Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

feat: add Qwen 1.8B demo #100

Merged
merged 3 commits into from
Jul 21, 2024
Merged

feat: add Qwen 1.8B demo #100

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
1184a3d
fix: add qwen 1.8B
yirongjie Jul 20, 2024
da8909a
feat:matmul
yirongjie Jul 21, 2024
00f29c5
fix: need_setup
yirongjie Jul 21, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion src/Module.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -145,7 +145,7 @@ class Module {
if(input.batch() == 0){
Tensor::gph_[input.name()] = input;
}
if(input.sequence()!=1 && !last_shape_bshd_.empty()){
if(inputs[0].sequence()!=1 && !last_shape_bshd_.empty()){
// if LLM/VLLM model, the `need_setup` should be `true`
if(input.batch() == last_shape_bshd_[i][0] &
input.sequence() == last_shape_bshd_[i][1] &
Expand Down
88 changes: 32 additions & 56 deletions src/backends/cpu/compute/Matmul.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -176,16 +176,13 @@ ErrorCode mat_mul_fp32(Tensor *src0, Tensor *src1, Tensor *dst, bool support_bia
Tensor *src0_cal = src0;
Tensor *src1_cal = src1;
const int64_t blck_0 = 16;
int is_0 = (src1->batch() == 1 && src1->head() == 1) ? 0 : 1;
#pragma omp parallel for collapse(4) num_threads(thread_count)
for (int b = 0; b < src0->batch(); b++) {
for (int h = 0; h < src0->head(); h++) {
const int b_1 = (src1->batch() == 1 && src1->head() == 1) ? 0 : b;
const int h_1 = (src1->batch() == 1 && src1->head() == 1) ? 0 : h;
for (int m = 0; m < M; m++) {
const int num_blocks = N / blck_0;
const int remainder = N % blck_0;
#pragma omp parallel for num_threads(thread_count)
for (int block = 0; block < num_blocks + 1; block++) {
for (int n = block * blck_0; n < (block + 1) * blck_0 & n < num_blocks * blck_0 + remainder; n++) {
for (int block = 0; block < N / blck_0 + 1; block++) {
for (int n = block * blck_0; n < (block + 1) * blck_0 & n < N; n++) {
int s_1, d_1;
int s_0, d_0;
if (!transpose0 && transpose1) {
Expand All @@ -197,15 +194,15 @@ ErrorCode mat_mul_fp32(Tensor *src0, Tensor *src1, Tensor *dst, bool support_bia
}
if(dst->dtypeAt(b,h,m,n) == MLLM_TYPE_F32) {
vec_dot_fp32(K, dst->ptrAt<float>(b, h, m, n),
src1_cal->hostPtr<float>() + src1_cal->offset(b_1, h_1, s_1, d_1),
src1_cal->hostPtr<float>() + src1_cal->offset(b*is_0, h*is_0, s_1, d_1),
src0_cal->hostPtr<float>() + src0_cal->offset(b, h, s_0, d_0));
if (support_bias) {
*dst->ptrAt<float>(b, h, m, n) += bias->dataAt<float>(0, 0, 0, n);
}
}else if (dst->dtypeAt(b,h,m,n) == MLLM_TYPE_F16) {
float tmp = 0;
vec_dot_fp32(K, &tmp,
src1_cal->hostPtr<float>() + src1_cal->offset(b_1, h_1, s_1, d_1),
src1_cal->hostPtr<float>() + src1_cal->offset(b*is_0, h*is_0, s_1, d_1),
src0_cal->hostPtr<float>() + src0_cal->offset(b, h, s_0, d_0));
