[LLM] Add pipeline and flashmask for Qwen2Moe and Deepseek (PaddlePaddle#9827)

* add modleing_pp

* add modleing_pp for qwen2moe

* add flashmask and pp for Qwen2MoE and Deepseek

* remove

* fix fast_tokenizer save

* update for topk_weight of noaux_tc

* fix for flashmask

* add use_expert_parallel for pretrain

* fix tokenizer test

Author: DrownFish19

llm/run_finetune.py

@@ -52,12 +52,18 @@
     AutoModelForCausalLM,
     AutoModelForCausalLMPipe,
     AutoTokenizer,
+    DeepseekV2ForCausalLM,
+    DeepseekV2ForCausalLMPipe,
+    DeepseekV3ForCausalLM,
+    DeepseekV3ForCausalLMPipe,
     Llama3Tokenizer,
     LlamaForCausalLM,
     LlamaForCausalLMPipe,
     LlamaTokenizer,
     Qwen2ForCausalLM,
     Qwen2ForCausalLMPipe,
+    Qwen2MoeForCausalLM,
+    Qwen2MoeForCausalLMPipe,
 )
 from paddlenlp.transformers.configuration_utils import LlmMetaConfig
 from paddlenlp.trl import DataConfig, ModelConfig, SFTConfig, SFTTrainer
@@ -74,7 +80,18 @@
 # Fine-tune Environment Variables to support sharding stage1 overlap optimization.
 os.environ["USE_CASUAL_MASK"] = "False"
 
-flash_mask_support_list = [LlamaForCausalLM, LlamaForCausalLMPipe, Qwen2ForCausalLM, Qwen2ForCausalLMPipe]
+flash_mask_support_list = [
+    DeepseekV2ForCausalLM,
+    DeepseekV2ForCausalLMPipe,
+    DeepseekV3ForCausalLM,
+    DeepseekV3ForCausalLMPipe,
+    LlamaForCausalLM,
+    LlamaForCausalLMPipe,
+    Qwen2ForCausalLM,
+    Qwen2ForCausalLMPipe,
+    Qwen2MoeForCausalLM,
+    Qwen2MoeForCausalLMPipe,
+]
 
 
 def paddlenlp_verison_check():
@@ -151,7 +168,11 @@ def main():
         quantization_config=quantization_config,
     )
 
-    if "Qwen2Moe" in str(model_config.architectures) and training_args.data_parallel_degree > 1:
+    architectures_to_check = {"Qwen2Moe", "DeepseekV2", "DeepseekV3"}
+    if (
+        any(architecture in str(model_config.architectures) for architecture in architectures_to_check)
+        and training_args.data_parallel_degree > 1
+    ):
         training_args.use_expert_parallel = True
 
     LlmMetaConfig.set_llm_config(model_config, training_args)
@@ -585,7 +606,12 @@ def create_peft_model(model_args, reft_args, training_args, dtype, model_config,
 def trans_dataset_to_ids(train_ds, dev_ds, test_ds, model_args, data_args, trans_func, eval_zero_padding):
     if train_ds is not None:
         train_ds = train_ds.map(
-            partial(trans_func, is_test=False, zero_padding=data_args.zero_padding, flash_mask=model_args.flash_mask)
+            partial(
+                trans_func,
+                is_test=False,
+                zero_padding=data_args.zero_padding,
+                flash_mask=model_args.flash_mask,
+            )
         )
     if dev_ds is not None:
         dev_ds = dev_ds.map(

llm/run_pretrain.py

@@ -478,7 +478,11 @@ def main():
             except:
                 print("Not register llama pp reshard information.")
 
-    if "Qwen2Moe" in str(config.architectures) and training_args.data_parallel_degree > 1:
+    architectures_to_check = {"Qwen2Moe", "DeepseekV2", "DeepseekV3"}
+    if (
+        any(architecture in str(config.architectures) for architecture in architectures_to_check)
+        and training_args.data_parallel_degree > 1
+    ):
         training_args.use_expert_parallel = True
 
     if model_args.continue_training:

llm/utils/data.py

@@ -59,11 +59,13 @@ def get_convert_example(model):
         "gpt",
         "yuan",
         "jamba",
+        "deepseek_v2",
+        "deepseek_v3",
     ]:
         return convert_example_common
     else:
         raise ValueError(
-            f"Unknown base_model_prefix: {model.base_model_prefix}. Supported base_model_prefix list: chatglm, bloom, llama, qwen, mixtral, gemma, qwen2, qwen2_moe, yuan, jamba",
+            f"Unknown base_model_prefix: {model.base_model_prefix}. Supported base_model_prefix list: chatglm, bloom, llama, qwen, mixtral, gemma, qwen2, qwen2_moe, yuan, jamba,deepseek_v2, deepseek_v3",
         )
 
 

paddlenlp/transformers/__init__.py

@@ -143,11 +143,8 @@
 from .deberta_v2.configuration import *
 from .deberta_v2.modeling import *
 from .deberta_v2.tokenizer import *
-from .deepseek_v2.configuration import *
-from .deepseek_v2.modeling import *
-from .deepseek_v2.tokenizer_fast import *
-from .deepseek_v3.configuration import *
-from .deepseek_v3.modeling import *
+from .deepseek_v2 import *
+from .deepseek_v3 import *
 from .distilbert.configuration import *
 from .distilbert.modeling import *
 from .distilbert.tokenizer import *

paddlenlp/transformers/deepseek_v2/__init__.py

@@ -14,4 +14,5 @@
 
 from .configuration import *
 from .modeling import *
+from .modeling_pp import *
 from .tokenizer_fast import *