Add safe tensor support to convert-llama.py

DifferentialityDevelopment commented 1 month ago

I haven't yet updated the other model conversion scripts yet, but this allows you to convert any llama model that uses safetensor.

b4rtaz commented 1 month ago

Please update also docs/LLAMA.md.

DifferentialityDevelopment commented 1 month ago

Please update also docs/LLAMA.md.

I updated the usage a bit, though could probably mention that it would work with the hugging face repo for Llama as well.

b4rtaz commented 1 month ago

@DifferentialityDevelopment I'm wondering about this part:

        with safetensors.safe_open(model_file, framework="pt") as f:
            for layer in f.keys():
                layers.append({
                    "name" : layer,
                    "file" : model_file
                })

Are you sure that the source model has all layers in the correct order that is expected by Distributed Llama?

DifferentialityDevelopment commented 1 month ago

@DifferentialityDevelopment I'm wondering about this part:
        with safetensors.safe_open(model_file, framework="pt") as f:
            for layer in f.keys():
                layers.append({
                    "name" : layer,
                    "file" : model_file
                })
Are you sure that the source model has all layers in the correct order that is expected by Distributed Llama?

DId not check yet, will do a full convert on llama-3 8B Instruct, do a test with distributed llama and report back.

DifferentialityDevelopment commented 1 month ago

The convert process itself does seem to work fine, but will test once it finishes

python converter/convert-llama.py J:\Llama-3\Meta-Llama-3-8B-Instruct J:\Llama-3\Meta-Llama-3-8B-Instruct-Distributed q40 Model name: Meta-Llama-3-8B-Instruct Target float type: q40 Target file: dllama_meta-llama-3-8b-instruct_q40.bin Total layers: 291 Total chunks: 7 Unknown header key: head_size {'head_size': 128.0, 'n_layers': 32, 'n_heads': 32, 'n_kv_heads': 8, 'max_seq_len': 8192, 'rope_theta': 500000, 'arch_type': 11259136, 'n_experts': 0, 'n_active_experts': 0} 💿 Chunking model 1/7... Loading tensors for model.embed_tokens.weight from: model-00001-of-00004.safetensors 🔶 Exporting model.embed_tokens.weight torch.Size([128256, 4096])... Saved q40 tensor in 123.95s, 295501824 bytes Loading tensors for model.layers.0.input_layernorm.weight from: model-00001-of-00004.safetensors 🔶 Exporting model.layers.0.input_layernorm.weight torch.Size([4096])... Saved q40 tensor in 0.00s, 2304 bytes Loading tensors for model.layers.0.mlp.down_proj.weight from: model-00001-of-00004.safetensors 🔶 Exporting model.layers.0.mlp.down_proj.weight torch.Size([4096, 14336])... Saved q40 tensor in 14.69s, 33030144 bytes Loading tensors for model.layers.0.mlp.gate_proj.weight from: model-00001-of-00004.safetensors 🔶 Exporting model.layers.0.mlp.gate_proj.weight torch.Size([14336, 4096])... Saved q40 tensor in 14.96s, 33030144 bytes Loading tensors for model.layers.0.mlp.up_proj.weight from: model-00001-of-00004.safetensors 🔶 Exporting model.layers.0.mlp.up_proj.weight torch.Size([14336, 4096])... Saved q40 tensor in 14.95s, 33030144 bytes Loading tensors for model.layers.0.post_attention_layernorm.weight from: model-00001-of-00004.safetensors 🔶 Exporting model.layers.0.post_attention_layernorm.weight torch.Size([4096])... Saved q40 tensor in 0.00s, 2304 bytes Loading tensors for model.layers.0.self_attn.k_proj.weight from: model-00001-of-00004.safetensors 🔶 Exporting model.layers.0.self_attn.k_proj.weight torch.Size([1024, 4096])... Saved q40 tensor in 1.08s, 2359296 bytes Loading tensors for model.layers.0.self_attn.o_proj.weight from: model-00001-of-00004.safetensors 🔶 Exporting model.layers.0.self_attn.o_proj.weight torch.Size([4096, 4096])... Saved q40 tensor in 4.37s, 9437184 bytes Loading tensors for model.layers.0.self_attn.q_proj.weight from: model-00001-of-00004.safetensors 🔶 Exporting model.layers.0.self_attn.q_proj.weight torch.Size([4096, 4096])... Saved q40 tensor in 4.27s, 9437184 bytes Loading tensors for model.layers.0.self_attn.v_proj.weight from: model-00001-of-00004.safetensors 🔶 Exporting model.layers.0.self_attn.v_proj.weight torch.Size([1024, 4096])... Saved q40 tensor in 1.05s, 2359296 bytes Loading tensors for model.layers.1.input_layernorm.weight from: model-00001-of-00004.safetensors 🔶 Exporting model.layers.1.input_layernorm.weight torch.Size([4096])... Saved q40 tensor in 0.00s, 2304 bytes Loading tensors for model.layers.1.mlp.down_proj.weight from: model-00001-of-00004.safetensors 🔶 Exporting model.layers.1.mlp.down_proj.weight torch.Size([4096, 14336])... Saved q40 tensor in 14.91s, 33030144 bytes Loading tensors for model.layers.1.mlp.gate_proj.weight from: model-00001-of-00004.safetensors 🔶 Exporting model.layers.1.mlp.gate_proj.weight torch.Size([14336, 4096])... Saved q40 tensor in 14.76s, 33030144 bytes