if (support_bias) {
*dst->ptrAt<mllm_fp16_t>(b, h, m, n) = MLLM_FP32_TO_FP16(tmp + bias->dataAt<float>(0, 0, 0, n));
Expand All @@ -228,9 +225,9 @@ ErrorCode mat_mul_fp32_fp16(Tensor *src0_, Tensor *src1, Tensor *dst, bool suppo
src0_qf16.setBackend(src0_->backend());
src0_qf16.setDtype(MLLM_TYPE_F16);
src0_qf16.alloc();
#pragma omp parallel for collapse(3) num_threads(thread_count)
for (int b = 0; b < src0_->batch(); b++) {
for (int h = 0; h < src0_->head(); h++) {
#pragma omp parallel for num_threads(thread_count)
for (int s = 0; s < src0_->sequence(); s++) {
mllm_fp32_to_fp16_row(src0_->hostPtr<float>() + src0_->offset(b, h, s, 0),
src0_qf16.hostPtr<mllm_fp16_t>() + src0_qf16.offset(b, h, s, 0),
Expand All @@ -239,30 +236,19 @@ ErrorCode mat_mul_fp32_fp16(Tensor *src0_, Tensor *src1, Tensor *dst, bool suppo
}
}
auto *src0 = &src0_qf16;
// for(int b=0; b<src0->dimension(); b++) {
// std::cout<<MLLM_COMPUTE_FP16_TO_FP32(*src0->ptrAt<mllm_fp16_t>(0, 0, 0, b))<<" ";
// }
// std::cout<<std::endl;
// for(int b=0; b<src1->dimension(); b++) {
// std::cout<<MLLM_COMPUTE_FP16_TO_FP32(*src1->ptrAt<mllm_fp16_t>(0, 0, 0, b))<<" ";
// }
// std::cout<<std::endl;
const int M = transpose0 ? src0->dimension() : src0->sequence();
const int K = transpose0 ? src0->sequence() : src0->dimension();
const int N = transpose1 ? src1->sequence() : src1->dimension();
Tensor *src0_cal = src0;
Tensor *src1_cal = src1;
const int64_t blck_0 = 16;
int is_0 = (src1->batch() == 1 && src1->head() == 1) ? 0 : 1;
#pragma omp parallel for collapse(4) num_threads(thread_count)
for (int b = 0; b < src0->batch(); b++) {
for (int h = 0; h < src0->head(); h++) {
const int b_1 = (src1->batch() == 1 && src1->head() == 1) ? 0 : b;
const int h_1 = (src1->batch() == 1 && src1->head() == 1) ? 0 : h;
for (int m = 0; m < M; m++) {
const int num_blocks = N / blck_0;
const int remainder = N % blck_0;
#pragma omp parallel for num_threads(thread_count)
for (int block = 0; block < num_blocks + 1; block++) {
for (int n = block * blck_0; n < (block + 1) * blck_0 & n < num_blocks * blck_0 + remainder; n++) {
for (int block = 0; block < N / blck_0 + 1; block++) {
for (int n = block * blck_0; n < (block + 1) * blck_0 & n < N; n++) {
int s_1, d_1;
int s_0, d_0;
if (!transpose0 && transpose1) {
Expand All @@ -273,7 +259,7 @@ ErrorCode mat_mul_fp32_fp16(Tensor *src0_, Tensor *src1, Tensor *dst, bool suppo
s_1 = 0; d_1 = n; s_0 = 0; d_0 = m;
}
vec_dot_fp16(K, dst->ptrAt<float>(b, h, m, n),
src1_cal->hostPtr<mllm_fp16_t>() + src1_cal->offset(b_1, h_1, s_1, d_1),
src1_cal->hostPtr<mllm_fp16_t>() + src1_cal->offset(b*is_0, h*is_0, s_1, d_1),
src0_cal->hostPtr<mllm_fp16_t>() + src0_cal->offset(b, h, s_0, d_0));
if (support_bias) {
