implement chameleon graph
This commit is contained in:
nopperl
parent
fc09437496
commit
0453f7d114
2 changed files with 235 additions and 0 deletions
|
|
@ -92,6 +92,7 @@ extern "C" {
|
||||||
LLAMA_VOCAB_PRE_TYPE_CHATGLM4 = 17,
|
LLAMA_VOCAB_PRE_TYPE_CHATGLM4 = 17,
|
||||||
LLAMA_VOCAB_PRE_TYPE_VIKING = 18,
|
LLAMA_VOCAB_PRE_TYPE_VIKING = 18,
|
||||||
LLAMA_VOCAB_PRE_TYPE_JAIS = 19,
|
LLAMA_VOCAB_PRE_TYPE_JAIS = 19,
|
||||||
|
LLAMA_VOCAB_PRE_TYPE_CHAMELEON = 20,
|
||||||
};
|
};
|
||||||
|
|
||||||
// note: these values should be synchronized with ggml_rope
|
// note: these values should be synchronized with ggml_rope
|
||||||
|
|
|
||||||
234
src/llama.cpp
234
src/llama.cpp
|
|
@ -239,6 +239,7 @@ enum llm_arch {
|
||||||
LLM_ARCH_BITNET,
|
LLM_ARCH_BITNET,
|
||||||
LLM_ARCH_T5,
|
LLM_ARCH_T5,
|
||||||
LLM_ARCH_JAIS,
|
LLM_ARCH_JAIS,
|
||||||
|
LLM_ARCH_CHAMELEON,
|
||||||
LLM_ARCH_UNKNOWN,
|
LLM_ARCH_UNKNOWN,
|
||||||
};
|
};
|
||||||
|
|
||||||
|
|
@ -283,6 +284,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
|
||||||
{ LLM_ARCH_BITNET, "bitnet" },
|
{ LLM_ARCH_BITNET, "bitnet" },
|
||||||
{ LLM_ARCH_T5, "t5" },
|
{ LLM_ARCH_T5, "t5" },
|
||||||
{ LLM_ARCH_JAIS, "jais" },
|
{ LLM_ARCH_JAIS, "jais" },
|
||||||
|
{ LLM_ARCH_CHAMELEON, "chameleon" },
|
||||||
{ LLM_ARCH_UNKNOWN, "(unknown)" },
|
{ LLM_ARCH_UNKNOWN, "(unknown)" },
|
||||||
};
|
};
|
||||||
|
|
||||||
|
|
@ -1296,6 +1298,25 @@ static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NA
|
||||||
{ LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" },
|
{ LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" },
|
||||||
},
|
},
|
||||||
},
|
},
|
||||||
|
{
|
||||||
|
LLM_ARCH_CHAMELEON,
|
||||||
|
{
|
||||||
|
{ LLM_TENSOR_TOKEN_EMBD, "token_embd" },
|
||||||
|
{ LLM_TENSOR_OUTPUT_NORM, "output_norm" },
|
||||||
|
{ LLM_TENSOR_OUTPUT, "output" },
|
||||||
|
{ 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_TENSOR_ATTN_Q_NORM, "blk.%d.attn_q_norm" },
|
||||||
|
{ LLM_TENSOR_ATTN_K_NORM, "blk.%d.attn_k_norm" },
|
||||||
|
},
|
||||||
|
},
|
||||||
{
|
{
|
||||||
LLM_ARCH_UNKNOWN,
|
LLM_ARCH_UNKNOWN,
|
||||||
{
|
{
|
||||||
|
|
@ -5217,6 +5238,17 @@ static void llm_load_hparams(
|
||||||
default: model.type = e_model::MODEL_UNKNOWN;
|
default: model.type = e_model::MODEL_UNKNOWN;
|
||||||
}
|
}
|
||||||
} break;
|
} break;
|
||||||
|
case LLM_ARCH_CHAMELEON:
|
||||||
|
{
|
||||||
|
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
|
||||||
|
hparams.f_norm_eps = 1e-5; // eps for qk-norm, torch default
|
||||||
|
|
||||||
|
switch (hparams.n_layer) {
|
||||||
|
case 32: model.type = e_model::MODEL_7B; break;
|
||||||
|
case 48: model.type = e_model::MODEL_34B; break;
|
||||||
|
default: model.type = e_model::MODEL_UNKNOWN;
|
||||||
|
}
|
||||||
|
} break;
|
||||||
default: (void)0;
|
default: (void)0;
|
||||||
}
|
}
|
||||||
|
|
||||||
|
|
@ -5451,6 +5483,11 @@ static void llm_load_vocab(
|
||||||
} else if (
|
} else if (
|
||||||
tokenizer_pre == "jais") {
|
tokenizer_pre == "jais") {
|
||||||
vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_JAIS;
|
vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_JAIS;
|
||||||
|
} else if (
|
||||||
|
tokenizer_pre == "chameleon") {
|
||||||
|
vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_CHAMELEON;
|
||||||
|
vocab.tokenizer_add_bos = true;
|
||||||
|
vocab.tokenizer_clean_spaces = false;
|
||||||
} else {
|
} else {
|
||||||
throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str()));
|
throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str()));
|
||||||
}
|
}
|
||||||
|
|
@ -7498,6 +7535,45 @@ static bool llm_load_tensors(
|
||||||
layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd});
|
layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd});
|
||||||
}
|
}
|
||||||
} break;
|
} break;
|
||||||
|
case LLM_ARCH_CHAMELEON:
|
||||||
|
{
|
||||||
|
