Mirrors/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-26 14:20:31 +01:00
Code Issues Packages Projects Releases Wiki Activity

add q8_0 q4_0 tests

Browse Source
This commit is contained in:
Johannes Gäßler 2024-05-26 22:30:46 +02:00
parent 3194a01058
commit f08776041d
2 changed files with 16 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
ggml-cuda.cu
View File

@ -2889,10 +2889,14 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
return op->src[0]->ne[0] == 64 || op->src[0]->ne[0] == 128;
#else
if (op->src[0]->ne[0] == 64 || op->src[0]->ne[0] == 128) {
if (op->src[0]->ne[0] == 128) {
return true;
}
return ggml_cuda_info().devices[cuda_ctx->device].cc >= CC_VOLTA;
if (op->src[0]->ne[0] == 64 && op->src[1]->type == GGML_TYPE_F16) {
return true;
}
return ggml_cuda_info().devices[cuda_ctx->device].cc >= CC_VOLTA &&
op->src[1]->type == GGML_TYPE_F16 && op->src[2]->type == GGML_TYPE_F16;
#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
default:
return false;

16
tests/test-backend-ops.cpp
View File

@ -1540,21 +1540,23 @@ struct test_flash_attn_ext : public test_case {
const float max_bias; // ALiBi
const ggml_type type_KV;
std::string vars() override {
return VARS_TO_STR6(hs, nh, kv, nb, mask, max_bias);
return VARS_TO_STR7(hs, nh, kv, nb, mask, max_bias, 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)
: hs(hs), nh(nh), kv(kv), nb(nb), mask(mask), max_bias(max_bias) {}
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) {}
ggml_tensor * build_graph(ggml_context * ctx) override {
ggml_tensor * q = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, hs, nb, nh, 1);
ggml_tensor * k = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, hs, kv, nh, 1);
ggml_tensor * v = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, hs, kv, nh, 1);
ggml_tensor * k = ggml_new_tensor_4d(ctx, type_KV, hs, kv, nh, 1);
ggml_tensor * v = ggml_new_tensor_4d(ctx, type_KV, hs, 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);
return out;
@ -2238,7 +2240,9 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
for (int nh : { 32, }) {
for (int kv : { 512, 1024, }) {
for (int nb : { 1, 2, 4, 8, }) {
test_cases.emplace_back(new test_flash_attn_ext(hs, nh, kv, nb, mask, max_bias));
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));
}
}
}
}