tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-07-02 05:15:47 +00:00
Code Issues Packages Projects Releases Wiki Activity

ggml : full ALiBi support (#7192)

Browse Source 
* ggml : full ALiBi support

* ggml : update ggml_soft_max_ext() CUDA, SYCL

* ggml : ggml_flash_attn_ext() support ALiBi (CPU)

* ggml : ggml_flash_attn_ext() support ALiBi (Metal)

* ggml : fix warning

* ggml : ggml_flash_attn_ext() support ALiBi (CUDA)

ggml-ci

* ggml : fix assert message

* vulkan : add dev notes

* ggml : require mask when using ALiBi

ggml-ci

* convert : fix convert for refact models
This commit is contained in:
Georgi Gerganov
2024-05-11 10:32:41 +03:00
committed by GitHub
parent e849648888
commit 9cb317f77e
16 changed files with 350 additions and 825 deletions
Download Patch File Download Diff File Expand all files Collapse all files

178
llama.cpp
View File

@ -1845,7 +1845,7 @@ struct llama_hparams {
float f_logit_scale = 0.0f;
bool causal_attn = true;
bool use_alibi = false; // currently, we need KQ_pos data for ALiBi-based models
bool use_alibi = false;
enum llama_pooling_type pooling_type = LLAMA_POOLING_TYPE_NONE;
enum llama_rope_type rope_type = LLAMA_ROPE_TYPE_NONE;
@ -2317,7 +2317,6 @@ struct llama_context {
struct ggml_tensor * inp_pos; // I32 [n_batch]
struct ggml_tensor * inp_out_ids; // I32 [n_outputs]
struct ggml_tensor * inp_KQ_mask; // F32 [kv_size, n_batch]
struct ggml_tensor * inp_KQ_pos; // F32 [n_kv]
struct ggml_tensor * inp_K_shift; // I32 [kv_size]
struct ggml_tensor * inp_mean; // F32 [n_batch, n_batch]
struct ggml_tensor * inp_cls; // I32 [n_batch]
@ -6500,7 +6499,6 @@ static struct ggml_tensor * llm_build_kqv(
struct ggml_tensor * wo_b,
struct ggml_tensor * q_cur,
struct ggml_tensor * kq_mask,
struct ggml_tensor * kq_pos,
int32_t n_tokens,
int32_t n_kv,
float kq_scale,
@ -6530,10 +6528,6 @@ static struct ggml_tensor * llm_build_kqv(
GGML_UNUSED(model);
GGML_UNUSED(n_ctx);
// note: if this assert triggers, then some check has failed earlier
// the idea is to detect during context creation that ALiBi would be used and disable Flash Attention
GGML_ASSERT(kq_pos == nullptr && "ALiBi is not yet supported with Flash Attention");
// split cached v into n_head heads (not transposed)
struct ggml_tensor * v =
ggml_view_3d(ctx, kv.v_l[il],
@ -6543,7 +6537,7 @@ static struct ggml_tensor * llm_build_kqv(
0);
cb(v, "v", il);
cur = ggml_flash_attn_ext(ctx, q, k, v, kq_mask, kq_scale);
cur = ggml_flash_attn_ext(ctx, q, k, v, kq_mask, kq_scale, hparams.f_max_alibi_bias);
if (model.arch == LLM_ARCH_PHI2 || model.arch == LLM_ARCH_PHI3) {
ggml_flash_attn_ext_set_prec(cur, GGML_PREC_F32);
@ -6574,28 +6568,8 @@ static struct ggml_tensor * llm_build_kqv(
kq = ggml_scale(ctx, kq, 30);
}
#if defined(GGML_USE_KOMPUTE)
#pragma message("TODO: ALiBi support in ggml_soft_max_ext is not implemented for Kompute")
#pragma message(" Falling back to ggml_alibi(). Will become an error in Mar 2024")
#pragma message("ref: https://github.com/ggerganov/llama.cpp/pull/5488")
if (hparams.use_alibi) {
kq = ggml_scale(ctx, kq, kq_scale);
cb(kq, "kq_scaled", il);
kq = ggml_alibi(ctx, kq, /*n_past*/ 0, n_head, hparams.f_max_alibi_bias);
cb(kq, "kq_scaled_alibi", il);
kq = ggml_add(ctx, kq, kq_mask);
cb(kq, "kq_masked", il);
kq = ggml_soft_max(ctx, kq);
cb(kq, "kq_soft_max", il);
} else
#endif
{
kq = ggml_soft_max_ext(ctx, kq, kq_mask, kq_pos, kq_scale, hparams.f_max_alibi_bias);
cb(kq, "kq_soft_max_ext", il);
}
kq = ggml_soft_max_ext(ctx, kq, kq_mask, kq_scale, hparams.f_max_alibi_bias);
cb(kq, "kq_soft_max_ext", il);
GGML_ASSERT(kv.size == n_ctx);
@ -6645,7 +6619,6 @@ static struct ggml_tensor * llm_build_kv(
struct ggml_tensor * v_cur,
struct ggml_tensor * q_cur,
struct ggml_tensor * kq_mask,
struct ggml_tensor * kq_pos,
int32_t n_tokens,
int32_t kv_head,
int32_t n_kv,
@ -6664,7 +6637,7 @@ static struct ggml_tensor * llm_build_kv(
struct ggml_tensor * cur;
cur = llm_build_kqv(ctx, model, hparams, cparams, kv, graph, wo, wo_b,
q_cur, kq_mask, kq_pos, n_tokens, n_kv, kq_scale, cb, il);
q_cur, kq_mask, n_tokens, n_kv, kq_scale, cb, il);
cb(cur, "kqv_out", il);
