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

added support for gpt2

Browse source
This commit is contained in:
manikbhandari 2023-12-16 09:17:45 -05:00
parent 29e3645501
commit f5c0184bf3
4 changed files with 278 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

66
convert-hf-to-gguf.py
View file

@ -13,6 +13,7 @@ from pathlib import Path
from typing import TYPE_CHECKING, Any, ContextManager, Iterator, cast, Optional
import numpy as np
import pdb
import torch
if TYPE_CHECKING:
@ -182,6 +183,8 @@ class Model:
return QwenModel
if model_architecture == "MixtralForCausalLM":
return MixtralModel
if model_architecture == "GPT2LMHeadModel":
return GPT2Model
return Model
def _is_model_safetensors(self) -> bool:
@ -221,6 +224,8 @@ class Model:
return gguf.MODEL_ARCH.QWEN
if arch == "MixtralForCausalLM":
return gguf.MODEL_ARCH.LLAMA
if arch == "GPT2LMHeadModel":
return gguf.MODEL_ARCH.GPT2
raise NotImplementedError(f'Architecture "{arch}" not supported!')
@ -980,6 +985,67 @@ class QwenModel(Model):
print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
self.gguf_writer.add_tensor(new_name, data)
class GPT2Model(Model):
def set_gguf_parameters(self):
self.gguf_writer.add_name(self.dir_model.name)
self.gguf_writer.add_block_count(self.hparams["n_layer"])
self.gguf_writer.add_context_length(self.hparams["n_ctx"])
self.gguf_writer.add_embedding_length(self.hparams["n_embd"])
self.gguf_writer.add_feed_forward_length(4 * self.hparams["n_embd"])
self.gguf_writer.add_head_count(self.hparams["n_head"])
self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_epsilon"])
self.gguf_writer.add_file_type(self.ftype)
def write_tensors(self):
block_count = self.hparams.get("n_layers", self.hparams.get("num_hidden_layers", self.hparams.get("n_layer")))
tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
# import pdb; pdb.set_trace()
for name, data_torch in self.get_tensors():
# we don't need these
if name.endswith((".attention.masked_bias", ".attention.bias", ".attention.rotary_emb.inv_freq", ".attn.bias")):
continue
if name.endswith((".c_attn.weight", ".c_proj.weight", ".c_fc.weight", ".c_proj.weight")):
data_torch = data_torch.transpose(1, 0)
old_dtype = data_torch.dtype
# convert any unsupported data types to float32
if data_torch.dtype not in (torch.float16, torch.float32):
data_torch = data_torch.to(torch.float32)
data = data_torch.squeeze().numpy()
# map tensor names
new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
if new_name is None:
print(f"Can not map tensor {name!r}")
sys.exit()
n_dims = len(data.shape)
data_dtype = data.dtype
# if f32 desired, convert any float16 to float32
if self.ftype == 0 and data_dtype == np.float16:
data = data.astype(np.float32)
# TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
data = data.astype(np.float32)
# if f16 desired, convert any float32 2-dim weight tensors to float16
if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
data = data.astype(np.float16)
print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
self.gguf_writer.add_tensor(new_name, data)
# note: GPT2 output is tied to (same as) wte in original model
if new_name == "token_embd.weight":
print(f"output.weight, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
self.gguf_writer.add_tensor("output.weight", data)
###### CONVERSION LOGIC ######

11
gguf-py/gguf/constants.py
View file

@ -348,7 +348,16 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.FFN_UP,
],
MODEL_ARCH.GPT2: [
# TODO
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.POS_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
MODEL_TENSOR.OUTPUT,
MODEL_TENSOR.ATTN_NORM,
MODEL_TENSOR.ATTN_QKV,
MODEL_TENSOR.ATTN_OUT,
MODEL_TENSOR.FFN_NORM,
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP,
],
# TODO
}