b4rtaz commented 1 month ago

Please consider also that, some models may have a different layers order by some reason.

DifferentialityDevelopment commented 1 month ago

Please consider also that, some models may have a different layers order by some reason.

I would think the order of the keys when loading a .safetensor model is the same as from the .pth file but I could be wrong, will do a bit of research.

DifferentialityDevelopment commented 1 month ago

Your absolutely right, the layers are not necessarily in the right order, see below output of their keys, I noticed that layer 9 only appears after layer 20. So I will need to fix the ordering. I'm not entirely sure where to place lm_head.weight and model.norm.weight, they appear near the end of the list. The other thing I have trouble with is that I'm not sure which of the layers is the feed_forward layer? Which is what you use in the pth conversion to get the hidden_dim size

Additionally they have a different name convention, so I had to change a few more things. Is this correct: [safetensor] model.embed_tokens.weight -> [pth] tok_embeddings.weight [safetensor] model.layers.0.mlp.gate_proj.weight -> [pth] layers.0.feed_forward.w1.weight [safetensor] model.layers.0.mlp.up_proj.weight -> [pth] layers.0.feed_forward.w2.weight [safetensor] model.layers.0.post_attention_layernorm.weight -> [pth] layers.0.attention_norm.weight [safetensor] model.norm.weight -> [pth] norm.weight

Keys:

