nicolas
/
ik_llama_opt
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Faster long context TG on CUDA for GLM-4.5/4.6/4.7/AIR (#1183) 

* Similar hack to #1182 for GLM-4.5/6/7

* Refinements

* Disable when the KV cache is not f16
This commit is contained in:
Kawrakow 2026-01-24 09:39:29 +02:00 committed by GitHub
parent f0fb76da64
commit 04beeffa4e
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 103 additions and 38 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

141
src/llama-build-context.cpp
View File

@ -1337,6 +1337,40 @@ llm_expert_gating_func_type gating_op,
return cur; return cur;
} }
static ggml_tensor * build_glm45_fa(ggml_context * ctx, ggml_tensor * q, ggml_tensor * k, ggml_tensor * v,
ggml_tensor * kq_mask, float kq_scale, bool should_use_f32_precision) {
auto ne1 = 8*v->ne[0];
auto ne2 = 4*v->ne[0];
ggml_tensor *q1, *q2;
if (q->ne[1] == 1 && k->ne[2] == 1) {
q1 = ggml_view_3d(ctx, q, q->ne[0], 1, 8, q->nb[1], q->nb[2], 0);
q2 = ggml_view_3d(ctx, q, q->ne[0], 1, 4, q->nb[1], q->nb[2], 8*q->ne[0]*ggml_element_size(q));
} else {
q1 = ggml_view_3d(ctx, q, q->ne[0], 8, k->ne[2]*q->ne[1], q->nb[2], q->nb[1]/k->ne[2], 0);
q2 = ggml_view_3d(ctx, q, q->ne[0], 4, k->ne[2]*q->ne[1], q->nb[2], q->nb[1]/k->ne[2], 8*q->ne[0]*ggml_element_size(q));
q1 = ggml_reshape_3d(ctx, ggml_cont(ctx, q1), q->ne[0], 8*k->ne[2], q->ne[1]);
q2 = ggml_reshape_3d(ctx, ggml_cont(ctx, q2), q->ne[0], 4*k->ne[2], q->ne[1]);
q1 = ggml_permute(ctx, q1, 0, 2, 1, 3);
q2 = ggml_permute(ctx, q2, 0, 2, 1, 3);
}
auto fa1 = ggml_flash_attn_ext(ctx, q1, k, v, kq_mask, kq_scale, 0.0f, 0.0f);
if (should_use_f32_precision) {
ggml_flash_attn_ext_set_prec(fa1, GGML_PREC_F32);
}
fa1 = ggml_reshape_2d(ctx, fa1, ne1, ggml_nelements(fa1)/ne1);
auto fa2 = ggml_flash_attn_ext(ctx, q2, k, v, kq_mask, kq_scale, 0.0f, 0.0f);
if (should_use_f32_precision) {
ggml_flash_attn_ext_set_prec(fa2, GGML_PREC_F32);
}
fa2 = ggml_reshape_2d(ctx, fa2, ne2, ggml_nelements(fa2)/ne2);
return ggml_concat(ctx, fa1, fa2, 0);
}
static ggml_tensor * llm_build_kqv( static ggml_tensor * llm_build_kqv(
struct ggml_context * ctx, struct ggml_context * ctx,
struct llama_context & lctx, struct llama_context & lctx,
@ -1381,6 +1415,17 @@ static ggml_tensor * llm_build_kqv(
constexpr bool use_f32_precision = false; constexpr bool use_f32_precision = false;
#endif #endif
bool should_use_f32_precision = use_f32_precision
|| model.arch == LLM_ARCH_PHI2
|| model.arch == LLM_ARCH_PHI3
|| model.arch == LLM_ARCH_GPTNEOX
|| model.arch == LLM_ARCH_QWEN2
|| model.arch == LLM_ARCH_COHERE2
|| model.arch == LLM_ARCH_GLM4
|| model.arch == LLM_ARCH_GLM4_MOE
|| model.arch == LLM_ARCH_MIMO2;
// || (model.arch == LLM_ARCH_DEEPSEEK2 && q->ne[1] <= 8);
struct ggml_tensor * cur; struct ggml_tensor * cur;
if (cparams.flash_attn) { if (cparams.flash_attn) {
@ -1396,21 +1441,27 @@ static ggml_tensor * llm_build_kqv(
0); 0);
cb(v, "v", il); cb(v, "v", il);
cur = ggml_flash_attn_ext(ctx, q, k, v, kq_mask, kq_scale, hparams.f_max_alibi_bias, if (q->ne[1] == 1 && k->ne[1] >= 8192 && q->ne[2] / k->ne[2] == 12 && !sinks && n_swa == 0 &&
hparams.attn_soft_cap ? hparams.f_attn_logit_softcapping : 0.0f); k->view_src && k->view_src->buffer && !ggml_backend_buffer_is_host(k->view_src->buffer) &&
