llama : add Qwen support (#4281)
* enable qwen to llama.cpp * llama : do not GPU split bias tensors --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
This commit is contained in:
Shijie
GitHub
parent
880f57973b
commit
37c746d687
5 changed files with 372 additions and 9 deletions
211
llama.cpp
211
llama.cpp
|
|
@ -192,6 +192,7 @@ enum llm_arch {
|
|||
LLM_ARCH_REFACT,
|
||||
LLM_ARCH_BLOOM,
|
||||
LLM_ARCH_STABLELM,
|
||||
LLM_ARCH_QWEN,
|
||||
LLM_ARCH_UNKNOWN,
|
||||
};
|
||||
|
||||
|
|
@ -208,6 +209,7 @@ static std::map<llm_arch, std::string> LLM_ARCH_NAMES = {
|
|||
{ LLM_ARCH_REFACT, "refact" },
|
||||
{ LLM_ARCH_BLOOM, "bloom" },
|
||||
{ LLM_ARCH_STABLELM, "stablelm" },
|
||||
{ LLM_ARCH_QWEN, "qwen" },
|
||||
};
|
||||
|
||||
enum llm_kv {
|
||||
|
|
@ -518,6 +520,22 @@ static std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NAMES =
|
|||
{ LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
|
||||
},
|
||||
},
|
||||
{
|
||||
LLM_ARCH_QWEN,
|
||||
{
|
||||
{ LLM_TENSOR_TOKEN_EMBD, "token_embd" },
|
||||
{ LLM_TENSOR_OUTPUT_NORM, "output_norm" },
|
||||
{ LLM_TENSOR_OUTPUT, "output" },
|
||||
{ LLM_TENSOR_ROPE_FREQS, "rope_freqs" },
|
||||
{ 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_GATE, "blk.%d.ffn_gate" },
|
||||
{ LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" },
|
||||
{ LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
|
||||
},
|
||||
},
|
||||
|
||||
{
|
||||
LLM_ARCH_UNKNOWN,
|
||||
|
|
@ -2347,6 +2365,15 @@ static void llm_load_hparams(
|
|||
default: model.type = e_model::MODEL_UNKNOWN;
|
||||
}
|
||||
} break;
|
||||
case LLM_ARCH_QWEN:
|
||||
{
|
||||
GGUF_GET_KEY(ctx, hparams.f_norm_rms_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS));
|
||||
switch (hparams.n_layer) {
|
||||
case 32: model.type = e_model::MODEL_7B; break;
|
||||
case 40: model.type = e_model::MODEL_13B; break;
|
||||
default: model.type = e_model::MODEL_UNKNOWN;
|
||||
}
|
||||
} break;
|
||||
|
||||
default: (void)0;
|
||||
}
|
||||
|
|
@ -3310,6 +3337,71 @@ static void llm_load_tensors(
|
|||
}
|
||||
}
|
||||
} break;
|
||||
case LLM_ARCH_QWEN:
|
||||
{
|
||||
model.tok_embd = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU);
|
||||
{
|
||||
ggml_backend_type backend_norm;
|
||||
ggml_backend_type backend_output;
|
||||
|
||||
if (n_gpu_layers > int(n_layer)) {
|
||||
// norm is not performance relevant on its own but keeping it in VRAM reduces data copying
|
||||
// on Windows however this is detrimental unless everything is on the GPU
|
||||
#ifndef _WIN32
|
||||
backend_norm = llama_backend_offload;
|
||||
#else
|
||||
backend_norm = n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : llama_backend_offload;
|
||||
#endif // _WIN32
|
||||
|
||||
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 = 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);
|
||||
}
|
||||
if (backend_output == GGML_BACKEND_GPU_SPLIT) {
|
||||
vram_weights += ggml_nbytes(model.output);
|
||||
}
|
||||
}
|
||||
|
||||
const uint32_t n_ff = hparams.n_ff / 2;
|
||||
|
||||
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.wqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd * 3}, backend_split);
|
||||
layer.bqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd * 3}, backend);
|
||||
layer.wo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, backend_split);
|
||||
|
||||
layer.ffn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, backend);
|
||||
|
||||
layer.ffn_gate = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, backend_split);
|
||||
layer.ffn_down = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, backend_split);
|
||||
