[Bugfix] fix convert part_file_size not read from YAML config-MindStudio-ModelSlim-AtomGit

Mmsmodelslim/core/convert/config.py+9-0

	@@ -121,6 +121,8 @@ class ConvertConfig(BaseModel):
121	save_path: str	121	save_path: str
122	model_family: Optional[str] = None	122	model_family: Optional[str] = None
123	dst_format: str = "ascendv1"	123	dst_format: str = "ascendv1"
		124	+ # 分片文件大小（GB）；0 表示不分片。来自 YAML ``spec.save[].part_file_size``。
		125	+ part_file_size: int = 4
124	defaults: ConvertDefaults = Field(default_factory=ConvertDefaults)	126	defaults: ConvertDefaults = Field(default_factory=ConvertDefaults)
125	preprocess_rules: List[WeightMappingRule] = Field(default_factory=list)	127	preprocess_rules: List[WeightMappingRule] = Field(default_factory=list)
126	module_rules: List[ModuleRule] = Field(default_factory=list)	128	module_rules: List[ModuleRule] = Field(default_factory=list)
	@@ -134,6 +136,13 @@ class ConvertConfig(BaseModel):
134	raise ValueError("path must be non-empty")	136	raise ValueError("path must be non-empty")
135	return v	137	return v
136		138
		139	+ @field_validator("part_file_size", mode="after")
		140	+ @classmethod
		141	+ def _non_negative_part_file_size(cls, v: int) -> int:
		142	+ if v < 0:
		143	+ raise ValueError("part_file_size must be >= 0 (0 means no sharding)")
		144	+ return v
		145	+
137	@model_validator(mode="after")	146	@model_validator(mode="after")
138	def _mxfp8_requires_ascendv1(self) -> ConvertConfig:	147	def _mxfp8_requires_ascendv1(self) -> ConvertConfig:
139	"""任意 convert_rule 目标为 W8A8_MXFP8 时，dst_format 必须为 ascendv1。"""	148	"""任意 convert_rule 目标为 W8A8_MXFP8 时，dst_format 必须为 ascendv1。"""

Mmsmodelslim/core/quant_service/modelslim_convert/config_mapper.py+8-0

	@@ -252,6 +252,13 @@ def _resolve_dst_format(save: List[SaveConfig], defaults: ConvertDefaults) -> st
252	return defaults.dst_format	252	return defaults.dst_format
253		253
254		254
		255	+def _resolve_part_file_size(save: List[SaveConfig]) -> int:
		256	+ """从 YAML ``spec.save[0].part_file_size`` 读取分片大小；未配置时默认 4GB。"""
		257	+ if save:
		258	+ return save[0].part_file_size
		259	+ return 4
		260	+
		261	+
255	def spec_to_convert_config(	262	def spec_to_convert_config(
256	spec: Union[ModelslimConvertServiceConfig, Dict[str, Any]],	263	spec: Union[ModelslimConvertServiceConfig, Dict[str, Any]],
257	model_path: str,	264	model_path: str,
	@@ -279,6 +286,7 @@ def spec_to_convert_config(
279	save_path=save_path,	286	save_path=save_path,
280	model_family=model_family,	287	model_family=model_family,
281	dst_format=_resolve_dst_format(spec.save, spec.defaults),	288	dst_format=_resolve_dst_format(spec.save, spec.defaults),
		289	+ part_file_size=_resolve_part_file_size(spec.save),
282	defaults=spec.defaults,	290	defaults=spec.defaults,
283	preprocess_rules=_preprocess_to_rules(spec),	291	preprocess_rules=_preprocess_to_rules(spec),
284	module_rules=module_rules,	292	module_rules=module_rules,

