Public/koboldcpp
1
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-09-20 01:31:42 +02:00
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'upstream' into concedo_experimental

Browse Source 
# Conflicts:
#	.github/workflows/build.yml
#	CMakeLists.txt
#	cmake/common.cmake
#	docs/backend/SYCL.md
#	examples/main/README.md
#	examples/speculative/speculative.cpp
#	ggml/CMakeLists.txt
#	ggml/src/CMakeLists.txt
#	ggml/src/ggml-cpu/CMakeLists.txt
#	ggml/src/ggml-musa/CMakeLists.txt
#	ggml/src/ggml-sycl/CMakeLists.txt
#	ggml/src/ggml-vulkan/vulkan-shaders/CMakeLists.txt
#	tests/test-backend-ops.cpp
This commit is contained in:
Concedo
2025-03-19 19:27:11 +08:00
parent ddaa8d5a38 0fd8487b14
commit 0c90d2ebcf
58 changed files with 4222 additions and 1537 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/llama-arch.cpp
+102 -16
View File
@@ -59,6 +59,8 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
{ LLM_ARCH_EXAONE, "exaone" },
{ LLM_ARCH_RWKV6, "rwkv6" },
{ LLM_ARCH_RWKV6QWEN2, "rwkv6qwen2" },
{ LLM_ARCH_RWKV7, "rwkv7" },
{ LLM_ARCH_ARWKV7, "arwkv7" },
{ LLM_ARCH_GRANITE, "granite" },
{ LLM_ARCH_GRANITE_MOE, "granitemoe" },
{ LLM_ARCH_CHAMELEON, "chameleon" },
@@ -110,22 +112,26 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
{ LLM_KV_EMBEDDING_SCALE, "%s.embedding_scale" },
{ LLM_KV_TOKEN_SHIFT_COUNT, "%s.token_shift_count" },
{ LLM_KV_ATTENTION_HEAD_COUNT, "%s.attention.head_count" },
{ LLM_KV_ATTENTION_HEAD_COUNT_KV, "%s.attention.head_count_kv" },
{ LLM_KV_ATTENTION_MAX_ALIBI_BIAS, "%s.attention.max_alibi_bias" },
{ LLM_KV_ATTENTION_CLAMP_KQV, "%s.attention.clamp_kqv" },
{ LLM_KV_ATTENTION_KEY_LENGTH, "%s.attention.key_length" },
{ LLM_KV_ATTENTION_VALUE_LENGTH, "%s.attention.value_length" },
{ LLM_KV_ATTENTION_LAYERNORM_EPS, "%s.attention.layer_norm_epsilon" },
{ LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, "%s.attention.layer_norm_rms_epsilon" },
{ LLM_KV_ATTENTION_GROUPNORM_EPS, "%s.attention.group_norm_epsilon" },
{ LLM_KV_ATTENTION_GROUPNORM_GROUPS, "%s.attention.group_norm_groups" },
{ LLM_KV_ATTENTION_CAUSAL, "%s.attention.causal" },
{ LLM_KV_ATTENTION_Q_LORA_RANK, "%s.attention.q_lora_rank" },
{ LLM_KV_ATTENTION_KV_LORA_RANK, "%s.attention.kv_lora_rank" },
{ LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT, "%s.attention.relative_buckets_count" },
{ LLM_KV_ATTENTION_SLIDING_WINDOW, "%s.attention.sliding_window" },
{ LLM_KV_ATTENTION_SCALE, "%s.attention.scale" },
{ LLM_KV_ATTENTION_HEAD_COUNT, "%s.attention.head_count" },
{ LLM_KV_ATTENTION_HEAD_COUNT_KV, "%s.attention.head_count_kv" },
{ LLM_KV_ATTENTION_MAX_ALIBI_BIAS, "%s.attention.max_alibi_bias" },
{ LLM_KV_ATTENTION_CLAMP_KQV, "%s.attention.clamp_kqv" },
{ LLM_KV_ATTENTION_KEY_LENGTH, "%s.attention.key_length" },
{ LLM_KV_ATTENTION_VALUE_LENGTH, "%s.attention.value_length" },
{ LLM_KV_ATTENTION_LAYERNORM_EPS, "%s.attention.layer_norm_epsilon" },
{ LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, "%s.attention.layer_norm_rms_epsilon" },
{ LLM_KV_ATTENTION_GROUPNORM_EPS, "%s.attention.group_norm_epsilon" },
{ LLM_KV_ATTENTION_GROUPNORM_GROUPS, "%s.attention.group_norm_groups" },
{ LLM_KV_ATTENTION_CAUSAL, "%s.attention.causal" },
{ LLM_KV_ATTENTION_Q_LORA_RANK, "%s.attention.q_lora_rank" },
{ LLM_KV_ATTENTION_KV_LORA_RANK, "%s.attention.kv_lora_rank" },
{ LLM_KV_ATTENTION_DECAY_LORA_RANK, "%s.attention.decay_lora_rank" },
{ LLM_KV_ATTENTION_ICLR_LORA_RANK, "%s.attention.iclr_lora_rank" },
{ LLM_KV_ATTENTION_VALUE_RESIDUAL_MIX_LORA_RANK, "%s.attention.value_residual_mix_lora_rank" },
{ LLM_KV_ATTENTION_GATE_LORA_RANK, "%s.attention.gate_lora_rank" },
{ LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT, "%s.attention.relative_buckets_count" },
{ LLM_KV_ATTENTION_SLIDING_WINDOW, "%s.attention.sliding_window" },
{ LLM_KV_ATTENTION_SCALE, "%s.attention.scale" },
{ LLM_KV_ROPE_DIMENSION_COUNT, "%s.rope.dimension_count" },
{ LLM_KV_ROPE_DIMENSION_SECTIONS, "%s.rope.dimension_sections" },
@@ -1238,6 +1244,74 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
{ LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
},
},
{
LLM_ARCH_RWKV7,
{
{ LLM_TENSOR_TOKEN_EMBD, "token_embd" },
{ LLM_TENSOR_TOKEN_EMBD_NORM, "token_embd_norm" },
{ LLM_TENSOR_OUTPUT_NORM, "output_norm" },
{ LLM_TENSOR_OUTPUT, "output" },
{ LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" },
{ LLM_TENSOR_ATTN_NORM_2, "blk.%d.attn_norm_2" },
{ LLM_TENSOR_TIME_MIX_W0, "blk.%d.time_mix_w0" },
{ LLM_TENSOR_TIME_MIX_W1, "blk.%d.time_mix_w1" },
{ LLM_TENSOR_TIME_MIX_W2, "blk.%d.time_mix_w2" },
{ LLM_TENSOR_TIME_MIX_A0, "blk.%d.time_mix_a0" },
{ LLM_TENSOR_TIME_MIX_A1, "blk.%d.time_mix_a1" },
{ LLM_TENSOR_TIME_MIX_A2, "blk.%d.time_mix_a2" },
{ LLM_TENSOR_TIME_MIX_V0, "blk.%d.time_mix_v0" },
{ LLM_TENSOR_TIME_MIX_V1, "blk.%d.time_mix_v1" },
{ LLM_TENSOR_TIME_MIX_V2, "blk.%d.time_mix_v2" },
{ LLM_TENSOR_TIME_MIX_G1, "blk.%d.time_mix_g1" },
{ LLM_TENSOR_TIME_MIX_G2, "blk.%d.time_mix_g2" },
{ LLM_TENSOR_TIME_MIX_K_K, "blk.%d.time_mix_k_k" },
{ LLM_TENSOR_TIME_MIX_K_A, "blk.%d.time_mix_k_a" },
{ LLM_TENSOR_TIME_MIX_R_K, "blk.%d.time_mix_r_k" },
{ LLM_TENSOR_TIME_MIX_LERP_FUSED, "blk.%d.time_mix_lerp_fused" },
{ LLM_TENSOR_TIME_MIX_KEY, "blk.%d.time_mix_key" },
{ LLM_TENSOR_TIME_MIX_VALUE, "blk.%d.time_mix_value" },
{ LLM_TENSOR_TIME_MIX_RECEPTANCE, "blk.%d.time_mix_receptance" },
{ LLM_TENSOR_TIME_MIX_LN, "blk.%d.time_mix_ln" },
{ LLM_TENSOR_TIME_MIX_OUTPUT, "blk.%d.time_mix_output" },
{ LLM_TENSOR_CHANNEL_MIX_LERP_K, "blk.%d.channel_mix_lerp_k" },
{ LLM_TENSOR_CHANNEL_MIX_KEY, "blk.%d.channel_mix_key" },
{ LLM_TENSOR_CHANNEL_MIX_VALUE, "blk.%d.channel_mix_value" },
},
},
{
LLM_ARCH_ARWKV7,
{
{ LLM_TENSOR_TOKEN_EMBD, "token_embd" },
{ LLM_TENSOR_TOKEN_EMBD_NORM, "token_embd_norm" },
{ LLM_TENSOR_OUTPUT_NORM, "output_norm" },
{ LLM_TENSOR_OUTPUT, "output" },
{ LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" },
{ LLM_TENSOR_TIME_MIX_W0, "blk.%d.time_mix_w0" },
{ LLM_TENSOR_TIME_MIX_W1, "blk.%d.time_mix_w1" },
{ LLM_TENSOR_TIME_MIX_W2, "blk.%d.time_mix_w2" },
{ LLM_TENSOR_TIME_MIX_A0, "blk.%d.time_mix_a0" },
{ LLM_TENSOR_TIME_MIX_A1, "blk.%d.time_mix_a1" },
{ LLM_TENSOR_TIME_MIX_A2, "blk.%d.time_mix_a2" },
{ LLM_TENSOR_TIME_MIX_V0, "blk.%d.time_mix_v0" },
{ LLM_TENSOR_TIME_MIX_V1, "blk.%d.time_mix_v1" },
{ LLM_TENSOR_TIME_MIX_V2, "blk.%d.time_mix_v2" },
{ LLM_TENSOR_TIME_MIX_G1, "blk.%d.time_mix_g1" },
{ LLM_TENSOR_TIME_MIX_G2, "blk.%d.time_mix_g2" },
{ LLM_TENSOR_TIME_MIX_K_K, "blk.%d.time_mix_k_k" },
{ LLM_TENSOR_TIME_MIX_K_A, "blk.%d.time_mix_k_a" },
{ LLM_TENSOR_TIME_MIX_R_K, "blk.%d.time_mix_r_k" },
{ LLM_TENSOR_TIME_MIX_LERP_FUSED, "blk.%d.time_mix_lerp_fused" },
{ LLM_TENSOR_TIME_MIX_KEY, "blk.%d.time_mix_key" },
{ LLM_TENSOR_TIME_MIX_VALUE, "blk.%d.time_mix_value" },
{ LLM_TENSOR_TIME_MIX_RECEPTANCE, "blk.%d.time_mix_receptance" },
{ LLM_TENSOR_TIME_MIX_LN, "blk.%d.time_mix_ln" },
{ LLM_TENSOR_TIME_MIX_OUTPUT, "blk.%d.time_mix_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_ARCH_GRANITE,
{
@@ -1397,6 +1471,12 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
{LLM_TENSOR_SSM_OUT, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_TIME_MIX_W1, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_TIME_MIX_W2, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_TIME_MIX_A1, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_TIME_MIX_A2, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_TIME_MIX_V1, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_TIME_MIX_V2, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_TIME_MIX_G1, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_TIME_MIX_G2, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_TIME_MIX_DECAY_W1, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_TIME_MIX_DECAY_W2, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_TIME_MIX_KEY, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
@@ -1415,6 +1495,9 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
{LLM_TENSOR_TIME_MIX_LN, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
{LLM_TENSOR_CHANNEL_MIX_LERP_K, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
{LLM_TENSOR_CHANNEL_MIX_LERP_R, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
{LLM_TENSOR_TIME_MIX_K_K, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
{LLM_TENSOR_TIME_MIX_K_A, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
{LLM_TENSOR_TIME_MIX_R_K, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
{LLM_TENSOR_TIME_MIX_LERP_W, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
{LLM_TENSOR_TIME_MIX_LERP_K, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
{LLM_TENSOR_TIME_MIX_LERP_V, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
@@ -1422,6 +1505,9 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