return cur;
@ -6771,18 +6744,17 @@ struct llm_build_context {
ctx0 = ggml_init(params);
lctx.inp_tokens = nullptr;
lctx.inp_embd = nullptr;
lctx.inp_pos = nullptr;
lctx.inp_tokens = nullptr;
lctx.inp_embd = nullptr;
lctx.inp_pos = nullptr;
lctx.inp_out_ids = nullptr;
lctx.inp_KQ_mask = nullptr;
lctx.inp_KQ_pos = nullptr;
lctx.inp_K_shift = nullptr;
lctx.inp_mean = nullptr;
lctx.inp_cls = nullptr;
lctx.inp_s_copy = nullptr;
lctx.inp_s_mask = nullptr;
lctx.inp_s_seq = nullptr;
lctx.inp_mean = nullptr;
lctx.inp_cls = nullptr;
lctx.inp_s_copy = nullptr;
lctx.inp_s_mask = nullptr;
lctx.inp_s_seq = nullptr;
}
void free() {
@ -6932,19 +6904,6 @@ struct llm_build_context {
return flash_attn ? ggml_cast(ctx0, lctx.inp_KQ_mask, GGML_TYPE_F16) : lctx.inp_KQ_mask;
}
struct ggml_tensor * build_inp_KQ_pos(bool causal = true) {
if (causal) {
lctx.inp_KQ_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, n_kv);
} else {
// TODO: this will be needed for ALiBi-based BERT models
// https://github.com/ggerganov/llama.cpp/pull/6826
lctx.inp_KQ_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, n_tokens);
}
cb(lctx.inp_KQ_pos, "KQ_pos", -1);
ggml_set_input(lctx.inp_KQ_pos);
return flash_attn ? ggml_cast(ctx0, lctx.inp_KQ_pos, GGML_TYPE_F16) : lctx.inp_KQ_pos;
}
struct ggml_tensor * build_inp_mean() {
lctx.inp_mean = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_tokens, n_tokens);
cb(lctx.inp_mean, "inp_mean", -1);
@ -7050,7 +7009,7 @@ struct llm_build_context {
cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
model.layers[il].wo, model.layers[il].bo,
Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
}
if (il == n_layer - 1) {
@ -7143,9 +7102,6 @@ struct llm_build_context {
// KQ_mask (mask for 1 head, it will be broadcasted to all heads)
struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
// positions of the tokens in the KV cache
struct ggml_tensor * KQ_pos = build_inp_KQ_pos();
for (int il = 0; il < n_layer; ++il) {
struct ggml_tensor * inpSA = inpL;
@ -7190,7 +7146,7 @@ struct llm_build_context {
cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
model.layers[il].wo, NULL,
Kcur, Vcur, Qcur, KQ_mask, KQ_pos, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
}
if (il == n_layer - 1) {
@ -7260,9 +7216,6 @@ struct llm_build_context {
// KQ_mask (mask for 1 head, it will be broadcasted to all heads)
struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
// positions of the tokens in the KV cache
struct ggml_tensor * KQ_pos = build_inp_KQ_pos();
for (int il = 0; il < n_layer; ++il) {
struct ggml_tensor * inpSA = inpL;
@ -7297,7 +7250,7 @@ struct llm_build_context {
cb(Kcur, "Kcur", il);
cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
model.layers[il].wo, NULL,
Kcur, Vcur, Qcur, KQ_mask, KQ_pos, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
}
if (il == n_layer - 1) {
@ -7417,7 +7370,7 @@ struct llm_build_context {
cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
model.layers[il].wo, NULL,
Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
}
if (il == n_layer - 1) {
@ -7542,7 +7495,7 @@ struct llm_build_context {
cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
model.layers[il].wo, model.layers[il].bo,
Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f, cb, il);
Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f, cb, il);
}
if (il == n_layer - 1) {
@ -7694,7 +7647,7 @@ struct llm_build_context {
cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
model.layers[il].wo, NULL,
Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
}
if (il == n_layer - 1) {
@ -7806,7 +7759,7 @@ struct llm_build_context {
cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
model.layers[il].wo, model.layers[il].bo,
Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
}
if (il == n_layer - 1) {
@ -8010,7 +7963,7 @@ struct llm_build_context {
cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
model.layers[il].wo, model.layers[il].bo,
Kcur, Vcur, Q, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
Kcur, Vcur, Q, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
}
if (il == n_layer - 1) {
