[bugfix] 解决了PD分离场景下开启Fused MoE后的精度问题-MindIE-LLM-AtomGit

已合并

[bugfix] 解决了PD分离场景下开启Fused MoE后的精度问题 #914

Dawn952创建于 4月17日从Dawn952/MindIE-LLM_open： T0005-bugfix 合入到 Ascend/MindIE-LLM： dev

[bugfix] 解决了PD分离场景下开启Fused MoE后的精度问题 #914

已合并

Dawn952创建于 4月17日

从Dawn952/MindIE-LLM_open T0005-bugfix合入到Ascend/MindIE-LLM dev

共 3 个文件变更+18-9

Mmindie_llm/runtime/layers/fused_moe/moe_comm_strategy.py+1-0

	@@ -202,6 +202,7 @@ class AllGatherStrategy(MoECommStrategyBase):
202		202
203	# Strategy selection order: first applicable strategy is used	203	# Strategy selection order: first applicable strategy is used
204	MOE_COMM_STRATEGIES = [	204	MOE_COMM_STRATEGIES = [
		205	+ FusedMC2Strategy, # P0: Optimal (910C Fused MC2)
205	MC2Strategy, # P1: High perf (large cluster/decode)	206	MC2Strategy, # P1: High perf (large cluster/decode)
206	All2AllStrategy, # P2: Fallback for prefill/specific cases	207	All2AllStrategy, # P2: Fallback for prefill/specific cases
207	AllGatherStrategy, # P3: Universal fallback	208	AllGatherStrategy, # P3: Universal fallback

