remove dsv2

Author: Difers

examples/experiments/deepseek_v3_pretrain/modeling.py

@@ -71,13 +71,13 @@
     StateDictNameMapping,
     init_name_mappings,
 )
-from paddleformers.transformers.deepseek_v2 import fp8_linear as linear_utils
-from paddleformers.transformers.deepseek_v2 import (
+from paddleformers.transformers.deepseek_v3 import fp8_linear as linear_utils
+from paddleformers.transformers.deepseek_v3 import (
     rotate_half,
     scaled_dot_product_attention,
     yarn_get_mscale,
 )
-from paddleformers.transformers.deepseek_v2.fp8_linear import Linear as Linear_
+from paddleformers.transformers.deepseek_v3.fp8_linear import Linear as Linear_
 from paddleformers.transformers.fp8_utils import (
     FP8KeepXLinear,
     FP8Linear,
@@ -1221,7 +1221,7 @@ class DeepseekV2PretrainedModelFast(PretrainedModel):
     _no_split_modules = ["DeepseekV2DecoderLayer"]
 
     def _get_model_flops(self, batch_size=1, seq_length=None, **kwargs):
-        from paddleformers.transformers.deepseek_v2.mfu_utils import DeepSeekProjection
+        from paddleformers.transformers.deepseek_v3.mfu_utils import DeepSeekProjection
 
         # self._
         mfu_cal_proj = DeepSeekProjection(self.config)

examples/experiments/deepseek_v3_pretrain/run_pretrain.py

@@ -48,7 +48,7 @@
     LinearAnnealingWithWarmupDecay,
 )
 from paddleformers.transformers.configuration_utils import LlmMetaConfig, llmmetaclass
-from paddleformers.transformers.deepseek_v2 import DeepseekV2ForCausalLM
+from paddleformers.transformers.deepseek_v3 import DeepseekV2ForCausalLM
 from paddleformers.utils.batch_sampler import DistributedBatchSampler
 from paddleformers.utils.log import logger
 

paddleformers/cli/train/pretrain/workflow.py

@@ -53,7 +53,7 @@
     LinearAnnealingWithWarmupDecay,
 )
 from paddleformers.transformers.configuration_utils import LlmMetaConfig, llmmetaclass
-from paddleformers.transformers.deepseek_v2 import DeepseekV2ForCausalLM
+from paddleformers.transformers.deepseek_v3 import DeepseekV3ForCausalLM
 from paddleformers.utils.batch_sampler import DistributedBatchSampler
 from paddleformers.utils.log import logger
 
@@ -485,7 +485,7 @@ def run_dsv3_pretrain(model_args, data_args, generating_args, training_args):
         if training_args.bf16:
             dtype = "bfloat16"
 
-    model_class = DeepseekV2ForCausalLM
+    model_class = DeepseekV3ForCausalLM
     if training_args.pipeline_parallel_degree > 1:
         model_class = DeepseekV2ForCausalLMPipe
         if "LLama" in str(config.architectures):

paddleformers/transformers/__init__.py

@@ -366,7 +366,6 @@
     from .auto.processing import *
     from .auto.tokenizer import *
     from .auto.video_processing import *
-    from .deepseek_v2 import *
     from .deepseek_v3 import *
     from .ernie4_5 import *
     from .ernie4_5_moe import *

paddleformers/transformers/auto/modeling.py

@@ -53,7 +53,6 @@
 MAPPING_NAMES = OrderedDict(
     [
         ("Bert", "bert"),
-        ("DeepseekV2", "deepseek_v2"),
         ("DeepseekV3", "deepseek_v3"),
         ("Ernie4_5", "ernie4_5"),
         ("Ernie4_5_Moe", "ernie4_5_moe"),