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});
|
||||||
|
model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED);
|
||||||
|
|
||||||
|
// if output is NULL, init from the input tok embed
|
||||||
|
if (model.output == NULL) {
|
||||||
|
model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED);
|
||||||
|
}
|
||||||
|
}
|
||||||
|
|
||||||
|
for (int 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.attn_q_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k, n_head});
|
||||||
|
layer.attn_k_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k, n_head_kv});
|
||||||
|
layer.attn_q_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_NORM, "bias", i), {n_embd_head_k, n_head}, llama_model_loader::TENSOR_NOT_REQUIRED);
|
||||||
|
layer.attn_k_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K_NORM, "bias", i), {n_embd_head_k, n_head}, llama_model_loader::TENSOR_NOT_REQUIRED);
|
||||||
|
|
||||||
|
layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd});
|
||||||
|
layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa});
|
||||||
|
layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa});
|
||||||
|
layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, 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_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd});
|
||||||
|
layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff});
|
||||||
|
}
|
||||||
|
} break;
|
||||||
default:
|
default:
|
||||||
throw std::runtime_error("unknown architecture");
|
throw std::runtime_error("unknown architecture");
|
||||||
}
|
}
|
||||||
|
|
@ -13606,6 +13682,149 @@ struct llm_build_context {
|
||||||
|
|
||||||
return gf;
|
return gf;
|
||||||
}
|
}
|
||||||
|
|
||||||
|
// ref: https://github.com/facebookresearch/chameleon
|
||||||
|
// based on the original build_llama() function, changes:
|
||||||
|
// * qk-norm
|
||||||
|
// * swin-norm (TODO)
|
||||||
|
// * removed bias
|
||||||
|
// * removed MoE
|
||||||
|
struct ggml_cgraph * build_chameleon() {
|
||||||
|
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
|
||||||
|
|
||||||
|
// mutable variable, needed during the last layer of the computation to skip unused tokens
|
||||||
|
int32_t n_tokens = this->n_tokens;
|
||||||
|
|
||||||
|
const int64_t n_embd_head = hparams.n_embd_head_v;
|
||||||
|
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
|
||||||
|
GGML_ASSERT(n_embd_head == hparams.n_rot);
|
||||||
|
|
||||||
|
struct ggml_tensor * cur;
|
||||||
|
struct ggml_tensor * inpL;
|
||||||
|
|
||||||
|
inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
|
||||||
|
|
||||||
|
// inp_pos - contains the positions
|
||||||
|
struct ggml_tensor * inp_pos = build_inp_pos();
|
||||||
|
|
||||||
|
// KQ_mask (mask for 1 head, it will be broadcasted to all heads)
|
||||||
|
struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
|
||||||
|
|
||||||
|
for (int il = 0; il < n_layer; ++il) {
|
||||||
|
struct ggml_tensor * inpSA = inpL;
|
||||||
|
|
||||||
|
// 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);
|
||||||
|
|
||||||
|
if (model.layers[il].attn_q_norm) {
|
||||||
|
Qcur = ggml_view_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens,
|
||||||
|
ggml_element_size(Qcur) * n_embd_head,
|
||||||
|
ggml_element_size(Qcur) * n_embd_head * n_head,
|
||||||
|
0);
|
||||||
|
cb(Qcur, "Qcur", il);
|
||||||
|
Kcur = ggml_view_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens,
|
||||||
|
ggml_element_size(Kcur) * n_embd_head,
|
||||||
|
ggml_element_size(Kcur) * n_embd_head * n_head_kv,
|
||||||
|
0);
|
||||||
|
cb(Kcur, "Kcur", il);
|
||||||
|
|
||||||
|
Qcur = llm_build_norm(ctx0, Qcur, hparams,
|
||||||
|
model.layers[il].attn_q_norm,
|
||||||
|
model.layers[il].attn_q_norm_b,
|
||||||
|
LLM_NORM, cb, il);
|
||||||
|
cb(Qcur, "Qcur", il);
|
||||||
|
|
||||||
|
Kcur = llm_build_norm(ctx0, Kcur, hparams,
|
||||||
|
model.layers[il].attn_k_norm,
