Public/koboldcpp
1
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-09-19 01:05:09 +02:00
Code Issues Packages Projects Releases Wiki Activity

Merge commit '2948e6049a4ad0f96a4ab15246db2d2086b80703' into concedo_experimental

Browse Source 
# Conflicts:
#	.github/workflows/build.yml
#	CONTRIBUTING.md
#	docs/backend/VirtGPU/development.md
#	docs/ops.md
#	docs/ops/WebGPU.csv
#	embd_res/templates/GigaChat3-10B-A1.8B.jinja
#	embd_res/templates/GigaChat3.1-10B-A1.8B.jinja
#	ggml/src/ggml-hip/CMakeLists.txt
#	ggml/src/ggml-opencl/CMakeLists.txt
#	ggml/src/ggml-opencl/ggml-opencl.cpp
#	ggml/src/ggml-webgpu/ggml-webgpu-shader-lib.hpp
#	ggml/src/ggml-webgpu/ggml-webgpu.cpp
#	scripts/sync_vendor.py
#	tests/CMakeLists.txt
#	tests/test-backend-ops.cpp
#	tests/test-chat.cpp
#	tests/test-grammar-integration.cpp
#	tests/test-quantize-fns.cpp
This commit is contained in:
Concedo
2026-03-15 11:21:24 +08:00
parent 67c9798d0b 2948e6049a
commit b1c500ae2b
72 changed files with 2338 additions and 430 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/llama-context.cpp
+92 -25
View File
@@ -7,6 +7,7 @@
#include "llama-memory.h"
#include "llama-mmap.h"
#include "llama-model.h"
#include "llama-ext.h"
#include <cinttypes>
#include <cmath>
@@ -154,7 +155,8 @@ llama_context::llama_context(
cparams.auto_fa = params.flash_attn_type == LLAMA_FLASH_ATTN_TYPE_AUTO;
cparams.fused_gdn_ar = true;
cparams.fused_gdn_ch = false; // TODO: implement
cparams.fused_gdn_ch = true;
cparams.auto_fgdn = true;
// with causal attention, the batch size is limited by the context size
cparams.n_batch = cparams.causal_attn ? std::min(cparams.n_ctx, params.n_batch) : params.n_batch;
@@ -348,6 +350,14 @@ llama_context::llama_context(
if (cparams.pipeline_parallel) {
LLAMA_LOG_INFO("%s: pipeline parallelism enabled\n", __func__);
if (!graph_reuse_disable) {
// TODO: figure out a way to make graph reuse work with pipeline parallelism
// ref: https://github.com/ggml-org/llama.cpp/pull/20463
LLAMA_LOG_WARN("%s: graph reuse is currently not compatible with pipeline parallelism - disabling\n", __func__);
graph_reuse_disable = true;
}
}
sched_reserve();
@@ -471,37 +481,81 @@ void llama_context::sched_reserve() {
cparams.auto_fa = false;
}
if (cparams.fused_gdn_ar) {
auto * gf = graph_reserve(1, n_seqs, n_outputs, mctx.get(), true);
if (!gf) {
throw std::runtime_error("failed to reserve graph for fused Gated Delta Net check");
}
if (cparams.auto_fgdn) {
LLAMA_LOG_INFO("%s: resolving fused Gated Delta Net support:\n", __func__);
const size_t prefix_len = strlen(LLAMA_TENSOR_NAME_FGDNAR) + 1;
bool gdn_device_mismatch = false;
for (int i = 0; i < ggml_graph_n_nodes(gf); i++) {
ggml_tensor * n = ggml_graph_node(gf, i);
if (n->op != GGML_OP_GATED_DELTA_NET) {
continue;
if (cparams.fused_gdn_ar) {
auto * gf = graph_reserve(1, n_seqs, n_outputs, mctx.get(), true);
if (!gf) {
throw std::runtime_error("failed to reserve graph for fused Gated Delta Net check (autoregressive)");
}
ggml_backend_dev_t device_gdn = ggml_backend_get_device(ggml_backend_sched_get_tensor_backend(sched.get(), n));
GGML_ASSERT(strncmp(n->name, LLAMA_TENSOR_NAME_FGDNAR "-", prefix_len) == 0);
const int il = std::stoi(n->name + prefix_len);
ggml_backend_dev_t device_kv = model.dev_layer(il);
if (device_gdn != device_kv) {
LLAMA_LOG_WARN("%s: layer %d is assigned to device %s but the fused Gated Delta Net tensor "
"is assigned to device %s (usually due to missing support)\n",
__func__, il, ggml_backend_dev_name(device_kv), ggml_backend_dev_name(device_gdn));
gdn_device_mismatch = true;
break;
const size_t prefix_len = strlen(LLAMA_TENSOR_NAME_FGDN_AR) + 1;
bool gdn_device_mismatch = false;
for (int i = 0; i < ggml_graph_n_nodes(gf); i++) {
ggml_tensor * n = ggml_graph_node(gf, i);
if (n->op != GGML_OP_GATED_DELTA_NET) {
continue;
}
ggml_backend_dev_t device_gdn = ggml_backend_get_device(ggml_backend_sched_get_tensor_backend(sched.get(), n));
GGML_ASSERT(strncmp(n->name, LLAMA_TENSOR_NAME_FGDN_AR "-", prefix_len) == 0);
const int il = std::stoi(n->name + prefix_len);
ggml_backend_dev_t device_kv = model.dev_layer(il);
if (device_gdn != device_kv) {
LLAMA_LOG_WARN("%s: layer %d is assigned to device %s but the fused Gated Delta Net tensor "
"is assigned to device %s (usually due to missing support)\n",
__func__, il, ggml_backend_dev_name(device_kv), ggml_backend_dev_name(device_gdn));
gdn_device_mismatch = true;
break;
}
}
if (gdn_device_mismatch) {
cparams.fused_gdn_ar = false;
LLAMA_LOG_WARN("%s: fused Gated Delta Net (autoregressive) not supported, set to disabled\n", __func__);
