vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2025-09-11 09:34:37 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Merge commit 'd4e0d95cf5' into concedo_experimental

Browse source 
# Conflicts:
#	.github/workflows/build.yml
#	common/CMakeLists.txt
#	ggml/src/CMakeLists.txt
#	ggml/src/ggml-opencl/CMakeLists.txt
#	ggml/src/ggml-opencl/ggml-opencl.cpp
#	ggml/src/ggml-rpc/ggml-rpc.cpp
#	scripts/sync-ggml.last
#	tests/CMakeLists.txt
This commit is contained in:
Concedo 2025-06-14 01:58:53 +08:00
commit 69e4a32ca2
18 changed files with 870 additions and 538 deletions
Download patch file Download diff file Expand all files Collapse all files

49
src/llama-model.cpp
View file

@ -2320,8 +2320,8 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
layer.attn_norm_2 = create_tensor(tn(LLM_TENSOR_ATTN_NORM_2, "weight", i), {n_embd}, TENSOR_NOT_REQUIRED);
layer.attn_norm_2_b = create_tensor(tn(LLM_TENSOR_ATTN_NORM_2, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, 0);
layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0);
layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, TENSOR_NOT_REQUIRED);
layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, layer.ffn_gate ? n_ff : n_ff * 2}, 0);
layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, 0);
@ -6143,7 +6143,7 @@ struct llm_build_bert : public llm_graph_context {
model.layers[il].ffn_gate, NULL, NULL,
model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
NULL,
LLM_FFN_GELU, LLM_FFN_PAR, il);
model.layers[il].ffn_gate ? LLM_FFN_GELU : LLM_FFN_GEGLU, LLM_FFN_PAR, il);
cb(cur, "ffn_out", il);
} else {
cur = build_ffn(cur,
@ -8957,7 +8957,6 @@ struct llm_build_mamba : public llm_graph_context {
inpL = build_inp_embd(model.tok_embd);
ggml_tensor * state_copy = build_inp_s_copy();
ggml_tensor * state_mask = build_inp_s_mask();
for (int il = 0; il < n_layer; ++il) {
// norm
@ -8966,8 +8965,7 @@ struct llm_build_mamba : public llm_graph_context {
LLM_NORM_RMS, il);
cb(cur, "attn_norm", il);
//cur = build_mamba_layer(gf, cur, state_copy, state_mask, il);
cur = build_mamba_layer(gf, cur, state_copy, state_mask, ubatch, il);
cur = build_mamba_layer(gf, cur, state_copy, ubatch, il);
if (il == n_layer - 1) {
// skip computing output for unused tokens
@ -9008,7 +9006,6 @@ struct llm_build_mamba : public llm_graph_context {
ggml_cgraph * gf,
ggml_tensor * cur,
ggml_tensor * state_copy,
ggml_tensor * state_mask,
const llama_ubatch & ubatch,
int il) const {
const auto * kv_state = static_cast<const llama_kv_cache_recurrent_state *>(mstate);
@ -9035,12 +9032,12 @@ struct llm_build_mamba : public llm_graph_context {
ggml_tensor * ssm_states_all = kv_state->get_v_l(il);
// (ab)using the KV cache to store the states
ggml_tensor * conv = build_copy_mask_state(
gf, conv_states_all, state_copy, state_mask,
ggml_tensor * conv = build_recurrent_state(
gf, conv_states_all, state_copy,
hparams.n_embd_k_s(), n_seqs);
conv = ggml_reshape_3d(ctx0, conv, d_conv - 1, d_inner, n_seqs);
ggml_tensor * ssm = build_copy_mask_state(
