vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-05-14 16:43:28 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 4

unstable merge

Browse source
This commit is contained in:
Concedo 2025-12-28 23:03:03 +08:00
commit 82d562ad7b
34 changed files with 1025 additions and 371 deletions
Download patch file Download diff file Expand all files Collapse all files

22
src/llama-arch.cpp
View file

@ -115,6 +115,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
{ LLM_ARCH_RND1, "rnd1" },
{ LLM_ARCH_PANGU_EMBED, "pangu-embedded" },
{ LLM_ARCH_MISTRAL3, "mistral3" },
{ LLM_ARCH_MIMO2, "mimo2" },
{ LLM_ARCH_LLAMA_EMBED, "llama-embed" },
{ LLM_ARCH_UNKNOWN, "(unknown)" },
};
@ -2190,6 +2191,27 @@ static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
LLM_TENSOR_VISEXP_FFN_DOWN,
LLM_TENSOR_VISEXP_FFN_UP,
};
case LLM_ARCH_MIMO2:
return {
LLM_TENSOR_TOKEN_EMBD,
LLM_TENSOR_OUTPUT_NORM,
LLM_TENSOR_OUTPUT,
LLM_TENSOR_ATTN_NORM,
LLM_TENSOR_ATTN_Q,
LLM_TENSOR_ATTN_K,
LLM_TENSOR_ATTN_V,
LLM_TENSOR_ATTN_SINKS,
LLM_TENSOR_ATTN_OUT,
LLM_TENSOR_FFN_NORM,
LLM_TENSOR_FFN_GATE,
LLM_TENSOR_FFN_DOWN,
LLM_TENSOR_FFN_UP,
LLM_TENSOR_FFN_GATE_INP,
LLM_TENSOR_FFN_GATE_EXPS,
LLM_TENSOR_FFN_DOWN_EXPS,
LLM_TENSOR_FFN_UP_EXPS,
LLM_TENSOR_FFN_EXP_PROBS_B,
};
case LLM_ARCH_GPTJ:
case LLM_ARCH_UNKNOWN:
return {

1
src/llama-arch.h
View file

@ -119,6 +119,7 @@ enum llm_arch {
LLM_ARCH_RND1,
LLM_ARCH_PANGU_EMBED,
LLM_ARCH_MISTRAL3,
LLM_ARCH_MIMO2,
LLM_ARCH_LLAMA_EMBED,
LLM_ARCH_UNKNOWN,
};

7
src/llama-hparams.h
View file

@ -123,10 +123,11 @@ struct llama_hparams {
llama_swa_type swa_type = LLAMA_SWA_TYPE_NONE;
// the size of the sliding window (0 - no SWA)
uint32_t n_swa = 0;
// if swa_layers[il] == true, then layer il is SWA
// if swa_layers[il] == false, then layer il is dense (i.e. non-SWA)
// if swa_layers[il] == 1, then layer il is SWA
// if swa_layers[il] == 0, then layer il is dense (i.e. non-SWA)
// by default, all layers are dense
std::array<bool, LLAMA_MAX_LAYERS> swa_layers;
// note: using uint32_t type for compatibility reason
std::array<uint32_t, LLAMA_MAX_LAYERS> swa_layers;
// for State Space Models
uint32_t ssm_d_conv = 0;

61
src/llama-model.cpp
View file

@ -79,6 +79,7 @@
#include "models/llama-iswa.cpp"
#include "models/llama.cpp"
#include "models/mamba.cpp"
#include "models/mimo2-iswa.cpp"
#include "models/minicpm3.cpp"
#include "models/minimax-m2.cpp"
#include "models/modern-bert.cpp"
@ -236,6 +237,7 @@ const char * llm_type_name(llm_type type) {
case LLM_TYPE_230B_A10B: return "230B.A10B";
case LLM_TYPE_235B_A22B: return "235B.A22B";
case LLM_TYPE_300B_A47B: return "300B.A47B";
case LLM_TYPE_310B_A15B: return "310B.A15B";
case LLM_TYPE_355B_A32B: return "355B.A32B";
case LLM_TYPE_E2B: return "E2B";
case LLM_TYPE_E4B: return "E4B";
@ -2445,6 +2447,22 @@ void llama_model::load_hparams(llama_model_loader & ml) {
default: type = LLM_TYPE_UNKNOWN;
}
} break;
case LLM_ARCH_MIMO2:
{
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp);
ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa);
ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa);
ml.get_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, hparams.swa_layers, hparams.n_layer);
switch (hparams.n_layer) {
case 48: type = LLM_TYPE_310B_A15B; break;
default: type = LLM_TYPE_UNKNOWN;
}
} break;
default: throw std::runtime_error("unsupported model architecture");
}
@ -6807,6 +6825,44 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), { hparams.n_ff_shexp, n_embd }, 0);
}
} break;
case LLM_ARCH_MIMO2:
{
tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
// output
output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, 0);
for (int i = 0; i < n_layer; ++i) {
auto & layer = layers[i];
uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(i);
uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(i);
uint32_t n_head = hparams.n_head(i);
layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), { n_embd, n_embd_head_k * n_head }, 0);
layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), { n_embd, n_embd_k_gqa }, 0);
layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), { n_embd, n_embd_v_gqa }, 0);
layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd_head_v * n_head, n_embd }, 0);
layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
layer.attn_sinks = create_tensor(tn(LLM_TENSOR_ATTN_SINKS, "weight", i), {n_head}, TENSOR_NOT_REQUIRED);
layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
// non-MoE branch
layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, TENSOR_NOT_REQUIRED);
layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, TENSOR_NOT_REQUIRED);
layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, TENSOR_NOT_REQUIRED);
// MoE branch
int64_t n_ff_exp = hparams.n_ff_exp;
layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, TENSOR_NOT_REQUIRED);
layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff_exp, n_expert}, TENSOR_NOT_REQUIRED);
layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp, n_embd, n_expert}, TENSOR_NOT_REQUIRED);
layer.ffn_up_exps = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), {n_embd, n_ff_exp, n_expert}, TENSOR_NOT_REQUIRED);
layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, TENSOR_NOT_REQUIRED);
}
} break;
default:
throw std::runtime_error("unknown architecture");
}
@ -7873,6 +7929,10 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
{
llm = std::make_unique<llm_build_mistral3>(*this, params);
} break;
case LLM_ARCH_MIMO2:
{
llm = std::make_unique<llm_build_mimo2_iswa>(*this, params);
} break;
default:
GGML_ABORT("fatal error");
}
@ -8103,6 +8163,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
case LLM_ARCH_PANGU_EMBED:
case LLM_ARCH_AFMOE:
case LLM_ARCH_QWEN3NEXT:
case LLM_ARCH_MIMO2:
return LLAMA_ROPE_TYPE_NEOX;
case LLM_ARCH_QWEN2VL:

1
src/llama-model.h
View file

@ -123,6 +123,7 @@ enum llm_type {
LLM_TYPE_230B_A10B, // Minimax M2
LLM_TYPE_235B_A22B,
LLM_TYPE_300B_A47B, // Ernie MoE big
LLM_TYPE_310B_A15B, // /MiMo-V2-Flash
LLM_TYPE_355B_A32B, // GLM-4.5
LLM_TYPE_E2B,
LLM_TYPE_E4B,

