Skip to content

PTQ-Level Efficiency, QAT-Level Accuracy | 训练后量化的效率,量化感知训练的精度

feat: Add support for deterministic algorithms flags in CLI and update logging behavior (#2261)

Xin Hed72933bceaab
diff --git a/auto_round/autoround.py b/auto_round/autoround.py
index 9370a5268..951e3e20d 100644
--- a/auto_round/autoround.py
+++ b/auto_round/autoround.py
@@ -332,7 +332,6 @@ _ENTRY_KWARG_OWNERS = {
"model_dtype": "base",
"trust_remote_code": "base",
"amp": "base",
- "disable_deterministic_algorithms": "base",
"enable_deterministic_algorithms": "base",
"static_kv_dtype": "base",
"static_attention_dtype": "base",
diff --git a/auto_round/cli/main.py b/auto_round/cli/main.py
index e52bad935..6eedfe4b6 100644
--- a/auto_round/cli/main.py
+++ b/auto_round/cli/main.py
@@ -62,6 +62,7 @@ def _build_entry_base_kwargs(args, *, low_cpu_mem_usage, enable_torch_compile, l
"low_cpu_mem_usage": low_cpu_mem_usage,
"device_map": args.device_map,
"enable_torch_compile": enable_torch_compile,
+ "enable_deterministic_algorithms": args.enable_deterministic_algorithms,
"seed": args.seed,
"layer_config": layer_config,
"model_dtype": args.model_dtype,
@@ -320,12 +321,6 @@ def tune(args):
if scheme not in PRESET_SCHEMES:
raise ValueError(f"{scheme} is not supported. only {PRESET_SCHEMES.keys()} are supported ")
- if args.disable_deterministic_algorithms:
- logger.warning(
- "default not use deterministic_algorithms. disable_deterministic_algorithms is deprecated,"
- " please use enable_deterministic_algorithms instead. "
- )
-
from auto_round.utils import parse_layer_config_arg
layer_config = {}
diff --git a/auto_round/cli/parser.py b/auto_round/cli/parser.py
index 570fe52b8..d0926e8ee 100644
--- a/auto_round/cli/parser.py
+++ b/auto_round/cli/parser.py
@@ -251,14 +251,10 @@ def build_quantize_parser(*, prog: str = "auto_round quantize") -> argparse.Argu
help="Ignore scale and zero-point overhead when computing AutoScheme target bits.",
)
compat.add_argument("--disable_amp", action="store_true", help="Disable AMP during tuning.")
- compat.add_argument(
- "--disable_deterministic_algorithms",
- action="store_true",
- help="Deprecated flag to disable deterministic algorithms.",
- )
compat.add_argument(
"--enable_deterministic_algorithms",
action="store_true",
+ default=None,
help="Enable deterministic algorithms for reproducible runs.",
)
compat.add_argument("--model_free", action="store_true", help="Force model-free quantization mode.")
diff --git a/auto_round/compressors/base.py b/auto_round/compressors/base.py
index d935bac2c..94f6bab58 100644
--- a/auto_round/compressors/base.py
+++ b/auto_round/compressors/base.py
@@ -309,7 +309,6 @@ class BaseOrchestrator(object):
kwargs.pop("vlm", None)
amp = kwargs.pop("amp", True)
nblocks = kwargs.pop("nblocks", 1)
- disable_deterministic_algorithms = kwargs.pop("disable_deterministic_algorithms", True)
enable_deterministic_algorithms = kwargs.pop("enable_deterministic_algorithms", False)
self._offloader = OffloadManager(enabled=low_cpu_mem_usage, mode="offload", offload_dir_prefix="compressor")
@@ -344,17 +343,9 @@ class BaseOrchestrator(object):
)
if "CUBLAS_WORKSPACE_CONFIG" not in os.environ:
os.environ["CUBLAS_WORKSPACE_CONFIG"] = ":4096:8"
- # Deprecated, default not to use torch.use_deterministic_algorithms
- if not disable_deterministic_algorithms or enable_deterministic_algorithms:
- if not disable_deterministic_algorithms:
- logger.warning(
- "default not use deterministic_algorithms. disable_deterministic_algorithms is deprecated,"
- " please use enable_deterministic_algorithms instead. "
- )
-
+ if enable_deterministic_algorithms:
+ logger.info("Deterministic algorithms are enabled.")
torch.use_deterministic_algorithms(True, warn_only=False)
- else:
- torch.use_deterministic_algorithms(True, warn_only=True)
# XPU SDPA workaround: drop pure causal masks so FLASH backend is used,
# and set torch.use_deterministic_algorithms(False)
diff --git a/test/unit/common/core/test_init.py b/test/unit/common/core/test_init.py
index 5c8598a6f..fdc16cea5 100644
--- a/test/unit/common/core/test_init.py
+++ b/test/unit/common/core/test_init.py
@@ -1,8 +1,11 @@
import inspect
import logging
+import pytest
+
from auto_round import AutoRound
from auto_round.auto_scheme import AutoScheme
+from auto_round.cli.main import _to_autoround_kwargs
from auto_round.cli.parser import build_quantize_parser
from auto_round.compressors.base import MIN_ITERS_FOR_TORCH_COMPILE, BaseOrchestrator
from auto_round.logger import logger
@@ -26,6 +29,54 @@ def test_cli_torch_compile_flags():
assert parser.parse_args(["--model", "test-model", "--disable_torch_compile"]).enable_torch_compile is False
+def test_cli_deterministic_algorithms_flags_are_forwarded():
+ for flag, expected in (
+ (None, None),
+ ("--enable_deterministic_algorithms", True),
+ ):
+ argv = ["--model", "test-model"]
+ if flag is not None:
+ argv.append(flag)
+ args = build_quantize_parser().parse_args(argv)
+ kwargs = _to_autoround_kwargs(
+ args,
+ low_cpu_mem_usage=True,
+ enable_torch_compile=None,
+ layer_config={},
+ )
+
+ assert kwargs["enable_deterministic_algorithms"] is expected
+
+
+def test_deterministic_algorithms_rejects_legacy_flag():
+ parser = build_quantize_parser()
+ with pytest.raises(SystemExit):
+ parser.parse_args(["--model", "test-model", "--disable_deterministic_algorithms"])
+
+
+def test_deterministic_algorithms_runtime_logging(monkeypatch, caplog, tiny_opt_model_path):
+ calls = []
+ monkeypatch.setattr("torch.use_deterministic_algorithms", lambda mode, warn_only: calls.append((mode, warn_only)))
+ monkeypatch.setattr(logger, "propagate", True)
+
+ with caplog.at_level(logging.INFO):
+ AutoRound(model=tiny_opt_model_path, scheme="W4A16", iters=0, nsamples=1)
+ assert calls == []
+ assert "Deterministic algorithms are enabled." not in caplog.text
+
+ caplog.clear()
+ with caplog.at_level(logging.INFO):
+ AutoRound(
+ model=tiny_opt_model_path,
+ scheme="W4A16",
+ iters=0,
+ nsamples=1,
+ enable_deterministic_algorithms=True,
+ )
+ assert calls[-1] == (True, False)
+ assert "Deterministic algorithms are enabled." in caplog.text
+
+
def test_cli_dataset_tracks_explicit_value():
parser = build_quantize_parser()
assert parser.parse_args(["--model", "test-model"]).dataset is None