@ -8076,9 +8029,6 @@ struct llm_build_context {
// KQ_mask (mask for 1 head, it will be broadcasted to all heads)
struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
// positions of the tokens in the KV cache
struct ggml_tensor * KQ_pos = build_inp_KQ_pos();
for (int il = 0; il < n_layer; ++il) {
struct ggml_tensor * inpSA = inpL;
@ -8106,7 +8056,7 @@ struct llm_build_context {
cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
model.layers[il].wo, NULL,
Kcur, Vcur, Qcur, KQ_mask, KQ_pos, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
}
if (il == n_layer - 1) {
@ -8246,7 +8196,7 @@ struct llm_build_context {
struct ggml_tensor * kq = ggml_mul_mat(ctx0, k, q);
cb(kq, "kq", il);
kq = ggml_soft_max_ext(ctx0, kq, KQ_mask, nullptr, 1.0f/sqrtf(float(n_embd_head)), hparams.f_max_alibi_bias);
kq = ggml_soft_max_ext(ctx0, kq, KQ_mask, 1.0f/sqrtf(float(n_embd_head)), hparams.f_max_alibi_bias);
cb(kq, "kq_soft_max_ext", il);
struct ggml_tensor * v = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_reshape_2d(ctx0, Vcur, n_embd_gqa, n_tokens)));
@ -8363,9 +8313,6 @@ struct llm_build_context {
// KQ_mask (mask for 1 head, it will be broadcasted to all heads)
struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
// positions of the tokens in the KV cache
struct ggml_tensor * KQ_pos = build_inp_KQ_pos();
inpL = llm_build_norm(ctx0, inpL, hparams,
model.tok_norm,
model.tok_norm_b,
@ -8399,7 +8346,7 @@ struct llm_build_context {
cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
model.layers[il].wo, model.layers[il].bo,
Kcur, Vcur, Qcur, KQ_mask, KQ_pos, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
}
if (il == n_layer - 1) {
@ -8464,9 +8411,6 @@ struct llm_build_context {
// KQ_mask (mask for 1 head, it will be broadcasted to all heads)
struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
// positions of the tokens in the KV cache
struct ggml_tensor * KQ_pos = build_inp_KQ_pos();
if (model.pos_embd) {
// inp_pos - contains the positions
struct ggml_tensor * inp_pos = build_inp_pos();
@ -8530,13 +8474,13 @@ struct llm_build_context {
cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
model.layers[il].wo, model.layers[il].bo,
Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
} else {
Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
model.layers[il].wo, model.layers[il].bo,
Kcur, Vcur, Qcur, KQ_mask, KQ_pos, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
}
}
@ -8680,7 +8624,7 @@ struct llm_build_context {
cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
model.layers[il].wo, NULL,
Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
}
if (il == n_layer - 1) {
@ -8798,7 +8742,7 @@ struct llm_build_context {
cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
model.layers[il].wo, NULL,
Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
}
if (il == n_layer - 1) {
@ -8911,7 +8855,7 @@ struct llm_build_context {
cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
model.layers[il].wo, model.layers[il].bo,
Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
}
if (il == n_layer - 1) {
@ -9025,7 +8969,7 @@ struct llm_build_context {
cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
model.layers[il].wo, model.layers[il].bo,
Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
}
if (il == n_layer - 1) {
@ -9180,7 +9124,7 @@ struct llm_build_context {
cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
model.layers[il].wo, model.layers[il].bo,
Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f, cb, il);
Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f, cb, il);
}
if (il == n_layer - 1) {
@ -9297,7 +9241,7 @@ struct llm_build_context {
cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
model.layers[il].wo, model.layers[il].bo,
Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f, cb, il);
Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f, cb, il);
}
if (il == n_layer - 1) {
@ -9410,7 +9354,7 @@ struct llm_build_context {
cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
model.layers[il].wo, NULL,
Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
}
