add single_model network and use intermediate api

llm/auto_parallel/gpt-3/gpt_with_intermediate.sh

@@ -0,0 +1,113 @@
+# Copyright (c) 2024 PaddlePaddle Authors. All Rights Reserved.
+# PaddlePaddle version 3.0-beta2 or higher is required, please upgrade your PaddlePaddle to 3.0-beta2 or other higher version.
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+# 
+#     http://www.apache.org/licenses/LICENSE-2.0
+# 
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+unset PADDLE_ELASTIC_JOB_ID
+unset PADDLE_TRAINER_ENDPOINTS
+unset DISTRIBUTED_TRAINER_ENDPOINTS
+unset FLAGS_START_PORT
+unset PADDLE_ELASTIC_TIMEOUT
+
+export NNODES=1
+export PADDLE_TRAINERS_NUM=1
+export FLAGS_benchmark=false
+
+export GLOG_v=0
+export FLAGS_print_ir=0
+#this optional is for compare train precison
+export FLAGS_cudnn_deterministic=1
+export FLAGS_embedding_deterministic=1 
+# export NVIDIA_TF32_OVERRIDE=0
+
+export FLAGS_call_stack_level=3
+export FLAGS_enable_pir_api=1
+export PYTHONPATH=../../../:$PYTHONPATH
+
+set -x
+unset CUDA_VISIBLE_DEVICES
+
+task_name="gpt3_13b_atuo_perf"
+log_dir="log/$task_name"
+rm -rf $log_dir
+to_static=1
+# export PYTHONPATH=../../../:$PYTHONPATH
+
+python -u -m paddle.distributed.launch \
+    --gpus "0,1,2,3" \
+    --log_dir ${log_dir} \
+    run_pretrain_auto.py \
+    --model_name_or_path gpt3-13B-en \
+    --tokenizer_name_or_path gpt3-13B-en \
+    --to_static ${to_static} \
+    --enable_auto_parallel 1 \
+    --input_dir "../gpt_data" \
+    --output_dir "output/$task_name" \
+    --split 949,50,1 \
+    --max_seq_length 4096 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --scale_loss 1024 \
+    --learning_rate 0.00001 \
+    --min_learning_rate 0.000001 \
+    --max_steps 10000 \
+    --save_steps 50000 \
+    --weight_decay 0.01 \
+    --warmup_ratio 0.01 \
+    --logging_steps 1\
+    --continue_training 0\
+    --dataloader_num_workers 4 \
+    --eval_steps 100000 \
+    --report_to "visualdl" \
+    --disable_tqdm true \
+    --do_train \
+    --do_eval \
+    --device "gpu" \
+    --model_type "gpt_network" \
+    --use_intermediate_api true \
+    --sharding "stage1" \
+    --tensor_parallel_degree 1 \
+    --pipeline_parallel_degree 2 \
+    --virtual_pp_degree 2 \
+    --pipeline_schedule_mode "1F1B" \
+    --virtual_pipeline_seg_method 'GPTDecoderLayerAuto' \
+    --sequence_parallel 0 \
+    --use_flash_attention 1 \
+    --fused_linear 1 \
+    --fuse_attention_ffn 1 \
+    --fuse_attention_qkv 1 \
+    --fused_linear_param_grad_add 1 \
+    --recompute 0 \
+    --recompute_use_reentrant true \
+    --recompute_granularity "full" \
+    --pp_recompute_interval 1 \
+    --gradient_accumulation_steps 32 \
+    --max_grad_norm 0 \
+    --bf16 1 \
+    --fp16_opt_level "O2"  \
+    --amp_master_grad true \
+    --attention_probs_dropout_prob 0.1 \
+    --hidden_dropout_prob 0.1 \
+    --sharding_parallel_config "enable_stage1_tensor_fusion enable_stage1_overlap" \
+    --tensor_parallel_config "enable_mp_async_allreduce" \
+    --data_parallel_config "enable_allreduce_avg_in_gradinent_scale gradient_sync_after_accumulate" \
+    --pipeline_parallel_config "enable_send_recv_overlap enable_split_backward" \
+    --num_hidden_layers 2 \
+    # --auto_parallel_resume_form_hybrid_parallel true \
+    # --resume_from_checkpoint "output/gpt3_13b_hand_perf/checkpoint-1" \
+    # --fused_linear 1 \
+    # --use_fast_layer_norm 1 \
+    # --use_fused_dropout_add 1 \
+    # --pipeline_parallel_config "enable_send_recv_overlap enable_split_backward" \
+
+

llm/auto_parallel/gpt-3/run_pretrain_auto.py

@@ -38,13 +38,16 @@
     CosineAnnealingWithWarmupDecay,
     GPTConfig,
     GPTForCausalLMAuto,
+    GPTForCausalLMNet,
     GPTPretrainingCriterionAuto,
+    GPTPretrainingCriterionNet,
     LinearAnnealingWithWarmupDecay,
 )
 from paddlenlp.utils.log import logger
 