model.embed_tokens.weight => 128256 model.layers.0.input_layernorm.weight => 4096 model.layers.0.mlp.down_proj.weight => 4096 model.layers.0.mlp.gate_proj.weight => 14336 model.layers.0.mlp.up_proj.weight => 14336 model.layers.0.post_attention_layernorm.weight => 4096 model.layers.0.self_attn.k_proj.weight => 1024 model.layers.0.self_attn.o_proj.weight => 4096 model.layers.0.self_attn.q_proj.weight => 4096 model.layers.0.self_attn.v_proj.weight => 1024 model.layers.1.input_layernorm.weight => 4096 model.layers.1.mlp.down_proj.weight => 4096 model.layers.1.mlp.gate_proj.weight => 14336 model.layers.1.mlp.up_proj.weight => 14336 model.layers.1.post_attention_layernorm.weight => 4096 model.layers.1.self_attn.k_proj.weight => 1024 model.layers.1.self_attn.o_proj.weight => 4096 model.layers.1.self_attn.q_proj.weight => 4096 model.layers.1.self_attn.v_proj.weight => 1024 model.layers.2.input_layernorm.weight => 4096 model.layers.2.mlp.down_proj.weight => 4096 model.layers.2.mlp.gate_proj.weight => 14336 model.layers.2.mlp.up_proj.weight => 14336 model.layers.2.post_attention_layernorm.weight => 4096 model.layers.2.self_attn.k_proj.weight => 1024 model.layers.2.self_attn.o_proj.weight => 4096 model.layers.2.self_attn.q_proj.weight => 4096 model.layers.2.self_attn.v_proj.weight => 1024 model.layers.3.input_layernorm.weight => 4096 model.layers.3.mlp.down_proj.weight => 4096 model.layers.3.mlp.gate_proj.weight => 14336 model.layers.3.mlp.up_proj.weight => 14336 model.layers.3.post_attention_layernorm.weight => 4096 model.layers.3.self_attn.k_proj.weight => 1024 model.layers.3.self_attn.o_proj.weight => 4096 model.layers.3.self_attn.q_proj.weight => 4096 model.layers.3.self_attn.v_proj.weight => 1024 model.layers.4.input_layernorm.weight => 4096 model.layers.4.mlp.down_proj.weight => 4096 model.layers.4.mlp.gate_proj.weight => 14336 model.layers.4.mlp.up_proj.weight => 14336 model.layers.4.post_attention_layernorm.weight => 4096 model.layers.4.self_attn.k_proj.weight => 1024 model.layers.4.self_attn.o_proj.weight => 4096 model.layers.4.self_attn.q_proj.weight => 4096 model.layers.4.self_attn.v_proj.weight => 1024 model.layers.5.input_layernorm.weight => 4096 model.layers.5.mlp.down_proj.weight => 4096 model.layers.5.mlp.gate_proj.weight => 14336 model.layers.5.mlp.up_proj.weight => 14336 model.layers.5.post_attention_layernorm.weight => 4096 model.layers.5.self_attn.k_proj.weight => 1024 model.layers.5.self_attn.o_proj.weight => 4096 model.layers.5.self_attn.q_proj.weight => 4096 model.layers.5.self_attn.v_proj.weight => 1024 model.layers.6.input_layernorm.weight => 4096 model.layers.6.mlp.down_proj.weight => 4096 model.layers.6.mlp.gate_proj.weight => 14336 model.layers.6.mlp.up_proj.weight => 14336 model.layers.6.post_attention_layernorm.weight => 4096 model.layers.6.self_attn.k_proj.weight => 1024 model.layers.6.self_attn.o_proj.weight => 4096 model.layers.6.self_attn.q_proj.weight => 4096 model.layers.6.self_attn.v_proj.weight => 1024 model.layers.7.input_layernorm.weight => 4096 model.layers.7.mlp.down_proj.weight => 4096 model.layers.7.mlp.gate_proj.weight => 14336 model.layers.7.mlp.up_proj.weight => 14336 model.layers.7.post_attention_layernorm.weight => 4096 model.layers.7.self_attn.k_proj.weight => 1024 model.layers.7.self_attn.o_proj.weight => 4096 model.layers.7.self_attn.q_proj.weight => 4096 model.layers.7.self_attn.v_proj.weight => 1024 model.layers.8.input_layernorm.weight => 4096 model.layers.8.mlp.down_proj.weight => 4096 model.layers.8.mlp.gate_proj.weight => 14336 model.layers.8.mlp.up_proj.weight => 14336 model.layers.8.post_attention_layernorm.weight => 4096 model.layers.8.self_attn.k_proj.weight => 1024 model.layers.8.self_attn.o_proj.weight => 4096 model.layers.8.self_attn.q_proj.weight => 4096 model.layers.8.self_attn.v_proj.weight => 1024 model.layers.10.input_layernorm.weight => 4096 model.layers.10.mlp.down_proj.weight => 4096 model.layers.10.mlp.gate_proj.weight => 14336 model.layers.10.mlp.up_proj.weight => 14336 model.layers.10.post_attention_layernorm.weight => 4096 model.layers.10.self_attn.k_proj.weight => 1024 model.layers.10.self_attn.o_proj.weight => 4096 model.layers.10.self_attn.q_proj.weight => 4096 model.layers.10.self_attn.v_proj.weight => 1024 model.layers.11.input_layernorm.weight => 4096 model.layers.11.mlp.down_proj.weight => 4096 model.layers.11.mlp.gate_proj.weight => 14336 model.layers.11.mlp.up_proj.weight => 14336 model.layers.11.post_attention_layernorm.weight => 4096 model.layers.11.self_attn.k_proj.weight => 1024 model.layers.11.self_attn.o_proj.weight => 4096 model.layers.11.self_attn.q_proj.weight => 4096 model.layers.11.self_attn.v_proj.weight => 1024 model.layers.12.input_layernorm.weight => 4096 model.layers.12.mlp.down_proj.weight => 4096 model.layers.12.mlp.gate_proj.weight => 14336 model.layers.12.mlp.up_proj.weight => 14336 model.layers.12.post_attention_layernorm.weight => 4096 model.layers.12.self_attn.k_proj.weight => 1024 model.layers.12.self_attn.o_proj.weight => 4096 model.layers.12.self_attn.q_proj.weight => 4096 model.layers.12.self_attn.v_proj.weight => 1024 model.layers.13.input_layernorm.weight => 4096 model.layers.13.mlp.down_proj.weight => 4096 model.layers.13.mlp.gate_proj.weight => 14336 model.layers.13.mlp.up_proj.weight => 14336 model.layers.13.post_attention_layernorm.weight => 4096 model.layers.13.self_attn.k_proj.weight => 1024 model.layers.13.self_attn.o_proj.weight => 4096 model.layers.13.self_attn.q_proj.weight => 4096 model.layers.13.self_attn.v_proj.weight => 1024 model.layers.14.input_layernorm.weight => 4096 model.layers.14.mlp.down_proj.weight => 4096 model.layers.14.mlp.gate_proj.weight => 14336 model.layers.14.mlp.up_proj.weight => 14336 model.layers.14.post_attention_layernorm.weight => 4096 model.layers.14.self_attn.k_proj.weight => 1024 model.layers.14.self_attn.o_proj.weight => 4096 model.layers.14.self_attn.q_proj.weight => 4096 model.layers.14.self_attn.v_proj.weight => 1024 model.layers.15.input_layernorm.weight => 4096 model.layers.15.mlp.down_proj.weight => 4096 model.layers.15.mlp.gate_proj.weight => 14336 model.layers.15.mlp.up_proj.weight => 14336 model.layers.15.post_attention_layernorm.weight => 4096 model.layers.15.self_attn.k_proj.weight => 1024 model.layers.15.self_attn.o_proj.weight => 4096 model.layers.15.self_attn.q_proj.weight => 4096 model.layers.15.self_attn.v_proj.weight => 1024 model.layers.16.input_layernorm.weight => 4096 model.layers.16.mlp.down_proj.weight => 4096 model.layers.16.mlp.gate_proj.weight => 14336 model.layers.16.mlp.up_proj.weight => 14336 model.layers.16.post_attention_layernorm.weight => 4096 model.layers.16.self_attn.k_proj.weight => 1024 model.layers.16.self_attn.o_proj.weight => 4096 model.layers.16.self_attn.q_proj.weight => 4096 model.layers.16.self_attn.v_proj.weight => 1024 model.layers.17.input_layernorm.weight => 4096 model.layers.17.mlp.down_proj.weight => 4096 model.layers.17.mlp.gate_proj.weight => 14336 model.layers.17.mlp.up_proj.weight => 14336 model.layers.17.post_attention_layernorm.weight => 4096 model.layers.17.self_attn.k_proj.weight => 1024 