} else {
LLAMA_LOG_INFO("%s: fused Gated Delta Net (autoregressive) enabled\n", __func__);
}
}
if (gdn_device_mismatch) {
cparams.fused_gdn_ar = false;
LLAMA_LOG_WARN("%s: fused Gated Delta Net not supported, set to disabled\n", __func__);
if (cparams.fused_gdn_ch) {
// more than one token in the batch per sequence in order to take the chunked path
auto * gf = graph_reserve(16*n_seqs, n_seqs, n_outputs, mctx.get(), true);
if (!gf) {
throw std::runtime_error("failed to reserve graph for fused Gated Delta Net check (chunked)");
}
const size_t prefix_len = strlen(LLAMA_TENSOR_NAME_FGDN_CH) + 1;
bool gdn_device_mismatch = false;
for (int i = 0; i < ggml_graph_n_nodes(gf); i++) {
ggml_tensor * n = ggml_graph_node(gf, i);
if (n->op != GGML_OP_GATED_DELTA_NET) {
continue;
}
ggml_backend_dev_t device_gdn = ggml_backend_get_device(ggml_backend_sched_get_tensor_backend(sched.get(), n));
GGML_ASSERT(strncmp(n->name, LLAMA_TENSOR_NAME_FGDN_CH "-", prefix_len) == 0);
const int il = std::stoi(n->name + prefix_len);
ggml_backend_dev_t device_kv = model.dev_layer(il);
if (device_gdn != device_kv) {
LLAMA_LOG_WARN("%s: layer %d is assigned to device %s but the fused Gated Delta Net tensor "
"is assigned to device %s (usually due to missing support)\n",
__func__, il, ggml_backend_dev_name(device_kv), ggml_backend_dev_name(device_gdn));
gdn_device_mismatch = true;
break;
}
}
if (gdn_device_mismatch) {
cparams.fused_gdn_ch = false;
LLAMA_LOG_WARN("%s: fused Gated Delta Net (chunked) not supported, set to disabled\n", __func__);
} else {
LLAMA_LOG_INFO("%s: fused Gated Delta Net (chunked) enabled\n", __func__);
}
}
cparams.auto_fgdn = false;
}
// reserve worst-case graph
@@ -3094,6 +3148,19 @@ uint32_t llama_get_sampled_probs_count_ith(llama_context * ctx, int32_t i) {
return static_cast<uint32_t>(ctx->get_sampled_probs_count(i));
}
struct ggml_cgraph * llama_graph_reserve(
struct llama_context * ctx,
uint32_t n_tokens,
uint32_t n_seqs,
uint32_t n_outputs) {
auto * memory = ctx->get_memory();
llama_memory_context_ptr mctx;
if (memory) {
mctx = memory->init_full();
}
return ctx->graph_reserve(n_tokens, n_seqs, n_outputs, mctx.get());
}
// llama adapter API
int32_t llama_set_adapters_lora(
src/llama-cparams.h
+1
View File
@@ -33,6 +33,7 @@ struct llama_cparams {
bool auto_fa;
bool fused_gdn_ar; // use fused gated delta net (autoregressive)
bool fused_gdn_ch; // use fused gated delta net (chunked)
bool auto_fgdn;
bool no_perf;
bool warmup;
bool op_offload;
src/llama-ext.h
+12
View File
@@ -0,0 +1,12 @@
#pragma once
#include "llama-context.h"
#include "ggml.h"
#include "stdint.h"
// Reserve a new compute graph. It is valid until the next call to llama_graph_reserve.
LLAMA_API struct ggml_cgraph * llama_graph_reserve(
struct llama_context * ctx,
uint32_t n_tokens,
uint32_t n_seqs,
uint32_t n_outputs);
src/llama-grammar.cpp
+5 -5
View File
@@ -1220,13 +1220,13 @@ struct llama_grammar * llama_grammar_init_impl(
// if there is a grammar, parse it
// rules will be empty (default) if there are parse errors
if (!parser.parse(grammar_str) || parser.rules.empty()) {
fprintf(stderr, "%s: failed to parse grammar\n", __func__);
LLAMA_LOG_ERROR("failed to parse grammar\n");
return nullptr;
}
// Ensure that there is a "root" node.
if (parser.symbol_ids.find("root") == parser.symbol_ids.end()) {
fprintf(stderr, "%s: grammar does not contain a 'root' symbol\n", __func__);
// Ensure that the grammar contains the start symbol
if (parser.symbol_ids.find(grammar_root) == parser.symbol_ids.end()) {
LLAMA_LOG_ERROR("grammar does not contain a '%s' symbol\n", grammar_root);
return nullptr;
}
@@ -1255,7 +1255,7 @@ struct llama_grammar * llama_grammar_init_impl(
continue;
}
if (llama_grammar_detect_left_recursion(vec_rules, i, &rules_visited, &rules_in_progress, &rules_may_be_empty)) {
LLAMA_LOG_ERROR("unsupported grammar, left recursion detected for nonterminal at index %zu", i);
LLAMA_LOG_ERROR("unsupported grammar, left recursion detected for nonterminal at index %zu\n", i);
return nullptr;
}
}
src/llama-graph.cpp
+55 -4
View File
@@ -900,7 +900,8 @@ ggml_tensor * llm_graph_context::build_cvec(
ggml_tensor * llm_graph_context::build_lora_mm(
ggml_tensor * w,
ggml_tensor * cur) const {
ggml_tensor * cur,
ggml_tensor * w_s) const {
ggml_tensor * res = ggml_mul_mat(ctx0, w, cur);
for (const auto & lora : *loras) {
@@ -921,6 +922,10 @@ ggml_tensor * llm_graph_context::build_lora_mm(
res = ggml_add(ctx0, res, ab_cur);
}
if (w_s) {
res = ggml_mul(ctx0, res, w_s);
}
return res;
}
@@ -1166,7 +1171,10 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
