vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

mpt : standardized all tensor names to follow GGUF spec

Browse source
This commit is contained in:
Jan Ploski 2023-10-01 01:43:39 +02:00
parent 1be89c4002
commit 26c253eda2
3 changed files with 19 additions and 20 deletions
Download patch file Download diff file Expand all files Collapse all files

2
convert-mpt-hf-to-gguf.py
View file

@ -247,7 +247,7 @@ for part_name in part_names:
# note: MPT output is tied to (same as) wte in original model; # note: MPT output is tied to (same as) wte in original model;
# for easier implementation in llama.cpp it's duplicated in GGUF, though :/ # for easier implementation in llama.cpp it's duplicated in GGUF, though :/
if new_name == "wte.weight": if new_name == "token_embd.weight":
gguf_writer.add_tensor("output.weight", data) gguf_writer.add_tensor("output.weight", data)
print("gguf: write header") print("gguf: write header")

16
gguf-py/gguf/gguf.py
View file

@ -186,17 +186,17 @@ MODEL_TENSOR_NAMES: dict[MODEL_ARCH, dict[MODEL_TENSOR, str]] = {
MODEL_TENSOR.FFN_UP: "blk.{bid}.ffn_up", MODEL_TENSOR.FFN_UP: "blk.{bid}.ffn_up",
}, },
MODEL_ARCH.MPT: { MODEL_ARCH.MPT: {
MODEL_TENSOR.TOKEN_EMBD: "wte", MODEL_TENSOR.TOKEN_EMBD: "token_embd",
MODEL_TENSOR.OUTPUT_NORM: "norm_f", MODEL_TENSOR.OUTPUT_NORM: "output_norm",
# note: MPT output is tied to (same as) wte in original model; # note: MPT output is tied to (same as) wte in original model;
# for easier implementation in llama.cpp it's duplicated in GGUF, though :/ # for easier implementation in llama.cpp it's duplicated in GGUF, though :/
MODEL_TENSOR.OUTPUT: "output", MODEL_TENSOR.OUTPUT: "output",
MODEL_TENSOR.ATTN_NORM: "blk.{bid}.norm_1", MODEL_TENSOR.ATTN_NORM: "blk.{bid}.attn_norm",
MODEL_TENSOR.FFN_NORM: "blk.{bid}.norm_2", MODEL_TENSOR.FFN_NORM: "blk.{bid}.ffn_norm",
MODEL_TENSOR.ATTN_QKV: "blk.{bid}.attn.Wqkv", MODEL_TENSOR.ATTN_QKV: "blk.{bid}.attn_qkv",
MODEL_TENSOR.ATTN_OUT: "blk.{bid}.attn.out_proj", MODEL_TENSOR.ATTN_OUT: "blk.{bid}.attn_output",
MODEL_TENSOR.FFN_DOWN: "blk.{bid}.ffn.down_proj", MODEL_TENSOR.FFN_DOWN: "blk.{bid}.ffn_down",
MODEL_TENSOR.FFN_UP: "blk.{bid}.ffn.up_proj", MODEL_TENSOR.FFN_UP: "blk.{bid}.ffn_up",
}, },
MODEL_ARCH.GPT2: { MODEL_ARCH.GPT2: {
# TODO # TODO

21
llama.cpp
View file

@ -377,15 +377,15 @@ static std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NAMES =
{ {
LLM_ARCH_MPT, LLM_ARCH_MPT,
{ {
{ LLM_TENSOR_TOKEN_EMBD, "wte" }, { LLM_TENSOR_TOKEN_EMBD, "token_embd" },
{ LLM_TENSOR_OUTPUT_NORM, "norm_f" }, { LLM_TENSOR_OUTPUT_NORM, "output_norm" },
{ LLM_TENSOR_OUTPUT, "output" }, { LLM_TENSOR_OUTPUT, "output" },
{ LLM_TENSOR_ATTN_NORM, "blk.%d.norm_1" }, { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" },
{ LLM_TENSOR_FFN_NORM, "blk.%d.norm_2" }, { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" },
{ LLM_TENSOR_ATTN_QKV, "blk.%d.attn.Wqkv" }, { LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" },
{ LLM_TENSOR_ATTN_OUT, "blk.%d.attn.out_proj" }, { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" },
{ LLM_TENSOR_FFN_DOWN, "blk.%d.ffn.down_proj" }, { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" },
{ LLM_TENSOR_FFN_UP, "blk.%d.ffn.up_proj" }, { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
}, },
}, },
{ {
@ -6516,11 +6516,10 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
// TODO: avoid hardcoded tensor names - use the TN_* constants // TODO: avoid hardcoded tensor names - use the TN_* constants
if (name.find("attn_v.weight") != std::string::npos || if (name.find("attn_v.weight") != std::string::npos ||
name.find("attn.Wqkv.weight") != std::string::npos) { name.find("attn_qkv.weight") != std::string::npos) {
++n_attention_wv; ++n_attention_wv;
} }
else if (name.find("ffn_down.weight") != std::string::npos || else if (name.find("ffn_down.weight") != std::string::npos) {
name.find("ffn.down_proj.weight") != std::string::npos) {
++n_feed_forward_w2; ++n_feed_forward_w2;
} }
} }