llama : add PLaMo model (#3557)
* add plamo mock * add tensor loading * plamo convert * update norm * able to compile * fix norm_rms_eps hparam * runnable * use inp_pos * seems ok * update kqv code * remove develop code * update README * shuffle attn_q.weight and attn_output.weight for broadcasting * remove plamo_llm_build_kqv and use llm_build_kqv * fix style * update * llama : remove obsolete KQ_scale * plamo : fix tensor names for correct GPU offload --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
This commit is contained in:
Shintarou Okada
GitHub
parent
5bf3953d7e
commit
753be377b6
5 changed files with 307 additions and 15 deletions
181
llama.cpp
181
llama.cpp
|
|
@ -198,6 +198,7 @@ enum llm_arch {
|
|||
LLM_ARCH_STABLELM,
|
||||
LLM_ARCH_QWEN,
|
||||
LLM_ARCH_PHI2,
|
||||
LLM_ARCH_PLAMO,
|
||||
LLM_ARCH_UNKNOWN,
|
||||
};
|
||||
|
||||
|
|
@ -216,6 +217,7 @@ static std::map<llm_arch, std::string> LLM_ARCH_NAMES = {
|
|||
{ LLM_ARCH_STABLELM, "stablelm" },
|
||||
{ LLM_ARCH_QWEN, "qwen" },
|
||||
{ LLM_ARCH_PHI2, "phi2" },
|
||||
{ LLM_ARCH_PLAMO, "plamo" },
|
||||
};
|
||||
|
||||
enum llm_kv {
|
||||
|
|
@ -567,6 +569,24 @@ static std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NAMES =
|
|||
{ LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
|
||||
},
|
||||
},
|
||||
{
|
||||
LLM_ARCH_PLAMO,
|
||||
{
|
||||
{ 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_Q, "blk.%d.attn_q" },
|
||||
{ LLM_TENSOR_ATTN_K, "blk.%d.attn_k" },
|
||||
{ LLM_TENSOR_ATTN_V, "blk.%d.attn_v" },
|
||||
{ LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" },
|
||||
{ LLM_TENSOR_ATTN_ROT_EMBD, "blk.%d.attn_rot_embd" },
|
||||
{ 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,
|
||||
|
|
@ -2749,6 +2769,15 @@ static void llm_load_hparams(
|
|||
default: model.type = e_model::MODEL_UNKNOWN;
|
||||
}
|
||||
} break;
|
||||
case LLM_ARCH_PLAMO:
|
||||
{
|
||||
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
|
||||
|
||||
switch (hparams.n_layer) {
|
||||
case 40: model.type = e_model::MODEL_13B; break;
|
||||
default: model.type = e_model::MODEL_UNKNOWN;
|
||||
}
|
||||
} break;
|
||||
|
||||
default: (void)0;
|
||||
}
|
||||
|
|
@ -3630,6 +3659,51 @@ static bool llm_load_tensors(
|
|||
layer.ffn_up_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff}, backend);
|
||||
}
|
||||
} break;
|
||||
case LLM_ARCH_PLAMO:
|
||||
{
|
||||
model.tok_embd = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 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 = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, backend_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.wq = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd}, backend_split);
|
||||
layer.wk = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa}, backend_split);
|
||||
layer.wv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa}, backend_split);
|
||||
layer.wo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, backend_split);
|
||||
|
||||
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);
|
||||
}
|
||||
} break;
|
||||
default:
|
||||
throw std::runtime_error("unknown architecture");
|
||||
}
|
||||
|
|
@ -5555,6 +5629,109 @@ struct llm_build_context {
|
|||
|
||||
return gf;
|
||||
}
|
||||
|
||||
struct ggml_cgraph * build_plamo() {
|
||||
struct ggml_cgraph * gf = ggml_new_graph(ctx0);
|
||||
|
||||
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_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);
|
||||
|
||||
// shift the entire K-cache if needed
|
||||
if (do_rope_shift) {
|
||||
llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE, n_ctx, n_embd_head, freq_base, freq_scale, cb);
|
||||
}
|
||||
|
||||
for (int il = 0; il < n_layer; ++il) {
|
||||
|
||||
// norm
|
||||
cur = llm_build_norm(ctx0, inpL, hparams,
|
||||
model.layers[il].attn_norm, NULL,
|
||||
LLM_NORM_RMS, cb, il);
|
||||
cb(cur, "attn_norm", il);
|
||||
|
||||
struct ggml_tensor * attention_norm = cur;
|
||||
|
||||
// self-attention
|
||||
{
|
||||
// compute Q and K and RoPE them
|
||||
struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
|
||||
cb(Qcur, "Qcur", il);
|
||||
|
||||
struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur);
|
||||
cb(Kcur, "Kcur", il);
|
||||
|
||||
struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur);
|
||||
cb(Vcur, "Vcur", il);
|
||||
|
||||
Qcur = ggml_rope_custom(
|
||||
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), 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, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), 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, model, hparams, kv_self,
|
||||
model.layers[il].wo, NULL,
|
||||
Qcur, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
|
||||
cb(cur, "kqv_out", il);
|
||||
}
|
||||
struct ggml_tensor * sa_out = cur;
|
||||
|
||||
cur = attention_norm;
|
||||
|
||||
// feed-forward network
|
||||
{
|
||||
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, sa_out);
|
||||
cb(cur, "l_out", il);
|
||||
|
||||
cur = ggml_add(ctx0, cur, inpL);
|
||||
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;
|
||||
}
|
||||
};
|
||||
|
||||
//
|
||||
|
|
@ -6065,6 +6242,10 @@ static struct ggml_cgraph * llama_build_graph(
|
|||
{
|
||||
result = llm.build_phi2();
|
||||
} break;
|
||||
case LLM_ARCH_PLAMO:
|
||||
{
|
||||
result = llm.build_plamo();
|
||||
} break;
|
||||
default:
|
||||
GGML_ASSERT(false);
|
||||
}
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue