count-only counter (`fast` feature) and the codec each parsed the same
artifact: TokenCounter took the Anthropic JSON and the tiktoken rank files
from Python while Tokenizer loaded them again. One parse now serves both.

- FastTokenizer builds from a model another loader holds: `from_shared`
  takes the Arc<tokenizers::Tokenizer> the HF codec keeps, and
  `from_*_pairs` take the ranks the tiktoken vocabulary already parsed.
- `FastCounter::fast_counter` in the core crate derives it from either codec;
  encodings the fast scanner does not reproduce are refused.
- Native `Tokenizer.count(text, fast=False)` opts into that counter, built
  once per tokenizer on first use; `TokenCounter.from_tokenizer(tokenizer,
  fast=False)` replaces the JSON and rank-file constructors.
- The Python route counts over the native tokenizers the codec path shares
  (`native_encoding`, `native_anthropic`) and no longer reads rank files;
  the packaged Anthropic tokenizer has one loader, `tokenizer_dispatch.anthropic`.
- Public wrappers gain `count(text, fast=False)`.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

---------

Co-authored-by: Yujong Lee <yujong@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
2026-09-22 04:41:11 +00:00

293 lines
13 KiB
Python

from __future__ import annotations
import importlib.util
import os
import re
import subprocess
import sys
import zipfile
from collections.abc import Callable, Mapping, Sequence
from itertools import product
from pathlib import Path, PurePosixPath
from types import MappingProxyType, ModuleType
from typing import Final, Protocol
EXPECTED_PYTHON_TAG: Final = "cp310"
EXPECTED_ABI_TAG: Final = "abi3"
EXPECTED_PLATFORM_TAG: Final = "linux_x86_64"
class CommandRunner(Protocol):
def __call__(
self,
command: tuple[str, ...],
*,
check: bool,
capture_output: bool,
text: bool,
) -> subprocess.CompletedProcess[str]: ...
def _run_command(
command: tuple[str, ...],
*,
check: bool,
capture_output: bool,
text: bool,
) -> subprocess.CompletedProcess[str]:
return subprocess.run(command, check=check, capture_output=capture_output, text=text)
def _dist_info_directory(member: zipfile.ZipInfo) -> str | None:
parts: Final = PurePosixPath(member.filename).parts
if not parts or not parts[0].endswith(".dist-info"):
return None
return parts[0]
def _wheel_metadata_tags(archive: zipfile.ZipFile, members: tuple[zipfile.ZipInfo, ...]) -> tuple[str, ...]:
if len(members) != 1:
return ()
lines: Final = archive.read(members[0]).splitlines()
return tuple(line.removeprefix(b"Tag:").strip().decode("ascii") for line in lines if line.startswith(b"Tag:"))
def _load_native_module(native_path: Path) -> ModuleType | None:
module_spec: Final = importlib.util.spec_from_file_location("litellm.rust_bridge._native", native_path)
if module_spec is None or module_spec.loader is None:
return None
try:
native_module: Final = importlib.util.module_from_spec(module_spec)
module_spec.loader.exec_module(native_module)
except Exception as error: # noqa: BLE001 # native module initialization can raise arbitrary exceptions
sys.stderr.write(f"native module load failed: {error}\n")
return None
return native_module
def main(
argv: Sequence[str] | None = None,
environment: Mapping[str, str] | None = None,
load_native_module: Callable[[Path], ModuleType | None] = _load_native_module,
run_command: CommandRunner = _run_command,
) -> int:
arguments: Final = tuple(sys.argv if argv is None else argv)
resolved_environment: Final = os.environ if environment is None else environment
if len(arguments) != 2:
sys.stderr.write(f"usage: {Path(arguments[0]).name} WHEEL\n")
return 2
wheel: Final = Path(arguments[1])
wheel_tags: Final = wheel.stem.rsplit("-", maxsplit=3)
if len(wheel_tags) != 4:
sys.stderr.write(f"cannot parse wheel tags from {wheel.name}\n")
return 1
wheel_identity: Final = wheel_tags[0].split("-")
if len(wheel_identity) != 2 or wheel_identity[0] != "litellm" or not wheel_identity[1]:
sys.stderr.write(f"unexpected wheel identity: {wheel_tags[0]}\n")
return 1
