tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-08-16 05:02:58 -04:00
Code Issues Packages Projects Releases Wiki Activity

llama : add gpt-oss (#15091)

Browse Source 
* oai moe

* compat with new checkpoint

* add attn sink impl

* add rope scaling yarn

* logits match with latest transformers code

* wip chat template

* rm trailing space

* use ggml_scale_bias

* rm redundant is_swa_all

* convert interleaved gate_up

* graph : fix activation function to match reference (#7)

* vocab : handle o200k_harmony special tokens

* ggml : add attention sinks support (#1)

* llama : add attn sinks

* ggml : add attn sinks

* cuda : add attn sinks

* vulkan : add support for sinks in softmax

remove unnecessary return

* ggml : add fused swiglu_oai op (#11)

* ggml : add fused swiglu_oai op

* Update ggml/src/ggml-cpu/ops.cpp

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

* update CUDA impl

* cont : metal impl

* add vulkan impl

* test-backend-ops : more test cases, clean up

* llama : remove unfused impl

* remove extra lines

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

---------

Co-authored-by: slaren <slarengh@gmail.com>

* repack mxfp4 upon conversion

* clean up a bit

* enable thinking

* add quick hack to render only some special tokens

* fix bf16 conversion

* remove vocab hack

* webui ok

* support chat parsing for gpt-oss

* fix webui

* direct mapping mxfp4, FINALLY

* force using mxfp4

* properly use lazy tensor

* ggml : add mxfp4

ggml : use e8m0 conversion instead of powf

Co-authored-by: Diego Devesa <slarengh@gmail.com>

change kvalues_mxfp4 table to match e2m1 (#6)

metal : remove quantization for now (not used)

cuda : fix disabled CUDA graphs due to ffn moe bias

vulkan : add support for mxfp4

cont : add cm2 dequant

* ggml : add ggml_add_id (#13)

* ggml : add ggml_add_id

* add cuda impl

* llama : add weight support check for add_id

* perf opt

* add vulkan impl

* rename cuda files

* add metal impl

* allow in-place ggml_add_id

* llama : keep biases on CPU with --cpu-moe

* llama : fix compile error

ggml-ci

* cuda : add fallback for __nv_cvt_e8m0_to_bf16raw

ggml-ci

* cleanup

ggml-ci

* sycl : fix supports_op for MXFP4

ggml-ci

* fix Unknown reasoning format

* ggml-cpu : fix AVX build

ggml-ci

* fix hip build

ggml-ci

* cuda : add mxfp4 dequantization support for cuBLAS

ggml-ci

* ggml-cpu : fix mxfp4 fallback definitions for some architectures

ggml-ci

* cuda : fix version required for __nv_cvt_e8m0_to_bf16raw

---------

Co-authored-by: Xuan Son Nguyen <son@huggingface.co>
Co-authored-by: slaren <slarengh@gmail.com>
This commit is contained in:
Georgi Gerganov
2025-08-05 22:10:36 +03:00
committed by GitHub
parent f324a3b715
commit fd1234cb46
83 changed files with 2942 additions and 227 deletions
Download Patch File Download Diff File Expand all files Collapse all files

