Fintune part of a whole embeding parameters.

[CongHan0808]

I add 100 new tokens to the vocabulary and corresponding embedings. I try to only fintune these embeddings and fix raw tokens with pretrained weights. I follow #4192, then use safe_get_full_grad and safe_set_full_fp32_param to get and modify the grad of the parameter. But all weights of the parameters are updated
There are my code:

model_engine.backward(total_loss)
textembeds_masks = torch.zeros_like(model_engine.in_adaptor.text_embed.weight).to(device=model_engine.local_rank)
textembeds_masks[VOCAB_SIZE_SRC+1,:] = 1
with torch.no_grad():
    for p_name,param in model_engine.named_parameters():
        if "in_adaptor.text_embed.weight" in p_name:
            if param.grad is not None:
           
                hp_grad = safe_get_full_grad(param)
                exp_avg = safe_get_full_optimizer_state(param, "exp_avg")
                exp_avg_sq = safe_get_full_optimizer_state(param, "exp_avg_sq")
                # hp_grad.copy_(hp_grad.data*textembeds_masks)
                
                safe_set_full_fp32_param(param, hp_grad.data*textembeds_masks)
                safe_set_full_optimizer_state(param, exp_avg.data * textembeds_masks, "exp_avg")
                safe_set_full_optimizer_state(param, exp_avg_sq.data * textembeds_masks, "exp_avg_sq")
model_engine.step()
scheduler(step)

After some checkpoints, the raw tokens' weights of in_adaptor.text_embed.weight in different cks are different. How should I change my code to keep the raw tokens' weights the same and only fintune the new tokens' weights.

[AuroraZengfh]

hi, I'm also searching for solutions to finetuning part of embeddings in deepspeed framework. Did you get a convenient way to solve the problem?