mikhail/llama.cpp-mtp-turboquant
1
0
Fork 0
Code Issues Pull Requests Actions 28 Packages Projects Releases Wiki Activity

model: move load_hparams and load_tensors to per-model definition (#22004)

Browse Source 
* git-friendly migration

* add build_graph

* nits

* exclude old code from build

* wip

* add llm_arch_model_i

* prepare downstream functions

* nits

* nits

* wip

* wip

* add back create_tensor_qkv

* fix files missing include

* enforce one llm_build per arch

* cmake: use glob

* missing model params

* nits

* wip

* wip (2)

* wip (3)

* test-llama-archs is happy

* improve switch case

* move more stuff into llm_arch_model_i

* fix downstream code

* nits

* nits (2)

* fix order

* llama_model_base

* LLAMA_LOAD_LOCALS

* small fix

* fix build errors

* auto

* rm migration script and ifdef
This commit is contained in:
Xuan-Son Nguyen
2026-05-04 12:36:59 +02:00
committed by GitHub
parent c84e6d6db5
commit 994118a183
129 changed files with 10667 additions and 8117 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/models/qwen35moe.cpp
+108 -7
View File
@@ -1,8 +1,109 @@
#include "models.h"
#include "llama-memory-recurrent.h"
llm_build_qwen35moe::llm_build_qwen35moe(const llama_model & model, const llm_graph_params & params) :
void llama_model_qwen35moe::load_arch_hparams(llama_model_loader & ml) {
ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp, false);
ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, false);
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
ml.get_key_or_arr(LLM_KV_ROPE_DIMENSION_SECTIONS, hparams.rope_sections, 4, true);
// Load linear attention (gated delta net) parameters
ml.get_key(LLM_KV_SSM_CONV_KERNEL, hparams.ssm_d_conv);
ml.get_key(LLM_KV_SSM_INNER_SIZE, hparams.ssm_d_inner);
ml.get_key(LLM_KV_SSM_STATE_SIZE, hparams.ssm_d_state);
ml.get_key(LLM_KV_SSM_TIME_STEP_RANK, hparams.ssm_dt_rank);
ml.get_key(LLM_KV_SSM_GROUP_COUNT, hparams.ssm_n_group);
// Mark recurrent layers (linear attention layers)
{
uint32_t full_attn_interval = 4;
ml.get_key(LLM_KV_FULL_ATTENTION_INTERVAL, full_attn_interval, false);
for (uint32_t i = 0; i < hparams.n_layer; ++i) {
hparams.recurrent_layer_arr[i] = ((i + 1) % full_attn_interval != 0);
}
}
switch (hparams.n_layer) {
case 40: type = LLM_TYPE_35B_A3B; break;
case 48: type = LLM_TYPE_122B_A10B; break;
case 60: type = LLM_TYPE_397B_A17B; break;
default: type = LLM_TYPE_UNKNOWN;
}
}
void llama_model_qwen35moe::load_arch_tensors(llama_model_loader &) {
LLAMA_LOAD_LOCALS;
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 }, TENSOR_NOT_REQUIRED);
// if output is NULL, init from the input tok embed
if (output == NULL) {
output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, TENSOR_DUPLICATED);
}
const int64_t n_ff_exp = hparams.n_ff_exp ? hparams.n_ff_exp : n_ff / n_expert_used;
// Calculate dimensions from hyperparameters
const int64_t head_k_dim = hparams.ssm_d_state;
const int64_t head_v_dim = hparams.ssm_d_state;
const int64_t n_k_heads = hparams.ssm_n_group;
const int64_t n_v_heads = hparams.ssm_dt_rank;
const int64_t key_dim = head_k_dim * n_k_heads;
const int64_t value_dim = head_v_dim * n_v_heads;
const int64_t conv_dim = key_dim * 2 + value_dim;
for (int i = 0; i < n_layer; ++i) {
auto & layer = layers[i];
layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), { n_embd }, 0);
layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), { n_embd }, 0);
if (!hparams.is_recurrent(i)) {
// Attention layers
create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head * 2, n_embd_k_gqa, n_embd_v_gqa, 0);
layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd_head_k * n_head, n_embd }, 0);
// Q/K normalization for attention layers
layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), { n_embd_head_k }, 0);
layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), { n_embd_head_k }, 0);
} else {
// Linear attention (gated delta net) specific tensors
// Create tensors with calculated dimensions
layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), { n_embd, key_dim * 2 + value_dim }, TENSOR_NOT_REQUIRED);
layer.wqkv_gate = create_tensor(tn(LLM_TENSOR_ATTN_GATE, "weight", i), { n_embd, value_dim }, TENSOR_NOT_REQUIRED);
layer.ssm_conv1d = create_tensor(tn(LLM_TENSOR_SSM_CONV1D, "weight", i), { hparams.ssm_d_conv, conv_dim }, 0);
layer.ssm_dt = create_tensor(tn(LLM_TENSOR_SSM_DT, "bias", i), { hparams.ssm_dt_rank }, 0);
layer.ssm_a = create_tensor(tn(LLM_TENSOR_SSM_A_NOSCAN, i), { hparams.ssm_dt_rank }, 0);
layer.ssm_beta = create_tensor(tn(LLM_TENSOR_SSM_BETA, "weight", i), { n_embd, n_v_heads }, 0);
layer.ssm_alpha = create_tensor(tn(LLM_TENSOR_SSM_ALPHA, "weight", i), { n_embd, n_v_heads }, 0);
layer.ssm_norm = create_tensor(tn(LLM_TENSOR_SSM_NORM, "weight", i), { head_v_dim }, 0);
layer.ssm_out = create_tensor(tn(LLM_TENSOR_SSM_OUT, "weight", i), { value_dim, n_embd }, 0);
}
layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), { n_embd, n_expert }, 0);
layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), { n_ff_exp, n_embd, n_expert }, 0);
create_tensor_gate_up_exps(layer, i, n_embd, n_ff_exp, n_expert, 0);
// Shared experts
const int64_t n_ff_shexp = hparams.n_ff_shexp ? hparams.n_ff_shexp : n_ff;
layer.ffn_gate_inp_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP_SHEXP, "weight", i), { n_embd }, 0);
layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), { n_embd, n_ff_shexp }, 0);
layer.ffn_up_shexp = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP, "weight", i), { n_embd, n_ff_shexp }, 0);
layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), { n_ff_shexp, n_embd }, 0);
}
}
std::unique_ptr<llm_graph_context> llama_model_qwen35moe::build_arch_graph(const llm_graph_params & params) const {
return std::make_unique<graph>(*this, params);
}
llama_model_qwen35moe::graph::graph(const llama_model & model, const llm_graph_params & params) :
llm_build_delta_net_base(params), model(model) {
const int64_t n_embd_head = hparams.n_embd_head_v();
@@ -87,7 +188,7 @@ llm_build_qwen35moe::llm_build_qwen35moe(const llama_model & model, const llm_gr
ggml_build_forward_expand(gf, cur);
}
std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen35moe::build_qkvz(
std::pair<ggml_tensor *, ggml_tensor *> llama_model_qwen35moe::graph::build_qkvz(
ggml_tensor * input,
int il) {
const int64_t n_seqs = ubatch.n_seqs;
@@ -103,7 +204,7 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen35moe::build_qkvz(
return { qkv_mixed, z };
}
ggml_tensor * llm_build_qwen35moe::build_norm_gated(
ggml_tensor * llama_model_qwen35moe::graph::build_norm_gated(
ggml_tensor * input,
ggml_tensor * weights,
ggml_tensor * gate,
@@ -114,7 +215,7 @@ ggml_tensor * llm_build_qwen35moe::build_norm_gated(
return ggml_mul(ctx0, normalized, gated_silu);
}
ggml_tensor * llm_build_qwen35moe ::build_layer_attn(
ggml_tensor * llama_model_qwen35moe::graph::build_layer_attn(
llm_graph_input_attn_kv * inp,
ggml_tensor * cur,
ggml_tensor * inp_pos,
@@ -195,7 +296,7 @@ ggml_tensor * llm_build_qwen35moe ::build_layer_attn(
return cur;
}
ggml_tensor * llm_build_qwen35moe ::build_layer_attn_linear(
ggml_tensor * llama_model_qwen35moe::graph::build_layer_attn_linear(
llm_graph_input_rs * inp,
ggml_tensor * cur,
int il) {
@@ -369,7 +470,7 @@ ggml_tensor * llm_build_qwen35moe ::build_layer_attn_linear(
return cur;
}
ggml_tensor * llm_build_qwen35moe ::build_layer_ffn(ggml_tensor * cur, const int il) {
ggml_tensor * llama_model_qwen35moe::graph::build_layer_ffn(ggml_tensor * cur, const int il) {
// Check if this is an MoE layer
GGML_ASSERT(model.layers[il].ffn_gate_inp != nullptr);