llama_expert_gating_func_type gating_op,
int il,
ggml_tensor * probs_in,
ggml_tensor * gate_up_exps) const {
ggml_tensor * gate_up_exps,
ggml_tensor * up_exps_s,
ggml_tensor * gate_exps_s,
ggml_tensor * down_exps_s) const {
return build_moe_ffn(
cur,
gate_inp, /* gate_inp_b */ nullptr,
@@ -1182,7 +1190,11 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
gating_op,
il,
probs_in,
gate_up_exps
gate_up_exps,
/* gate_up_exps_b */ nullptr,
up_exps_s,
gate_exps_s,
down_exps_s
);
}
@@ -1206,7 +1218,10 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
int il,
ggml_tensor * probs_in,
ggml_tensor * gate_up_exps,
ggml_tensor * gate_up_exps_b) const {
ggml_tensor * gate_up_exps_b,
ggml_tensor * up_exps_s,
ggml_tensor * gate_exps_s,
ggml_tensor * down_exps_s) const {
const int64_t n_embd = cur->ne[0];
const int64_t n_tokens = cur->ne[1];
const bool weight_before_ffn = arch == LLM_ARCH_LLAMA4; // for llama4, we apply the sigmoid-ed weights before the FFN
@@ -1358,6 +1373,15 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
cb(gate_up, "ffn_moe_gate_up_biased", il);
}
// apply per-expert scale2 to merged gate_up (use up_exps_s since gate and up are fused)
if (up_exps_s) {
ggml_tensor * s = ggml_reshape_3d(ctx0, up_exps_s, 1, n_expert, 1);
s = ggml_repeat_4d(ctx0, s, 1, n_expert, n_tokens, 1);
s = ggml_get_rows(ctx0, s, selected_experts); // [1, n_expert_used, n_tokens]
gate_up = ggml_mul(ctx0, gate_up, s);
cb(gate_up, "ffn_moe_gate_up_scaled", il);
}
const int64_t n_ff = gate_up->ne[0] / 2;
cur = ggml_view_3d(ctx0, gate_up, n_ff, gate_up->ne[1], gate_up->ne[2], gate_up->nb[1], gate_up->nb[2], 0);
cb(cur, "ffn_moe_gate", il);
@@ -1373,6 +1397,15 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
cb(up, "ffn_moe_up_biased", il);
}
// apply per-expert scale2 to up
if (up_exps_s) {
ggml_tensor * s = ggml_reshape_3d(ctx0, up_exps_s, 1, n_expert, 1);
s = ggml_repeat_4d(ctx0, s, 1, n_expert, n_tokens, 1);
s = ggml_get_rows(ctx0, s, selected_experts); // [1, n_expert_used, n_tokens]
up = ggml_mul(ctx0, up, s);
cb(up, "ffn_moe_up_scaled", il);
}
if (gate_exps) {
cur = build_lora_mm_id(gate_exps, cur, selected_experts); // [n_ff, n_expert_used, n_tokens]
cb(cur, "ffn_moe_gate", il);
@@ -1384,6 +1417,15 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
cur = ggml_add_id(ctx0, cur, gate_exps_b, selected_experts);
cb(cur, "ffn_moe_gate_biased", il);
}
// apply per-expert scale2 to gate
if (gate_exps_s) {
ggml_tensor * s = ggml_reshape_3d(ctx0, gate_exps_s, 1, n_expert, 1);
s = ggml_repeat_4d(ctx0, s, 1, n_expert, n_tokens, 1);
s = ggml_get_rows(ctx0, s, selected_experts); // [1, n_expert_used, n_tokens]
cur = ggml_mul(ctx0, cur, s);
cb(cur, "ffn_moe_gate_scaled", il);
}
}
const bool has_gate = gate_exps || gate_up_exps;
@@ -1463,6 +1505,15 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
cb(experts, "ffn_moe_down_biased", il);
}
// apply per-expert scale2 to down
if (down_exps_s) {
ggml_tensor * s = ggml_reshape_3d(ctx0, down_exps_s, 1, n_expert, 1);
s = ggml_repeat_4d(ctx0, s, 1, n_expert, n_tokens, 1);
s = ggml_get_rows(ctx0, s, selected_experts); // [1, n_expert_used, n_tokens]
experts = ggml_mul(ctx0, experts, s);
cb(experts, "ffn_moe_down_scaled", il);
}
if (!weight_before_ffn) {
experts = ggml_mul(ctx0, experts, weights);
cb(cur, "ffn_moe_weighted", il);
src/llama-graph.h
+11 -4
View File
@@ -764,10 +764,11 @@ struct llm_graph_context {
ggml_tensor * cur,
int il) const;
// do mat_mul, while optionally apply lora
// do mat_mul, while optionally apply lora and per-tensor scale
ggml_tensor * build_lora_mm(
ggml_tensor * w,
ggml_tensor * cur) const;
ggml_tensor * cur,
ggml_tensor * w_s = nullptr) const;
// do mat_mul_id, while optionally apply lora
ggml_tensor * build_lora_mm_id(
@@ -814,7 +815,10 @@ struct llm_graph_context {
llama_expert_gating_func_type gating_op,
int il,
ggml_tensor * probs_in = nullptr,
ggml_tensor * gate_up_exps = nullptr) const;
ggml_tensor * gate_up_exps = nullptr,
ggml_tensor * up_exps_s = nullptr,
ggml_tensor * gate_exps_s = nullptr,
ggml_tensor * down_exps_s = nullptr) const;
ggml_tensor * build_moe_ffn(
ggml_tensor * cur,
@@ -836,7 +840,10 @@ struct llm_graph_context {
int il,
ggml_tensor * probs_in = nullptr,
ggml_tensor * gate_up_exps = nullptr,
ggml_tensor * gate_up_exps_b = nullptr) const;
ggml_tensor * gate_up_exps_b = nullptr,
ggml_tensor * up_exps_s = nullptr,
ggml_tensor * gate_exps_s = nullptr,
ggml_tensor * down_exps_s = nullptr) const;
//
// inputs
src/llama-impl.h
+3 -3
View File
@@ -70,6 +70,6 @@ std::string llama_format_tensor_shape(const struct ggml_tensor * t);