expected_dist_info_directory: Final = f"{wheel_tags[0]}.dist-info"
expected_dist_info_directories: Final = frozenset((expected_dist_info_directory,))
python_tag: Final = wheel_tags[1]
abi_tag: Final = wheel_tags[2]
platform_tag: Final = wheel_tags[3]
expanded_filename_tags: Final = frozenset(
"-".join(tag) for tag in product(python_tag.split("."), abi_tag.split("."), platform_tag.split("."))
)
with zipfile.ZipFile(wheel) as archive:
wheel_members: Final = archive.infolist()
dist_info_directories: Final = frozenset(
directory for member in wheel_members if (directory := _dist_info_directory(member)) is not None
)
required_dist_info_files: Final = ("METADATA", "RECORD", "WHEEL")
dist_info_file_counts: Final = MappingProxyType(
{
filename: sum(
member.filename == f"{expected_dist_info_directory}/{filename}" for member in wheel_members
)
for filename in required_dist_info_files
}
)
wheel_metadata_members: Final = tuple(
member for member in wheel_members if member.filename == f"{expected_dist_info_directory}/WHEEL"
)
wheel_metadata_tags: Final = _wheel_metadata_tags(archive, wheel_metadata_members)
native_members: Final = tuple(
member
for member in wheel_members
if member.filename.startswith("litellm/rust_bridge/_native.") and member.filename.endswith(".so")
)
if len(native_members) != 1:
sys.stderr.write(f"expected one native extension, found {len(native_members)}\n")
return 1
unexpected_members: Final = tuple(
member.filename
for member in wheel_members
if member.filename.endswith((".pdb", ".dwp", ".rlib", ".rmeta", "Cargo.toml", "Cargo.lock"))
or any(part.endswith(".dSYM") for part in PurePosixPath(member.filename).parts)
)
native_member: Final = native_members[0]
uncompressed_wheel_size: Final = sum(member.file_size for member in wheel_members)
native_path: Final = wheel.parent / "native" / Path(native_member.filename).name
native_path.parent.mkdir(parents=True, exist_ok=True)
native_bytes: Final = archive.read(native_member)
native_path.write_bytes(native_bytes)
duplicated_vocabularies: Final = tuple(
member.filename
for member in wheel_members
if member.filename.startswith("litellm/litellm_core_utils/tokenizers/")
and re.fullmatch(r"[0-9a-f]{40}", PurePosixPath(member.filename).name)
and member.file_size > 0
and archive.read(member) in native_bytes
)
wheel_metadata_tags_match: Final = (
len(wheel_metadata_tags) == len(expanded_filename_tags)
and frozenset(wheel_metadata_tags) == expanded_filename_tags
)
commit_sha: Final = resolved_environment.get(
"RELEASE_WHEEL_COMMIT_SHA", resolved_environment.get("GITHUB_SHA", "unknown")
)
rustc_version: Final = run_command(
("rustc", "--version"),
check=True,
capture_output=True,
text=True,
).stdout.strip()
pyproject: Final = (Path(__file__).parents[2] / "pyproject.toml").read_text()
maturin_match: Final = re.search(r'"maturin==([^";]+)', pyproject)
if maturin_match is None:
sys.stderr.write("build-system does not pin an exact Maturin version\n")
return 1
maturin_version: Final = maturin_match.group(1)
native_percentage: Final = native_member.file_size / uncompressed_wheel_size * 100
size_report: Final = "\n".join(
(
"## Native wheel build report",
"",
"| Build | Value |",
"| --- | --- |",
f"| Commit | `{commit_sha}` |",
f"| Platform | `{platform_tag}` |",
f"| Python ABI | `{python_tag}-{abi_tag}` |",
f"| Rust compiler | `{rustc_version}` |",
f"| Maturin | `{maturin_version}` |",
"| Cargo profile | `release` |",
"",
"| Artifact | Size |",
"| --- | ---: |",
f"| Compressed wheel | {wheel.stat().st_size / 1_000_000:.2f} MB |",
f"| Uncompressed wheel | {uncompressed_wheel_size / 1_000_000:.2f} MB |",
f"| Native extension | {native_member.file_size / 1_000_000:.2f} MB |",
f"| Native share | {native_percentage:.2f}% |",
"",
)
)
summary_path: Final = resolved_environment.get("GITHUB_STEP_SUMMARY")
if summary_path is None:
sys.stdout.write(size_report)
else:
Path(summary_path).write_text(size_report)
sections: Final = run_command(
("readelf", "--sections", "--wide", str(native_path)),
check=True,
capture_output=True,
text=True,
).stdout
