Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

ggml-webgpu: add mulmat with overlapping src0/src1 (e.g., for minimax-01) (#27321)

Browse Source
This commit is contained in:
Masashi Yoshimura
2026-08-19 16:29:33 +09:00
committed by GitHub
parent 0adcc3bb57
commit 5112b9738b
10 changed files with 189 additions and 92 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/test-backend-ops.cpp
+17 -3
View File
@@ -4470,9 +4470,10 @@ struct test_mul_mat : public test_case {
const std::array<int64_t, 4> per; // permutation of dimensions
const int64_t k_v; // size of k in memory, resulting in a non-contiguous view for k_v > k, no view for k_v == 0
const uint32_t o; // number of outputs
const bool src_overlap; // a and b are overlapping views of the same tensor
std::string vars() override {
return VARS_TO_STR10(type_a, type_b, m, n, k, bs, nr, per, k_v, o);
return VARS_TO_STR11(type_a, type_b, m, n, k, bs, nr, per, k_v, o, src_overlap);
}
double max_nmse_err() override {
@@ -4501,8 +4502,8 @@ struct test_mul_mat : public test_case {
std::array<int64_t, 2> bs = {10, 10},
std::array<int64_t, 2> nr = {2, 2},
std::array<int64_t, 4> per = {0, 1, 2, 3},
int64_t k_v = 0, uint32_t o = 1)
: type_a(type_a), type_b(type_b), m(m), n(n), k(k), bs(bs), nr(nr), per(per), k_v(k_v), o(o) {}
int64_t k_v = 0, uint32_t o = 1, bool src_overlap = false)
: type_a(type_a), type_b(type_b), m(m), n(n), k(k), bs(bs), nr(nr), per(per), k_v(k_v), o(o), src_overlap(src_overlap) {}
ggml_tensor * build_graph(ggml_context * ctx) override {
// C^T = A * B^T: (k, m) * (k, n) => (m, n)
@@ -4535,6 +4536,18 @@ struct test_mul_mat : public test_case {
b = ggml_permute(ctx, b, per[0], per[1], per[2], per[3]);
ggml_set_name(a, "a_permuted");
ggml_set_name(b, "b_permuted");
} else if (src_overlap) {
GGML_ASSERT(type_a == type_b);
GGML_ASSERT(k_v == 0);
// a and b are interleaved views of the same tensor: (e.g. fused QKV in MiniMax-01)
ggml_tensor * base = ggml_new_tensor_4d(ctx, type_a, 2*k, std::max(m, n), bs[0]*nr[0], bs[1]*nr[1]);
ggml_set_name(base, "base");
a = ggml_view_4d(ctx, base, k, m, bs[0], bs[1], base->nb[1], base->nb[2], base->nb[3], 0);
b = ggml_view_4d(ctx, base, k, n, bs[0]*nr[0], bs[1]*nr[1], base->nb[1], base->nb[2], base->nb[3], k*ggml_type_size(type_a));
ggml_set_name(a, "a");
ggml_set_name(b, "b");
} else {
const int64_t k_physical = k_v == 0 ? k : k_v;
a = ggml_new_tensor_4d(ctx, type_a, k_physical, m, bs[0], bs[1]);
@@ -9243,6 +9256,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 1056, 1, 67, {1, 1}, {4, 1}, {0, 2, 1, 3}));
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F32, GGML_TYPE_F32, 16, 32, 32, { 1, 1}, {1, 1}, {0, 1, 2, 3}, 64, 3));
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F32, GGML_TYPE_F32, 64, 77, 77, {12,1}, {1,1}));
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F32, GGML_TYPE_F32, 32, 4, 96, {3, 2}, {1, 1}, {0, 1, 2, 3}, 0, 1, true));
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_Q4_0, GGML_TYPE_F32, 576, 512, 576, {1,1}, {1,1}));
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_Q4_0, GGML_TYPE_F32, 1, 2048, 8192, {1, 1}, {1, 1}));
tests/test-llama-archs.cpp
+1 -1
View File
@@ -456,7 +456,7 @@ static bool arch_supported(const llm_arch arch) {
// FIXME: these hit scheduler/view-backed-output issues with WebGPU on CI.
#ifdef GGML_USE_WEBGPU
if (arch == LLM_ARCH_DEEPSEEK32 || arch == LLM_ARCH_GLM_DSA || arch == LLM_ARCH_MINIMAX_01) {
if (arch == LLM_ARCH_DEEPSEEK32 || arch == LLM_ARCH_GLM_DSA) {
return false;
}
#endif // GGML_USE_WEBGPU