gf, ssm_states_all, state_copy, state_mask,
ggml_tensor * ssm = build_recurrent_state(
gf, ssm_states_all, state_copy,
hparams.n_embd_v_s(), n_seqs);
ssm = ggml_reshape_3d(ctx0, ssm, d_state, d_inner, n_seqs);
@ -11756,7 +11753,6 @@ struct llm_build_rwkv6_base : public llm_graph_context {
ggml_tensor * cur,
ggml_tensor * x_prev,
ggml_tensor * state_copy,
ggml_tensor * state_mask,
const llama_ubatch & ubatch,
int il) const {
const auto * kv_state = static_cast<const llama_kv_cache_recurrent_state *>(mstate);
@ -11880,8 +11876,8 @@ struct llm_build_rwkv6_base : public llm_graph_context {
k = ggml_sub(ctx0, k, ggml_mul(ctx0, k, w));
}
ggml_tensor * wkv_state = build_copy_mask_state(
gf, kv_state->get_v_l(il), state_copy, state_mask,
ggml_tensor * wkv_state = build_recurrent_state(
gf, kv_state->get_v_l(il), state_copy,
hparams.n_embd_v_s(), n_seqs);
ggml_tensor * wkv_output;
@ -11937,7 +11933,6 @@ struct llm_build_rwkv6 : public llm_build_rwkv6_base {
inpL = build_norm(inpL, model.tok_norm, model.tok_norm_b, LLM_NORM, -1);
ggml_tensor * state_copy = build_inp_s_copy();
ggml_tensor * state_mask = build_inp_s_mask();
const auto n_embd = hparams.n_embd;
const auto n_seq_tokens = ubatch.n_seq_tokens;
@ -11948,7 +11943,7 @@ struct llm_build_rwkv6 : public llm_build_rwkv6_base {
inpL = ggml_reshape_3d(ctx0, inpL, n_embd, n_seq_tokens, n_seqs);
ggml_tensor * token_shift = build_rwkv_token_shift_load(
gf, state_copy, state_mask, ubatch, il
gf, state_copy, ubatch, il
);
ggml_tensor * att_shift = ggml_view_3d(ctx0, token_shift, n_embd, 1, n_seqs, token_shift->nb[1], token_shift->nb[2], 0);
@ -11964,7 +11959,7 @@ struct llm_build_rwkv6 : public llm_build_rwkv6_base {
1
);
cur = build_rwkv6_time_mix(gf, att_norm, x_prev, state_copy, state_mask, ubatch, il);
cur = build_rwkv6_time_mix(gf, att_norm, x_prev, state_copy, ubatch, il);
ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
cb(ffn_inp, "ffn_inp", il);
@ -12035,7 +12030,6 @@ struct llm_build_rwkv6qwen2 : public llm_build_rwkv6_base {
inpL = build_inp_embd(model.tok_embd);
ggml_tensor * state_copy = build_inp_s_copy();
ggml_tensor * state_mask = build_inp_s_mask();
const auto n_embd = hparams.n_embd;
const auto n_seq_tokens = ubatch.n_seq_tokens;
@ -12046,7 +12040,7 @@ struct llm_build_rwkv6qwen2 : public llm_build_rwkv6_base {
inpL = ggml_reshape_3d(ctx0, inpL, n_embd, n_seq_tokens, n_seqs);
ggml_tensor * token_shift = build_rwkv_token_shift_load(
gf, state_copy, state_mask, ubatch, il
gf, state_copy, ubatch, il
);
ggml_tensor * att_norm = build_norm(inpL, layer->attn_norm, layer->attn_norm_b, LLM_NORM_RMS, il);
@ -12059,7 +12053,7 @@ struct llm_build_rwkv6qwen2 : public llm_build_rwkv6_base {
1
);
cur = build_rwkv6_time_mix(gf, att_norm, x_prev, state_copy, state_mask, ubatch, il);
cur = build_rwkv6_time_mix(gf, att_norm, x_prev, state_copy, ubatch, il);
token_shift = ggml_view_3d(ctx0, att_norm, n_embd, 1, n_seqs, att_norm->nb[1], att_norm->nb[2], (n_seq_tokens-1)*n_embd*ggml_element_size(att_norm));
ggml_build_forward_expand(gf, build_rwkv_token_shift_store(token_shift, ubatch, il));