{LLM_TENSOR_TIME_MIX_LERP_G, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
{LLM_TENSOR_TIME_MIX_LERP_FUSED, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
{LLM_TENSOR_TIME_MIX_DECAY, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
{LLM_TENSOR_TIME_MIX_W0, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
{LLM_TENSOR_TIME_MIX_A0, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
{LLM_TENSOR_TIME_MIX_V0, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
{LLM_TENSOR_TIME_MIX_FIRST, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_RWKV_WKV6}},
{LLM_TENSOR_ATTN_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
{LLM_TENSOR_ATTN_NORM_2, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
src/llama-arch.h
+18
View File
@@ -63,6 +63,8 @@ enum llm_arch {
LLM_ARCH_EXAONE,
LLM_ARCH_RWKV6,
LLM_ARCH_RWKV6QWEN2,
LLM_ARCH_RWKV7,
LLM_ARCH_ARWKV7,
LLM_ARCH_GRANITE,
LLM_ARCH_GRANITE_MOE,
LLM_ARCH_CHAMELEON,
@@ -127,6 +129,10 @@ enum llm_kv {
LLM_KV_ATTENTION_CAUSAL,
LLM_KV_ATTENTION_Q_LORA_RANK,
LLM_KV_ATTENTION_KV_LORA_RANK,
LLM_KV_ATTENTION_DECAY_LORA_RANK,
LLM_KV_ATTENTION_ICLR_LORA_RANK,
LLM_KV_ATTENTION_VALUE_RESIDUAL_MIX_LORA_RANK,
LLM_KV_ATTENTION_GATE_LORA_RANK,
LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT,
LLM_KV_ATTENTION_SLIDING_WINDOW,
LLM_KV_ATTENTION_SCALE,
@@ -250,8 +256,20 @@ enum llm_tensor {
LLM_TENSOR_SSM_A,
LLM_TENSOR_SSM_D,
LLM_TENSOR_SSM_OUT,
LLM_TENSOR_TIME_MIX_W0,
LLM_TENSOR_TIME_MIX_W1,
LLM_TENSOR_TIME_MIX_W2,
LLM_TENSOR_TIME_MIX_A0,
LLM_TENSOR_TIME_MIX_A1,
LLM_TENSOR_TIME_MIX_A2,
LLM_TENSOR_TIME_MIX_V0,
LLM_TENSOR_TIME_MIX_V1,
LLM_TENSOR_TIME_MIX_V2,
LLM_TENSOR_TIME_MIX_G1,
LLM_TENSOR_TIME_MIX_G2,
LLM_TENSOR_TIME_MIX_K_K,
LLM_TENSOR_TIME_MIX_K_A,
LLM_TENSOR_TIME_MIX_R_K,
LLM_TENSOR_TIME_MIX_LERP_X,
LLM_TENSOR_TIME_MIX_LERP_W,
LLM_TENSOR_TIME_MIX_LERP_K,
src/llama-context.cpp
+19 -2
View File
@@ -285,11 +285,15 @@ llama_context::llama_context(
// reserve worst-case graph
if (!hparams.vocab_only) {
uint32_t n_seqs = 1; // TODO: worst-case number of sequences
uint32_t n_tokens = std::min(cparams.n_ctx, cparams.n_ubatch);
const uint32_t n_seqs = 1; // TODO: worst-case number of sequences
const uint32_t n_tokens = std::min(cparams.n_ctx, cparams.n_ubatch);
llama_token token = model.vocab.token_bos(); // not actually used by llama_build_graph, but required to choose between token and embedding inputs graph
// restore later
// TODO: something cleaner
const auto n_outputs_save = n_outputs;
// max number of outputs
n_outputs = n_tokens;
@@ -341,6 +345,8 @@ llama_context::llama_context(
}
}
n_outputs = n_outputs_save;
for (size_t i = 0; i < backend_ptrs.size(); ++i) {
ggml_backend_t backend = backend_ptrs[i];
ggml_backend_buffer_type_t buft = backend_buft[i];
@@ -1052,6 +1058,13 @@ int llama_context::encode(llama_batch & inp_batch) {
ggml_backend_sched_reset(sched.get());
ggml_backend_sched_set_eval_callback(sched.get(), cparams.cb_eval, cparams.cb_eval_user_data);
