vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

remove superfluous moe, bias and rope tensors

Browse source
This commit is contained in:
nopperl 2024-04-18 21:34:46 +02:00
parent 9855bb6d6c
commit 5993a97eb8
2 changed files with 1 additions and 39 deletions
Download patch file Download diff file Expand all files Collapse all files

10
gguf-py/gguf/constants.py
View file

@ -698,19 +698,13 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_ARCH.OLMO: [ MODEL_ARCH.OLMO: [
MODEL_TENSOR.TOKEN_EMBD, MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT, MODEL_TENSOR.OUTPUT,
MODEL_TENSOR.ROPE_FREQS,
MODEL_TENSOR.ATTN_Q, MODEL_TENSOR.ATTN_Q,
MODEL_TENSOR.ATTN_K, MODEL_TENSOR.ATTN_K,
MODEL_TENSOR.ATTN_V, MODEL_TENSOR.ATTN_V,
MODEL_TENSOR.ATTN_OUT, MODEL_TENSOR.ATTN_OUT,
MODEL_TENSOR.ATTN_ROT_EMBD,
MODEL_TENSOR.FFN_GATE_INP,
MODEL_TENSOR.FFN_GATE, MODEL_TENSOR.FFN_GATE,
MODEL_TENSOR.FFN_DOWN, MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP, MODEL_TENSOR.FFN_UP,
MODEL_TENSOR.FFN_GATE_EXP,
MODEL_TENSOR.FFN_DOWN_EXP,
MODEL_TENSOR.FFN_UP_EXP,
], ],
# TODO # TODO
} }
@ -748,10 +742,6 @@ MODEL_TENSOR_SKIP: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.ROPE_FREQS, MODEL_TENSOR.ROPE_FREQS,
MODEL_TENSOR.ATTN_ROT_EMBD, MODEL_TENSOR.ATTN_ROT_EMBD,
], ],
MODEL_ARCH.OLMO: [
MODEL_TENSOR.ROPE_FREQS,
MODEL_TENSOR.ATTN_ROT_EMBD,
],
} }
# #

30
llama.cpp
View file

@ -997,22 +997,13 @@ static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NA
{ {
{ LLM_TENSOR_TOKEN_EMBD, "token_embd" }, { LLM_TENSOR_TOKEN_EMBD, "token_embd" },
{ LLM_TENSOR_OUTPUT, "output" }, { LLM_TENSOR_OUTPUT, "output" },
{ LLM_TENSOR_ROPE_FREQS, "rope_freqs" },
{ LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" },
{ LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" },
{ LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" },
{ LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" },
{ LLM_TENSOR_ATTN_ROT_EMBD, "blk.%d.attn_rot_embd" },
{ LLM_TENSOR_FFN_GATE_INP, "blk.%d.ffn_gate_inp" },
{ LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" },
{ LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" },
{ LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
{ LLM_TENSOR_FFN_GATE_EXP, "blk.%d.ffn_gate.%d" },
{ LLM_TENSOR_FFN_DOWN_EXP, "blk.%d.ffn_down.%d" },
{ LLM_TENSOR_FFN_UP_EXP, "blk.%d.ffn_up.%d" },
{ LLM_TENSOR_FFN_GATE_EXPS, "blk.%d.ffn_gate_exps" },
{ LLM_TENSOR_FFN_DOWN_EXPS, "blk.%d.ffn_down_exps" },
{ LLM_TENSOR_FFN_UP_EXPS, "blk.%d.ffn_up_exps" },
}, },
}, },
{ {
@ -5711,7 +5702,6 @@ static bool llm_load_tensors(
} }
for (int i = 0; i < n_layer; ++i) { 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); ggml_context * ctx_split = ctx_for_layer_split(i);
auto & layer = model.layers[i]; auto & layer = model.layers[i];
@ -5721,12 +5711,6 @@ static bool llm_load_tensors(
layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V, "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.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
// optional bias tensors
layer.bq = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q, "bias", i), {n_embd}, false);
layer.bk = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K, "bias", i), {n_embd_gqa}, false);
layer.bv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_V, "bias", i), {n_embd_gqa}, false);
layer.bo = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, false);
layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}); 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_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd});
@ -10152,10 +10136,6 @@ struct llm_build_context {
// compute Q and K and RoPE them // compute Q and K and RoPE them
struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur); struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
cb(Qcur, "Qcur", il); cb(Qcur, "Qcur", il);
if (model.layers[il].bq) {
Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
cb(Qcur, "Qcur", il);
}
if (hparams.f_clamp_kqv > 0.0f) { if (hparams.f_clamp_kqv > 0.0f) {
Qcur = ggml_clamp(ctx0, Qcur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv); Qcur = ggml_clamp(ctx0, Qcur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv);
cb(Qcur, "Qcur", il); cb(Qcur, "Qcur", il);
@ -10163,10 +10143,6 @@ struct llm_build_context {
struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur); struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur);
cb(Kcur, "Kcur", il); cb(Kcur, "Kcur", il);
if (model.layers[il].bk) {
Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
cb(Kcur, "Kcur", il);
}
if (hparams.f_clamp_kqv > 0.0f) { if (hparams.f_clamp_kqv > 0.0f) {
Kcur = ggml_clamp(ctx0, Kcur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv); Kcur = ggml_clamp(ctx0, Kcur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv);
cb(Kcur, "Kcur", il); cb(Kcur, "Kcur", il);
@ -10174,10 +10150,6 @@ struct llm_build_context {
struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur); struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur);
cb(Vcur, "Vcur", il); cb(Vcur, "Vcur", il);
if (model.layers[il].bv) {
Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
cb(Vcur, "Vcur", il);
}
if (hparams.f_clamp_kqv > 0.0f) { if (hparams.f_clamp_kqv > 0.0f) {
Vcur = ggml_clamp(ctx0, Vcur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv); Vcur = ggml_clamp(ctx0, Vcur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv);
cb(Vcur, "Vcur", il); cb(Vcur, "Vcur", il);
@ -10198,7 +10170,7 @@ struct llm_build_context {
cb(Kcur, "Kcur", il); cb(Kcur, "Kcur", il);
cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo, nullptr,
Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
} }