vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2025-09-11 01:24:36 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Merge branch 'master' into concedo_experimental

Browse source 
# Conflicts:
#	.devops/nix/nixpkgs-instances.nix
#	.devops/nix/package.nix
#	.devops/nix/scope.nix
#	.github/workflows/build.yml
#	.github/workflows/nix-ci.yml
#	CMakeLists.txt
#	flake.nix
#	ggml.c
This commit is contained in:
Concedo 2024-01-22 22:31:22 +08:00
commit f96f29be7b
20 changed files with 1807 additions and 162 deletions
Download patch file Download diff file Expand all files Collapse all files

86
llama.cpp
View file

@ -1330,8 +1330,10 @@ static llama_state g_state;
// available llama models
enum e_model {
MODEL_UNKNOWN,
MODEL_0_5B,
MODEL_1B,
MODEL_3B,
MODEL_4B,
MODEL_7B,
MODEL_8B,
MODEL_13B,
@ -2670,6 +2672,7 @@ static std::string llama_model_ftype_name(llama_ftype ftype) {
case LLAMA_FTYPE_MOSTLY_Q6_K: return "Q6_K";
case LLAMA_FTYPE_MOSTLY_IQ2_XXS:return "IQ2_XSS - 2.0625 bpw";
case LLAMA_FTYPE_MOSTLY_IQ2_XS: return "IQ2_XS - 2.3125 bpw";
case LLAMA_FTYPE_MOSTLY_Q3_K_XS:return "Q3_K - Extra small";
default: return "unknown, may not work";
}
@ -2885,6 +2888,7 @@ static void llm_load_hparams(
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
switch (hparams.n_layer) {
case 24: model.type = e_model::MODEL_1B; break;
case 32: model.type = e_model::MODEL_3B; break;
default: model.type = e_model::MODEL_UNKNOWN;
}
@ -2903,9 +2907,9 @@ static void llm_load_hparams(
{
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
switch (hparams.n_layer) {
case 24: model.type = e_model::MODEL_1B; break;
case 24: model.type = hparams.n_embd == 1024 ? e_model::MODEL_0_5B : e_model::MODEL_1B; break;
case 32: model.type = e_model::MODEL_7B; break;
case 40: model.type = e_model::MODEL_13B; break;
case 40: model.type = hparams.n_head == 20 ? e_model::MODEL_4B : e_model::MODEL_13B; break;
case 80: model.type = e_model::MODEL_70B; break;
default: model.type = e_model::MODEL_UNKNOWN;
}
@ -3727,6 +3731,11 @@ 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.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
// optional bias tensors, present in Stable LM 2 1.6B
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.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
layer.ffn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "bias", i), {n_embd});
@ -5625,12 +5634,24 @@ struct llm_build_context {
// compute Q and K and RoPE them
struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
cb(Qcur, "Qcur", il);
if (model.layers[il].bq) {
Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
cb(Qcur, "Qcur", il);
}
struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur);
cb(Kcur, "Kcur", il);
if (model.layers[il].bk) {
Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
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].bv) {
Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
cb(Vcur, "Vcur", il);
}
Qcur = ggml_rope_custom(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos,
@ -9100,9 +9121,13 @@ struct quantize_state_internal {
const llama_model_quantize_params * params;
int n_attention_wv = 0;
int n_feed_forward_w2 = 0;
int n_ffn_down = 0;
int n_ffn_gate = 0;
int n_ffn_up = 0;
int i_attention_wv = 0;
int i_feed_forward_w2 = 0;
int i_ffn_down = 0;
int i_ffn_gate = 0;
int i_ffn_up = 0;
int n_k_quantized = 0;
int n_fallback = 0;
@ -9205,8 +9230,8 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
++qs.i_attention_wv;
}
else if (name.find("ffn_down") != std::string::npos) {
if (qs.i_feed_forward_w2 < qs.n_feed_forward_w2/8) new_type = GGML_TYPE_Q2_K;
++qs.i_feed_forward_w2;
if (qs.i_ffn_down < qs.n_ffn_down/8) new_type = GGML_TYPE_Q2_K;
++qs.i_ffn_down;
}
else if (name == "token_embd.weight") new_type = GGML_TYPE_Q2_K;
} else if (name.find("attn_v.weight") != std::string::npos) {
@ -9243,18 +9268,21 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
// TODO: explore better strategies
new_type = GGML_TYPE_Q8_0;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_XS) {
new_type = GGML_TYPE_Q2_K;
}
} else if (name.find("ffn_down") != std::string::npos) {
const int n_expert = std::max(1, (int)qs.model.hparams.n_expert);
int i_layer, n_layer;