std::string gguf_kv_to_str(const struct gguf_context * ctx_gguf, int i);
#define LLAMA_TENSOR_NAME_FATTN "__fattn__"
#define LLAMA_TENSOR_NAME_FGDNAR "__fgdnar__"
#define LLAMA_TENSOR_NAME_FGDNCH "__fgdnch__"
#define LLAMA_TENSOR_NAME_FATTN "__fattn__"
#define LLAMA_TENSOR_NAME_FGDN_AR "__fgdn_ar__"
#define LLAMA_TENSOR_NAME_FGDN_CH "__fgdn_ch__"
src/llama-model-loader.cpp
+2
View File
@@ -42,6 +42,7 @@ static std::string llama_model_ftype_name(llama_ftype ftype) {
case LLAMA_FTYPE_MOSTLY_Q5_1: return "Q5_1";
case LLAMA_FTYPE_MOSTLY_Q8_0: return "Q8_0";
case LLAMA_FTYPE_MOSTLY_MXFP4_MOE: return "MXFP4 MoE";
case LLAMA_FTYPE_MOSTLY_NVFP4: return "NVFP4";
case LLAMA_FTYPE_MOSTLY_Q2_K: return "Q2_K - Medium";
case LLAMA_FTYPE_MOSTLY_Q2_K_S: return "Q2_K - Small";
case LLAMA_FTYPE_MOSTLY_Q3_K_S: return "Q3_K - Small";
@@ -725,6 +726,7 @@ llama_model_loader::llama_model_loader(
case GGML_TYPE_IQ4_NL: ftype = LLAMA_FTYPE_MOSTLY_IQ4_NL; break;
case GGML_TYPE_IQ4_XS: ftype = LLAMA_FTYPE_MOSTLY_IQ4_XS; break;
case GGML_TYPE_IQ3_S: ftype = LLAMA_FTYPE_MOSTLY_IQ3_S; break;
case GGML_TYPE_NVFP4: ftype = LLAMA_FTYPE_MOSTLY_NVFP4; break;
default:
{
LLAMA_LOG_WARN("%s: unknown type %s\n", __func__, ggml_type_name(type_max));
src/llama-model.cpp
+49 -7
View File
@@ -5168,23 +5168,23 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
layer.attn_sub_norm = create_tensor(tn(LLM_TENSOR_ATTN_SUB_NORM, "weight", i), {n_embd}, 0);
layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd}, 0);
layer.wq_scale = create_tensor(tn(LLM_TENSOR_ATTN_Q, "scale", i), {1}, TENSOR_NOT_REQUIRED);
layer.wq_s = create_tensor(tn(LLM_TENSOR_ATTN_Q, "scale", i), {1}, TENSOR_NOT_REQUIRED);
layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa}, 0);
layer.wk_scale = create_tensor(tn(LLM_TENSOR_ATTN_K, "scale", i), {1}, TENSOR_NOT_REQUIRED);
layer.wk_s = create_tensor(tn(LLM_TENSOR_ATTN_K, "scale", i), {1}, TENSOR_NOT_REQUIRED);
layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa}, 0);
layer.wv_scale = create_tensor(tn(LLM_TENSOR_ATTN_V, "scale", i), {1}, TENSOR_NOT_REQUIRED);
layer.wv_s = create_tensor(tn(LLM_TENSOR_ATTN_V, "scale", i), {1}, TENSOR_NOT_REQUIRED);
layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
layer.wo_scale = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "scale", i), {1}, TENSOR_NOT_REQUIRED);
layer.wo_s = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "scale", i), {1}, TENSOR_NOT_REQUIRED);
layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
layer.ffn_sub_norm = create_tensor(tn(LLM_TENSOR_FFN_SUB_NORM, "weight", i), {n_ff}, 0);
layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0);
layer.ffn_gate_scale = create_tensor(tn(LLM_TENSOR_FFN_GATE, "scale", i), {1}, TENSOR_NOT_REQUIRED);
layer.ffn_gate_s = create_tensor(tn(LLM_TENSOR_FFN_GATE, "scale", i), {1}, TENSOR_NOT_REQUIRED);
layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
layer.ffn_down_scale = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "scale", i), {1}, TENSOR_NOT_REQUIRED);
layer.ffn_down_s = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "scale", i), {1}, TENSOR_NOT_REQUIRED);
layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, 0);
layer.ffn_up_scale = create_tensor(tn(LLM_TENSOR_FFN_UP, "scale", i), {1}, TENSOR_NOT_REQUIRED);
layer.ffn_up_s = create_tensor(tn(LLM_TENSOR_FFN_UP, "scale", i), {1}, TENSOR_NOT_REQUIRED);
}
} break;
case LLM_ARCH_T5:
@@ -7601,6 +7601,48 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
default:
throw std::runtime_error("unknown architecture");
}
// generic pass: load optional per-tensor/per-expert ".scale" tensors (e.g. NVFP4 scale2)
// this avoids having to add scale loading to every architecture
for (int i = 0; i < n_layer; ++i) {
auto & layer = layers[i];
// attention weight scales (per-tensor, shape {1})
if (!layer.wq_s && layer.wq) {
layer.wq_s = create_tensor(tn(LLM_TENSOR_ATTN_Q, "scale", i), {1}, TENSOR_NOT_REQUIRED);
}
if (!layer.wk_s && layer.wk) {
layer.wk_s = create_tensor(tn(LLM_TENSOR_ATTN_K, "scale", i), {1}, TENSOR_NOT_REQUIRED);
}
if (!layer.wv_s && layer.wv) {
layer.wv_s = create_tensor(tn(LLM_TENSOR_ATTN_V, "scale", i), {1}, TENSOR_NOT_REQUIRED);
}
if (!layer.wo_s && layer.wo) {
layer.wo_s = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "scale", i), {1}, TENSOR_NOT_REQUIRED);
}
// dense FFN weight scales (per-tensor, shape {1})
if (!layer.ffn_gate_s && layer.ffn_gate) {
layer.ffn_gate_s = create_tensor(tn(LLM_TENSOR_FFN_GATE, "scale", i), {1}, TENSOR_NOT_REQUIRED);
}
if (!layer.ffn_down_s && layer.ffn_down) {