model.layers.17.self_attn.o_proj.weight => 4096 model.layers.17.self_attn.q_proj.weight => 4096 model.layers.17.self_attn.v_proj.weight => 1024 model.layers.18.input_layernorm.weight => 4096 model.layers.18.mlp.down_proj.weight => 4096 model.layers.18.mlp.gate_proj.weight => 14336 model.layers.18.mlp.up_proj.weight => 14336 model.layers.18.post_attention_layernorm.weight => 4096 model.layers.18.self_attn.k_proj.weight => 1024 model.layers.18.self_attn.o_proj.weight => 4096 model.layers.18.self_attn.q_proj.weight => 4096 model.layers.18.self_attn.v_proj.weight => 1024 model.layers.19.input_layernorm.weight => 4096 model.layers.19.mlp.down_proj.weight => 4096 model.layers.19.mlp.gate_proj.weight => 14336 model.layers.19.mlp.up_proj.weight => 14336 model.layers.19.post_attention_layernorm.weight => 4096 model.layers.19.self_attn.k_proj.weight => 1024 model.layers.19.self_attn.o_proj.weight => 4096 model.layers.19.self_attn.q_proj.weight => 4096 model.layers.19.self_attn.v_proj.weight => 1024 model.layers.20.mlp.gate_proj.weight => 14336 model.layers.20.self_attn.k_proj.weight => 1024 model.layers.20.self_attn.o_proj.weight => 4096 model.layers.20.self_attn.q_proj.weight => 4096 model.layers.20.self_attn.v_proj.weight => 1024 model.layers.9.input_layernorm.weight => 4096 model.layers.9.mlp.down_proj.weight => 4096 model.layers.9.mlp.gate_proj.weight => 14336 model.layers.9.mlp.up_proj.weight => 14336 model.layers.9.post_attention_layernorm.weight => 4096 model.layers.9.self_attn.k_proj.weight => 1024 model.layers.9.self_attn.o_proj.weight => 4096 model.layers.9.self_attn.q_proj.weight => 4096 model.layers.9.self_attn.v_proj.weight => 1024 model.layers.20.input_layernorm.weight => 4096 model.layers.20.mlp.down_proj.weight => 4096 model.layers.20.mlp.up_proj.weight => 14336 model.layers.20.post_attention_layernorm.weight => 4096 model.layers.21.input_layernorm.weight => 4096 model.layers.21.mlp.down_proj.weight => 4096 model.layers.21.mlp.gate_proj.weight => 14336 model.layers.21.mlp.up_proj.weight => 14336 model.layers.21.post_attention_layernorm.weight => 4096 model.layers.21.self_attn.k_proj.weight => 1024 model.layers.21.self_attn.o_proj.weight => 4096 model.layers.21.self_attn.q_proj.weight => 4096 model.layers.21.self_attn.v_proj.weight => 1024 model.layers.22.input_layernorm.weight => 4096 model.layers.22.mlp.down_proj.weight => 4096 model.layers.22.mlp.gate_proj.weight => 14336 model.layers.22.mlp.up_proj.weight => 14336 model.layers.22.post_attention_layernorm.weight => 4096 model.layers.22.self_attn.k_proj.weight => 1024 model.layers.22.self_attn.o_proj.weight => 4096 model.layers.22.self_attn.q_proj.weight => 4096 model.layers.22.self_attn.v_proj.weight => 1024 model.layers.23.input_layernorm.weight => 4096 model.layers.23.mlp.down_proj.weight => 4096 model.layers.23.mlp.gate_proj.weight => 14336 model.layers.23.mlp.up_proj.weight => 14336 model.layers.23.post_attention_layernorm.weight => 4096 model.layers.23.self_attn.k_proj.weight => 1024 model.layers.23.self_attn.o_proj.weight => 4096 model.layers.23.self_attn.q_proj.weight => 4096 model.layers.23.self_attn.v_proj.weight => 1024 model.layers.24.input_layernorm.weight => 4096 model.layers.24.mlp.down_proj.weight => 4096 model.layers.24.mlp.gate_proj.weight => 14336 model.layers.24.mlp.up_proj.weight => 14336 model.layers.24.post_attention_layernorm.weight => 4096 model.layers.24.self_attn.k_proj.weight => 1024 model.layers.24.self_attn.o_proj.weight => 4096 model.layers.24.self_attn.q_proj.weight => 4096 model.layers.24.self_attn.v_proj.weight => 1024 