 MODEL_CLASSES = {
     "gpt": (GPTConfig, GPTForCausalLMAuto, GPTPretrainingCriterionAuto),
+    "gpt_network": (GPTConfig, GPTForCausalLMNet, GPTPretrainingCriterionNet),
 }
 
 from paddlenlp.data.causal_dataset import (
@@ -91,6 +94,10 @@ class PreTrainingArguments(AutoTrainingArguments):
         default=False,
         metadata={"help": "Weather to run benchmark by autotuner. True for from_scratch and pad_max_length."},
     )
+    use_intermediate_api: bool = field(
+        default=False,
+        metadata={"help": "Weather to use auto_parallel intermediate api"},
+    )
 
     def __post_init__(self):
         super().__post_init__()
@@ -529,9 +536,9 @@ def main():
             dtype = "float16"
         if training_args.bf16:
             dtype = "bfloat16"
-
-    model = model_class.from_config(config, dtype=dtype)
-    criterion = criterion_class(config)
+    with paddle.LazyGuard():
+        model = model_class.from_config(config, dtype=dtype)
+        criterion = criterion_class(config)
     if training_args.recompute:
 
         def fn(layer):

llm/auto_parallel/llama/llama_with_api.sh

@@ -0,0 +1,86 @@
+# Copyright (c) 2024 PaddlePaddle Authors. All Rights Reserved.
+#PaddlePaddle version 3.0-beta2 or higher is required, please upgrade your PaddlePaddle to 3.0-beta2 or other higher version.
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+set -x
+
+export FLAGS_cudnn_deterministic=1
+export FLAGS_embedding_deterministic=1
+export FLAGS_max_inplace_grad_add=65536
+
+task_name="llama_auto"
+rm -rf output
+rm -rf single
+
+export PYTHONPATH=../../../:$PYTHONPATH
+
+python -u  -m paddle.distributed.launch \
+    --gpus "0,1,2,3,4,5,6,7" \
+    --log_dir  "single" \
+    ./run_pretrain_auto.py \
+    --model_name_or_path "facebook/llama-7b" \
+    --tokenizer_name_or_path "facebook/llama-7b" \
+    --input_dir "./data" \
+    --output_dir "./output" \
+    --split 949,50,1 \
+    --to_static false \
+    --seed 1234 \
+    --pipeline_parallel_degree 2 \
+    --sharding_parallel_degree 2 \
+    --tensor_parallel_degree 2 \
+    --virtual_pp_degree 1 \
+    --weight_decay 0.01 \
+    --warmup_ratio 0.01 \
+    --max_grad_norm 1.0 \
+    --learning_rate 3e-05 \
+    --min_learning_rate 3e-06 \
+    --max_steps 2000 \
+    --logging_steps 1 \
+    --eval_steps 10000 \
+    --save_steps 2000 \
+    --auto_parallel_resume_form_hybrid_parallel true \
+    --continue_training 0 \
+    --do_train true \
+    --do_eval false \
+    --do_predict false \
+    --disable_tqdm true \
+    --save_total_limit 2 \
+    --device gpu \
+    --dataloader_num_workers 4 \
+    --distributed_dataloader 0 \
+    --enable_auto_parallel 1 \
+    --per_device_train_batch_size 1 \
+    --gradient_accumulation_steps 8 \
+    --per_device_eval_batch_size 1 \
+    --recompute false \
+    --recompute_use_reentrant true \
+    --skip_profile_timer true \
+    --recompute_granularity full \
+    --pp_recompute_interval 0 \
+    --bf16 true \
+    --fp16_opt_level "O2"  \
+    --amp_master_grad true \
+    --fuse_attention_ffn false \
+    --fuse_attention_qkv true \
+    --use_flash_attention true \
+    --use_fused_rope true \
+    --use_fused_rms_norm true \
+    --max_seq_length 4096 \
+    --sequence_parallel true \
+    --sharding "stage1" \
+    --sharding_parallel_config "enable_stage1_tensor_fusion enable_stage1_overlap" \
+    --tensor_parallel_config "enable_mp_async_allreduce" \
+    --model_type "llama_network" \
+    --ignore_load_lr_and_optim true \
+    --ignore_save_lr_and_optim true \
+    --use_intermediate_api true \

llm/auto_parallel/llama/run_pretrain_auto.py

@@ -41,12 +41,15 @@
     LinearAnnealingWithWarmupDecay,
     LlamaConfig,
     LlamaForCausalLM3DAuto,
+    LlamaForCausalLMNet,
     LlamaPretrainingCriterion3DAuto,
+    LlamaPretrainingCriterionNet,
 )
 from paddlenlp.utils.log import logger
 
 MODEL_CLASSES = {
     "llama": (LlamaConfig, LlamaForCausalLM3DAuto, LlamaPretrainingCriterion3DAuto),
+    "llama_network": (LlamaConfig, LlamaForCausalLMNet, LlamaPretrainingCriterionNet),
 }
 