ggml_flash_attn_ext_add_sinks(cur, sinks); k->type == GGML_TYPE_F16 && v->type == GGML_TYPE_F16) {
if (n_swa > 0) { cur = build_glm45_fa(ctx, q, k, v, kq_mask, kq_scale, should_use_f32_precision);
((int32_t *)cur->op_params)[4] = n_swa; } else {
}
// Some models produced NaNs/gibberish when FA is computed with f16 precision on CUDA cur = ggml_flash_attn_ext(ctx, q, k, v, kq_mask, kq_scale, hparams.f_max_alibi_bias,
// For DeepSeek-2, it is perfectly fine with fp16 for PP, but I get gibberish when uding fp16 for TG. hparams.attn_soft_cap ? hparams.f_attn_logit_softcapping : 0.0f);
// Not sure if it is really a matter of insufficient precision, or I have made a mistake in the fattn-vec-f16 kernel. ggml_flash_attn_ext_add_sinks(cur, sinks);
if (use_f32_precision || model.arch == LLM_ARCH_PHI2 || model.arch == LLM_ARCH_PHI3 || model.arch == LLM_ARCH_GPTNEOX || if (n_swa > 0) {
(model.arch == LLM_ARCH_DEEPSEEK2 && q->ne[1] <= 8) || model.arch == LLM_ARCH_COHERE2 || model.arch == LLM_ARCH_GLM4 || model.arch == LLM_ARCH_GLM4_MOE) { ((int32_t *)cur->op_params)[4] = n_swa;
ggml_flash_attn_ext_set_prec(cur, GGML_PREC_F32); }
// Some models produced NaNs/gibberish when FA is computed with f16 precision on CUDA
// For DeepSeek-2, it is perfectly fine with fp16 for PP, but I get gibberish when uding fp16 for TG.
// Not sure if it is really a matter of insufficient precision, or I have made a mistake in the fattn-vec-f16 kernel.
if (should_use_f32_precision) {
ggml_flash_attn_ext_set_prec(cur, GGML_PREC_F32);
}
//ggml_flash_attn_ext_set_prec(cur, GGML_PREC_F32);
} }
//ggml_flash_attn_ext_set_prec(cur, GGML_PREC_F32);
cur = ggml_reshape_2d(ctx, cur, n_embd_head_v*n_head, n_tokens); cur = ggml_reshape_2d(ctx, cur, n_embd_head_v*n_head, n_tokens);
} else { } else {
@ -1431,8 +1482,7 @@ static ggml_tensor * llm_build_kqv(
//ggml_mul_mat_set_prec(kq, GGML_PREC_F32); //ggml_mul_mat_set_prec(kq, GGML_PREC_F32);
if (use_f32_precision || model.arch == LLM_ARCH_PHI2 || model.arch == LLM_ARCH_PHI3 || model.arch == LLM_ARCH_GPTNEOX || model.arch == LLM_ARCH_QWEN2 || if (should_use_f32_precision) {
model.arch == LLM_ARCH_COHERE2 || model.arch == LLM_ARCH_GLM4 || model.arch == LLM_ARCH_GLM4_MOE || model.arch == LLM_ARCH_MIMO2) {
// for this arch, we need to perform the KQ multiplication with F32 precision, otherwise we get NaNs // for this arch, we need to perform the KQ multiplication with F32 precision, otherwise we get NaNs
// ref: https://github.com/ggerganov/llama.cpp/pull/4490#issuecomment-1859055847 // ref: https://github.com/ggerganov/llama.cpp/pull/4490#issuecomment-1859055847
ggml_mul_mat_set_prec(kq, GGML_PREC_F32); ggml_mul_mat_set_prec(kq, GGML_PREC_F32);
@ -9182,6 +9232,22 @@ ggml_tensor * llm_build_context::build_std_attention(ggml_cgraph * gf, ggml_tens
float freq_base_l = n_swa > 0 ? hparams.rope_freq_base_train_swa : cparams.rope_freq_base; float freq_base_l = n_swa > 0 ? hparams.rope_freq_base_train_swa : cparams.rope_freq_base;
float freq_scale_l = n_swa > 0 ? hparams.rope_freq_scale_train_swa : hparams.rope_freq_scale_train; float freq_scale_l = n_swa > 0 ? hparams.rope_freq_scale_train_swa : hparams.rope_freq_scale_train;
#ifdef GGML_USE_VULKAN
