@fugood/llama.node 0.3.15 → 0.3.17

package/src/llama.cpp/tests/test-backend-ops.cpp

@@ -271,6 +271,14 @@ static std::string var_to_str(ggml_op_pool pool) {
     }
 }
 
+static std::string var_to_str(ggml_scale_mode mode) {
+    switch (mode) {
+        case GGML_SCALE_MODE_NEAREST:  return "nearest";
+        case GGML_SCALE_MODE_BILINEAR: return "bilinear";
+        default:                      return std::to_string(mode);
+    }
+}
+
 #define VAR_TO_STR(x) (#x "=" + var_to_str(x))
 
 #define VARS_TO_STR1(a) VAR_TO_STR(a)
@@ -1463,11 +1471,13 @@ struct test_cpy : public test_case {
     const ggml_type type_src;
     const ggml_type type_dst;
     const std::array<int64_t, 4> ne;
-    const std::array<int64_t, 4> permute;
+    const std::array<int64_t, 4> permute_src;
+    const std::array<int64_t, 4> permute_dst;
     bool _src_use_permute;
+    bool _dst_use_permute;
 
     std::string vars() override {
-        return VARS_TO_STR4(type_src, type_dst, ne, permute);
+        return VARS_TO_STR5(type_src, type_dst, ne, permute_src, permute_dst);
     }
 
     double max_nmse_err() override {
@@ -1480,9 +1490,11 @@ struct test_cpy : public test_case {
 
     test_cpy(ggml_type type_src = GGML_TYPE_F32, ggml_type type_dst = GGML_TYPE_F32,
             std::array<int64_t, 4> ne = {10, 10, 10, 1},
-            std::array<int64_t, 4> permute = {0, 0, 0, 0})
-        : type_src(type_src), type_dst(type_dst), ne(ne), permute(permute),
-          _src_use_permute(permute[0] + permute[1] + permute[2] + permute[3] > 0) {}
+            std::array<int64_t, 4> permute_src = {0, 0, 0, 0},
+            std::array<int64_t, 4> permute_dst = {0, 0, 0, 0})
+        : type_src(type_src), type_dst(type_dst), ne(ne), permute_src(permute_src), permute_dst(permute_dst),
+          _src_use_permute(permute_src[0] + permute_src[1] + permute_src[2] + permute_src[3] > 0),
+          _dst_use_permute(permute_dst[0] + permute_dst[1] + permute_dst[2] + permute_dst[3] > 0) {}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * src = ggml_new_tensor(ctx, type_src, 4, ne.data());
@@ -1490,13 +1502,18 @@ struct test_cpy : public test_case {
         ggml_set_name(src, "src");
 
         if (_src_use_permute) {
-            src = ggml_permute(ctx, src, permute[0], permute[1], permute[2], permute[3]);
+            src = ggml_permute(ctx, src, permute_src[0], permute_src[1], permute_src[2], permute_src[3]);
             ggml_set_name(src, "src_permuted");
         }
 
-        ggml_tensor* dst = ggml_new_tensor(ctx, type_dst, 4, src->ne);
+        ggml_tensor * dst = ggml_new_tensor(ctx, type_dst, 4, src->ne);
         ggml_set_name(dst, "dst");
 
+        if (_dst_use_permute) {
+            dst = ggml_permute(ctx, dst, permute_dst[0], permute_dst[1], permute_dst[2], permute_dst[3]);
+            ggml_set_name(dst, "dst_permuted");
+        }
+
         ggml_tensor * out = ggml_cpy(ctx, src, dst);
         ggml_set_name(out, "out");
 
@@ -1964,9 +1981,10 @@ struct test_mul_mat : public test_case {
     const std::array<int64_t, 2> bs;  // dims 3 and 4
     const std::array<int64_t, 2> nr;  // repeat in dims 3 and 4
     const std::array<int64_t, 4> per; // permutation of dimensions
+    const bool v; // whether a is a non-contiguous view
 
     std::string vars() override {
-        return VARS_TO_STR8(type_a, type_b, m, n, k, bs, nr, per);
+        return VARS_TO_STR9(type_a, type_b, m, n, k, bs, nr, per, v);
     }
 