@ -12151,7 +12145,6 @@ struct llm_build_rwkv7_base : public llm_graph_context {
ggml_tensor * cur,
ggml_tensor * x_prev,
ggml_tensor * state_copy,
ggml_tensor * state_mask,
ggml_tensor *& first_layer_value,
const llama_ubatch & ubatch,
int il) const {
@ -12234,8 +12227,8 @@ struct llm_build_rwkv7_base : public llm_graph_context {
v = ggml_reshape_3d(ctx0, v, head_size, head_count, n_tokens);
a = ggml_reshape_3d(ctx0, a, head_size, head_count, n_tokens);
ggml_tensor * wkv_state = build_copy_mask_state(
gf, kv_state->get_v_l(il), state_copy, state_mask,
ggml_tensor * wkv_state = build_recurrent_state(
gf, kv_state->get_v_l(il), state_copy,
hparams.n_embd_v_s(), n_seqs);
ggml_tensor * wkv_output = ggml_rwkv_wkv7(ctx0, r, w, k, v, ggml_neg(ctx0, kk), ggml_mul(ctx0, kk, a), wkv_state);
@ -12293,7 +12286,6 @@ struct llm_build_rwkv7 : public llm_build_rwkv7_base {
inpL = build_norm(inpL, model.tok_norm, model.tok_norm_b, LLM_NORM, -1);
ggml_tensor * state_copy = build_inp_s_copy();
ggml_tensor * state_mask = build_inp_s_mask();
const auto n_embd = hparams.n_embd;
const auto n_seq_tokens = ubatch.n_seq_tokens;
@ -12304,7 +12296,7 @@ struct llm_build_rwkv7 : public llm_build_rwkv7_base {
inpL = ggml_reshape_3d(ctx0, inpL, n_embd, n_seq_tokens, n_seqs);
ggml_tensor * token_shift = build_rwkv_token_shift_load(
gf, state_copy, state_mask, ubatch, il
gf, state_copy, ubatch, il
);
ggml_tensor * att_shift = ggml_view_3d(ctx0, token_shift, n_embd, 1, n_seqs, token_shift->nb[1], token_shift->nb[2], 0);
@ -12320,7 +12312,7 @@ struct llm_build_rwkv7 : public llm_build_rwkv7_base {
1
);
cur = build_rwkv7_time_mix(gf, att_norm, x_prev, state_copy, state_mask, v_first, ubatch, il);
cur = build_rwkv7_time_mix(gf, att_norm, x_prev, state_copy, v_first, ubatch, il);
ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
cb(ffn_inp, "ffn_inp", il);
@ -12387,7 +12379,6 @@ struct llm_build_arwkv7 : public llm_build_rwkv7_base {
inpL = build_inp_embd(model.tok_embd);
ggml_tensor * state_copy = build_inp_s_copy();
ggml_tensor * state_mask = build_inp_s_mask();
const auto n_embd = hparams.n_embd;
const auto n_seq_tokens = ubatch.n_seq_tokens;
@ -12398,7 +12389,7 @@ struct llm_build_arwkv7 : public llm_build_rwkv7_base {
inpL = ggml_reshape_3d(ctx0, inpL, n_embd, n_seq_tokens, n_seqs);
ggml_tensor * token_shift = build_rwkv_token_shift_load(
gf, state_copy, state_mask, ubatch, il
gf, state_copy, ubatch, il
);
ggml_tensor * att_norm = build_norm(inpL, layer->attn_norm, layer->attn_norm_b, LLM_NORM_RMS, il);
@ -12411,7 +12402,7 @@ struct llm_build_arwkv7 : public llm_build_rwkv7_base {
1
);
cur = build_rwkv7_time_mix(gf, att_norm, x_prev, state_copy, state_mask, v_first, ubatch, il);
cur = build_rwkv7_time_mix(gf, att_norm, x_prev, state_copy, v_first, ubatch, il);
token_shift = ggml_view_3d(ctx0, att_norm, n_embd, 1, n_seqs, att_norm->nb[1], att_norm->nb[2], (n_seq_tokens-1)*n_embd*ggml_element_size(att_norm));
ggml_build_forward_expand(gf, build_rwkv_token_shift_store(token_shift, ubatch, il));