PTQ-Level Efficiency, QAT-Level Accuracy | 训练后量化的效率,量化感知训练的精度
fix: CLI ignore --scheme when bits/act_bits are unset (#2019)
Heng Guoc70e99f63048
diff --git a/auto_round/algorithms/quantization/sign_round/config.py b/auto_round/algorithms/quantization/sign_round/config.pyindex a3af1799c..442cf2775 100644--- a/auto_round/algorithms/quantization/sign_round/config.py+++ b/auto_round/algorithms/quantization/sign_round/config.py@@ -79,16 +79,10 @@ class SignRoundConfig(QuantizationConfig):logger.warning("`iters` must be non-negative, reset it to 200")self.iters = 200- if not lr:- # TODO need to check 4 bits lr setting for auto-round-best, 3bits only validate on small models- if self.iters >= 1000 and self.bits is not None and self.bits <= 3:- self.lr = 2.0 / self.iters- logger.info("set the lr to 2.0/iters for better accuracy")- else:- self.lr = 1.0 / self.iters- else:- self.lr = lr- self.minmax_lr = minmax_lr or self.lr+ # lr/minmax_lr depend on `bits`, which may still be unresolved here+ # (e.g. only `scheme=` was given) -- finalize_scheme() fills them in.+ self.lr = lr+ self.minmax_lr = minmax_lrself.lr_scheduler = lr_schedulerself.nblocks = nblocks@@ -108,6 +102,17 @@ class SignRoundConfig(QuantizationConfig):self.optimizer = optimizerself.enable_adam = enable_adam+ def finalize_scheme(self) -> None:+ """Resolve lr/minmax_lr once `bits` is known (low-bit schemes use a higher lr)."""+ if self.lr is None:+ # TODO need to check 4 bits lr setting for auto-round-best, 3bits only validate on small models+ if self.iters >= 1000 and self.bits is not None and self.bits <= 3:+ self.lr = 2.0 / self.iters+ logger.info("set the lr to 2.0/iters for better accuracy")+ else:+ self.lr = 1.0 / self.iters+ self.minmax_lr = self.minmax_lr or self.lr+def check_configs(self) -> None:"""Checks if the configurations are valid.diff --git a/auto_round/compressors/entry.py b/auto_round/compressors/entry.pyindex 380f1dc4f..7abf12767 100644--- a/auto_round/compressors/entry.py+++ b/auto_round/compressors/entry.py@@ -251,13 +251,18 @@ def _build_model_type_ctor_kwargs(model, base_kwargs, mllm_kwargs, diffusion_kwadef _select_rtn_compressor_base_cls(quant_config: RTNConfig, scheme, format, base_kwargs) -> type:enable_imatrix = False- resolved_attrs = {}disable_opt_rtn = getattr(quant_config, "disable_opt_rtn", False)- # If disable_opt_rtn was not explicitly set and scheme is W8A16/W8A8,- # auto-disable optimization to improve efficiency.+ # Preview resolved scheme attrs once (authoritative resolution happens later).+ resolved_attrs = _preview_resolved_attrs(quant_config, scheme)++ # Auto-disable rtn optimization for W8A16/W8A8-equivalent resolved schemes,+ # unless the user already set disable_opt_rtn explicitly.if getattr(quant_config, "orig_disable_opt_rtn", None) is None:- if isinstance(scheme, str) and scheme.upper() in ["W8A16", "W8A8"]:+ bits = resolved_attrs.get("bits", getattr(quant_config, "bits", None))+ act_bits = resolved_attrs.get("act_bits", getattr(quant_config, "act_bits", None))+ data_type = resolved_attrs.get("data_type", getattr(quant_config, "data_type", None))+ if bits is not None and bits >= 8 and act_bits is not None and act_bits >= 8 and data_type == "int":logger.warning("`disable_opt_rtn` is turned on for W8A16/W8A8 quantization to improve efficiency.")disable_opt_rtn = Truequant_config.disable_opt_rtn = True@@ -267,9 +272,6 @@ def _select_rtn_compressor_base_cls(quant_config: RTNConfig, scheme, format, basif has_gguf_k:enable_imatrix = Trueelse:- # Resolve scheme attrs for routing. SchemeMixin will do the authoritative- # resolution later; this preview only chooses the compressor class.