layer.ffn_down_s = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "scale", i), {1}, TENSOR_NOT_REQUIRED);
}
if (!layer.ffn_up_s && layer.ffn_up) {
layer.ffn_up_s = create_tensor(tn(LLM_TENSOR_FFN_UP, "scale", i), {1}, TENSOR_NOT_REQUIRED);
}
// MoE expert weight scales (per-expert, shape {n_expert})
if (!layer.ffn_gate_exps_s && layer.ffn_gate_exps) {
layer.ffn_gate_exps_s = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "scale", i), {n_expert}, TENSOR_NOT_REQUIRED);
}
if (!layer.ffn_down_exps_s && layer.ffn_down_exps) {
layer.ffn_down_exps_s = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "scale", i), {n_expert}, TENSOR_NOT_REQUIRED);
}
if (!layer.ffn_up_exps_s && layer.ffn_up_exps) {
layer.ffn_up_exps_s = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "scale", i), {n_expert}, TENSOR_NOT_REQUIRED);
}
}
}
ml.done_getting_tensors();
src/llama-model.h
+12 -7
View File
@@ -295,6 +295,11 @@ struct llama_layer {
struct ggml_tensor * ffn_up_exps_b = nullptr;
struct ggml_tensor * ffn_gate_up_exps_b = nullptr;
// ff MoE per-expert scales (NVFP4 per-tensor scale2)
struct ggml_tensor * ffn_gate_exps_s = nullptr;
struct ggml_tensor * ffn_down_exps_s = nullptr;
struct ggml_tensor * ffn_up_exps_s = nullptr;
// ff MoE latent proj
struct ggml_tensor * ffn_latent_down = nullptr;
struct ggml_tensor * ffn_latent_up = nullptr;
@@ -392,13 +397,13 @@ struct llama_layer {
struct ggml_tensor * rope_freqs = nullptr;
// bitnet scale
struct ggml_tensor * wq_scale = nullptr;
struct ggml_tensor * wk_scale = nullptr;
struct ggml_tensor * wv_scale = nullptr;
struct ggml_tensor * wo_scale = nullptr;
struct ggml_tensor * ffn_gate_scale = nullptr;
struct ggml_tensor * ffn_up_scale = nullptr;
struct ggml_tensor * ffn_down_scale = nullptr;
struct ggml_tensor * wq_s = nullptr;
struct ggml_tensor * wk_s = nullptr;
struct ggml_tensor * wv_s = nullptr;
struct ggml_tensor * wo_s = nullptr;
struct ggml_tensor * ffn_gate_s = nullptr;
struct ggml_tensor * ffn_up_s = nullptr;
struct ggml_tensor * ffn_down_s = nullptr;
// altup & laurel
struct ggml_tensor * per_layer_inp_gate = nullptr;
src/models/bitnet.cpp
+7 -22
View File
@@ -29,10 +29,7 @@ llm_build_bitnet::llm_build_bitnet(const llama_model & model, const llm_graph_pa
// self-attention
{
// compute Q and K and RoPE them
ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
if (model.layers[il].wq_scale) {
Qcur = ggml_mul(ctx0, Qcur, model.layers[il].wq_scale);
}
ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur, model.layers[il].wq_s);
cb(Qcur, "Qcur", il);
if (model.layers[il].bq) {
Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
@@ -40,10 +37,7 @@ llm_build_bitnet::llm_build_bitnet(const llama_model & model, const llm_graph_pa
}
// B1.K
ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
if (model.layers[il].wk_scale) {
Kcur = ggml_mul(ctx0, Kcur, model.layers[il].wk_scale);
}
ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur, model.layers[il].wk_s);
cb(Kcur, "Kcur", il);
if (model.layers[il].bk) {
Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
@@ -51,10 +45,7 @@ llm_build_bitnet::llm_build_bitnet(const llama_model & model, const llm_graph_pa
}
// B1.V
ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
if (model.layers[il].wv_scale) {
Vcur = ggml_mul(ctx0, Vcur, model.layers[il].wv_scale);
}
ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur, model.layers[il].wv_s);
cb(Vcur, "Vcur", il);
if (model.layers[il].bv) {
Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
@@ -90,10 +81,7 @@ llm_build_bitnet::llm_build_bitnet(const llama_model & model, const llm_graph_pa
LLM_NORM_RMS, il);
cb(cur, "attn_sub_norm", il);
cur = build_lora_mm(model.layers[il].wo, cur);
if (model.layers[il].wo_scale) {
cur = ggml_mul(ctx0, cur, model.layers[il].wo_scale);
}
cur = build_lora_mm(model.layers[il].wo, cur, model.layers[il].wo_s);
if (model.layers[il].bo) {
cur = ggml_add(ctx0, cur, model.layers[il].bo);
}
@@ -115,8 +103,8 @@ llm_build_bitnet::llm_build_bitnet(const llama_model & model, const llm_graph_pa
cb(cur, "ffn_norm", il);
cur = build_ffn(cur,
model.layers[il].ffn_up, NULL, model.layers[il].ffn_up_scale,
model.layers[il].ffn_gate, NULL, model.layers[il].ffn_gate_scale,
model.layers[il].ffn_up, NULL, model.layers[il].ffn_up_s,
model.layers[il].ffn_gate, NULL, model.layers[il].ffn_gate_s,
NULL, NULL, NULL,
NULL,
LLM_FFN_SILU, LLM_FFN_PAR, il);
@@ -127,10 +115,7 @@ llm_build_bitnet::llm_build_bitnet(const llama_model & model, const llm_graph_pa
LLM_NORM_RMS, il);
cb(cur, "ffn_sub_norm", il);