     double max_nmse_err() override {
@@ -1986,8 +2004,9 @@ struct test_mul_mat : public test_case {
             int64_t m = 32, int64_t n = 32, int64_t k = 32,
             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})
-        : type_a(type_a), type_b(type_b), m(m), n(n), k(k), bs(bs), nr(nr), per(per) {}
+            std::array<int64_t, 4> per = {0, 1, 2, 3},
+            bool v = false)
+        : type_a(type_a), type_b(type_b), m(m), n(n), k(k), bs(bs), nr(nr), per(per), v(v) {}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         // C^T = A * B^T: (k, m) * (k, n) => (m, n)
@@ -1997,6 +2016,7 @@ struct test_mul_mat : public test_case {
         const int npermuted = (per[0] != 0) + (per[1] != 1) + (per[2] != 2) + (per[3] != 3);
         if (npermuted > 0) {
             GGML_ASSERT(npermuted == 2);
+            GGML_ASSERT(!v); // not handled
             GGML_ASSERT(!ggml_is_quantized(type_a) || per[0] == 0);
             GGML_ASSERT(!ggml_is_quantized(type_b) || per[0] == 0);
 
@@ -2020,7 +2040,13 @@ struct test_mul_mat : public test_case {
             ggml_set_name(a, "a_permuted");
             ggml_set_name(b, "b_permuted");
         } else {
-            a = ggml_new_tensor_4d(ctx, type_a, k, m, bs[0],       bs[1]);
+
+            if (v) {
+                a = ggml_new_tensor_4d(ctx, type_a, k*2, m, bs[0], bs[1]);
+                a = ggml_view_4d(ctx, a, k, m, bs[0], bs[1], a->nb[1], a->nb[2], a->nb[3], 0);
+            } else {
+                a = ggml_new_tensor_4d(ctx, type_a, k, m, bs[0],       bs[1]);
+            }
             b = ggml_new_tensor_4d(ctx, type_b, k, n, bs[0]*nr[0], bs[1]*nr[1]);
             if (!ggml_is_quantized(type_a)) {
                 if (bs[1] == 1 && nr[1] == 1) {
@@ -2045,7 +2071,7 @@ struct test_mul_mat_id : public test_case {
     const ggml_type type_b;
     const int n_mats;
     const int n_used;
-    const bool b; // brodcast b matrix
+    const bool b; // broadcast b matrix
     const int64_t m;
     const int64_t n;
     const int64_t k;
@@ -2580,6 +2606,8 @@ struct test_rope : public test_case {
             } else {
                 out = ggml_rope_ext_back(ctx, a, pos, freq, n_dims, mode, 0, 10000.0f, fs, ef, af, 1.0f, 1.0f);
             }
+
+            // TODO: add test with a non-contiguous view as input ; this case is needed for build_rope_2d in clip.cpp
         }
         ggml_set_name(out, "out");
 
@@ -2930,15 +2958,16 @@ struct test_upscale : public test_case {
     const std::array<int64_t, 4> ne;
     const int32_t scale_factor;
     const bool transpose;
+    const ggml_scale_mode mode;
 
     std::string vars() override {
-        return VARS_TO_STR4(type, ne, scale_factor, transpose);
+        return VARS_TO_STR5(type, ne, scale_factor, mode, transpose);
     }
 
     test_upscale(ggml_type type = GGML_TYPE_F32,
             std::array<int64_t, 4> ne = {512, 512, 3, 1},
-            int32_t scale_factor = 2, bool transpose = false)
-        : type(type), ne(ne), scale_factor(scale_factor), transpose(transpose) {}
+            int32_t scale_factor = 2, ggml_scale_mode mode = GGML_SCALE_MODE_NEAREST, bool transpose = false)
+        : type(type), ne(ne), scale_factor(scale_factor), transpose(transpose), mode(mode) {}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
@@ -2949,7 +2978,7 @@ struct test_upscale : public test_case {
             ggml_set_name(a, "a_transposed");
         }
 
-        ggml_tensor * out = ggml_upscale(ctx, a, scale_factor);
+        ggml_tensor * out = ggml_upscale(ctx, a, scale_factor, mode);
         ggml_set_name(out, "out");
 
         return out;
@@ -2961,21 +2990,23 @@ struct test_upscale_ext : public test_case {
     const ggml_type type;
     const std::array<int64_t, 4> ne;
     const std::array<int64_t, 4> ne_tgt;
+    const ggml_scale_mode mode = GGML_SCALE_MODE_NEAREST;
 
     std::string vars() override {
-        return VARS_TO_STR3(type, ne, ne_tgt);
+        return VARS_TO_STR4(type, ne, ne_tgt, mode);
     }
 
     test_upscale_ext(ggml_type type = GGML_TYPE_F32,
             std::array<int64_t, 4> ne     = {2, 5,  7, 11},
-            std::array<int64_t, 4> ne_tgt = {5, 7, 11, 13})
-        : type(type), ne(ne), ne_tgt(ne_tgt) {}
+            std::array<int64_t, 4> ne_tgt = {5, 7, 11, 13},
+            ggml_scale_mode mode = GGML_SCALE_MODE_NEAREST)
+        : type(type), ne(ne), ne_tgt(ne_tgt), mode(mode) {}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
         ggml_set_name(a, "a");
 