Mmsmodelslim/core/quant_service/modelslim_convert/impl/save_adapter.py+14-4

	@@ -83,7 +83,8 @@ class SaveProcessorAdapter(ISaveProcessorAdapter):
83	save_dir: str,	83	save_dir: str,
84	adapter: IModel,	84	adapter: IModel,
85	) -> None:	85	) -> None:
86	- format_cfg = parse_format_config({"type": "compressed_tensors", "part_file_size": 4})	86	+ part_file_size = context.config.part_file_size
		87	+ format_cfg = parse_format_config({"type": "compressed_tensors", "part_file_size": part_file_size})
87	cfg = QuantSaveProcessorConfig(type="saver", format=format_cfg)	88	cfg = QuantSaveProcessorConfig(type="saver", format=format_cfg)
88	cfg.set_save_directory(save_dir)	89	cfg.set_save_directory(save_dir)
89	_lazy_init_unsaved_modules(context, tree)	90	_lazy_init_unsaved_modules(context, tree)
	@@ -95,7 +96,11 @@ class SaveProcessorAdapter(ISaveProcessorAdapter):
95	if name:	96	if name:
96	saver.postprocess(BatchProcessRequest(name=name, module=module, datas=None, outputs=None))	97	saver.postprocess(BatchProcessRequest(name=name, module=module, datas=None, outputs=None))
97	saver.post_run()	98	saver.post_run()
98	- logger.info("Saved HF/compressed_tensors checkpoint to %s", save_dir)	99	+ logger.info(
		100	+ "Saved HF/compressed_tensors checkpoint to %s (part_file_size=%s)",
		101	+ save_dir,
		102	+ part_file_size,
		103	+ )
99		104
100	@staticmethod	105	@staticmethod
101	def _save_ascendv1(	106	def _save_ascendv1(
	@@ -106,9 +111,14 @@ class SaveProcessorAdapter(ISaveProcessorAdapter):
106	) -> None:	111	) -> None:
107	from msmodelslim.core.quant_service.modelslim_v1.save.ascendv1 import AscendV1Config, AscendV1Saver	112	from msmodelslim.core.quant_service.modelslim_v1.save.ascendv1 import AscendV1Config, AscendV1Saver
108		113
		114	+ part_file_size = context.config.part_file_size
109	_lazy_init_unsaved_modules(context, tree)	115	_lazy_init_unsaved_modules(context, tree)
110	- cfg = AscendV1Config(save_directory=save_dir, part_file_size=4)	116	+ cfg = AscendV1Config(save_directory=save_dir, part_file_size=part_file_size)
111	saver = AscendV1Saver(model=tree, config=cfg, adapter=adapter)	117	saver = AscendV1Saver(model=tree, config=cfg, adapter=adapter)
112	saver.pre_run()	118	saver.pre_run()
113	saver.post_run()	119	saver.post_run()
114	- logger.info("Saved AscendV1 checkpoint to %s", save_dir)	120	+ logger.info(
		121	+ "Saved AscendV1 checkpoint to %s (part_file_size=%s)",
		122	+ save_dir,
		123	+ part_file_size,
		124	+ )