const auto causal_attn_org = cparams.causal_attn;
// always use non-causal attention for encoder graphs
// TODO: this is a tmp solution until we have a proper way to support enc-dec models
// ref: https://github.com/ggml-org/llama.cpp/pull/12181#issuecomment-2730451223
cparams.causal_attn = false;
auto * gf = graph_init();
auto res = graph_build(ctx_compute.get(), gf, ubatch, LLM_GRAPH_TYPE_ENCODER);
@@ -1059,6 +1072,8 @@ int llama_context::encode(llama_batch & inp_batch) {
res->set_inputs(&ubatch);
cparams.causal_attn = causal_attn_org;
const auto compute_status = graph_compute(gf, n_tokens > 1);
switch (compute_status) {
case GGML_STATUS_SUCCESS:
@@ -1129,6 +1144,8 @@ int llama_context::encode(llama_batch & inp_batch) {
if (model.arch == LLM_ARCH_T5 && t_embd) {
//cross.t_embd = t_embd;
synchronize();
cross.n_embd = t_embd->ne[0];
cross.n_enc = t_embd->ne[1];
cross.v_embd.resize(cross.n_embd*cross.n_enc);
src/llama-graph.cpp
+1 -1
View File
@@ -1378,7 +1378,7 @@ ggml_tensor * llm_graph_context::build_attn(
// note: storing RoPE-ed version of K in the KV cache
ggml_build_forward_expand(gf, ggml_cpy(ctx0, k_cur, k_cache_view));
assert(v_cur->ne[0] == n_embd_v_gqa && v_cur->ne[1] == n_tokens);
v_cur = ggml_reshape_2d(ctx0, v_cur, n_embd_v_gqa, n_tokens);
ggml_tensor * v_cache_view = nullptr;
src/llama-graph.h
+12 -12
View File
@@ -487,9 +487,9 @@ struct llm_graph_context {
ggml_tensor * build_attn_mha(
ggml_cgraph * gf,
ggml_tensor * q,
ggml_tensor * k,
ggml_tensor * v,
ggml_tensor * q, // [n_embd_head_q, n_tokens, n_head_q]
ggml_tensor * k, // [n_embd_head_k, n_tokens, n_head_k]
ggml_tensor * v, // [n_embd_head_v, n_tokens, n_head_v] (v_trans == false)
ggml_tensor * kq_b,
ggml_tensor * kq_mask,
bool v_trans,
@@ -502,9 +502,9 @@ struct llm_graph_context {
ggml_cgraph * gf,
ggml_tensor * wo,
ggml_tensor * wo_b,
ggml_tensor * q_cur,
ggml_tensor * k_cur,
ggml_tensor * v_cur,
ggml_tensor * q_cur, // [n_embd_head_q, n_head_q, n_tokens]
ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens]
ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens]
ggml_tensor * kq_b,
float kq_scale,
int il) const;
@@ -516,9 +516,9 @@ struct llm_graph_context {
ggml_cgraph * gf,
ggml_tensor * wo,
ggml_tensor * wo_b,
ggml_tensor * q_cur,
ggml_tensor * k_cur,
ggml_tensor * v_cur,
ggml_tensor * q_cur, // [n_embd_head_q, n_head_q, n_tokens]
ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens]
ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens]
ggml_tensor * kq_b,
float kq_scale,
int il) const;
@@ -530,9 +530,9 @@ struct llm_graph_context {
ggml_cgraph * gf,
ggml_tensor * wo,
ggml_tensor * wo_b,
ggml_tensor * q_cur,
ggml_tensor * k_cur,
ggml_tensor * v_cur,
ggml_tensor * q_cur, // [n_embd_head_q, n_head_q, n_tokens]
ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens]
ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens]
ggml_tensor * kq_b,
float kq_scale,
