From 948107acf7ef8813b8ec94fff7c5ab62d4aba9ae Mon Sep 17 00:00:00 2001 From: Xin He Date: Tue, 28 Jul 2026 20:37:55 +0800 Subject: [PATCH] [Bugfix] Enhance extra_config handling for layer name suffix matching (#48589) Signed-off-by: Xin He Co-authored-by: Kunshang Ji --- tests/quantization/test_auto_round.py | 49 +++++++++++++++++++ .../layers/quantization/inc/config_parser.py | 8 +++ 2 files changed, 57 insertions(+) diff --git a/tests/quantization/test_auto_round.py b/tests/quantization/test_auto_round.py index 732080fc96..59c6a1e326 100644 --- a/tests/quantization/test_auto_round.py +++ b/tests/quantization/test_auto_round.py @@ -243,6 +243,27 @@ def test_inc_config_parser_parallel_lm_head_defaults_to_unquantized() -> None: assert layer_config.bits == 16 +def test_inc_config_parser_suffix_match_for_lm_head() -> None: + """Short extra_config key should match fully-qualified lm_head layer name.""" + layer = object.__new__(ParallelLMHead) + config = make_config( + extra_config={ + "lm_head": { + "bits": 4, + "group_size": 128, + "sym": True, + } + } + ) + + layer_config = config.config_parser.resolve(layer, "model.language_model.lm_head") + + assert layer_config.quantized is True + assert layer_config.bits == 4 + assert layer_config.group_size == 128 + assert layer_config.sym is True + + def test_inc_config_parser_fused_moe_requires_consistent_configs() -> None: config = make_config( extra_config={ @@ -790,6 +811,34 @@ def test_inc_get_quant_method_linear_uses_resolved_scheme(monkeypatch) -> None: assert method is sentinel +def test_inc_get_quant_method_lm_head_uses_suffix_match(monkeypatch) -> None: + """lm_head extra_config should apply to fully-qualified prefix.""" + config = make_config( + extra_config={ + "lm_head": { + "bits": 4, + "group_size": 128, + "sym": True, + } + } + ) + layer = object.__new__(ParallelLMHead) + sentinel = object() + + class DummyScheme: + def get_linear_method(self, _config, _layer, _prefix, _layer_config): + return sentinel + + monkeypatch.setattr( + "vllm.model_executor.layers.quantization.inc.schemes.factory.resolve_scheme", + lambda _layer_config: DummyScheme(), + ) + + method = config.get_quant_method(layer, "model.language_model.lm_head") + + assert method is sentinel + + def test_inc_get_quant_method_moe_uses_resolved_scheme(monkeypatch) -> None: config = make_config() layer = object.__new__(RoutedExperts) diff --git a/vllm/model_executor/layers/quantization/inc/config_parser.py b/vllm/model_executor/layers/quantization/inc/config_parser.py index 603b80b7cd..6e94cad2cc 100644 --- a/vllm/model_executor/layers/quantization/inc/config_parser.py +++ b/vllm/model_executor/layers/quantization/inc/config_parser.py @@ -142,6 +142,14 @@ class INCConfigParser: if self._config.extra_config and layer_name in self._config.extra_config: return get_config(layer_name) + # Suffix match: handle cases where extra_config keys use short names + # (e.g. "lm_head") but the layer_name is fully qualified + # (e.g. "model.language_model.lm_head") due to model nesting. + if self._config.extra_config: + for cfg_key in self._config.extra_config: + if layer_name.endswith(f".{cfg_key}"): + return get_config(cfg_key) + quantized = not isinstance(layer, ParallelLMHead) if self._config.block_name_to_quantize: quantized = any(