layer.ffn_up = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split);
|
||||
|
||||
if (backend == GGML_BACKEND_GPU) {
|
||||
vram_weights +=
|
||||
ggml_nbytes(layer.attn_norm) + ggml_nbytes(layer.wqkv) + ggml_nbytes(layer.bqkv) +
|
||||
ggml_nbytes(layer.wo) + ggml_nbytes(layer.ffn_norm) + ggml_nbytes(layer.ffn_gate) +
|
||||
ggml_nbytes(layer.ffn_down) + ggml_nbytes(layer.ffn_up);
|
||||
}
|
||||
}
|
||||
} break;
|
||||
|
||||
default:
|
||||
throw std::runtime_error("unknown architecture");
|
||||
|
|
@ -4908,6 +5000,121 @@ struct llm_build_context {
|
|||
|
||||
return gf;
|
||||
}
|
||||
|
||||
struct ggml_cgraph * build_qwen() {
|
||||
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
|
||||
|
||||
struct ggml_tensor * cur;
|
||||
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 wil 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);
|
||||
|
||||
// shift the entire K-cache if needed
|
||||
if (do_rope_shift) {
|
||||
llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE_NEOX, n_ctx, n_embd_head, freq_base, freq_scale, cb);
|
||||
}
|
||||
|
||||
for (int il = 0; il < n_layer; ++il) {
|
||||
struct ggml_tensor * inpSA = inpL;
|
||||
|
||||
cur = llm_build_norm(ctx0, inpL, hparams,
|
||||
model.layers[il].attn_norm, NULL,
|
||||
LLM_NORM_RMS, 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, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)));
|
||||
struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 2*sizeof(float)*(n_embd)));
|
||||
|
||||
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);
|
||||
Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
|
||||
|
||||
// using mode = 2 for neox mode
|
||||
Qcur = ggml_rope_custom(
|
||||
ctx0, Qcur, inp_pos, n_embd_head, 2, 0, n_orig_ctx,
|
||||
freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
|
||||
);
|
||||
cb(Qcur, "Qcur", il);
|
||||
|
||||
Kcur = ggml_rope_custom(
|
||||
ctx0, Kcur, inp_pos, n_embd_head, 2, 0, n_orig_ctx,
|
||||
freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
|
||||
);
|
||||
cb(Kcur, "Kcur", il);
|
||||
|
||||
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, NULL,
|
||||
Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, cb, il);
|
||||
cb(cur, "kqv_out", il);
|
||||
}
|
||||
|
||||
struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
|
||||
cb(ffn_inp, "ffn_inp", il);
|
||||
|
||||
// feed-forward forward
|
||||
{
|
||||
cur = llm_build_norm(ctx0, ffn_inp, hparams,
|
||||
model.layers[il].ffn_norm, NULL,
|
||||
LLM_NORM_RMS, cb, il);
|
||||
cb(cur, "ffn_norm", il);
|
||||
|
||||
cur = llm_build_ffn(ctx0, cur,
|
||||
model.layers[il].ffn_up, NULL,
|
||||
model.layers[il].ffn_gate, NULL,
|
||||
model.layers[il].ffn_down, NULL,
|
||||
LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
|
||||
cb(cur, "ffn_out", il);
|
||||
}
|
||||
|
||||
cur = ggml_add(ctx0, cur, ffn_inp);
|
||||
cb(cur, "l_out", il);
|
||||
|
||||
// input for next layer
|
||||
inpL = cur;
|
||||
}
|
||||
|
||||
cur = inpL;
|
||||
|
||||
cur = llm_build_norm(ctx0, cur, hparams,
|
||||
model.output_norm, NULL,
|
||||
LLM_NORM_RMS, cb, -1);
|
||||
cb(cur, "result_norm", -1);
|
||||
|
||||
// lm_head
|
||||
cur = ggml_mul_mat(ctx0, model.output, cur);
|
||||
cb(cur, "result_output", -1);
|
||||
|
||||
ggml_build_forward_expand(gf, cur);
|
||||
|
||||
return gf;
|
||||
}
|
||||
};
|
||||
|
||||
//
|
||||
|
|
@ -5382,6 +5589,10 @@ static struct ggml_cgraph * llama_build_graph(
|
|||
{
|
||||
result = llm.build_stablelm();
|
||||
} break;
|
||||
case LLM_ARCH_QWEN:
|
||||
{
|
||||
result = llm.build_qwen();
|
||||
} break;
|
||||
default:
|
||||
GGML_ASSERT(false);
|
||||
}
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue