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

whisper broke

Browse Source
This commit is contained in:
Concedo
2026-02-26 16:45:04 +08:00
parent 44182ebefe 2943210c1e
commit 749a606374
27 changed files with 1369 additions and 637 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/llama-memory-recurrent.cpp
+1 -1
View File
@@ -163,7 +163,7 @@ bool llama_memory_recurrent::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos
const auto & cell = cells[tail_id];
// partial intersection is invalid if it includes the final pos
if (0 < p0 && p0 <= cell.pos && p1 > cell.pos) {
//printf("[DEBUG] inside `llama_memory_recurrent::seq_rm`: partial intersection is invalid, so returning false\n");
//printf("[DEBUG] inside `llama_memory_recurrent::seq_rm`: partial intersection is invalid, so returning false, p0 = %d, cell.pos = %d, p1 = %d\n", p0, cell.pos, p1);
return false;
}
// invalidate tails which will be cleared
src/llama-model.cpp
+6 -1
View File
@@ -236,6 +236,7 @@ const char * llm_type_name(llm_type type) {
case LLM_TYPE_8B_A1B: return "8B.A1B";
case LLM_TYPE_16B_A1B: return "16B.A1B";
case LLM_TYPE_21B_A3B: return "21B.A3B";
case LLM_TYPE_24B_A2B: return "24B.A2B";
case LLM_TYPE_30B_A3B: return "30B.A3B";
case LLM_TYPE_31B_A3_5B: return "31B.A3.5B";
case LLM_TYPE_35B_A3B: return "35B.A3B";
@@ -2494,7 +2495,11 @@ void llama_model::load_hparams(llama_model_loader & ml) {
hparams.recurrent_layer_arr[il] = hparams.n_head_kv(il) == 0;
}
type = LLM_TYPE_8B_A1B;
switch (hparams.n_layer) {
case 24: type = LLM_TYPE_8B_A1B; break;
case 40: type = LLM_TYPE_24B_A2B; break;
default: type = LLM_TYPE_UNKNOWN;
}
} break;
case LLM_ARCH_SMALLTHINKER:
{
src/llama-model.h
+1
View File
@@ -116,6 +116,7 @@ enum llm_type {
LLM_TYPE_8B_A1B, // lfm2moe
LLM_TYPE_16B_A1B,
LLM_TYPE_21B_A3B, // Ernie MoE small
LLM_TYPE_24B_A2B, // lfm2moe
LLM_TYPE_30B_A3B,
LLM_TYPE_31B_A3_5B,
LLM_TYPE_35B_A3B, // Qwen3.5
src/models/kimi-linear.cpp
+2
View File
@@ -116,6 +116,8 @@ llm_build_kimi_linear::llm_build_kimi_linear(const llama_model & model, const ll
cur = build_norm(inpL, layer.attn_norm, NULL, LLM_NORM_RMS, il);
cb(cur, "attn_norm", il);
ggml_build_forward_expand(gf, cur);
// Check layer type by checking which tensors exist
// KDA layers have ssm_a_log tensor, MLA layers have wkv_a_mqa tensor
bool is_kda = (layer.ssm_a != nullptr);
src/models/qwen35.cpp
+2 -1
View File
@@ -29,6 +29,8 @@ llm_build_qwen35::llm_build_qwen35(const llama_model & model, const llm_graph_pa
cur = build_norm(inpL, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il);
cb(cur, "attn_norm", il);
ggml_build_forward_expand(gf, cur);
// Determine layer type and build appropriate attention mechanism
if (hparams.is_recurrent(il)) {
// Linear attention layer (gated delta net)
@@ -269,7 +271,6 @@ ggml_tensor * llm_build_qwen35::build_layer_attn_linear(
cb(state_update_target, "state_update_target", il);
ggml_build_forward_expand(gf, ggml_cpy(ctx0, last_conv_states, state_update_target));
cb(conv_states_all, "conv_states_updated", il);
ggml_tensor * state = build_rs(inp, ssm_states_all, hparams.n_embd_s(), n_seqs);
state = ggml_reshape_4d(ctx0, state, head_v_dim, head_v_dim, num_v_heads, n_seqs);
src/models/qwen35moe.cpp
+2 -1
View File
@@ -29,6 +29,8 @@ llm_build_qwen35moe::llm_build_qwen35moe(const llama_model & model, const llm_gr
cur = build_norm(inpL, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il);
cb(cur, "attn_norm", il);
ggml_build_forward_expand(gf, cur);
// Determine layer type and build appropriate attention mechanism
if (hparams.is_recurrent(il)) {
// Linear attention layer (gated delta net)
@@ -269,7 +271,6 @@ ggml_tensor * llm_build_qwen35moe ::build_layer_attn_linear(
cb(state_update_target, "state_update_target", il);
ggml_build_forward_expand(gf, ggml_cpy(ctx0, last_conv_states, state_update_target));
cb(conv_states_all, "conv_states_updated", il);
ggml_tensor * state = build_rs(inp, ssm_states_all, hparams.n_embd_s(), n_seqs);
state = ggml_reshape_4d(ctx0, state, head_v_dim, head_v_dim, num_v_heads, n_seqs);
src/models/qwen3next.cpp
+2 -1
View File
@@ -21,6 +21,8 @@ llm_build_qwen3next::llm_build_qwen3next(const llama_model & model, const llm_gr
cur = build_norm(inpL, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il);
cb(cur, "attn_norm", il);
ggml_build_forward_expand(gf, cur);
// Determine layer type and build appropriate attention mechanism
if (hparams.is_recurrent(il)) {
// Linear attention layer (gated delta net)
@@ -355,7 +357,6 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear(
cb(state_update_target, "state_update_target", il);
ggml_build_forward_expand(gf, ggml_cpy(ctx0, last_conv_states, state_update_target));
cb(conv_states_all, "conv_states_updated", il);
ggml_tensor * state = build_rs(inp, ssm_states_all, hparams.n_embd_s(), n_seqs);
state = ggml_reshape_4d(ctx0, state, head_v_dim, head_v_dim, num_v_heads, n_seqs);