model.layers.25.input_layernorm.weight => 4096 model.layers.25.mlp.down_proj.weight => 4096 model.layers.25.mlp.gate_proj.weight => 14336 model.layers.25.mlp.up_proj.weight => 14336 model.layers.25.post_attention_layernorm.weight => 4096 model.layers.25.self_attn.k_proj.weight => 1024 model.layers.25.self_attn.o_proj.weight => 4096 model.layers.25.self_attn.q_proj.weight => 4096 model.layers.25.self_attn.v_proj.weight => 1024 model.layers.26.input_layernorm.weight => 4096 model.layers.26.mlp.down_proj.weight => 4096 model.layers.26.mlp.gate_proj.weight => 14336 model.layers.26.mlp.up_proj.weight => 14336 model.layers.26.post_attention_layernorm.weight => 4096 model.layers.26.self_attn.k_proj.weight => 1024 model.layers.26.self_attn.o_proj.weight => 4096 model.layers.26.self_attn.q_proj.weight => 4096 model.layers.26.self_attn.v_proj.weight => 1024 model.layers.27.input_layernorm.weight => 4096 model.layers.27.mlp.down_proj.weight => 4096 model.layers.27.mlp.gate_proj.weight => 14336 model.layers.27.mlp.up_proj.weight => 14336 model.layers.27.post_attention_layernorm.weight => 4096 model.layers.27.self_attn.k_proj.weight => 1024 model.layers.27.self_attn.o_proj.weight => 4096 model.layers.27.self_attn.q_proj.weight => 4096 model.layers.27.self_attn.v_proj.weight => 1024 model.layers.28.input_layernorm.weight => 4096 model.layers.28.mlp.down_proj.weight => 4096 model.layers.28.mlp.gate_proj.weight => 14336 model.layers.28.mlp.up_proj.weight => 14336 model.layers.28.post_attention_layernorm.weight => 4096 model.layers.28.self_attn.k_proj.weight => 1024 model.layers.28.self_attn.o_proj.weight => 4096 model.layers.28.self_attn.q_proj.weight => 4096 model.layers.28.self_attn.v_proj.weight => 1024 model.layers.29.input_layernorm.weight => 4096 model.layers.29.mlp.down_proj.weight => 4096 model.layers.29.mlp.gate_proj.weight => 14336 model.layers.29.mlp.up_proj.weight => 14336 model.layers.29.post_attention_layernorm.weight => 4096 model.layers.29.self_attn.k_proj.weight => 1024 model.layers.29.self_attn.o_proj.weight => 4096 model.layers.29.self_attn.q_proj.weight => 4096 model.layers.29.self_attn.v_proj.weight => 1024 model.layers.30.input_layernorm.weight => 4096 model.layers.30.mlp.down_proj.weight => 4096 model.layers.30.mlp.gate_proj.weight => 14336 model.layers.30.mlp.up_proj.weight => 14336 model.layers.30.post_attention_layernorm.weight => 4096 model.layers.30.self_attn.k_proj.weight => 1024 model.layers.30.self_attn.o_proj.weight => 4096 model.layers.30.self_attn.q_proj.weight => 4096 model.layers.30.self_attn.v_proj.weight => 1024 model.layers.31.mlp.gate_proj.weight => 14336 model.layers.31.mlp.up_proj.weight => 14336 model.layers.31.self_attn.k_proj.weight => 1024 model.layers.31.self_attn.o_proj.weight => 4096 model.layers.31.self_attn.q_proj.weight => 4096 model.layers.31.self_attn.v_proj.weight => 1024 lm_head.weight => 128256 model.layers.31.input_layernorm.weight => 4096 model.layers.31.mlp.down_proj.weight => 4096 model.layers.31.post_attention_layernorm.weight => 4096 model.norm.weight => 4096

b4rtaz commented 1 month ago

I reccomend to use the same appraoch as you can see in the convert_pth method. You should build a list with layer names, then you need to pass it to the loop. BTW: this loop could be extracted from these two functions.

b4rtaz commented 1 month ago

@DifferentialityDevelopment I'm closing this pull request. The convert-hf.py script introduced in the 0.7.0 version supports the safe tensor format and 3 model types.

b4rtaz / distributed-llama

Add safe tensor support to convert-llama.py #52