struct ggml_tensor * sa_out = cur;
@ -9513,7 +9457,7 @@ struct llm_build_context {
cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
model.layers[il].wo, model.layers[il].bo,
Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
}
if (il == n_layer - 1) {
@ -9620,7 +9564,7 @@ struct llm_build_context {
cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
model.layers[il].wo, model.layers[il].bo,
Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
}
if (il == n_layer - 1) {
@ -9736,7 +9680,7 @@ struct llm_build_context {
cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
model.layers[il].wo, NULL,
Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
}
if (il == n_layer - 1) {
@ -9853,7 +9797,7 @@ struct llm_build_context {
cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
model.layers[il].wo, model.layers[il].bo,
Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
}
if (il == n_layer - 1) {
@ -9983,7 +9927,7 @@ struct llm_build_context {
cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
model.layers[il].wo, model.layers[il].bo,
Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
}
if (il == n_layer - 1) {
@ -10104,7 +10048,7 @@ struct llm_build_context {
cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
model.layers[il].wo, NULL,
Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f, cb, il);
Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f, cb, il);
}
if (il == n_layer - 1) {
@ -10223,7 +10167,7 @@ struct llm_build_context {
cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
model.layers[il].wo, model.layers[il].bo,
Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
}
if (il == n_layer - 1) {
@ -10513,7 +10457,7 @@ struct llm_build_context {
cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
model.layers[il].wo, model.layers[il].bo,
Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
}
if (il == n_layer - 1) {
@ -10644,7 +10588,7 @@ struct llm_build_context {
cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
model.layers[il].wo, nullptr,
Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
}
if (il == n_layer - 1) {
@ -11032,11 +10976,21 @@ static void llama_set_inputs(llama_context & lctx, const llama_batch & batch) {
if (!lctx.kv_self.cells[i].has_seq_id(seq_id) || lctx.kv_self.cells[i].pos > pos) {
f = -INFINITY;
} else {
f = 0.0f;
if (hparams.use_alibi) {
f = -fabs(lctx.kv_self.cells[i].pos - pos);
} else {
f = 0.0f;
}
}
data[h*(n_kv*n_tokens) + j*n_kv + i] = f;
}
}
for (int i = n_tokens; i < GGML_PAD(n_tokens, GGML_KQ_MASK_PAD); ++i) {
for (int j = 0; j < n_kv; ++j) {
data[h*(n_kv*n_tokens) + i*n_kv + j] = -INFINITY;
}
}
}
} else {
// when using kv cache, the mask needs to match the kv cache size
@ -11055,7 +11009,11 @@ static void llama_set_inputs(llama_context & lctx, const llama_batch & batch) {
float f = -INFINITY;
for (int s = 0; s < batch.n_seq_id[i]; ++s) {
if (batch.seq_id[i][s] == seq_id) {
f = 0.0f;
if (hparams.use_alibi) {
f = -fabs(batch.pos[i] - batch.pos[j]);
} else {
f = 0.0f;
}
break;
}
}
@ -11071,21 +11029,6 @@ static void llama_set_inputs(llama_context & lctx, const llama_batch & batch) {
}
}
// ALiBi requires the KQ_pos tensor to provide the sequence position of each token in the batch
// this allows to process multiple sequences in parallel with ALiBi-based models
if (hparams.use_alibi) {
const int64_t n_kv = kv_self.n;
GGML_ASSERT(lctx.inp_KQ_pos);
GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_KQ_pos->buffer));
float * data = (float *) lctx.inp_KQ_pos->data;
for (int i = 0; i < n_kv; ++i) {
data[i] = float(lctx.kv_self.cells[i].pos);
}
}
if (cparams.pooling_type == LLAMA_POOLING_TYPE_MEAN) {
const int64_t n_tokens = batch.n_tokens;
@ -15509,11 +15452,6 @@ struct llama_context * llama_new_context_with_model(
}
}
if (cparams.flash_attn && hparams.use_alibi) {
LLAMA_LOG_WARN("%s: flash_attn is not yet compatible with ALiBi - forcing off\n", __func__);
cparams.flash_attn = false;
}
if (cparams.flash_attn && model->arch == LLM_ARCH_GROK) {
LLAMA_LOG_WARN("%s: flash_attn is not compatible with Grok - forcing off\n", __func__);
cparams.flash_attn = false;