summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
authorJohannes Gäßler <johannesg@5d6.de>2024-06-01 23:26:10 +0200
committerGitHub <noreply@github.com>2024-06-01 23:26:10 +0200
commite141ce624af57bdffbaf57014a044eb1d9689230 (patch)
treef41f3a1b08c3ce608de244e0e45ff1ed699d2ce9
parent2e666832e6ac78194edf030bd1c295e21bdb022c (diff)
Fix FlashAttention debug test, FP32 assert (#7684)
-rw-r--r--ggml-cuda/fattn-vec-f32.cuh4
-rw-r--r--tests/test-backend-ops.cpp8
2 files changed, 5 insertions, 7 deletions
diff --git a/ggml-cuda/fattn-vec-f32.cuh b/ggml-cuda/fattn-vec-f32.cuh
index ce23a4eb..ddf0c837 100644
--- a/ggml-cuda/fattn-vec-f32.cuh
+++ b/ggml-cuda/fattn-vec-f32.cuh
@@ -278,14 +278,10 @@ void ggml_cuda_flash_attn_ext_vec_f32_case_impl(ggml_backend_cuda_context & ctx,
template <int D, ggml_type type_K, ggml_type type_V>
void ggml_cuda_flash_attn_ext_vec_f32_case(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
- ggml_tensor * KQV = dst;
ggml_tensor * Q = dst->src[0];
ggml_tensor * K = dst->src[1];
ggml_tensor * V = dst->src[2];
- const int32_t precision = KQV->op_params[2];
- GGML_ASSERT(precision == GGML_PREC_DEFAULT);
-
GGML_ASSERT(K->type == type_K);
GGML_ASSERT(V->type == type_V);
diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp
index 77723012..8dc90a45 100644
--- a/tests/test-backend-ops.cpp
+++ b/tests/test-backend-ops.cpp
@@ -1584,9 +1584,11 @@ struct test_flash_attn_ext : public test_case {
: 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, type_KV, hs, kv, nh, 1);
- ggml_tensor * v = ggml_new_tensor_4d(ctx, type_KV, hs, kv, nh, 1);
+ const int64_t hs_padded = GGML_PAD(hs, ggml_blck_size(type_KV));
+
+ ggml_tensor * q = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, hs_padded, nb, nh, 1);
+ 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);
return out;