Questions about the amount of fine-tunable parameters #27
Description
Hello, author, first of all, thank you for your MoE-PEFT contact work. I encountered a problem when fine-tuning with your MoE-PEFT. The moe_peft.json configuration I used is as follows:
{
"cutoff_len": 2048,
"save_strategy": "epoch",
"train_lora_candidate_num": 2,
"train_lora_simultaneously_num": 2,
"train_strategy": "optim",
"lora": [
{
"name": "expert_0_name",
"task_name": "expert_0",
"optim": "adamw",
"scheduler_type": "constant",
"warmup_steps": 0,
"lr": 0.0002,
"batch_size": 1,
"micro_batch_size": 1,
"evaluate_batch_size": 1,
"num_epochs": 3,
"r": 4,
"lora_alpha": 16,
"lora_dropout": 0.05,
"target_modules": {
"q_proj": false,
"k_proj": false,
"v_proj": false,
"o_proj": false,
"gate_proj": true,
"down_proj": false,
"up_proj": false
},
"routing_strategy": "mixlora",
"num_experts": 7,
"top_k": 2,
"group_by_length": false
},
{
"name": "expert_1_name",
"task_name": "expert_1",
"optim": "adamw",
"scheduler_type": "constant",
"warmup_steps": 0,
"lr": 0.0002,
"batch_size": 1,
"micro_batch_size": 1,
"evaluate_batch_size": 1,
"num_epochs": 3,
"r": 4,
"lora_alpha": 16,
"lora_dropout": 0.05,
"target_modules": {
"q_proj": false,
"k_proj": false,
"v_proj": false,
"o_proj": false,
"gate_proj": true,
"down_proj": false,
"up_proj": false
},
"routing_strategy": "mixlora",
"num_experts": 7,
"top_k": 2,
"group_by_length": false
},
{
"name": "expert_2_name",
"task_name": "expert_2",
"optim": "adamw",
"scheduler_type": "constant",
"warmup_steps": 0,
"lr": 0.0002,
"batch_size": 1,
"micro_batch_size": 1,
"evaluate_batch_size": 1,
"num_epochs": 3,
"r": 4,
"lora_alpha": 16,
"lora_dropout": 0.05,
"target_modules": {
"q_proj": false,
"k_proj": false,
"v_proj": false,
"o_proj": false,
"gate_proj": true,
"down_proj": false,
"up_proj": false
},
"routing_strategy": "mixlora",
"num_experts": 7,
"top_k": 2,
"group_by_length": false
},
{
"name": "expert_3_name",
"task_name": "expert_3",
"optim": "adamw",
"scheduler_type": "constant",
"warmup_steps": 0,
"lr": 0.0002,
"batch_size": 1,
"micro_batch_size": 1,
"evaluate_batch_size": 1,
"num_epochs": 3,
"r": 4,
"lora_alpha": 16,
"lora_dropout": 0.05,
"target_modules": {
"q_proj": false,
"k_proj": false,
"v_proj": false,
"o_proj": false,
"gate_proj": true,
"down_proj": false,
"up_proj": false
},
"routing_strategy": "mixlora",
"num_experts": 7,
"top_k": 2,
"group_by_length": false
},
{
"name": "expert_4_name",
"task_name": "expert_4",
"optim": "adamw",
"scheduler_type": "constant",
"warmup_steps": 0,
"lr": 0.0002,
"batch_size": 1,
"micro_batch_size": 1,
"evaluate_batch_size": 1,
"num_epochs": 3,
"r": 4,
"lora_alpha": 16,
"lora_dropout": 0.05,
"target_modules": {
"q_proj": false,
"k_proj": false,
"v_proj": false,
"o_proj": false,
"gate_proj": true,
"down_proj": false,
"up_proj": false
},
"routing_strategy": "mixlora",
"num_experts": 7,
"top_k": 2,
"group_by_length": false
},
{
"name": "expert_5_name",
"task_name": "expert_5",
"optim": "adamw",
"scheduler_type": "constant",
"warmup_steps": 0,
"lr": 0.0002,
"batch_size": 1,
"micro_batch_size": 1,
"evaluate_batch_size": 1,
"num_epochs": 3,
"r": 4,
"lora_alpha": 16,
"lora_dropout": 0.05,
"target_modules": {
"q_proj": false,
"k_proj": false,
"v_proj": false,
"o_proj": false,
"gate_proj": true,
"down_proj": false,
"up_proj": false
},
"routing_strategy": "mixlora",
"num_experts": 7,
"top_k": 2,
"group_by_length": false
},
{
"name": "expert_6_name",
"task_name": "expert_6",
"optim": "adamw",
"scheduler_type": "constant",
"warmup_steps": 0,
"lr": 0.0002,
"batch_size": 1,
"micro_batch_size": 1,
"evaluate_batch_size": 1,
"num_epochs": 3,
"r": 4,
"lora_alpha": 16,
"lora_dropout": 0.05,
"target_modules": {