*dst->ptrAt<float>(b, h, m, n) += bias->dataAt<float>(0, 0, 0, n);
Expand All @@ -294,9 +280,9 @@ ErrorCode mat_mul_fp32_q4_0(Tensor *src0_, Tensor *src1, Tensor *dst, bool suppo
src0_q8.setDtype(MLLM_TYPE_Q8_0);
src0_q8.alloc();
if (src0_->dimension() % QK8_0 == 0) {
#pragma omp parallel for collapse(3) num_threads(thread_count)
for (int b = 0; b < src0_->batch(); b++) {
for (int h = 0; h < src0_->head(); h++) {
#pragma omp parallel for num_threads(thread_count)
for (int s = 0; s < src0_->sequence(); s++) {
quantize_row_q8_0(src0_->hostPtr<float>() + src0_->offset(b, h, s, 0),
src0_q8.hostPtr<block_q8_0>() + src0_q8.offset(b, h, s, 0) / QK8_0,
Expand All @@ -316,18 +302,15 @@ ErrorCode mat_mul_fp32_q4_0(Tensor *src0_, Tensor *src1, Tensor *dst, bool suppo
Tensor *src0_cal = src0;
Tensor *src1_cal = src1;
const int64_t blck_0 = 16;
int is_0 = (src1->batch() == 1 && src1->head() == 1) ? 0 : 1;
#pragma omp parallel for collapse(4) num_threads(thread_count)
for (int b = 0; b < src0->batch(); b++) {
for (int h = 0; h < src0->head(); h++) {
const int b_1 = (src1->batch() == 1 && src1->head() == 1) ? 0 : b;
const int h_1 = (src1->batch() == 1 && src1->head() == 1) ? 0 : h;
for (int m = 0; m < M; m++) {
int num_blocks = N / blck_0;
int remainder = N % blck_0;
#pragma omp parallel for num_threads(thread_count)
for (int block = 0; block < num_blocks + 1; block++) {
for (int n = block * blck_0; n < (block + 1) * blck_0 & n < num_blocks * blck_0 + remainder; n++) {
for (int block = 0; block < N / blck_0 + 1; block++) {
for (int n = block * blck_0; n < (block + 1) * blck_0 & n < N; n++) {
vec_dot_q4_0_q8_0(K, dst->ptrAt<float>(b, h, m, n),
src1_cal->hostPtr<block_q4_0>() + src1_cal->offset(b_1, h_1, n, 0) / QK4_0,
src1_cal->hostPtr<block_q4_0>() + src1_cal->offset(b*is_0, h*is_0, n, 0) / QK4_0,
src0_cal->hostPtr<block_q8_0>() + src0_cal->offset(b, h, m, 0) / QK8_0);
if (support_bias) {
*dst->ptrAt<float>(b, h, m, n) += bias->dataAt<float>(0, 0, 0, n);
Expand All @@ -348,9 +331,9 @@ ErrorCode mat_mul_fp32_q4_K(Tensor *src0_, Tensor *src1, Tensor *dst, bool suppo
src0_q8.setDtype(MLLM_TYPE_Q8_K);
src0_q8.alloc();
if (src0_->dimension() % QK_K == 0) {
#pragma omp parallel for collapse(3) num_threads(thread_count)
for (int b = 0; b < src0_->batch(); b++) {
for (int h = 0; h < src0_->head(); h++) {
#pragma omp parallel for num_threads(thread_count)
for (int s = 0; s < src0_->sequence(); s++) {
quantize_row_q8_K(src0_->hostPtr<float>() + src0_->offset(b, h, s, 0),
src0_q8.hostPtr<block_q8_K>() + src0_q8.offset(b, h, s, 0) / QK_K,
Expand All @@ -370,28 +353,24 @@ ErrorCode mat_mul_fp32_q4_K(Tensor *src0_, Tensor *src1, Tensor *dst, bool suppo
Tensor *src0_cal = src0;
Tensor *src1_cal = src1;
const int64_t blck_0 = 16;