- resolved_attrs = _preview_resolved_attrs(quant_config, scheme)sym = resolved_attrs.get("sym", getattr(quant_config, "sym", None))data_type = resolved_attrs.get("data_type", getattr(quant_config, "data_type", "") or "")bits = resolved_attrs.get("bits", getattr(quant_config, "bits", None))@@ -280,7 +282,6 @@ def _select_rtn_compressor_base_cls(quant_config: RTNConfig, scheme, format, baselif is_weight_scheme(scheme):enable_imatrix = True- resolved_attrs = resolved_attrs if not disable_opt_rtn else _preview_resolved_attrs(quant_config, scheme)act_bits = resolved_attrs.get("act_bits", getattr(quant_config, "act_bits", None))act_data_type = resolved_attrs.get("act_data_type", getattr(quant_config, "act_data_type", None))act_dynamic = resolved_attrs.get("act_dynamic", getattr(quant_config, "act_dynamic", None))diff --git a/test/test_cpu/core/test_pipeline_fail_fast.py b/test/test_cpu/core/test_pipeline_fail_fast.pyindex cc6779028..84b68ca10 100644--- a/test/test_cpu/core/test_pipeline_fail_fast.py+++ b/test/test_cpu/core/test_pipeline_fail_fast.py@@ -15,7 +15,9 @@ from auto_round.algorithms.quantization.rtn.quantizer import RTNQuantizerfrom auto_round.compressors.base import collect_user_scheme_overridesfrom auto_round.compressors.data_driven import DataDrivenCompressorfrom auto_round.compressors.entry import AutoRound as NewAutoRound+from auto_round.compressors.entry import _select_rtn_compressor_base_clsfrom auto_round.logger import logger+from auto_round.schemes import QuantizationSchemeclass PartialSharedConfig(RTNConfig):@@ -206,3 +208,53 @@ def test_user_scheme_overrides_reject_explicit_conflicts():collect_user_scheme_overrides([AWQConfig(bits=8), RTNConfig(bits=4)])with pytest.raises(ValueError, match="Conflicting shared scheme field 'bits'"):resolve_shared_config_values([AWQConfig(bits=8), RTNConfig(bits=4)])+++# ===========================================================================+# Scheme-dependent config heuristics must see resolved values, not just+# whatever (often None) bits/lr the config was constructed with directly.+# ===========================================================================+++ "scheme, expect_disable_opt_rtn",+ [+ ("W8A16", True),+ # "INT8" (bits=8, act_bits=8, data_type=int) is W8A8-equivalent but was+ # previously missed because routing only matched the literal strings+ # "W8A16"/"W8A8", not schemes reaching the same resolved values.+ ("INT8", True),+ ("W4A16", False),+ ({"bits": 8, "act_bits": 8, "data_type": "int", "sym": True}, True),+ ],+)+def test_rtn_routing_disable_opt_rtn_from_resolved_scheme(scheme, expect_disable_opt_rtn):+ config = RTNConfig()+ _select_rtn_compressor_base_cls(config, scheme, "auto_round", {})+ assert config.disable_opt_rtn is expect_disable_opt_rtn+++def test_rtn_routing_respects_explicit_enable_opt_rtn():+ """An explicit user choice must not be clobbered by the W8A16/W8A8 heuristic."""+ config = RTNConfig(enable_opt_rtn=True)+ _select_rtn_compressor_base_cls(config, "W8A16", "auto_round", {})+ assert config.disable_opt_rtn is False++[email protected]("bits, expected_lr", [(3, 2.0 / 1000), (4, 1.0 / 1000)])+def test_sign_round_finalize_scheme_lr_heuristic(bits, expected_lr):+ """The low-bit lr bump must apply once `bits` is resolved via the scheme,+ even though it was unset (None) at construction time (e.g. `scheme=` alone,+ no explicit `bits=`)."""+ config = SignRoundConfig(iters=1000)+ config.scheme = QuantizationScheme(bits=bits, act_bits=16, data_type="int")+ config.finalize_scheme()+ assert config.lr == expected_lr+++def test_sign_round_finalize_scheme_respects_explicit_lr():+ config = SignRoundConfig(iters=1000, lr=0.01, minmax_lr=0.05)+ config.scheme = QuantizationScheme(bits=2, act_bits=16, data_type="int")+ config.finalize_scheme()+ assert config.lr == 0.01+ assert config.minmax_lr == 0.05