10
gguf-py/gguf/tensor_mapping.py
View file

@ -17,6 +17,7 @@ class TensorNameMap:
"tok_embeddings", # llama-pth
"embeddings.word_embeddings", # bert
"language_model.embedding.word_embeddings", # persimmon
"wte", # gpt2
),
# Token type embeddings
@ -33,6 +34,7 @@ class TensorNameMap:
MODEL_TENSOR.POS_EMBD: (
"transformer.wpe", # gpt2
"embeddings.position_embeddings", # bert
"wpe", # gpt2
),
# Output
@ -51,7 +53,7 @@ class TensorNameMap:
"norm", # llama-pth
"embeddings.LayerNorm", # bert
"transformer.norm_f", # mpt
"ln_f", # refact bloom qwen
"ln_f", # refact bloom qwen gpt2
"language_model.encoder.final_layernorm", # persimmon
),
@ -75,6 +77,7 @@ class TensorNameMap:
"encoder.layer.{bid}.attention.output.LayerNorm", # bert
"language_model.encoder.layers.{bid}.input_layernorm", # persimmon
"model.layers.{bid}.ln1", # yi
"h.{bid}.ln_1", # gpt2
),
# Attention norm 2
@ -90,6 +93,7 @@ class TensorNameMap:
"transformer.h.{bid}.self_attention.query_key_value", # falcon
"h.{bid}.self_attention.query_key_value", # bloom
"language_model.encoder.layers.{bid}.self_attention.query_key_value", # persimmon
"h.{bid}.attn.c_attn", # gpt2
),
# Attention query
@ -128,6 +132,7 @@ class TensorNameMap:
"encoder.layer.{bid}.attention.output.dense", # bert
"transformer.h.{bid}.attn.out_proj", # gpt-j
"language_model.encoder.layers.{bid}.self_attention.dense", # persimmon
"h.{bid}.attn.c_proj", # gpt2
),
# Rotary embeddings
@ -147,6 +152,7 @@ class TensorNameMap:
"encoder.layer.{bid}.output.LayerNorm", # bert
"language_model.encoder.layers.{bid}.post_attention_layernorm", # persimmon
"model.layers.{bid}.ln2", # yi
"h.{bid}.ln_2", # gpt2
),
MODEL_TENSOR.FFN_GATE_INP: (
@ -167,6 +173,7 @@ class TensorNameMap:
"transformer.h.{bid}.mlp.fc_in", # gpt-j
"language_model.encoder.layers.{bid}.mlp.dense_h_to_4h", # persimmon
"transformer.h.{bid}.mlp.w1", # qwen
"h.{bid}.mlp.c_fc", # gpt2
),
MODEL_TENSOR.FFN_UP_EXP: (
@ -198,6 +205,7 @@ class TensorNameMap:
"encoder.layer.{bid}.output.dense", # bert
"transformer.h.{bid}.mlp.fc_out", # gpt-j
"language_model.encoder.layers.{bid}.mlp.dense_4h_to_h", # persimmon
"h.{bid}.mlp.c_proj", # gpt2
),
MODEL_TENSOR.FFN_DOWN_EXP: (