Mmindie_llm/runtime/layers/quantization/ms_model_slim/w8a8.py+4-9

	@@ -43,7 +43,6 @@ from mindie_llm.runtime.utils.distributed.utils import even_divide
43	from mindie_llm.utils.log.logging import logger	43	from mindie_llm.utils.log.logging import logger
44		44
45		45
46	-SUPPORT_NZ_NPU_LIST = ("Ascend910B3", "Ascend910B4-1", "Ascend910_9382", "Ascend910_9362")
47	MXFP8_GROUP_SIZE = 32	46	MXFP8_GROUP_SIZE = 32
48		47
49		48
	@@ -160,8 +159,7 @@ class W8A8PerTensorLinearMethod(LinearMethodBase):
160	)	159	)
161	layer.weight.data = layer.weight.data.transpose(0, 1).contiguous()	160	layer.weight.data = layer.weight.data.transpose(0, 1).contiguous()
162		161
163	- soc_name = get_npu_node_info().soc_name	162	+ if get_npu_node_info().is_nz_format_beneficial():
164	- if soc_name in SUPPORT_NZ_NPU_LIST:
165	layer.weight.data = torch_npu.npu_format_cast(layer.weight.data, 29)	163	layer.weight.data = torch_npu.npu_format_cast(layer.weight.data, 29)
166	logger.debug(	164	logger.debug(
167	"Convert weight to FRACTAL_NZ done, current format is %s", torch_npu.get_npu_format(layer.weight.data)	165	"Convert weight to FRACTAL_NZ done, current format is %s", torch_npu.get_npu_format(layer.weight.data)
	@@ -249,8 +247,7 @@ class W8A8PerTokenLinearMethod(LinearMethodBase):
249	layer.weight.data = layer.weight.data.transpose(0, 1).contiguous()	247	layer.weight.data = layer.weight.data.transpose(0, 1).contiguous()
250	layer.weight_scale.data = layer.weight_scale.data.flatten()	248	layer.weight_scale.data = layer.weight_scale.data.flatten()
251		249
252	- soc_name = get_npu_node_info().soc_name	250	+ if get_npu_node_info().is_nz_format_beneficial():
253	- if soc_name in SUPPORT_NZ_NPU_LIST:
254	layer.weight.data = torch_npu.npu_format_cast(layer.weight.data, 29)	251	layer.weight.data = torch_npu.npu_format_cast(layer.weight.data, 29)
255	logger.debug(	252	logger.debug(
256	"Convert weight to FRACTAL_NZ done, current format is %s", torch_npu.get_npu_format(layer.weight.data)	253	"Convert weight to FRACTAL_NZ done, current format is %s", torch_npu.get_npu_format(layer.weight.data)
	@@ -324,8 +321,7 @@ class W8A8MixLinearMethod(LinearMethodBase):
324	layer.weight.data = layer.weight.data.transpose(0, 1).contiguous()	321	layer.weight.data = layer.weight.data.transpose(0, 1).contiguous()
325	layer.weight_scale.data = layer.weight_scale.data.flatten()	322	layer.weight_scale.data = layer.weight_scale.data.flatten()
326		323
327	- soc_name = get_npu_node_info().soc_name	324	+ if get_npu_node_info().is_nz_format_beneficial():
328	- if soc_name in SUPPORT_NZ_NPU_LIST:
329	layer.weight.data = torch_npu.npu_format_cast(layer.weight.data, 29)	325	layer.weight.data = torch_npu.npu_format_cast(layer.weight.data, 29)
330	logger.debug(	326	logger.debug(
331	"Convert weight to FRACTAL_NZ done, current format is %s", torch_npu.get_npu_format(layer.weight.data)	327	"Convert weight to FRACTAL_NZ done, current format is %s", torch_npu.get_npu_format(layer.weight.data)
	@@ -534,8 +530,7 @@ class W8A8PerTokenFusedMoEMethod(FusedMoEMethodBase):
534	"fused_down_weight_scale", ScalerParameter(scale_from_float_to_int64(layer.down_weight_scale.data))	530	"fused_down_weight_scale", ScalerParameter(scale_from_float_to_int64(layer.down_weight_scale.data))
535	)	531	)
536		532
537	- soc_name = get_npu_node_info().soc_name	533	+ if get_npu_node_info().is_nz_format_beneficial():
538	- if soc_name in SUPPORT_NZ_NPU_LIST:
539	layer.gate_up_weight.data = torch_npu.npu_format_cast(layer.gate_up_weight.data, 29)	534	layer.gate_up_weight.data = torch_npu.npu_format_cast(layer.gate_up_weight.data, 29)
540	logger.debug(	535	logger.debug(
541	"Convert weight to FRACTAL_NZ done, current format is %s",	536	"Convert weight to FRACTAL_NZ done, current format is %s",

Mmindie_llm/runtime/utils/npu/device_utils.py+13-0

	@@ -142,6 +142,15 @@ class _NPUNodeInfo:
142	"Ascend310P5",	142	"Ascend310P5",
143	"Ascend310P7",	143	"Ascend310P7",
144	}	144	}
		145	+
		146	+ self._nz_beneficial_soc_set = {
		147	+ "Ascend910B3",
		148	+ "Ascend910B4-1",
		149	+ "Ascend910_9362",
		150	+ "Ascend910_9382",
		151	+ "Ascend910_9392",
		152	+ }
		153	+
145	if self.soc_name in nz_only_names:	154	if self.soc_name in nz_only_names:
146	self.only_supports_nz = True	155	self.only_supports_nz = True
147	self.need_nz = True	156	self.need_nz = True
	@@ -192,6 +201,10 @@ class _NPUNodeInfo:
192	self._visible_device_ids = devices	201	self._visible_device_ids = devices
193	return self._visible_device_ids	202	return self._visible_device_ids
194		203
		204	+ def is_nz_format_beneficial(self) -> bool:
		205	+ """Return True if current soc_name benefits from NZ format conversion."""
		206	+ return self.soc_name in self._nz_beneficial_soc_set
		207	+
195	def get_device_info_map(self) -> dict[int, _DeviceInfo]:	208	def get_device_info_map(self) -> dict[int, _DeviceInfo]:
196	"""Retrieve a mapping of logical NPU chip IDs to their device information.	209	"""Retrieve a mapping of logical NPU chip IDs to their device information.
197		210