cur = build_lora_mm(model.layers[il].ffn_down, cur);
if (model.layers[il].ffn_down_scale) {
cur = ggml_mul(ctx0, cur, model.layers[il].ffn_down_scale);
}
cur = build_lora_mm(model.layers[il].ffn_down, cur, model.layers[il].ffn_down_s);
cb(cur, "ffn_down", il);
cur = ggml_add(ctx0, cur, ffn_inp);
src/models/delta-net-base.cpp
+75 -27
View File
@@ -41,13 +41,6 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_delta_net_base::build_delta_ne
GGML_ASSERT(b->ne[0] == 1 && b->ne[1] == H_v && b->ne[2] == n_tokens && b->ne[3] == n_seqs);
GGML_ASSERT(s->ne[0] == S_v && s->ne[1] == S_v && s->ne[2] == H_v && s->ne[3] == n_seqs);
if (cparams.fused_gdn_ch) {
//ggml_tensor * result = ggml_gated_delta_net(ctx0, q, k, v, g, b, s);
//cb(result, LLAMA_TENSOR_NAME_FGDNCH, il);
GGML_ABORT("not implemented yet");
}
const float scale = 1.0f / sqrtf(S_k);
q = ggml_scale(ctx0, q, scale);
@@ -325,26 +318,6 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_delta_net_base::build_delta_ne
GGML_ASSERT(b->ne[0] == 1 && b->ne[1] == H_v && b->ne[2] == n_tokens && b->ne[3] == n_seqs);
GGML_ASSERT(s->ne[0] == S_v && s->ne[1] == S_v && s->ne[2] == H_v && s->ne[3] == n_seqs);
if (cparams.fused_gdn_ar) {
ggml_tensor * result = ggml_gated_delta_net(ctx0, q, k, v, g, b, s);
cb(result, LLAMA_TENSOR_NAME_FGDNAR, il);
ggml_tensor * output = ggml_view_4d(ctx0, result,
S_v, H_v, n_tokens, n_seqs,
ggml_row_size(result->type, S_v),
ggml_row_size(result->type, S_v * H_v),
ggml_row_size(result->type, S_v * H_v * n_tokens), 0);
ggml_tensor * new_state = ggml_view_4d(ctx0, result,
S_v, S_v, H_v, n_seqs,
ggml_row_size(result->type, S_v),
ggml_row_size(result->type, S_v * S_v),
ggml_row_size(result->type, S_v * S_v * H_v),
ggml_row_size(result->type, S_v * H_v * n_tokens * n_seqs));
return {output, new_state};
}
const float scale = 1.0f / sqrtf(S_k);
q = ggml_scale(ctx0, q, scale);
@@ -401,3 +374,78 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_delta_net_base::build_delta_ne
return {o, s};
}
std::pair<ggml_tensor *, ggml_tensor *> llm_build_delta_net_base::build_delta_net_fused(
ggml_tensor * q,
ggml_tensor * k,
ggml_tensor * v,
ggml_tensor * g,
ggml_tensor * b,
ggml_tensor * s,
int il) {
const int64_t S_k = q->ne[0];
const int64_t H_k = q->ne[1];
const int64_t n_tokens = q->ne[2];
const int64_t n_seqs = q->ne[3];
const int64_t S_v = v->ne[0];
const int64_t H_v = v->ne[1];
GGML_ASSERT(S_k == S_v);
GGML_ASSERT(H_v % H_k == 0);
GGML_ASSERT(q->ne[0] == S_k && q->ne[1] == H_k && q->ne[2] == n_tokens && q->ne[3] == n_seqs);
GGML_ASSERT(k->ne[0] == S_k && k->ne[1] == H_k && k->ne[2] == n_tokens && k->ne[3] == n_seqs);
GGML_ASSERT(v->ne[0] == S_v && v->ne[1] == H_v && v->ne[2] == n_tokens && v->ne[3] == n_seqs);
GGML_ASSERT(g->ne[0] == 1 || g->ne[0] == S_v);
GGML_ASSERT( g->ne[1] == H_v && g->ne[2] == n_tokens && g->ne[3] == n_seqs);
GGML_ASSERT(b->ne[0] == 1 && b->ne[1] == H_v && b->ne[2] == n_tokens && b->ne[3] == n_seqs);
GGML_ASSERT(s->ne[0] == S_v && s->ne[1] == S_v && s->ne[2] == H_v && s->ne[3] == n_seqs);
ggml_tensor * result = ggml_gated_delta_net(ctx0, q, k, v, g, b, s);
if (n_tokens == 1) {
cb(result, LLAMA_TENSOR_NAME_FGDN_AR, il);
} else {
cb(result, LLAMA_TENSOR_NAME_FGDN_CH, il);
}
ggml_tensor * output = ggml_view_4d(ctx0, result,
S_v, H_v, n_tokens, n_seqs,
ggml_row_size(result->type, S_v),
ggml_row_size(result->type, S_v * H_v),
ggml_row_size(result->type, S_v * H_v * n_tokens), 0);
ggml_tensor * new_state = ggml_view_4d(ctx0, result,
S_v, S_v, H_v, n_seqs,
ggml_row_size(result->type, S_v),
ggml_row_size(result->type, S_v * S_v),
ggml_row_size(result->type, S_v * S_v * H_v),
ggml_row_size(result->type, S_v * H_v * n_tokens * n_seqs));
return {output, new_state};
}
std::pair<ggml_tensor *, ggml_tensor *> llm_build_delta_net_base::build_delta_net(
ggml_tensor * q,
ggml_tensor * k,
ggml_tensor * v,
ggml_tensor * g,
ggml_tensor * b,
ggml_tensor * s,
int il) {
const int64_t n_seq_tokens = q->ne[2];
if (n_seq_tokens == 1) {
if (cparams.fused_gdn_ar) {
return build_delta_net_fused(q, k, v, g, b, s, il);
}
return build_delta_net_autoregressive(q, k, v, g, b, s, il);
}
if (cparams.fused_gdn_ch) {
return build_delta_net_fused(q, k, v, g, b, s, il);
}
return build_delta_net_chunking(q, k, v, g, b, s, il);
}
src/models/kimi-linear.cpp
+1 -3
View File
@@ -169,9 +169,7 @@ llm_build_kimi_linear::llm_build_kimi_linear(const llama_model & model, const ll
Kcur = ggml_l2_norm(ctx0, Kcur, eps_norm);
// Choose between build_delta_net_chunking and build_delta_net_recurrent based on n_tokens
std::pair<ggml_tensor *, ggml_tensor *> attn_out = n_seq_tokens == 1 ?