-        ggml_tensor * out = ggml_upscale_ext(ctx, a, ne_tgt[0], ne_tgt[1],ne_tgt[2], ne_tgt[3]);
+        ggml_tensor * out = ggml_upscale_ext(ctx, a, ne_tgt[0], ne_tgt[1],ne_tgt[2], ne_tgt[3], mode);
         ggml_set_name(out, "out");
 
         return out;
@@ -3199,7 +3230,8 @@ struct test_leaky_relu : public test_case {
 
 // GGML_OP_FLASH_ATTN_EXT
 struct test_flash_attn_ext : public test_case {
-    const int64_t hs; // head size
+    const int64_t hsk; // K head size
+    const int64_t hsv; // V head size
     const int64_t nh; // num heads
     const int64_t nr; // repeat in Q, tests for grouped-query attention
     const int64_t kv; // kv size
@@ -3215,7 +3247,7 @@ struct test_flash_attn_ext : public test_case {
     std::array<int32_t, 4> permute;
 
     std::string vars() override {
-        return VARS_TO_STR11(hs, nh, nr, kv, nb, mask, max_bias, logit_softcap, prec, type_KV, permute);
+        return VARS_TO_STR12(hsk, hsv, nh, nr, kv, nb, mask, max_bias, logit_softcap, prec, type_KV, permute);
     }
 
     double max_nmse_err() override {
@@ -3225,17 +3257,18 @@ struct test_flash_attn_ext : public test_case {
     uint64_t op_flops(ggml_tensor * t) override {
         GGML_UNUSED(t);
         // Just counting matmul costs:
-        // Q*K^T is nb x hs x kv, P*V is nb x kv x hs, per head
-        return 2 * 2 * nh*nr * nb * hs * kv;
+        // Q*K^T is nb x hsk x kv, P*V is nb x kv x hsv, per head
+        return 2 * nh*nr * nb * (hsk + hsv) * kv;
     }
 
-    test_flash_attn_ext(int64_t hs = 128, int64_t nh = 32, int64_t nr = 1, int64_t kv = 96, int64_t nb = 8,
+    test_flash_attn_ext(int64_t hsk = 128, int64_t hsv = 128, int64_t nh = 32, int64_t nr = 1, int64_t kv = 96, int64_t nb = 8,
                         bool mask = true, float max_bias = 0.0f, float logit_softcap = 0.0f, ggml_prec prec = GGML_PREC_F32,
                         ggml_type type_KV = GGML_TYPE_F16, std::array<int32_t, 4> permute = {0, 1, 2, 3})
-        : hs(hs), nh(nh), nr(nr), kv(kv), nb(nb), mask(mask), max_bias(max_bias), logit_softcap(logit_softcap), prec(prec), type_KV(type_KV), permute(permute) {}
+        : hsk(hsk), hsv(hsv), nh(nh), nr(nr), kv(kv), nb(nb), mask(mask), max_bias(max_bias), logit_softcap(logit_softcap), prec(prec), type_KV(type_KV), permute(permute) {}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
-        const int64_t hs_padded = GGML_PAD(hs, ggml_blck_size(type_KV));
+        const int64_t hsk_padded = GGML_PAD(hsk, ggml_blck_size(type_KV));
+        const int64_t hsv_padded = GGML_PAD(hsv, ggml_blck_size(type_KV));
 
         auto const &create_permuted = [&](ggml_type type, int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne3) -> ggml_tensor * {
             int64_t ne[4] = {ne0, ne1, ne2, ne3};
@@ -3250,13 +3283,13 @@ struct test_flash_attn_ext : public test_case {
             return t;
         };
 
-        ggml_tensor * q = create_permuted(GGML_TYPE_F32, hs_padded, nb, nh*nr, 1);
+        ggml_tensor * q = create_permuted(GGML_TYPE_F32, hsk_padded, nb, nh*nr, 1);
         ggml_set_name(q, "q");
 
-        ggml_tensor * k = create_permuted(type_KV,       hs_padded, kv, nh,    1);
+        ggml_tensor * k = create_permuted(type_KV,       hsk_padded, kv, nh,    1);
         ggml_set_name(k, "k");
 