 
@@ -100,6 +103,10 @@ class PreTrainingArguments(AutoTrainingArguments):
         default=False,
         metadata={"help": "Weather to run benchmark by autotuner. True for from_scratch and pad_max_length."},
     )
+    use_intermediate_api: bool = field(
+        default=False,
+        metadata={"help": "Weather to use auto_parallel intermediate api"},
+    )
 
     def __post_init__(self):
         super().__post_init__()
@@ -544,6 +551,7 @@ def main():
     config.use_recompute = training_args.recompute
     config.tensor_parallel_degree = training_args.tensor_parallel_degree
     config.tensor_parallel_rank = training_args.tensor_parallel_rank
+    config.sharding_parallel_degree = training_args.sharding_parallel_degree
 
     if training_args.strategy.pipeline.enable and config.virtual_pp_degree > 1:
         pipeline = training_args.strategy.pipeline
@@ -564,10 +572,6 @@ def main():
         model = model_class.from_config(config, dtype="float32")
         criterion = criterion_class(config)
 
-    for param in model.parameters():
-        assert not param._is_initialized()
-        param.initialize()
-
     if training_args.recompute:
 
         def fn(layer):
@@ -621,6 +625,7 @@ def fn(layer):
         eval_dataset=eval_dataset if training_args.do_eval else None,
         optimizers=(None, lr_scheduler),
         tokenizer=tokenizer,
+        model_args=model_args,
     )
 
     checkpoint = None

llm/auto_parallel/qwen/run_intermediate_api.sh

@@ -0,0 +1,99 @@
+# Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved.
+# PaddlePaddle version 3.0-beta2 or higher is required, please upgrade your PaddlePaddle to 3.0-beta2 or other higher version.
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+# 
+#     http://www.apache.org/licenses/LICENSE-2.0
+# 
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# just for debug
+
+set -x
+unset PADDLE_ELASTIC_JOB_ID
+unset PADDLE_TRAINER_ENDPOINTS
+unset DISTRIBUTED_TRAINER_ENDPOINTS
+unset FLAGS_START_PORT
+unset PADDLE_ELASTIC_TIMEOUT
+
+export NNODES=1
+export PADDLE_TRAINERS_NUM=1
+export FLAGS_call_stack_level=3
+export FLAGS_use_cuda_managed_memory=true
+
+task_name="llama_auto"
+rm -rf output/$task_name/
+rm -rf "output/$task_name""_log"
+
+export SOT_LOG_LEVEL=4
+export PYTHONPATH=../../../:$PYTHONPATH
+
+
+rm -rf ./log/auto_3d_auto
+
+export FLAGS_embedding_deterministic=1        
+export FLAGS_cudnn_deterministic=1
+export FLAGS_max_inplace_grad_add=65536
+export NVIDIA_TF32_OVERRIDE=0
+export FLAGS_enable_pir_in_executor=1
+export FLAGS_enable_pir_api=1
+
+
+python -u  -m paddle.distributed.launch \
+    --gpus "4,5" \
+    --log_dir "log/auto_3d_auto" \
+    run_pretrain_3D_auto.py \
+    --model_name_or_path "qwen/qwen-14b" \
+    --tokenizer_name_or_path "qwen/qwen-14b" \
+    --model_type "qwen_network" \
+    --use_intermediate_api true \
+    --input_dir "../data" \
+    --output_dir "./checkpoints/qwen_pretrain_ckpts" \
+    --per_device_train_batch_size 1\
+    --gradient_accumulation_steps 32\
+    --per_device_eval_batch_size 16\
+    --sharding "stage1" \
+    --sharding_parallel_degree 1\
+    --tensor_parallel_degree 2\
+    --pipeline_parallel_degree 1\
+    --virtual_pp_degree 1\
+    --use_flash_attention false\
+    --use_fused_rms_norm false\
+    --use_fused_rope false\
+    --max_seq_length 4096\
+    --learning_rate 3e-05\
+    --min_learning_rate 3e-06\
+    --scale_loss 1024\
+    --warmup_steps 30\
+    --logging_steps 1\
+    --max_steps 10000\
+    --save_steps 1000\
+    --eval_steps 10000\
+    --weight_decay 0.01\
+    --bf16 true\
+    --fp16_opt_level "O2"\
+    --amp_master_grad true \
+    --warmup_ratio 0.01\
+    --max_grad_norm 0.0\
+    --dataloader_num_workers 4\
+    --continue_training 0\
+    --do_train true\
+    --do_eval false\
+    --do_predict false \
+    --disable_tqdm true\
+    --recompute false\
+    --recompute_granularity "core_attn"\
+    --recompute_use_reentrant true\
+    --distributed_dataloader 0\
+    --save_total_limit 2\
+    --enable_auto_parallel 1\
+    --to_static 1 \
+    --num_hidden_layers 1 \
+    --attention_probs_dropout_prob 0 \
+    --hidden_dropout_prob 0 \
+    --auto_parallel_resume_form_hybrid_parallel true \