llama : add gemma
model (#5631)
There are couple things in this architecture: 1. Shared input and output embedding parameters. 2. Key length and value length are not derived from `n_embd`. More information about the models can be found at https://ai.google.dev/gemma. GGUFs can be downloaded from https://huggingface.co/google.
This commit is contained in:
parent
88c46cbdac
commit
580111d42b
|
@ -107,6 +107,7 @@ Typically finetunes of the base models below are supported as well.
|
|||
- [x] [Orion 14B](https://github.com/ggerganov/llama.cpp/pull/5118)
|
||||
- [x] [InternLM2](https://huggingface.co/models?search=internlm2)
|
||||
- [x] [CodeShell](https://github.com/WisdomShell/codeshell)
|
||||
- [x] [Gemma](https://ai.google.dev/gemma)
|
||||
|
||||
**Multimodal models:**
|
||||
|
||||
|
|
|
@ -111,6 +111,7 @@ class MODEL_ARCH(IntEnum):
|
|||
ORION = auto()
|
||||
INTERNLM2 = auto()
|
||||
MINICPM = auto()
|
||||
GEMMA = auto()
|
||||
|
||||
|
||||
class MODEL_TENSOR(IntEnum):
|
||||
|
@ -167,6 +168,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
|
|||
MODEL_ARCH.ORION: "orion",
|
||||
MODEL_ARCH.INTERNLM2: "internlm2",
|
||||
MODEL_ARCH.MINICPM: "minicpm",
|
||||
MODEL_ARCH.GEMMA: "gemma",
|
||||
}
|
||||
|
||||
TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
|
||||
|
@ -511,6 +513,19 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
|
|||
MODEL_TENSOR.FFN_DOWN_EXP,
|
||||
MODEL_TENSOR.FFN_UP_EXP,
|
||||
],
|
||||
MODEL_ARCH.GEMMA: [
|
||||
MODEL_TENSOR.TOKEN_EMBD,
|
||||
MODEL_TENSOR.OUTPUT_NORM,
|
||||
MODEL_TENSOR.ATTN_NORM,
|
||||
MODEL_TENSOR.ATTN_Q,
|
||||
MODEL_TENSOR.ATTN_K,
|
||||
MODEL_TENSOR.ATTN_V,
|
||||
MODEL_TENSOR.ATTN_OUT,
|
||||
MODEL_TENSOR.FFN_GATE,
|
||||
MODEL_TENSOR.FFN_DOWN,
|
||||
MODEL_TENSOR.FFN_UP,
|
||||
MODEL_TENSOR.FFN_NORM,
|
||||
],
|
||||
# TODO
|
||||
}
|
||||
|
||||
|
|
170
llama.cpp
170
llama.cpp
|
@ -208,6 +208,7 @@ enum llm_arch {
|
|||
LLM_ARCH_ORION,
|
||||
LLM_ARCH_INTERNLM2,
|
||||
LLM_ARCH_MINICPM,
|
||||
LLM_ARCH_GEMMA,
|
||||
LLM_ARCH_UNKNOWN,
|
||||
};
|
||||
|
||||
|
@ -234,6 +235,7 @@ static std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
|
|||
{ LLM_ARCH_ORION, "orion" },
|
||||
{ LLM_ARCH_INTERNLM2, "internlm2" },
|
||||
{ LLM_ARCH_MINICPM, "minicpm" },
|
||||
{ LLM_ARCH_GEMMA, "gemma" },
|
||||
};
|
||||
|
||||
enum llm_kv {
|
||||
|
@ -760,6 +762,22 @@ static std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NAMES =
|
|||
{ LLM_TENSOR_FFN_UP_EXP, "blk.%d.ffn_up.%d" },
|
||||
},
|
||||
},
|
||||
{
|
||||
LLM_ARCH_GEMMA,
|
||||
{
|
||||
{ LLM_TENSOR_TOKEN_EMBD, "token_embd" },
|
||||
{ LLM_TENSOR_OUTPUT_NORM, "output_norm" },
|
||||
{ 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_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,
|
||||
{
|
||||
|
@ -3243,6 +3261,16 @@ static void llm_load_hparams(
|
|||
default: model.type = e_model::MODEL_UNKNOWN;
|
||||
}
|
||||
} break;
|
||||
case LLM_ARCH_GEMMA:
|
||||
{
|
||||
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
|
||||
|
||||
switch (hparams.n_layer) {
|
||||
case 18: model.type = e_model::MODEL_2B; break;
|
||||
case 28: model.type = e_model::MODEL_7B; break;
|
||||
default: model.type = e_model::MODEL_UNKNOWN;
|
||||
}
|
||||
} break;
|
||||
default: (void)0;
|
||||
}
|
||||
|
||||
|
@ -4360,6 +4388,37 @@ static bool llm_load_tensors(
|
|||
layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff});
|
||||
}
|
||||
} break;
|
||||
case LLM_ARCH_GEMMA:
|
||||
{
|
||||
model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
|
||||
|
||||
// output
|
||||
model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
|
||||
|
||||
const int64_t n_ff = hparams.n_ff;
|
||||
const int64_t n_embd_head_k = hparams.n_embd_head_k;
|
||||
const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa();
|
||||
const int64_t n_embd_v_gqa = hparams.n_embd_v_gqa();
|
||||