int is_0 = (src1->batch() == 1 && src1->head() == 1) ? 0 : 1;
#pragma omp parallel for collapse(4) num_threads(thread_count)
for (int b = 0; b < src0->batch(); b++) {
for (int h = 0; h < src0->head(); h++) {
const int b_1 = (src1->batch() == 1 && src1->head() == 1) ? 0 : b;
const int h_1 = (src1->batch() == 1 && src1->head() == 1) ? 0 : h;
for (int m = 0; m < M; m++) {
int num_blocks = N / blck_0;
int remainder = N % blck_0;
#pragma omp parallel for num_threads(thread_count)
for (int block = 0; block < num_blocks + 1; block++) {
for (int n = block * blck_0; n < (block + 1) * blck_0 & n < num_blocks * blck_0 + remainder; n++) {
for (int block = 0; block < N / blck_0 + 1; block++) {
for (int n = block * blck_0; n < (block + 1) * blck_0 & n < N; n++) {
if(dst->dtypeAt(b,h,m,n) == MLLM_TYPE_F32) {
vec_dot_q4_K_q8_K(K, dst->ptrAt<float>(b, h, m, n),
src1_cal->hostPtr<block_q4_K>() + src1_cal->offset(b_1, h_1, n, 0) / QK_K,
src1_cal->hostPtr<block_q4_K>() + src1_cal->offset(b*is_0, h*is_0, n, 0) / QK_K,
src0_cal->hostPtr<block_q8_K>() + src0_cal->offset(b, h, m, 0) / QK_K);
if (support_bias) {
*dst->ptrAt<float>(b, h, m, n) += bias->dataAt<float>(0, 0, 0, n);
}
} else if (dst->dtypeAt(b,h,m,n) == MLLM_TYPE_F16) {
float tmp = 0;
vec_dot_q4_K_q8_K(K, &tmp,
src1_cal->hostPtr<block_q4_K>() + src1_cal->offset(b_1, h_1, n, 0) / QK_K,
src1_cal->hostPtr<block_q4_K>() + src1_cal->offset(b*is_0, h*is_0, n, 0) / QK_K,
src0_cal->hostPtr<block_q8_K>() + src0_cal->offset(b, h, m, 0) / QK_K);
if (support_bias) {
*dst->ptrAt<mllm_fp16_t>(b, h, m, n) = MLLM_FP32_TO_FP16(tmp + bias->dataAt<float>(0, 0, 0, n));
Expand All @@ -415,9 +394,9 @@ ErrorCode mat_mul_fp32_q6_K(Tensor *src0_, Tensor *src1, Tensor *dst, bool suppo
src0_q8.setDtype(MLLM_TYPE_Q8_K);
src0_q8.alloc();
if (src0_->dimension() % QK_K == 0) {
#pragma omp parallel for collapse(3) num_threads(thread_count)
for (int b = 0; b < src0_->batch(); b++) {
for (int h = 0; h < src0_->head(); h++) {
#pragma omp parallel for num_threads(thread_count)
for (int s = 0; s < src0_->sequence(); s++) {
quantize_row_q8_K(src0_->hostPtr<float>() + src0_->offset(b, h, s, 0),
src0_q8.hostPtr<block_q8_K>() + src0_q8.offset(b, h, s, 0) / QK_K,
Expand All @@ -437,27 +416,24 @@ ErrorCode mat_mul_fp32_q6_K(Tensor *src0_, Tensor *src1, Tensor *dst, bool suppo
Tensor *src0_cal = src0;
Tensor *src1_cal = src1;
const int64_t blck_0 = 16;
int is_0 = (src1->batch() == 1 && src1->head() == 1) ? 0 : 1;
#pragma omp parallel for collapse(4) num_threads(thread_count)
for (int b = 0; b < src0->batch(); b++) {
for (int h = 0; h < src0->head(); h++) {
const int b_1 = (src1->batch() == 1 && src1->head() == 1) ? 0 : b;
const int h_1 = (src1->batch() == 1 && src1->head() == 1) ? 0 : h;
for (int m = 0; m < M; m++) {
int num_blocks = N / blck_0;
int remainder = N % blck_0;
#pragma omp parallel for num_threads(thread_count)
for (int block = 0; block < num_blocks + 1; block++) {
for (int n = block * blck_0; n < (block + 1) * blck_0 & n < num_blocks * blck_0 + remainder; n++) {
for (int block = 0; block < N / blck_0 + 1; block++) {
for (int n = block * blck_0; n < (block + 1) * blck_0 & n < N; n++) {
if (dst->dtypeAt(n, h, m, n) == MLLM_TYPE_F32) {
vec_dot_q6_K_q8_K(K, dst->ptrAt<float>(b, h, m, n),
src1_cal->hostPtr<block_q6_K>() + src1_cal->offset(b_1, h_1, n, 0) / QK_K,
src1_cal->hostPtr<block_q6_K>() + src1_cal->offset(b*is_0, h*is_0, n, 0) / QK_K,
src0_cal->hostPtr<block_q8_K>() + src0_cal->offset(b, h, m, 0) / QK_K);
if (support_bias) {
*dst->ptrAt<float>(b, h, m, n) += bias->dataAt<float>(0, 0, 0, n);
}
} else if (dst->dtypeAt(n, h, m, n) == MLLM_TYPE_F16) {
float tmp = 0;
vec_dot_q6_K_q8_K(K, &tmp,
src1_cal->hostPtr<block_q6_K>() + src1_cal->offset(b_1, h_1, n, 0) / QK_K,
src1_cal->hostPtr<block_q6_K>() + src1_cal->offset(b*is_0, h*is_0, n, 0) / QK_K,
src0_cal->hostPtr<block_q8_K>() + src0_cal->offset(b, h, m, 0) / QK_K);