Mtest/cases/core/quant_service/modelslim_convert/impl/test_save_adapter.py+43-0

	@@ -72,3 +72,46 @@ class TestSaveProcessorAdapter:
72	cfg = mock_cls.call_args[0][1]	72	cfg = mock_cls.call_args[0][1]
73	assert isinstance(cfg.format, CompressedTensorsQuantFormatConfig)	73	assert isinstance(cfg.format, CompressedTensorsQuantFormatConfig)
74	assert cfg.format.part_file_size == 4	74	assert cfg.format.part_file_size == 4
		75	+
		76	+ def test_save_compressed_tensors_use_part_file_size_when_config_set(self):
		77	+ from msmodelslim.format.compressed_tensors_format.compressed_tensors import (
		78	+ CompressedTensorsQuantFormatConfig,
		79	+ )
		80	+
		81	+ tree = nn.Module()
		82	+ config = ConvertConfig(model_path="/m", save_path="/out", dst_format="huggingface", part_file_size=0)
		83	+ context = ConvertContext(config=config)
		84	+ context.reader = MagicMock()
		85	+
		86	+ with (
		87	+ patch.object(QuantSaveProcessor, "pre_run"),
		88	+ patch.object(QuantSaveProcessor, "postprocess"),
		89	+ patch.object(QuantSaveProcessor, "post_run"),
		90	+ patch(
		91	+ "msmodelslim.core.quant_service.modelslim_convert.impl.save_adapter.QuantSaveProcessor",
		92	+ ) as mock_cls,
		93	+ ):
		94	+ SaveProcessorAdapter().save(context, tree)
		95	+ cfg = mock_cls.call_args[0][1]
		96	+ assert isinstance(cfg.format, CompressedTensorsQuantFormatConfig)
		97	+ assert cfg.format.part_file_size == 0
		98	+
		99	+ def test_save_ascendv1_use_part_file_size_when_config_set(self):
		100	+ tree = nn.Module()
		101	+ config = ConvertConfig(model_path="/m", save_path="/out", dst_format="ascendv1", part_file_size=8)
		102	+ context = ConvertContext(config=config)
		103	+ context.reader = MagicMock()
		104	+
		105	+ with (
		106	+ patch("msmodelslim.core.quant_service.modelslim_v1.save.ascendv1.AscendV1Saver") as mock_saver_cls,
		107	+ patch(
		108	+ "msmodelslim.core.quant_service.modelslim_convert.impl.save_adapter._lazy_init_unsaved_modules",
		109	+ ),
		110	+ ):
		111	+ mock_saver = MagicMock()
		112	+ mock_saver_cls.return_value = mock_saver
		113	+ SaveProcessorAdapter().save(context, tree)
		114	+ saver_cfg = mock_saver_cls.call_args.kwargs["config"]
		115	+ assert saver_cfg.part_file_size == 8
		116	+ mock_saver.pre_run.assert_called_once()
		117	+ mock_saver.post_run.assert_called_once()

Mtest/cases/core/quant_service/modelslim_convert/test_config_mapper.py+17-0

	@@ -74,9 +74,26 @@ class TestSpecToConvertConfig:
74	assert len(cfg.convert_rules) == 2	74	assert len(cfg.convert_rules) == 2
75	assert cfg.convert_rules[0].target_ir == IRKind.W8A8_MXFP8	75	assert cfg.convert_rules[0].target_ir == IRKind.W8A8_MXFP8
76	assert cfg.dst_format == "ascendv1"	76	assert cfg.dst_format == "ascendv1"
		77	+ assert cfg.part_file_size == 4
77	assert cfg.parallel.max_workers == 8	78	assert cfg.parallel.max_workers == 8
78	assert cfg.model_family == "qwen3_5_moe"	79	assert cfg.model_family == "qwen3_5_moe"
79		80
		81	+ def test_spec_to_convert_config_map_part_file_size_when_save_given(self):
		82	+ spec = ModelslimConvertServiceConfig.model_validate(
		83	+ {
		84	+ "linears": [],
		85	+ "save": [{"type": "huggingface", "part_file_size": 0}],
		86	+ }
		87	+ )
		88	+ cfg = spec_to_convert_config(spec, model_path="/m", save_path="/o")
		89	+ assert cfg.dst_format == "huggingface"
		90	+ assert cfg.part_file_size == 0
		91	+
		92	+ def test_spec_to_convert_config_default_part_file_size_when_save_empty(self):
		93	+ spec = ModelslimConvertServiceConfig.model_validate({"linears": []})
		94	+ cfg = spec_to_convert_config(spec, model_path="/m", save_path="/o")
		95	+ assert cfg.part_file_size == 4
		96	+
80	def test_spec_to_convert_config_auto_route_infer_source_ir_from_catalog_later(self):	97	def test_spec_to_convert_config_auto_route_infer_source_ir_from_catalog_later(self):
81	spec = ModelslimConvertServiceConfig.model_validate(	98	spec = ModelslimConvertServiceConfig.model_validate(
82	{	99	{