119
src/llama-graph.cpp
View File

@@ -740,6 +740,8 @@ ggml_tensor * llm_graph_context::build_ffn(
cur = ggml_reglu(ctx0, cur);
cb(cur, "ffn_reglu", il);
} break;
default:
GGML_ABORT("fatal error");
}
if (gate && type_gate == LLM_FFN_PAR) {
@@ -787,6 +789,45 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
llama_expert_gating_func_type gating_op,
int il,
ggml_tensor * probs_in) const {
return build_moe_ffn(
cur,
gate_inp, /* gate_inp_b */ nullptr,
up_exps, /* up_exps_b */ nullptr,
gate_exps, /* gate_exps_b */ nullptr,
down_exps, /* down_exps_b */ nullptr,
exp_probs_b,
n_expert,
n_expert_used,
type_op,
norm_w,
scale_w,
w_scale,
gating_op,
il,
probs_in
);
}
ggml_tensor * llm_graph_context::build_moe_ffn(
ggml_tensor * cur,
ggml_tensor * gate_inp,
ggml_tensor * gate_inp_b,
ggml_tensor * up_exps,
ggml_tensor * up_exps_b,
ggml_tensor * gate_exps,
ggml_tensor * gate_exps_b,
ggml_tensor * down_exps,
ggml_tensor * down_exps_b,
ggml_tensor * exp_probs_b,
int64_t n_expert,
int64_t n_expert_used,
llm_ffn_op_type type_op,
bool norm_w,
bool scale_w,
float w_scale,
llama_expert_gating_func_type gating_op,
int il,
ggml_tensor * probs_in) 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
@@ -800,6 +841,11 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
logits = probs_in;
}
if (gate_inp_b) {
logits = ggml_add(ctx0, logits, gate_inp_b);
cb(logits, "ffn_moe_logits_biased", il);
}
ggml_tensor * probs = nullptr;
switch (gating_op) {
case LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX:
@@ -810,6 +856,10 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
{
probs = ggml_sigmoid(ctx0, logits); // [n_expert, n_tokens]
} break;
case LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX_WEIGHT:
{
probs = logits; // [n_expert, n_tokens]
} break;
default:
GGML_ABORT("fatal error");
}
@@ -838,6 +888,13 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
ggml_reshape_3d(ctx0, probs, 1, n_expert, n_tokens), selected_experts); // [1, n_expert_used, n_tokens]
cb(weights, "ffn_moe_weights", il);
if (gating_op == LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX_WEIGHT) {
weights = ggml_reshape_2d(ctx0, weights, n_expert_used, n_tokens);
weights = ggml_soft_max(ctx0, weights); // [n_expert_used, n_tokens]
weights = ggml_reshape_3d(ctx0, weights, 1, n_expert_used, n_tokens);
cb(weights, "ffn_moe_weights_softmax", il);
}
if (norm_w) {
weights = ggml_reshape_2d(ctx0, weights, n_expert_used, n_tokens);
@@ -866,6 +923,11 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
ggml_tensor * up = build_lora_mm_id(up_exps, cur, selected_experts); // [n_ff, n_expert_used, n_tokens]
cb(up, "ffn_moe_up", il);
if (up_exps_b) {
up = ggml_add_id(ctx0, up, up_exps_b, selected_experts);
cb(up, "ffn_moe_up_biased", il);
}
ggml_tensor * experts = nullptr;
if (gate_exps) {
cur = build_lora_mm_id(gate_exps, cur, selected_experts); // [n_ff, n_expert_used, n_tokens]
@@ -874,6 +936,11 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
cur = up;
}
if (gate_exps_b) {
cur = ggml_add_id(ctx0, cur, gate_exps_b, selected_experts);
cb(cur, "ffn_moe_gate_biased", il);
}
switch (type_op) {
case LLM_FFN_SILU:
if (gate_exps) {
@@ -891,6 +958,14 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
cur = ggml_gelu(ctx0, cur);
cb(cur, "ffn_moe_gelu", il);
} break;
case LLM_FFN_SWIGLU_OAI_MOE:
{
// TODO: move to hparams?
constexpr float alpha = 1.702f;
constexpr float limit = 7.0f;