if (support_bias) {
Expand Down
14 changes: 14 additions & 0 deletions src/models/qwen/configuration_qwen.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -102,6 +102,20 @@ struct QWenConfig {
sliding_window = 32768;
vocab_size = 151936;
tie_embedding_words = true;
} else if (billionsType == "1.8b") {
attention_dropout = 0.0;
std::string hidden_act = "silu";
hidden_size = 2048;
intermediate_size = 5504;
max_position_embeddings = 32768;
num_attention_heads = 16;
num_hidden_layers = 24;
num_key_value_heads = 16;
rms_norm_eps = 1e-6;
rope_theta = 1000000.0;
sliding_window = 32768;
vocab_size = 151936;
tie_embedding_words = false;
} else {
throw std::runtime_error("Unsupported model size");
}
Expand Down
5 changes: 5 additions & 0 deletions src/models/qwen/modeling_qwen.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -184,6 +184,8 @@ class QWenForCausalLM final : public Module {
// Others use nn.Linear()
if (tie_embedding_words) {
lm_head = Parameter(1, config.vocab_size, 1, config.hidden_size, names.token_embd_name + ".weight");
} else{
lm_head_layer = Linear(config.hidden_size, config.vocab_size, false, names.lm_head_name);
}
}

Expand All @@ -194,6 +196,8 @@ class QWenForCausalLM final : public Module {
auto outputs = model({x})[0];
if (tie_embedding_words) {
outputs = Tensor::mm(outputs, lm_head().transpose(Chl::SEQUENCE, Chl::DIMENSION));
} else {
outputs = lm_head_layer(outputs);
}
return {outputs};
}
Expand All @@ -203,6 +207,7 @@ class QWenForCausalLM final : public Module {
bool tie_embedding_words;
Layer embedding;
Parameter lm_head;
Layer lm_head_layer;
QWenModel model;
};

Expand Down
15 changes: 15 additions & 0 deletions src/quantizer/QuantWriter.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -40,6 +40,8 @@ vector<string> fp32_layers = {"norm", "rope", "bias","rotary_emb", "embed_tokens
"modality_preprocessors", "modality_heads", "modality_postprocessors", "pre_transformer_layer"};
vector<string> q6_layers = {"w2", "wv", "dense_h_to_4h", "v_proj", "down_proj"};

int tmp_hidden_dim = -1;

bool find_names(const string &name, const vector<string> &layer_names) {
for (const auto &layer : layer_names) {
if (name.find(layer) != std::string::npos) {
Expand All @@ -58,8 +60,21 @@ void QuantWriter::quantParams(DataType dataType) {
__exit(-1);
}
auto size = param_loader_->offsets_[name].second / sizeof(float);
if(find_names(name, {"input_layernorm"})) {
tmp_hidden_dim = size;
}
void *quant_ptr = nullptr;
std::pair<void *, uint64_t> block_t;
if (find_names(name, q6_layers)) {
if(tmp_hidden_dim>0 && (size/tmp_hidden_dim)%256!=0){
std::cout << "Quantize param " << name << " to " << DataTypeName(MLLM_TYPE_F32) << "\t";
const auto s = param_loader_->offsets_[name].second / sizeof(float);
const auto tsize = alloc_quant_block(s, MLLM_TYPE_F32).second;
writeParam(name, MLLM_TYPE_F32, param, tsize);
std::cout << " size:" << tsize << std::endl;
continue;
}
}
if(find_names(name, fp32_layers)) {
std::cout << "Quantize param " << name << " to " << DataTypeName(MLLM_TYPE_F32) << "\t";
const auto s = param_loader_->offsets_[name].second / sizeof(float);
Expand Down
Loading