constexpr bool use_f32_precision = true;
#else
constexpr bool use_f32_precision = false;
#endif
bool should_use_f32_precision = use_f32_precision
|| model.arch == LLM_ARCH_PHI2
|| model.arch == LLM_ARCH_PHI3
|| model.arch == LLM_ARCH_GPTNEOX
|| model.arch == LLM_ARCH_QWEN2
|| model.arch == LLM_ARCH_COHERE2
|| model.arch == LLM_ARCH_GLM4
// || model.arch == LLM_ARCH_GLM4_MOE
|| model.arch == LLM_ARCH_MIMO2;
// || (model.arch == LLM_ARCH_DEEPSEEK2 && q->ne[1] <= 8);
if (!model.layers[il].wqkv && !model.layers[il].wqk && cparams.flash_attn && if (!model.layers[il].wqkv && !model.layers[il].wqk && cparams.flash_attn &&
model.layers[il].wq->extra && model.layers[il].wk->extra && model.layers[il].wv->extra && model.layers[il].wo->extra) { model.layers[il].wq->extra && model.layers[il].wk->extra && model.layers[il].wv->extra && model.layers[il].wo->extra) {
@ -9324,30 +9390,29 @@ ggml_tensor * llm_build_context::build_std_attention(ggml_cgraph * gf, ggml_tens
ggml_row_size(split_vl->type, n_embd_head_v), 0); ggml_row_size(split_vl->type, n_embd_head_v), 0);
cb(v, "v", il_cb); cb(v, "v", il_cb);
#ifdef GGML_USE_VULKAN if (q->ne[1] == 1 && k->ne[1] >= 65536/k->ne[2] && q->ne[2] / k->ne[2] == 12 && !sinks && n_swa == 0 &&
constexpr bool use_f32_precision = true; k->view_src && k->view_src->buffer && !ggml_backend_buffer_is_host(k->view_src->buffer) &&
#else k->type == GGML_TYPE_F16 && v->type == GGML_TYPE_F16) {
constexpr bool use_f32_precision = false; cur = build_glm45_fa(ctx0, q, k, v, KQ_mask, KQ_scale, should_use_f32_precision);
#endif
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);
cb(cur, "flash_attn", il_cb);
if (model.layers[il].attn_sinks && model.layers[il].attn_sinks->extra) {
auto split = (ggml_split_tensor_t *)model.layers[il].attn_sinks->extra;
GGML_ASSERT(split->n_device == wq->n_device);
GGML_ASSERT(split->splits[id]);
ggml_flash_attn_ext_add_sinks(cur, split->splits[id]);
} else { } else {
ggml_flash_attn_ext_add_sinks(cur, sinks); 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);
if (n_swa > 0) { cb(cur, "flash_attn", il_cb);
((int32_t *)cur->op_params)[4] = n_swa; if (model.layers[il].attn_sinks && model.layers[il].attn_sinks->extra) {
} auto split = (ggml_split_tensor_t *)model.layers[il].attn_sinks->extra;
// Some models produced NaNs/gibberish when FA is computed with f16 precision on CUDA GGML_ASSERT(split->n_device == wq->n_device);
if (use_f32_precision || model.arch == LLM_ARCH_PHI2 || model.arch == LLM_ARCH_PHI3 || model.arch == LLM_ARCH_GPTNEOX || GGML_ASSERT(split->splits[id]);
(model.arch == LLM_ARCH_DEEPSEEK2 && q->ne[1] <= 8) || model.arch == LLM_ARCH_COHERE2 || model.arch == LLM_ARCH_GLM4 || ggml_flash_attn_ext_add_sinks(cur, split->splits[id]);
model.arch == LLM_ARCH_GLM4_MOE) { } else {
ggml_flash_attn_ext_set_prec(cur, GGML_PREC_F32); ggml_flash_attn_ext_add_sinks(cur, sinks);
}
if (n_swa > 0) {
((int32_t *)cur->op_params)[4] = n_swa;
}
// Some models produced NaNs/gibberish when FA is computed with f16 precision on CUDA
if (should_use_f32_precision) {
ggml_flash_attn_ext_set_prec(cur, GGML_PREC_F32);
}
} }
cur = ggml_reshape_2d(ctx0, cur, split_wo->ne[0], n_tokens); cur = ggml_reshape_2d(ctx0, cur, split_wo->ne[0], n_tokens);