cur = ggml_swiglu_oai(ctx0, cur, up, alpha, limit);
cb(cur, "ffn_moe_swiglu_oai", il);
} break;
case LLM_FFN_RELU:
if (gate_exps) {
cur = ggml_reglu_split(ctx0, cur, up);
@@ -906,6 +981,11 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
experts = build_lora_mm_id(down_exps, cur, selected_experts); // [n_embd, n_expert_used, n_tokens]
cb(experts, "ffn_moe_down", il);
if (down_exps_b) {
experts = ggml_add_id(ctx0, experts, down_exps_b, selected_experts);
cb(experts, "ffn_moe_down_biased", il);
}
if (!weight_before_ffn) {
experts = ggml_mul(ctx0, experts, weights);
cb(cur, "ffn_moe_weighted", il);
@@ -1144,6 +1224,7 @@ ggml_tensor * llm_graph_context::build_attn_mha(
ggml_tensor * kq_b,
ggml_tensor * kq_mask,
ggml_tensor * v_mla,
ggml_tensor * sinks,
float kq_scale) const {
const bool v_trans = v->nb[1] > v->nb[2];
@@ -1180,7 +1261,8 @@ ggml_tensor * llm_graph_context::build_attn_mha(
cur = ggml_flash_attn_ext(ctx0, q, k, v, kq_mask, kq_scale, hparams.f_max_alibi_bias,
hparams.attn_soft_cap ? hparams.f_attn_logit_softcapping : 0.0f);
ggml_flash_attn_ext_set_prec(cur, GGML_PREC_F32);
ggml_flash_attn_ext_add_sinks(cur, sinks);
ggml_flash_attn_ext_set_prec (cur, GGML_PREC_F32);
if (v_mla) {
#if 0
@@ -1228,6 +1310,7 @@ ggml_tensor * llm_graph_context::build_attn_mha(
}
kq = ggml_soft_max_ext(ctx0, kq, kq_mask, kq_scale, hparams.f_max_alibi_bias);
ggml_soft_max_add_sinks(kq, sinks);
if (!v_trans) {
// note: avoid this branch
@@ -1298,7 +1381,7 @@ ggml_tensor * llm_graph_context::build_attn(
ggml_tensor * k = k_cur;
ggml_tensor * v = v_cur;
ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, v_mla, kq_scale);
ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, v_mla, nullptr, kq_scale);
cb(cur, "kqv_out", il);
if (wo) {
@@ -1386,7 +1469,7 @@ ggml_tensor * llm_graph_context::build_attn(
ggml_tensor * k = mctx_cur->get_k(ctx0, il);
ggml_tensor * v = mctx_cur->get_v(ctx0, il);
ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, v_mla, kq_scale);
ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, v_mla, nullptr, kq_scale);
cb(cur, "kqv_out", il);
if (wo) {
@@ -1415,6 +1498,32 @@ ggml_tensor * llm_graph_context::build_attn(
ggml_tensor * v_mla,
float kq_scale,
int il) const {
return build_attn_with_sinks(
inp,
wo,
wo_b,
q_cur,
k_cur,
v_cur,
kq_b,
v_mla,
nullptr,
kq_scale,
il);
}
ggml_tensor * llm_graph_context::build_attn_with_sinks(
llm_graph_input_attn_kv_unified_iswa * inp,
ggml_tensor * wo,
ggml_tensor * wo_b,
ggml_tensor * q_cur,
ggml_tensor * k_cur,
ggml_tensor * v_cur,
ggml_tensor * kq_b,
ggml_tensor * v_mla,
ggml_tensor * sinks,
float kq_scale,
int il) const {
// these nodes are added to the graph together so that they are not reordered
// by doing so, the number of splits in the graph is reduced
ggml_build_forward_expand(gf, q_cur);
@@ -1452,7 +1561,7 @@ ggml_tensor * llm_graph_context::build_attn(
ggml_tensor * k = mctx_cur->get_k(ctx0, il);
ggml_tensor * v = mctx_cur->get_v(ctx0, il);
ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, v_mla, kq_scale);
ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, v_mla, sinks, kq_scale);
cb(cur, "kqv_out", il);
if (wo) {
@@ -1506,7 +1615,7 @@ ggml_tensor * llm_graph_context::build_attn(
ggml_tensor * k = k_cur;
ggml_tensor * v = v_cur;
ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, v_mla, kq_scale);
ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, v_mla, nullptr, kq_scale);
cb(cur, "kqv_out", il);
if (wo) {