|
||||||
|
model.layers[il].attn_k_norm_b,
|
||||||
|
LLM_NORM, cb, il);
|
||||||
|
cb(Kcur, "Kcur", il);
|
||||||
|
}
|
||||||
|
|
||||||
|
Qcur = ggml_rope_ext(
|
||||||
|
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
|
||||||
|
n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
|
||||||
|
ext_factor, attn_factor, beta_fast, beta_slow
|
||||||
|
);
|
||||||
|
cb(Qcur, "Qcur", il);
|
||||||
|
|
||||||
|
Kcur = ggml_rope_ext(
|
||||||
|
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
|
||||||
|
n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
|
||||||
|
ext_factor, attn_factor, beta_fast, beta_slow
|
||||||
|
);
|
||||||
|
cb(Kcur, "Kcur", il);
|
||||||
|
|
||||||
|
cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
|
||||||
|
model.layers[il].wo, nullptr,
|
||||||
|
Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
|
||||||
|
}
|
||||||
|
|
||||||
|
if (il == n_layer - 1) {
|
||||||
|
// skip computing output for unused tokens
|
||||||
|
struct ggml_tensor * inp_out_ids = build_inp_out_ids();
|
||||||
|
n_tokens = n_outputs;
|
||||||
|
cur = ggml_get_rows(ctx0, cur, inp_out_ids);
|
||||||
|
inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
|
||||||
|
}
|
||||||
|
|
||||||
|
struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
|
||||||
|
cb(ffn_inp, "ffn_inp", il);
|
||||||
|
|
||||||
|
// feed-forward network
|
||||||
|
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, NULL,
|
||||||
|
model.layers[il].ffn_gate, NULL, NULL,
|
||||||
|
model.layers[il].ffn_down, NULL, NULL,
|
||||||
|
NULL,
|
||||||
|
LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
|
||||||
|
cb(cur, "ffn_out", il);
|
||||||
|
|
||||||
|
cur = ggml_add(ctx0, cur, ffn_inp);
|
||||||
|
cb(cur, "ffn_out", il);
|
||||||
|
|
||||||
|
cur = lctx.cvec.apply_to(ctx0, cur, il);
|
||||||
|
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;
|
||||||
|
}
|
||||||
};
|
};
|
||||||
|
|
||||||
static struct ggml_cgraph * llama_build_graph_defrag(llama_context & lctx, const std::vector<uint32_t> & ids) {
|
static struct ggml_cgraph * llama_build_graph_defrag(llama_context & lctx, const std::vector<uint32_t> & ids) {
|
||||||
|
|
@ -13853,6 +14072,10 @@ static struct ggml_cgraph * llama_build_graph(
|
||||||
{
|
{
|
||||||
result = llm.build_jais();
|
result = llm.build_jais();
|
||||||
} break;
|
} break;
|
||||||
|
case LLM_ARCH_CHAMELEON:
|
||||||
|
{
|
||||||
|
result = llm.build_chameleon();
|
||||||
|
} break;
|
||||||
default:
|
default:
|
||||||
GGML_ASSERT(false);
|
GGML_ASSERT(false);
|
||||||
}
|
}
|
||||||
|
|
@ -15457,6 +15680,16 @@ struct llm_tokenizer_bpe {
|
||||||
"\\p{N}",
|
"\\p{N}",
|
||||||
};
|
};
|
||||||
break;
|
break;
|
||||||
|
case LLAMA_VOCAB_PRE_TYPE_CHAMELEON:
|
||||||
|
regex_exprs = {
|
||||||
|
"<sentinel:[0-9]+>", // Sentinel tokens
|
||||||
|
"(IMGIMG)((A|B|C|D|E|F|G|H|I){1,4})Z", // Image tokens
|
||||||
|
"([\t\n]| | )", // directly from tokenizer.json
|
||||||
|
"\\p{N}", // Individual digits
|
||||||
|
"[\\p{P}\\$\\+<=>\\^~\\|`]+", // Punctuation
|
||||||
|
"'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)",
|
||||||
|
};
|
||||||
|
break;
|
||||||
default:
|
default:
|
||||||
// default regex for BPE tokenization pre-processing
|
// default regex for BPE tokenization pre-processing
|
||||||
regex_exprs = {
|
regex_exprs = {
|
||||||
|
|
@ -19367,6 +19600,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
|
||||||
case LLM_ARCH_ARCTIC:
|
case LLM_ARCH_ARCTIC:
|
||||||
case LLM_ARCH_DEEPSEEK2:
|
case LLM_ARCH_DEEPSEEK2:
|
||||||
case LLM_ARCH_CHATGLM:
|
case LLM_ARCH_CHATGLM:
|
||||||
|
case LLM_ARCH_CHAMELEON:
|
||||||
return LLAMA_ROPE_TYPE_NORM;
|
return LLAMA_ROPE_TYPE_NORM;
|
||||||
|
|
||||||
// the pairs of head values are offset by n_rot/2
|
// the pairs of head values are offset by n_rot/2
|
||||||
|
|
|
||||||
Loading…
Add table
Add a link
Reference in a new issue