debug_sections: Final = tuple(section for section in (".debug_", ".zdebug_") if section in sections)
debug_sections_absent: Final = not debug_sections
static_symbol_table_absent: Final = ".symtab" not in sections
dynamic_symbols: Final = run_command(
("readelf", "--dyn-syms", "--wide", str(native_path)),
check=True,
capture_output=True,
text=True,
).stdout
extension_entry_point_present: Final = "PyInit__native" in dynamic_symbols
native_module: Final = load_native_module(native_path)
native_module_loads: Final = native_module is not None
panic_test_hook_absent: Final = native_module is not None and not hasattr(native_module, "_panic_for_test")
native_size_limit: Final = 35_000_000
native_size_within_limit: Final = native_member.file_size <= native_size_limit
validations: Final = (
(f"Python tag is {EXPECTED_PYTHON_TAG}", python_tag == EXPECTED_PYTHON_TAG),
(f"ABI tag is {EXPECTED_ABI_TAG}", abi_tag == EXPECTED_ABI_TAG),
(f"Platform tag is {EXPECTED_PLATFORM_TAG}", platform_tag == EXPECTED_PLATFORM_TAG),
("Wheel dist-info directory matches the filename", dist_info_directories == expected_dist_info_directories),
(
"Required dist-info files are present exactly once",
all(count == 1 for count in dist_info_file_counts.values()),
),
("Wheel metadata tags match the filename", wheel_metadata_tags_match),
("Debug sections are absent", debug_sections_absent),
("Static symbol table is absent", static_symbol_table_absent),
("Python extension entry point is present", extension_entry_point_present),
("Native module loads", native_module_loads),
("Production module omits the panic test hook", panic_test_hook_absent),
(f"Native extension does not exceed {native_size_limit / 1_000_000:.0f} MB", native_size_within_limit),
("Tokenizer vocabularies are not duplicated in the native extension", not duplicated_vocabularies),
("Wheel contents are valid", not unexpected_members),
)
verified_report: Final = size_report + "\n".join(
("", "| Validation | Expected | Result |", "| --- | --- | :---: |")
+ tuple(f"| {label} | Yes | {'O' if passed else 'X'} |" for label, passed in validations)
+ ("",)
)
if summary_path is not None:
Path(summary_path).write_text(verified_report)
invalid_dist_info_files: Final = any(count != 1 for count in dist_info_file_counts.values())
validation_errors: Final = tuple(
message
for failed, message in (
(bool(debug_sections), f"{native_member.filename} contains debug sections: {', '.join(debug_sections)}"),
(not static_symbol_table_absent, f"{native_member.filename} contains a static symbol table"),
(not extension_entry_point_present, "native extension does not export PyInit__native"),
(
python_tag != EXPECTED_PYTHON_TAG,
f"unexpected Python tag: expected {EXPECTED_PYTHON_TAG}, found {python_tag}",
),
(abi_tag != EXPECTED_ABI_TAG, f"unexpected ABI tag: expected {EXPECTED_ABI_TAG}, found {abi_tag}"),
(
platform_tag != EXPECTED_PLATFORM_TAG,
f"unexpected platform tag: expected {EXPECTED_PLATFORM_TAG}, found {platform_tag}",
),
(
dist_info_directories != expected_dist_info_directories,
f"unexpected dist-info directories: expected {expected_dist_info_directory}, "
f"found {', '.join(sorted(dist_info_directories))}",
),
(invalid_dist_info_files, f"required dist-info file counts are invalid: {dist_info_file_counts}"),
(
not invalid_dist_info_files and not wheel_metadata_tags_match,
f"WHEEL tags do not match filename: expected {', '.join(sorted(expanded_filename_tags))}, "
f"found {', '.join(sorted(wheel_metadata_tags))}",
),
(
native_module is not None and not panic_test_hook_absent,
"production native module exposes _panic_for_test",
),
(
not native_size_within_limit,
f"native extension exceeds {native_size_limit / 1_000_000:.0f} MB: "
f"{native_member.file_size / 1_000_000:.2f} MB",
),
(bool(unexpected_members), f"wheel contains unexpected build artifacts: {', '.join(unexpected_members)}"),
)
if failed
)
sys.stderr.write("".join(f"{message}\n" for message in validation_errors))
return 0 if all(passed for _, passed in validations) else 1
if __name__ == "__main__":
sys.exit(main())