193
llama.cpp
View file

@ -416,6 +416,15 @@ static std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NAMES =
LLM_ARCH_GPT2,
{
{ LLM_TENSOR_TOKEN_EMBD, "token_embd" },
{ LLM_TENSOR_POS_EMBD, "position_embd" },
{ LLM_TENSOR_OUTPUT_NORM, "output_norm" },
{ LLM_TENSOR_OUTPUT, "output" },
{ LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" },
{ LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" },
{ LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" },
{ LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" },
{ LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
{ LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" },
},
},
{
@ -1165,6 +1174,7 @@ enum e_model {
MODEL_40B,
MODEL_65B,
MODEL_70B,
MODEL_SMALL,
};
static const size_t kiB = 1024;
@ -2430,6 +2440,7 @@ static const char * llama_model_type_name(e_model type) {
case MODEL_40B: return "40B";
case MODEL_65B: return "65B";
case MODEL_70B: return "70B";
case MODEL_SMALL: return "0.1B";
default: return "?B";
}
}
@ -2630,6 +2641,14 @@ static void llm_load_hparams(
default: model.type = e_model::MODEL_UNKNOWN;
}
} break;
case LLM_ARCH_GPT2:
{
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
switch (hparams.n_layer) {
case 12: model.type = e_model::MODEL_SMALL; break;
default: model.type = e_model::MODEL_UNKNOWN;
}
} break;
default: (void)0;
}
@ -3625,6 +3644,78 @@ static void llm_load_tensors(
}
}
} break;
case LLM_ARCH_GPT2:
{
model.tok_embd = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU);
model.pos_embd = ml.create_tensor(ctx, tn(LLM_TENSOR_POS_EMBD, "weight"), {n_embd, hparams.n_ctx_train}, GGML_BACKEND_CPU);
// output
{
ggml_backend_type backend_norm;
ggml_backend_type backend_output;
if (n_gpu_layers > int(n_layer)) {
backend_norm = llama_backend_offload;
backend_output = llama_backend_offload_split;
} else {
backend_norm = GGML_BACKEND_CPU;
backend_output = GGML_BACKEND_CPU;
}
model.output_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, backend_norm);
model.output_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}, backend_norm);
model.output = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, backend_output);
if (backend_norm == GGML_BACKEND_GPU) {
vram_weights += ggml_nbytes(model.output_norm);
vram_weights += ggml_nbytes(model.output_norm_b);
}
if (backend_output == GGML_BACKEND_GPU_SPLIT) {
vram_weights += ggml_nbytes(model.output);
}
}
const uint32_t n_ff = hparams.n_ff;
const int i_gpu_start = n_layer - n_gpu_layers;
model.layers.resize(n_layer);
for (uint32_t i = 0; i < n_layer; ++i) {
const ggml_backend_type backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : llama_backend_offload; // NOLINT
const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : llama_backend_offload_split; // NOLINT
auto & layer = model.layers[i];
layer.attn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, backend);
layer.attn_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "bias", i), {n_embd}, backend);
layer.wqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, backend_split);
layer.bqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd + 2*n_embd_gqa}, backend);
layer.wo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, backend_split);
layer.bo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, backend);
layer.ffn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, backend);
layer.ffn_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "bias", i), {n_embd}, backend);
layer.ffn_down = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, backend_split);
layer.ffn_down_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, backend);
layer.ffn_up = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split);
layer.ffn_up_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff}, backend);
if (backend == GGML_BACKEND_GPU) {
vram_weights +=
ggml_nbytes(layer.attn_norm) + ggml_nbytes(layer.attn_norm_b) +
ggml_nbytes(layer.wqkv) + ggml_nbytes(layer.bqkv) +
ggml_nbytes(layer.wo) + ggml_nbytes(layer.bo) +
ggml_nbytes(layer.ffn_norm) + ggml_nbytes(layer.ffn_norm_b) +
ggml_nbytes(layer.ffn_down) + ggml_nbytes(layer.ffn_down_b) +
ggml_nbytes(layer.ffn_up) + ggml_nbytes(layer.ffn_up_b);
}
}
} break;
default:
throw std::runtime_error("unknown architecture");
@ -5417,6 +5508,104 @@ struct llm_build_context {
ggml_build_forward_expand(gf, cur);
return gf;
}
struct ggml_cgraph * build_gpt2() {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
struct ggml_tensor * cur;
struct ggml_tensor * pos;
struct ggml_tensor * inpL;
inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb);
cb(inpL, "inp_embd", -1);
// inp_pos - contains the positions
struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);
cb(inp_pos, "inp_pos", -1);
// KQ_scale
struct ggml_tensor * KQ_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1);
cb(KQ_scale, "KQ_scale", -1);
// KQ_mask (mask for 1 head, it will be broadcasted to all heads)
struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1);
cb(KQ_mask, "KQ_mask", -1);
pos = ggml_get_rows(ctx0, model.pos_embd, inp_pos);
cb(pos, "pos_embd", -1);
inpL = ggml_add(ctx0, inpL, pos);
cb(inpL, "inpL", -1);
for (int il = 0; il < n_layer; ++il) {
cur = llm_build_norm(ctx0, inpL, hparams,
model.layers[il].attn_norm,
model.layers[il].attn_norm_b,
LLM_NORM, cb, il);
cb(cur, "attn_norm", il);
// self-attention
{
cur = ggml_mul_mat(ctx0, model.layers[il].wqkv, cur);
cb(cur, "wqkv", il);
cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
cb(cur, "bqkv", il);
struct ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)));
struct ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)));
struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
cb(Qcur, "Qcur", il);
cb(Kcur, "Kcur", il);
cb(Vcur, "Vcur", il);
Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il);
cur = llm_build_kqv(ctx0, hparams, kv_self,
model.layers[il].wo, model.layers[il].bo,
Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, cb, il);
cb(cur, "kqv_out", il);
}
// add the input
struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
cb(ffn_inp, "ffn_inp", il);
// FF
{
cur = llm_build_norm(ctx0, ffn_inp, hparams,
model.layers[il].ffn_norm,
model.layers[il].ffn_norm_b,
LLM_NORM, cb, il);
cb(cur, "ffn_norm", il);
cur = llm_build_ffn(ctx0, cur,
model.layers[il].ffn_up, model.layers[il].ffn_up_b,
NULL, NULL,
model.layers[il].ffn_down, model.layers[il].ffn_down_b,
LLM_FFN_GELU, LLM_FFN_SEQ, cb, il);
cb(cur, "ffn_out", il);
}
inpL = ggml_add(ctx0, cur, ffn_inp);
cb(inpL, "l_out", il);
}
cur = llm_build_norm(ctx0, inpL, hparams,
model.output_norm,
model.output_norm_b,
LLM_NORM, cb, -1);
cb(cur, "result_norm", -1);
cur = ggml_mul_mat(ctx0, model.output, cur);
cb(cur, "result_output", -1);
ggml_build_forward_expand(gf, cur);
return gf;
}
};
@ -5917,6 +6106,10 @@ static struct ggml_cgraph * llama_build_graph(
{
result = llm.build_qwen();
} break;
case LLM_ARCH_GPT2:
{
result = llm.build_gpt2();
} break;
default:
GGML_ASSERT(false);
}