build_delta_net_autoregressive(Qcur, Kcur, Vcur, g1, beta, state, il) :
build_delta_net_chunking(Qcur, Kcur, Vcur, g1, beta, state, il);
auto attn_out = build_delta_net(Qcur, Kcur, Vcur, g1, beta, state, il);
ggml_tensor * output = ggml_cont(ctx0, attn_out.first);
ggml_tensor * new_state = attn_out.second;
src/models/llama.cpp
+14 -7
View File
@@ -43,19 +43,19 @@ llm_build_llama<embed>::llm_build_llama(const llama_model & model, const llm_gra
ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
// compute Q and K and RoPE them
ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur, model.layers[il].wq_s);
cb(Qcur, "Qcur", il);
if (model.layers[il].bq) {
Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
cb(Qcur, "Qcur", il);
}
ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur, model.layers[il].wk_s);
cb(Kcur, "Kcur", il);
if (model.layers[il].bk) {
Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
cb(Kcur, "Kcur", il);
}
ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur, model.layers[il].wv_s);
cb(Vcur, "Vcur", il);
if (model.layers[il].bv) {
Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
@@ -91,6 +91,9 @@ llm_build_llama<embed>::llm_build_llama(const llama_model & model, const llm_gra
cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
if (model.layers[il].wo_s) {
cur = ggml_mul(ctx0, cur, model.layers[il].wo_s);
}
cb(cur, "attn_out", il);
}
if (il == n_layer - 1 && inp_out_ids) {
@@ -109,9 +112,9 @@ llm_build_llama<embed>::llm_build_llama(const llama_model & model, const llm_gra
cb(cur, "ffn_norm", il);
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,
model.layers[il].ffn_up, model.layers[il].ffn_up_b, model.layers[il].ffn_up_s,
model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, model.layers[il].ffn_gate_s,
model.layers[il].ffn_down, model.layers[il].ffn_down_b, model.layers[il].ffn_down_s,
NULL,
LLM_FFN_SILU, LLM_FFN_PAR, il);
cb(cur, "ffn_out", il);
@@ -132,7 +135,11 @@ llm_build_llama<embed>::llm_build_llama(const llama_model & model, const llm_gra
LLM_FFN_SILU, true,
hparams.expert_weights_scale,
LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
il);
il,
nullptr, nullptr,
model.layers[il].ffn_up_exps_s,
model.layers[il].ffn_gate_exps_s,
model.layers[il].ffn_down_exps_s);
cb(cur, "ffn_moe_out", il);
}
cur = ggml_add(ctx0, cur, ffn_inp);
src/models/models.h
+20
View File
@@ -44,6 +44,26 @@ struct llm_build_delta_net_base : public llm_graph_context {
ggml_tensor * b,
ggml_tensor * s,
int il);
// use the ggml_gated_delta_net fused operator
std::pair<ggml_tensor *, ggml_tensor *> build_delta_net_fused(
ggml_tensor * q,
ggml_tensor * k,
ggml_tensor * v,
ggml_tensor * g,
ggml_tensor * b,
ggml_tensor * s,
int il);
// choose one of two implementations above based on the number of tokens
std::pair<ggml_tensor *, ggml_tensor *> build_delta_net(
ggml_tensor * q,
ggml_tensor * k,
ggml_tensor * v,
ggml_tensor * g,
ggml_tensor * b,
ggml_tensor * s,
int il);
};
struct llm_build_rwkv6_base : public llm_graph_context {
src/models/qwen3.cpp
+9 -6
View File
@@ -30,13 +30,13 @@ llm_build_qwen3::llm_build_qwen3(const llama_model & model, const llm_graph_para
// self-attention
{
// compute Q and K and RoPE them
ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur, model.layers[il].wq_s);
cb(Qcur, "Qcur", il);
ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur, model.layers[il].wk_s);
cb(Kcur, "Kcur", il);
ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur, model.layers[il].wv_s);
cb(Vcur, "Vcur", il);
Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
@@ -68,6 +68,9 @@ llm_build_qwen3::llm_build_qwen3(const llama_model & model, const llm_graph_para
cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
if (model.layers[il].wo_s) {
cur = ggml_mul(ctx0, cur, model.layers[il].wo_s);
}
}
if (il == n_layer - 1 && inp_out_ids) {
cur = ggml_get_rows(ctx0, cur, inp_out_ids);
@@ -83,9 +86,9 @@ llm_build_qwen3::llm_build_qwen3(const llama_model & model, const llm_graph_para
cb(cur, "ffn_norm", il);
cur = build_ffn(cur,
model.layers[il].ffn_up, NULL, NULL,
model.layers[il].ffn_gate, NULL, NULL,
model.layers[il].ffn_down, NULL, NULL,
model.layers[il].ffn_up, NULL, model.layers[il].ffn_up_s,
model.layers[il].ffn_gate, NULL, model.layers[il].ffn_gate_s,
model.layers[il].ffn_down, NULL, model.layers[il].ffn_down_s,
NULL,
LLM_FFN_SILU, LLM_FFN_PAR, il);
cb(cur, "ffn_out", il);
src/models/qwen35.cpp
+4 -8
View File
@@ -321,9 +321,9 @@ ggml_tensor * llm_build_qwen35::build_layer_attn_linear(
//v_conv = ggml_cont_4d(ctx0, v_conv, head_v_dim, num_v_heads, n_seq_tokens, n_seqs);
// if head keys and value keys are different, repeat to force tensors into matching shapes
if (num_k_heads != num_v_heads) {
// note: need explicit repeat only if we are not using the fused GDN
if (num_k_heads != num_v_heads && (!cparams.fused_gdn_ar || !cparams.fused_gdn_ch)) {