"q_proj": false,
"k_proj": false,
"v_proj": false,
"o_proj": false,
"gate_proj": true,
"down_proj": false,
"up_proj": false
},
"routing_strategy": "mixlora",
"num_experts": 7,
"top_k": 2,
"group_by_length": false
}
]
}
The model I chose is Meta-Llama-3-8B-Instruct, I printed the output of the model's trainable parameters as follows:
===== LoRA LAYER TYPE STATISTICS FOR expert_0_name =====
FFN: 17,432,576 parameters (100.00%)
Attention: 0 parameters (0.00%)
Other: 0 parameters (0.00%)
LoRA Total Parameters: 17,432,576
LoRA Rank: 4
LoRA Alpha: 16
Target Modules: ['gate_proj']
===== LoRA LAYER TYPE STATISTICS FOR expert_1_name =====
FFN: 17,432,576 parameters (100.00%)
Attention: 0 parameters (0.00%)
Other: 0 parameters (0.00%)
LoRA Total Parameters: 17,432,576
LoRA Rank: 4
LoRA Alpha: 16
Target Modules: ['gate_proj']
===== LoRA LAYER TYPE STATISTICS FOR expert_2_name =====
FFN: 17,432,576 parameters (100.00%)
Attention: 0 parameters (0.00%)
Other: 0 parameters (0.00%)
LoRA Total Parameters: 17,432,576
LoRA Rank: 4
LoRA Alpha: 16
Target Modules: ['gate_proj']
===== LoRA LAYER TYPE STATISTICS FOR expert_3_name =====
FFN: 17,432,576 parameters (100.00%)
Attention: 0 parameters (0.00%)
Other: 0 parameters (0.00%)
LoRA Total Parameters: 17,432,576
LoRA Rank: 4
LoRA Alpha: 16
Target Modules: ['gate_proj']
===== LoRA LAYER TYPE STATISTICS FOR expert_4_name =====
FFN: 17,432,576 parameters (100.00%)
Attention: 0 parameters (0.00%)
Other: 0 parameters (0.00%)
LoRA Total Parameters: 17,432,576
LoRA Rank: 4
LoRA Alpha: 16
Target Modules: ['gate_proj']
===== LoRA LAYER TYPE STATISTICS FOR expert_5_name =====
FFN: 17,432,576 parameters (100.00%)
Attention: 0 parameters (0.00%)
Other: 0 parameters (0.00%)
LoRA Total Parameters: 17,432,576
LoRA Rank: 4
LoRA Alpha: 16
Target Modules: ['gate_proj']
===== LoRA LAYER TYPE STATISTICS FOR expert_6_name =====
FFN: 17,432,576 parameters (100.00%)
Attention: 0 parameters (0.00%)
Other: 0 parameters (0.00%)
LoRA Total Parameters: 17,432,576
LoRA Rank: 4
LoRA Alpha: 16
Target Modules: ['gate_proj']
================================
Your code prints like this:
[2025-03-25 00:39:18,638] MoE-PEFT: expert_0_name total trainable params: 17432576
[2025-03-25 00:39:18,639] MoE-PEFT: expert_0_name total trainable params (except gates): 16515072
[2025-03-25 00:39:18,647] MoE-PEFT: expert_1_name total trainable params: 17432576
[2025-03-25 00:39:18,649] MoE-PEFT: expert_1_name total trainable params (except gates): 16515072
[2025-03-25 00:39:18,656] MoE-PEFT: expert_2_name total trainable params: 17432576
[2025-03-25 00:39:18,658] MoE-PEFT: expert_2_name total trainable params (except gates): 16515072
[2025-03-25 00:39:18,663] MoE-PEFT: expert_3_name total trainable params: 17432576
[2025-03-25 00:39:18,665] MoE-PEFT: expert_3_name total trainable params (except gates): 16515072
[2025-03-25 00:39:18,670] MoE-PEFT: expert_4_name total trainable params: 17432576
[2025-03-25 00:39:18,672] MoE-PEFT: expert_4_name total trainable params (except gates): 16515072
[2025-03-25 00:39:18,678] MoE-PEFT: expert_5_name total trainable params: 17432576
[2025-03-25 00:39:18,679] MoE-PEFT: expert_5_name total trainable params (except gates): 16515072
[2025-03-25 00:39:18,686] MoE-PEFT: expert_6_name total trainable params: 17432576
[2025-03-25 00:39:18,688] MoE-PEFT: expert_6_name total trainable params (except gates): 16515072
But the code printed in your trainer.py seems to print the total trainable parameters for each expert：
Under the above parameters, my model still exceeds the video memory on a single 24GB 4090, which seems a bit abnormal. Is there any good solution?
Looking forward to your answer and thank you！