|
||||
for (uint32_t i = 0; i < n_layer; ++i) {
|
||||
ggml_context * ctx_layer = ctx_for_layer(i);
|
||||
ggml_context * ctx_split = ctx_for_layer_split(i);
|
||||
|
||||
auto & layer = model.layers[i];
|
||||
|
||||
layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
|
||||
|
||||
layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_head_k * hparams.n_head});
|
||||
layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_k_gqa});
|
||||
layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_v_gqa});
|
||||
layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * hparams.n_head, n_embd});
|
||||
|
||||
layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
|
||||
layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff});
|
||||
layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff});
|
||||
layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd});
|
||||
}
|
||||
} break;
|
||||
default:
|
||||
throw std::runtime_error("unknown architecture");
|
||||
}
|
||||
|
@ -7366,6 +7425,113 @@ struct llm_build_context {
|
|||
|
||||
return gf;
|
||||
}
|
||||
|
||||
struct ggml_cgraph * build_gemma() {
|
||||
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
|
||||
|
||||
const int64_t n_embd_head_k = hparams.n_embd_head_k;
|
||||
|
||||
struct ggml_tensor * cur;
|
||||
struct ggml_tensor * inpL;
|
||||
|
||||
inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb);
|
||||
cb(inpL, "inp_embd", -1);
|
||||
inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
|
||||
cb(inpL, "inp_scaled", -1);
|
||||
|
||||
// inp_pos - contains the positions
|
||||
struct ggml_tensor * inp_pos = ggml_view_1d(ctx0, lctx.inp_pos, n_tokens, 0);
|
||||
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_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0);
|
||||
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, lctx.inp_K_shift, LLM_ROPE, n_ctx, 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);
|
||||
|
||||
// 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_k, n_head, n_tokens), inp_pos,
|
||||
n_embd_head_k, 2, 0, n_orig_ctx, freq_base, freq_scale,
|
||||
ext_factor, attn_factor, beta_fast, beta_slow);
|
||||
cb(Qcur, "Qcur", il);
|
||||
Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head_k)));
|
||||
cb(Qcur, "Qcur_scaled", il);
|
||||
|
||||
Kcur = ggml_rope_custom(
|
||||
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head_k, n_head_kv, n_tokens), inp_pos,
|
||||
n_embd_head_k, 2, 0, n_orig_ctx, freq_base, freq_scale,
|
||||
ext_factor, attn_factor, beta_fast, beta_slow);
|
||||
cb(Kcur, "Kcur", il);
|
||||
|
||||
cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
|
||||
model.layers[il].wo, NULL,
|
||||
Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f, cb, il);
|
||||
cb(cur, "kqv_out", il);
|
||||
}
|
||||
struct ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL);
|
||||
cb(sa_out, "sa_out", il);
|
||||
|
||||
cur = llm_build_norm(ctx0, sa_out, hparams,
|
||||
model.layers[il].ffn_norm, NULL,
|
||||
LLM_NORM_RMS, cb, il);
|
||||
cb(cur, "ffn_norm", il);
|
||||
|
||||
// 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,
|
||||
NULL,
|
||||
LLM_FFN_GELU, LLM_FFN_PAR, cb, il);
|
||||
cb(cur, "ffn_out", il);
|
||||
}
|
||||
|
||||
cur = ggml_add(ctx0, cur, sa_out);
|
||||
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.tok_embd, cur);
|
||||
cb(cur, "result_output", -1);
|
||||
|
||||
ggml_build_forward_expand(gf, cur);
|
||||
|
||||
return gf;
|
||||
}
|
||||
};
|
||||
|
||||
static struct ggml_cgraph * llama_build_graph(
|
||||
|
@ -7474,6 +7640,10 @@ static struct ggml_cgraph * llama_build_graph(
|
|||
{
|
||||
result = llm.build_minicpm();
|
||||
} break;
|
||||
case LLM_ARCH_GEMMA:
|
||||
{
|
||||
result = llm.build_gemma();
|
||||
} break;
|
||||
default:
|
||||
GGML_ASSERT(false);
|
||||
}
|
||||
|
|
Loading…
Reference in a new issue