45
src/llama.cpp
View file

@ -205,12 +205,11 @@ static void llama_params_fit_impl(
}
}
int64_t sum_total = 0;
int64_t sum_free = 0;
int64_t sum_projected_free = 0;
int64_t min_projected_free = INT64_MAX;
int64_t sum_projected_used = 0;
int64_t sum_projected_model = 0;
int64_t sum_projected_ctx = 0;
if (nd > 1) {
LLAMA_LOG_INFO("%s: projected memory use with initial parameters [MiB]:\n", __func__);
@ -221,12 +220,11 @@ static void llama_params_fit_impl(
const int64_t projected_used = dmd.mb.total();
const int64_t projected_free = dmd.free - projected_used;
sum_total += dmd.total;
sum_free += dmd.free;
sum_projected_used += projected_used;
sum_projected_free += projected_free;
min_projected_free = std::min(min_projected_free, projected_free);
sum_projected_model += dmd.mb.model;
sum_projected_ctx += dmd.mb.context;
if (nd > 1) {
LLAMA_LOG_INFO("%s: - %s: %6" PRId64 " total, %6" PRId64 " used, %6" PRId64 " %s\n",
@ -234,10 +232,9 @@ static void llama_params_fit_impl(
projected_free >= 0 ? "surplus" : "deficit");
}
}
assert(sum_total >= 0 && sum_projected_used >= 0 && sum_projected_ctx >= 0);
assert(sum_projected_used >= sum_projected_ctx);
assert(sum_free >= 0 && sum_projected_used >= 0);
LLAMA_LOG_INFO("%s: projected to use %" PRId64 " MiB of device memory vs. %" PRId64 " MiB of free device memory\n",
__func__, sum_projected_used/MiB, sum_total/MiB);
__func__, sum_projected_used/MiB, sum_free/MiB);
if (min_projected_free >= margin) {
if (nd == 1) {
LLAMA_LOG_INFO("%s: will leave %" PRId64 " >= %" PRId64 " MiB of free device memory, no changes needed\n",
@ -260,9 +257,7 @@ static void llama_params_fit_impl(
__func__, margin/MiB, -global_surplus/MiB);
if (cparams->n_ctx == 0) {
if (hp_nct > n_ctx_min) {
const int64_t bytes_per_ctx = sum_projected_ctx / hp_nct;
int64_t memory_reduction = -global_surplus;
int64_t sum_used_target = sum_free - nd*margin_s;
if (nd > 1) {
// for multiple devices we need to be more conservative in terms of how much context we think can fit:
// - for dense models only whole layers can be assigned to devices
@ -270,24 +265,34 @@ static void llama_params_fit_impl(
// - on average we expect a waste of 0.5 layers/tensors per device
// - use slightly more than the expected average for nd devices to be safe
const int64_t model_per_layer = sum_projected_model / std::min(uint32_t(mparams->n_gpu_layers), hp_ngl);
memory_reduction += (nd + 1) * model_per_layer / (hp_nex == 0 ? 2 : 6);
sum_used_target -= (nd + 1) * model_per_layer / (hp_nex == 0 ? 2 : 6);
}
uint32_t ctx_reduction = std::min(uint32_t((memory_reduction + bytes_per_ctx - 1) / bytes_per_ctx), hp_nct - n_ctx_min);
cparams->n_ctx = hp_nct - ctx_reduction;
cparams->n_ctx = std::max(cparams->n_ctx - cparams->n_ctx % 256, n_ctx_min); // round down context for CUDA backend
int64_t sum_projected_used_min_ctx = 0;
cparams->n_ctx = n_ctx_min;
const dmds_t dmds_min_ctx = llama_get_device_memory_data(path_model, mparams, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
for (const auto & dmd : dmds_min_ctx) {
sum_projected_used_min_ctx += dmd.mb.total();
}
if (sum_used_target > sum_projected_used_min_ctx) {
// linear interpolation between minimum and maximum context size:
cparams->n_ctx += (hp_nct - n_ctx_min) * (sum_used_target - sum_projected_used_min_ctx)
/ (sum_projected_used - sum_projected_used_min_ctx);
cparams->n_ctx = std::max(cparams->n_ctx - cparams->n_ctx % 256, n_ctx_min); // round down context for CUDA backend
ctx_reduction = hp_nct - cparams->n_ctx;
memory_reduction = ctx_reduction * bytes_per_ctx;
global_surplus += memory_reduction;
LLAMA_LOG_INFO("%s: context size reduced from %" PRIu32 " to %" PRIu32 " -> need %" PRId64 " MiB less memory in total\n",
__func__, hp_nct, cparams->n_ctx, memory_reduction/MiB);
if (global_surplus >= 0) {
const int64_t bytes_per_ctx = (sum_projected_used - sum_projected_used_min_ctx) / (hp_nct - n_ctx_min);
const int64_t memory_reduction = (hp_nct - cparams->n_ctx) * bytes_per_ctx;
LLAMA_LOG_INFO("%s: context size reduced from %" PRIu32 " to %" PRIu32 " -> need %" PRId64 " MiB less memory in total\n",
__func__, hp_nct, cparams->n_ctx, memory_reduction/MiB);
if (nd == 1) {
LLAMA_LOG_INFO("%s: entire model can be fit by reducing context\n", __func__);
return;
}
LLAMA_LOG_INFO("%s: entire model should be fit across devices by reducing context\n", __func__);
} else {
const int64_t memory_reduction = sum_projected_used - sum_projected_used_min_ctx;
LLAMA_LOG_INFO("%s: context size reduced from %" PRIu32 " to %" PRIu32 " -> need %" PRId64 " MiB less memory in total\n",
__func__, hp_nct, cparams->n_ctx, memory_reduction/MiB);
}
} else {
LLAMA_LOG_INFO("%s: default model context size is %" PRIu32 " which is <= the min. context size of %" PRIu32 " -> no change\n",