GGML_ASSERT(num_v_heads % num_k_heads == 0);
// TODO: try to avoid these explicit repeats by utilizing op broadcast
q_conv = ggml_repeat_4d(ctx0, q_conv, head_k_dim, num_v_heads, n_seq_tokens, n_seqs);
k_conv = ggml_repeat_4d(ctx0, k_conv, head_k_dim, num_v_heads, n_seq_tokens, n_seqs);
}
@@ -332,12 +332,8 @@ ggml_tensor * llm_build_qwen35::build_layer_attn_linear(
cb(k_conv, "k_conv_predelta", il);
cb(v_conv, "v_conv_predelta", il);
std::pair<ggml_tensor *, ggml_tensor *> attn_out;
if (n_seq_tokens == 1) {
attn_out = build_delta_net_autoregressive(q_conv, k_conv, v_conv, gate, beta, state, il);
} else {
attn_out = build_delta_net_chunking(q_conv, k_conv, v_conv, gate, beta, state, il);
}
auto attn_out = build_delta_net(q_conv, k_conv, v_conv, gate, beta, state, il);
ggml_tensor * output = attn_out.first;
ggml_tensor * new_state = attn_out.second;
cb(output, "attn_output", il);
src/models/qwen35moe.cpp
+4 -8
View File
@@ -321,9 +321,9 @@ ggml_tensor * llm_build_qwen35moe ::build_layer_attn_linear(
//v_conv = ggml_cont_4d(ctx0, v_conv, head_v_dim, num_v_heads, n_seq_tokens, n_seqs);
// if head keys and value keys are different, repeat to force tensors into matching shapes
if (num_k_heads != num_v_heads) {
// note: need explicit repeat only if we are not using the fused GDN
if (num_k_heads != num_v_heads && (!cparams.fused_gdn_ar || !cparams.fused_gdn_ch)) {
GGML_ASSERT(num_v_heads % num_k_heads == 0);
// TODO: try to avoid these explicit repeats by utilizing op broadcast
q_conv = ggml_repeat_4d(ctx0, q_conv, head_k_dim, num_v_heads, n_seq_tokens, n_seqs);
k_conv = ggml_repeat_4d(ctx0, k_conv, head_k_dim, num_v_heads, n_seq_tokens, n_seqs);
}
@@ -332,12 +332,8 @@ ggml_tensor * llm_build_qwen35moe ::build_layer_attn_linear(
cb(k_conv, "k_conv_predelta", il);
cb(v_conv, "v_conv_predelta", il);
std::pair<ggml_tensor *, ggml_tensor *> attn_out;
if (n_seq_tokens == 1) {
attn_out = build_delta_net_autoregressive(q_conv, k_conv, v_conv, gate, beta, state, il);
} else {
attn_out = build_delta_net_chunking(q_conv, k_conv, v_conv, gate, beta, state, il);
}
auto attn_out = build_delta_net(q_conv, k_conv, v_conv, gate, beta, state, il);
ggml_tensor * output = attn_out.first;
ggml_tensor * new_state = attn_out.second;
cb(output, "attn_output", il);
src/models/qwen3moe.cpp
+11 -4
View File
@@ -30,13 +30,13 @@ llm_build_qwen3moe::llm_build_qwen3moe(const llama_model & model, const llm_grap
// self_attention
{
// compute Q and K and RoPE them
ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur, model.layers[il].wq_s);
cb(Qcur, "Qcur", il);
ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur, model.layers[il].wk_s);
cb(Kcur, "Kcur", il);
ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur, model.layers[il].wv_s);
cb(Vcur, "Vcur", il);
Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
@@ -68,6 +68,9 @@ llm_build_qwen3moe::llm_build_qwen3moe(const llama_model & model, const llm_grap
cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
if (model.layers[il].wo_s) {
cur = ggml_mul(ctx0, cur, model.layers[il].wo_s);
}
}
if (il == n_layer - 1 && inp_out_ids) {
cur = ggml_get_rows(ctx0, cur, inp_out_ids);
@@ -93,7 +96,11 @@ llm_build_qwen3moe::llm_build_qwen3moe(const llama_model & model, const llm_grap
LLM_FFN_SILU, true,
hparams.expert_weights_scale,
LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
il);
il,
nullptr, nullptr,
model.layers[il].ffn_up_exps_s,
model.layers[il].ffn_gate_exps_s,
model.layers[il].ffn_down_exps_s);
cb(moe_out, "ffn_moe_out", il);
cur = moe_out;
src/models/qwen3next.cpp
+3 -7
View File
@@ -407,6 +407,7 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear(
//v_conv = ggml_cont_4d(ctx0, v_conv, head_v_dim, num_v_heads, n_seq_tokens, n_seqs);
// if head keys and value keys are different, repeat to force tensors into matching shapes
// TODO: avoid repeats for fused GDN, needs broadcast configuration for GDN op [TAG_GGML_GDN_BCAST]
if (num_k_heads != num_v_heads) {
GGML_ASSERT(num_v_heads % num_k_heads == 0);
int64_t repeat_factor = num_v_heads / num_k_heads;
@@ -432,13 +433,8 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear(
cb(k_conv, "k_conv_predelta", il);
cb(v_conv, "v_conv_predelta", il);
// Choose between build_delta_net_chunking, build_delta_net_recurrent, and build_delta_net_autoregressive based on n_tokens
std::pair<ggml_tensor *, ggml_tensor *> attn_out; // pair of (output, new_state)
if (n_seq_tokens == 1) {
attn_out = build_delta_net_autoregressive(q_conv, k_conv, v_conv, gate, beta, state, il);
} else {
attn_out = build_delta_net_chunking(q_conv, k_conv, v_conv, gate, beta, state, il);
}
auto attn_out = build_delta_net(q_conv, k_conv, v_conv, gate, beta, state, il);
ggml_tensor * output = attn_out.first;
ggml_tensor * new_state = attn_out.second;
cb(output, "attn_output", il);