int il) const;
src/llama-hparams.h
+4
View File
@@ -76,6 +76,10 @@ struct llama_hparams {
uint32_t time_decay_extra_dim = 0;
uint32_t wkv_head_size = 0;
uint32_t token_shift_count = 2;
uint32_t n_lora_decay = 0;
uint32_t n_lora_iclr = 0;
uint32_t n_lora_value_res_mix = 0;
uint32_t n_lora_gate = 0;
float rope_attn_factor = 1.0f;
float rope_freq_base_train;
src/llama-model.cpp
+1011 -285
View File
File diff suppressed because it is too large Load Diff
src/llama-model.h
+16
View File
@@ -29,6 +29,7 @@ enum llm_type {
LLM_TYPE_109M,
LLM_TYPE_137M,
LLM_TYPE_160M,
LLM_TYPE_190M,
LLM_TYPE_220M,
LLM_TYPE_250M,
LLM_TYPE_270M,
@@ -45,6 +46,7 @@ enum llm_type {
LLM_TYPE_1_6B,
LLM_TYPE_2B,
LLM_TYPE_2_8B,
LLM_TYPE_2_9B,
LLM_TYPE_3B,
LLM_TYPE_4B,
LLM_TYPE_6B,
@@ -260,6 +262,20 @@ struct llama_layer {
struct ggml_tensor * time_mix_receptance_b = nullptr;
struct ggml_tensor * time_mix_gate = nullptr;
// rwkv7
struct ggml_tensor * time_mix_w0 = nullptr;
struct ggml_tensor * time_mix_a0 = nullptr;
struct ggml_tensor * time_mix_a1 = nullptr;
struct ggml_tensor * time_mix_a2 = nullptr;
struct ggml_tensor * time_mix_v0 = nullptr;
struct ggml_tensor * time_mix_v1 = nullptr;
struct ggml_tensor * time_mix_v2 = nullptr;
struct ggml_tensor * time_mix_g1 = nullptr;
struct ggml_tensor * time_mix_g2 = nullptr;
struct ggml_tensor * time_mix_k_k = nullptr;
struct ggml_tensor * time_mix_k_a = nullptr;
struct ggml_tensor * time_mix_r_k = nullptr;
struct ggml_tensor * time_mix_ln = nullptr;
struct ggml_tensor * time_mix_ln_b = nullptr;
struct ggml_tensor * time_mix_output = nullptr;
src/llama-quant.cpp
+10 -1
View File
@@ -759,10 +759,19 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
// NOTE: can't use LLM_TN here because the layer number is not known
quantize &= name.find("ssm_conv1d.weight") == std::string::npos;
// do not quantize RWKV's time_mix_first tensors
// do not quantize RWKV's small yet 2D weights
quantize &= name.find("time_mix_first.weight") == std::string::npos;
quantize &= name.find("time_mix_w0.weight") == std::string::npos;
quantize &= name.find("time_mix_w1.weight") == std::string::npos;
quantize &= name.find("time_mix_w2.weight") == std::string::npos;
quantize &= name.find("time_mix_v0.weight") == std::string::npos;
quantize &= name.find("time_mix_v1.weight") == std::string::npos;
quantize &= name.find("time_mix_v2.weight") == std::string::npos;
quantize &= name.find("time_mix_a0.weight") == std::string::npos;
quantize &= name.find("time_mix_a1.weight") == std::string::npos;
quantize &= name.find("time_mix_a2.weight") == std::string::npos;
quantize &= name.find("time_mix_g1.weight") == std::string::npos;
quantize &= name.find("time_mix_g2.weight") == std::string::npos;
quantize &= name.find("time_mix_decay_w1.weight") == std::string::npos;
quantize &= name.find("time_mix_decay_w2.weight") == std::string::npos;
quantize &= name.find("time_mix_lerp_fused.weight") == std::string::npos;