if (n_expert == 1) {
i_layer = qs.i_feed_forward_w2;
n_layer = qs.n_feed_forward_w2;
i_layer = qs.i_ffn_down;
n_layer = qs.n_ffn_down;
} else {
// Believe it or not, "experts" in the FFN of Mixtral-8x7B are not consecutive, but iccasionally randomly
// sprinkled in the model. Hence, simply dividing i_feed_forward_w2 by n_expert does not work
// sprinkled in the model. Hence, simply dividing i_ffn_down by n_expert does not work
// for getting the current layer as I initially thought, and we need to resort to parsing the
// tensor name.
n_layer = qs.n_feed_forward_w2 / n_expert;
n_layer = qs.n_ffn_down / n_expert;
if (sscanf(name.c_str(), "blk.%d.ffn_down", &i_layer) != 1) {
throw std::runtime_error(format("Failed to determine layer for tensor %s", name.c_str()));
}
@ -9263,7 +9291,7 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
}
}
if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K;
else if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S) {
else if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_XS) {
if (i_layer < n_layer/8) new_type = GGML_TYPE_Q4_K;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) {
@ -9293,11 +9321,12 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
// same quantization as before imatrix stuff, and b) Q4_1/Q5_1 do go crazy on ffn_down without an imatrix.
new_type = ftype == LLAMA_FTYPE_MOSTLY_Q4_0 ? GGML_TYPE_Q4_1 : GGML_TYPE_Q5_1;
}
++qs.i_feed_forward_w2;
++qs.i_ffn_down;
} else if (name.find("attn_output.weight") != std::string::npos) {
if (arch != LLM_ARCH_FALCON) {
if (qs.model.hparams.n_expert == 8) {
if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M ||
if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_XS ||
ftype == LLAMA_FTYPE_MOSTLY_Q3_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M ||
ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M) {
new_type = GGML_TYPE_Q5_K;
}
@ -9315,6 +9344,20 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M) new_type = GGML_TYPE_Q5_K;
else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) new_type = GGML_TYPE_Q6_K;
}
else if (name.find("ffn_gate") != std::string::npos) {
if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_XS && !use_more_bits(qs.i_ffn_gate, qs.n_ffn_gate)) {
new_type = GGML_TYPE_Q2_K;
}
++qs.i_ffn_gate;
}
else if (name.find("ffn_up") != std::string::npos) {
if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_XS && !use_more_bits(qs.i_ffn_up, qs.n_ffn_up)) {
new_type = GGML_TYPE_Q2_K;
}
++qs.i_ffn_up;
}
// if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K;
//}
// IK: let's remove this, else Q2_K is almost the same as Q3_K_S
//else if (name.find("ffn_gate") != std::string::npos || name.find("ffn_up") != std::string::npos) {
// if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K;
@ -9369,8 +9412,9 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
case LLAMA_FTYPE_ALL_F32: quantized_type = GGML_TYPE_F32; break;
// K-quants
case LLAMA_FTYPE_MOSTLY_Q2_K_S:
case LLAMA_FTYPE_MOSTLY_Q2_K: quantized_type = GGML_TYPE_Q2_K; break;
case LLAMA_FTYPE_MOSTLY_Q2_K_S: quantized_type = GGML_TYPE_Q2_K; break;
case LLAMA_FTYPE_MOSTLY_Q3_K_XS:
case LLAMA_FTYPE_MOSTLY_Q3_K_S:
case LLAMA_FTYPE_MOSTLY_Q3_K_M:
case LLAMA_FTYPE_MOSTLY_Q3_K_L: quantized_type = GGML_TYPE_Q3_K; break;
@ -9438,12 +9482,18 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
++qs.n_attention_wv;
}
else if (name.find("ffn_down") != std::string::npos) {
++qs.n_feed_forward_w2;
++qs.n_ffn_down;
}
else if (name.find("ffn_gate") != std::string::npos) {
++qs.n_ffn_gate;
}
else if (name.find("ffn_up") != std::string::npos) {
++qs.n_ffn_up;
}
}
if (qs.n_attention_wv != qs.n_feed_forward_w2 || (uint32_t)qs.n_attention_wv != model.hparams.n_layer) {
LLAMA_LOG_WARN("%s ============ Strange model: n_attention_wv = %d, n_feed_forward_w2 = %d, hparams.n_layer = %d\n",
__func__, qs.n_attention_wv, qs.n_feed_forward_w2, model.hparams.n_layer);
if (qs.n_attention_wv != qs.n_ffn_down || (uint32_t)qs.n_attention_wv != model.hparams.n_layer) {
LLAMA_LOG_WARN("%s ============ Strange model: n_attention_wv = %d, n_ffn_down = %d, hparams.n_layer = %d\n",
__func__, qs.n_attention_wv, qs.n_ffn_down, model.hparams.n_layer);
}
size_t total_size_org = 0;