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

llama : Add support for DeepSeek V3 (#11049)

Browse source 
* convert : extend DEEPSEEK2 model architecture to support DeepseekV3ForCausalLM by adding EXPERT_WEIGHTS_NORM and EXPERT_GATING_FUNC model parameters and FFN_EXP_PROBS_B tensor type

* vocab : add DeepSeek V3 pre-tokenizer regexes

* unicode : handle ACCENT_MARK and SYMBOL categories in regex

* llama : add DeepSeek V3 chat template, handle new model parameters and tensor types

---------

Co-authored-by: Stanisław Szymczyk <sszymczy@gmail.com>
This commit is contained in:
fairydreaming 2025-01-04 21:06:11 +01:00 • committed by GitHub
parent f922a9c542
commit 9394bbd484
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
16 changed files with 162 additions and 5 deletions
Download patch file Download diff file Expand all files Collapse all files

45
src/llama.cpp
View file

@ -1857,6 +1857,7 @@ static bool llm_load_tensors(
layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, 0);
} else {
layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0);
layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, llama_model_loader::TENSOR_NOT_REQUIRED);
if (n_expert == 0) {
throw std::runtime_error("n_expert must be > 0");
@ -2837,12 +2838,14 @@ static struct ggml_tensor * llm_build_moe_ffn(
struct ggml_tensor * up_exps,
struct ggml_tensor * gate_exps,
struct ggml_tensor * down_exps,
struct ggml_tensor * exp_probs_b,
int64_t n_expert,
int64_t n_expert_used,
llm_ffn_op_type type_op,
bool norm_w,
bool scale_w,
float w_scale,
llama_expert_gating_func_type gating_op,
const llm_build_cb & cb,
int il) {
int64_t n_embd = cur->ne[0];
@ -2851,11 +2854,31 @@ static struct ggml_tensor * llm_build_moe_ffn(
ggml_tensor * logits = llm_build_lora_mm(lctx, ctx, gate_inp, cur); // [n_expert, n_tokens]
cb(logits, "ffn_moe_logits", il);
ggml_tensor * probs = ggml_soft_max(ctx, logits); // [n_expert, n_tokens]
ggml_tensor * probs = nullptr;
switch (gating_op) {
case LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX:
{
probs = ggml_soft_max(ctx, logits); // [n_expert, n_tokens]
} break;
case LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID:
{
probs = ggml_sigmoid(ctx, logits); // [n_expert, n_tokens]
} break;
default:
GGML_ABORT("fatal error");
}
cb(probs, "ffn_moe_probs", il);
// add experts selection bias - introduced in DeepSeek V3
// leave probs unbiased as it's later used to get expert weights
ggml_tensor * selection_probs = probs;
if (exp_probs_b != nullptr) {
selection_probs = ggml_add(ctx, probs, exp_probs_b);
cb(selection_probs, "ffn_moe_probs_biased", il);
}
// select experts
ggml_tensor * selected_experts = ggml_top_k(ctx, probs, n_expert_used); // [n_expert_used, n_tokens]
ggml_tensor * selected_experts = ggml_top_k(ctx, selection_probs, n_expert_used); // [n_expert_used, n_tokens]
cb(selected_experts->src[0], "ffn_moe_argsort", il);
cb(selected_experts, "ffn_moe_topk", il);
@ -3976,9 +3999,11 @@ struct llm_build_context {
model.layers[il].ffn_up_exps,
model.layers[il].ffn_gate_exps,
model.layers[il].ffn_down_exps,
nullptr,
n_expert, n_expert_used,
LLM_FFN_SILU, true,
false, 0.0,
LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
cb, il);
cb(cur, "ffn_moe_out", il);
}
@ -4628,9 +4653,11 @@ struct llm_build_context {
model.layers[il].ffn_up_exps,
model.layers[il].ffn_gate_exps,
model.layers[il].ffn_down_exps,
nullptr,
n_expert, n_expert_used,
LLM_FFN_GELU, true,
false, 0.0,
LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
cb, il);
cb(cur, "ffn_moe_out", il);
@ -4769,9 +4796,11 @@ struct llm_build_context {
model.layers[il].ffn_up_exps,
model.layers[il].ffn_gate_exps,
model.layers[il].ffn_down_exps,
nullptr,
n_expert, n_expert_used,
LLM_FFN_SILU, true,
false, 0.0,
LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
cb, il);
cb(cur, "ffn_moe_out", il);
@ -6017,9 +6046,11 @@ struct llm_build_context {
model.layers[il].ffn_up_exps,
model.layers[il].ffn_gate_exps,
model.layers[il].ffn_down_exps,
nullptr,
n_expert, n_expert_used,
LLM_FFN_SILU, false,
false, 0.0,
LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
cb, il);
cb(cur, "ffn_moe_out", il);
@ -8142,9 +8173,11 @@ struct llm_build_context {
model.layers[il].ffn_up_exps,
model.layers[il].ffn_gate_exps,
model.layers[il].ffn_down_exps,
nullptr,
n_expert, n_expert_used,
LLM_FFN_SILU, false,
false, 0.0,
LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
cb, il);
cb(cur, "ffn_moe_out", il);
@ -8539,9 +8572,11 @@ struct llm_build_context {
model.layers[il].ffn_up_exps,
model.layers[il].ffn_gate_exps,
model.layers[il].ffn_down_exps,
nullptr,
n_expert, n_expert_used,
LLM_FFN_SILU, true,
false, 0.0,
LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
cb, il);
cb(cur, "ffn_moe_out", il);
@ -8680,9 +8715,11 @@ struct llm_build_context {
model.layers[il].ffn_up_exps,
model.layers[il].ffn_gate_exps,
model.layers[il].ffn_down_exps,
nullptr,
n_expert, n_expert_used,
LLM_FFN_SILU, false,
false, hparams.expert_weights_scale,
LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
cb, il);
cb(moe_out, "ffn_moe_out", il);
@ -8909,9 +8946,11 @@ struct llm_build_context {
model.layers[il].ffn_up_exps,
model.layers[il].ffn_gate_exps,
model.layers[il].ffn_down_exps,
model.layers[il].ffn_exp_probs_b,
n_expert, n_expert_used,
LLM_FFN_SILU, false,
LLM_FFN_SILU, hparams.expert_weights_norm,
true, hparams.expert_weights_scale,
(enum llama_expert_gating_func_type) hparams.expert_gating_func,
cb, il);
cb(moe_out, "ffn_moe_out", il);