123
src/models/mimo2-iswa.cpp Normal file
View file

@ -0,0 +1,123 @@
#include "models.h"
llm_build_mimo2_iswa::llm_build_mimo2_iswa(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
ggml_tensor * cur;
ggml_tensor * inpL;
inpL = build_inp_embd(model.tok_embd);
ggml_tensor * inp_pos = build_inp_pos();
auto * inp_attn = build_attn_inp_kv_iswa();
ggml_tensor * inp_out_ids = build_inp_out_ids();
for (int il = 0; il < n_layer; ++il) {
ggml_tensor * inpSA = inpL;
uint32_t n_head_l = hparams.n_head(il);
uint32_t n_head_kv_l = hparams.n_head_kv(il);
const float freq_base_l = model.get_rope_freq_base(cparams, il);
const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
cur = inpL;
// self_attention
{
cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
cb(cur, "attn_norm", il);
// compute Q and K and RoPE them
ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
cb(Qcur, "Qcur", il);
ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
cb(Kcur, "Kcur", il);
ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
cb(Vcur, "Vcur", il);
Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head_k, n_head_l, n_tokens);
Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head_k, n_head_kv_l, n_tokens);
Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head_v, n_head_kv_l, n_tokens);
Qcur = ggml_rope_ext(
ctx0, Qcur, inp_pos, nullptr,
n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
ext_factor, attn_factor, beta_fast, beta_slow
);
Kcur = ggml_rope_ext(
ctx0, Kcur, inp_pos, nullptr,
n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
ext_factor, attn_factor, beta_fast, beta_slow
);
cb(Qcur, "Qcur", il);
cb(Kcur, "Kcur", il);
cb(Vcur, "Vcur", il);
ggml_tensor * sinks = model.layers[il].attn_sinks;
cur = build_attn(inp_attn,
model.layers[il].wo, NULL,
Qcur, Kcur, Vcur, nullptr, sinks, nullptr, 1.0f/sqrtf(float(n_embd_head_k)), il);
}
if (il == n_layer - 1 && inp_out_ids) {
cur = ggml_get_rows(ctx0, cur, inp_out_ids);
inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
}
ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
cb(ffn_inp, "ffn_inp", il);
cur = build_norm(ffn_inp,
model.layers[il].ffn_norm, NULL,
LLM_NORM_RMS, il);
cb(cur, "ffn_norm", il);
// feed-forward network
if (model.layers[il].ffn_gate_inp == nullptr) {
// dense branch
cur = build_ffn(cur,
model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL,
model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
NULL,
LLM_FFN_SILU, LLM_FFN_PAR, il);
cb(cur, "ffn_out", il);
} else {
// MoE branch
cur = build_moe_ffn(cur, model.layers[il].ffn_gate_inp, 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, true, false,
0.0, LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID, il);
cb(cur, "ffn_moe_out", il);
}
cur = ggml_add(ctx0, cur, ffn_inp);
cur = build_cvec(cur, il);
cb(cur, "l_out", il);
// input for next layer
inpL = cur;
}
cur = inpL;
cur = build_norm(cur,
model.output_norm, NULL,
LLM_NORM_RMS, -1);
cb(cur, "result_norm", -1);
res->t_embd = cur;
// lm_head
cur = build_lora_mm(model.output, cur);
cb(cur, "result_output", -1);
res->t_logits = cur;
ggml_build_forward_expand(gf, cur);
}

4
src/models/models.h
View file

@ -316,6 +316,10 @@ struct llm_build_mamba : public llm_graph_context_mamba {
llm_build_mamba(const llama_model & model, const llm_graph_params & params);
};
struct llm_build_mimo2_iswa : public llm_graph_context {
llm_build_mimo2_iswa(const llama_model & model, const llm_graph_params & params);
};
struct llm_build_minicpm3 : public llm_graph_context {
llm_build_minicpm3(const llama_model & model, const llm_graph_params & params);
};