vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

CPU/CUDA: Gemma 2 FlashAttention support (#8542)

Browse source 
* CPU/CUDA: Gemma 2 FlashAttention support

* apply logit_softcap to scale in kernel

* disable logit softcapping tests on Metal

* remove metal check
This commit is contained in:
Johannes Gäßler 2024-08-24 21:34:59 +02:00 • committed by GitHub
parent 8f824ffe8e
commit e11bd856d5
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
12 changed files with 319 additions and 79 deletions
Download patch file Download diff file Expand all files Collapse all files

22
tests/test-backend-ops.cpp
View file

@ -1652,19 +1652,20 @@ struct test_flash_attn_ext : public test_case {
const bool mask; // use mask
const float max_bias; // ALiBi
const float logit_softcap; // Gemma 2
const ggml_type type_KV;
std::string vars() override {
return VARS_TO_STR7(hs, nh, kv, nb, mask, max_bias, type_KV);
return VARS_TO_STR8(hs, nh, kv, nb, mask, max_bias, logit_softcap, type_KV);
}
double max_nmse_err() override {
return 5e-4;
}
test_flash_attn_ext(int64_t hs = 128, int64_t nh = 32, int64_t kv = 96, int64_t nb = 8, bool mask = true, float max_bias = 0.0f, ggml_type type_KV = GGML_TYPE_F16)
: hs(hs), nh(nh), kv(kv), nb(nb), mask(mask), max_bias(max_bias), type_KV(type_KV) {}
test_flash_attn_ext(int64_t hs = 128, int64_t nh = 32, int64_t kv = 96, int64_t nb = 8, bool mask = true, float max_bias = 0.0f, float logit_softcap = 0.0f, ggml_type type_KV = GGML_TYPE_F16)
: hs(hs), nh(nh), kv(kv), nb(nb), mask(mask), max_bias(max_bias), logit_softcap(logit_softcap), type_KV(type_KV) {}
ggml_tensor * build_graph(ggml_context * ctx) override {
const int64_t hs_padded = GGML_PAD(hs, ggml_blck_size(type_KV));
@ -1673,7 +1674,7 @@ struct test_flash_attn_ext : public test_case {
ggml_tensor * k = ggml_new_tensor_4d(ctx, type_KV, hs_padded, kv, nh, 1);
ggml_tensor * v = ggml_new_tensor_4d(ctx, type_KV, hs_padded, kv, nh, 1);
ggml_tensor * m = mask ? ggml_new_tensor_4d(ctx, GGML_TYPE_F16, kv, GGML_PAD(nb, GGML_KQ_MASK_PAD), 1, 1) : nullptr;
ggml_tensor * out = ggml_flash_attn_ext(ctx, q, k, v, m, 1.0f/sqrtf(hs), max_bias);
ggml_tensor * out = ggml_flash_attn_ext(ctx, q, k, v, m, 1.0f/sqrtf(hs), max_bias, logit_softcap);
return out;
}
};
@ -2437,11 +2438,14 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
for (bool mask : { true, false } ) {
for (float max_bias : { 0.0f, 8.0f }) {
if (!mask && max_bias > 0.0f) continue;
for (int nh : { 32, }) {
for (int kv : { 512, 1024, }) {
for (int nb : { 1, 2, 4, 8, }) {
for (ggml_type type_KV : {GGML_TYPE_F16, GGML_TYPE_Q8_0, GGML_TYPE_Q4_0}) {
test_cases.emplace_back(new test_flash_attn_ext(hs, nh, kv, nb, mask, max_bias, type_KV));
for (float logit_softcap : {0.0f, 10.0f}) {
if (hs != 128 && logit_softcap != 0.0f) continue;
for (int nh : { 32, }) {
for (int kv : { 512, 1024, }) {
for (int nb : { 1, 2, 4, 8, }) {
for (ggml_type type_KV : {GGML_TYPE_F16, GGML_TYPE_Q8_0, GGML_TYPE_Q4_0}) {
test_cases.emplace_back(new test_flash_attn_ext(hs, nh, kv, nb, mask, max_bias, logit_softcap, type_KV));
}
}
}
}