-        ggml_tensor * v = create_permuted(type_KV,       hs_padded, kv, nh,    1);
+        ggml_tensor * v = create_permuted(type_KV,       hsv_padded, kv, nh,    1);
         ggml_set_name(v, "v");
 
         ggml_tensor * m = nullptr;
@@ -3265,7 +3298,7 @@ struct test_flash_attn_ext : public test_case {
             ggml_set_name(m, "m");
         }
 
-        ggml_tensor * out = ggml_flash_attn_ext(ctx, q, k, v, m, 1.0f/sqrtf(hs), max_bias, logit_softcap);
+        ggml_tensor * out = ggml_flash_attn_ext(ctx, q, k, v, m, 1.0f/sqrtf(hsk), max_bias, logit_softcap);
         ggml_flash_attn_ext_set_prec(out, prec);
         ggml_set_name(out, "out");
 
@@ -3995,14 +4028,25 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
         test_cases.emplace_back(new test_set(GGML_TYPE_I32, GGML_TYPE_I32, {6, 5, 4, 3}, dim));
     }
 
-    for (ggml_type type_src : {GGML_TYPE_F16, GGML_TYPE_F32}) {
+    // same-type copy
+    for (ggml_type type : all_types) {
+        const auto nk = ggml_blck_size(type);
+
+        for (int k = 1; k < 4; ++k) {
+            test_cases.emplace_back(new test_cpy(type, type, {k*nk, 2, 3, 4}));
+            test_cases.emplace_back(new test_cpy(type, type, {k*nk, 2, 3, 4}, {0, 2, 1, 3}));
+            test_cases.emplace_back(new test_cpy(type, type, {k*nk, 2, 3, 4}, {0, 3, 1, 2}, {0, 2, 1, 3}));
+        }
+    }
+
+    for (ggml_type type_src : {GGML_TYPE_F16, GGML_TYPE_BF16, GGML_TYPE_F32}) {
         for (ggml_type type_dst : all_types) {
             test_cases.emplace_back(new test_cpy(type_src, type_dst, {256, 4, 4, 4}));
             test_cases.emplace_back(new test_cpy(type_src, type_dst, {256, 2, 3, 4}, {0, 2, 1, 3})); // cpy by rows
         }
     }
-    for (ggml_type type_dst : {GGML_TYPE_F32}) {
-        for (ggml_type type_src : all_types) {
+    for (ggml_type type_src : all_types) {
+        for (ggml_type type_dst : {GGML_TYPE_F32}) {
             test_cases.emplace_back(new test_cpy(type_src, type_dst, {256, 4, 4, 4}));
             test_cases.emplace_back(new test_cpy(type_src, type_dst, {256, 2, 3, 4}, {0, 2, 1, 3})); // cpy by rows
         }
@@ -4140,6 +4184,11 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
             test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {2, 3}, {1, 1}, {0, 2, 1, 3}));
             test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {2, 3}, {1, 1}, {0, 1, 3, 2}));
             test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {2, 3}, {1, 1}, {0, 3, 2, 1}));
+
+            // test cases with large ne00/ne10 to cover stream-k fixup
+            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  1, 1024, {3, 2}, {1, 1}));
+            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  8, 1024, {3, 2}, {1, 1}));
+            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 1024, {3, 2}, {1, 1}));
         }
     }
     for (ggml_type type_a : other_types) {
@@ -4175,6 +4224,19 @@ 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,  83, 2,   64, { 8,  1}, {4, 1}));
     test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32,  64, 45, 128, { 8,  1}, {4, 1}));
     test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 128, 45,  64, { 8,  1}, {4, 1}));
+    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 1056, 1, 193, {1,  1}, {4, 1}, {0, 2, 1, 3}));
+    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}));
+
+    for (auto bs : {1,2,4,8}) {
+        for (auto nr : {1,4}) {
+            for (uint32_t m = 0; m < 2; ++m) {
+                for (uint32_t k = 0; k < 2; ++k) {
+                    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 1056 + m, 1, 128 + k,  {bs,  1}, {nr, 1}, {0, 2, 1, 3}));
+                    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 128 + m,  1, 1056 + k, {bs,  1}, {nr, 1}, {0, 1, 2, 3}, true));
+                }
+            }
+        }
+    }
 
     // sycl backend will limit task global_range < MAX_INT
     // test case for f16-type-convert-to-fp32 kernel with large k under fp32 compute dtype (occurs in stable-diffusion)
@@ -4355,12 +4417,15 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
         test_cases.emplace_back(new test_argsort(GGML_TYPE_F32, {60, 10, 10, 10}, order)); // qwen
     }
 
+    for (ggml_scale_mode mode : {GGML_SCALE_MODE_NEAREST, GGML_SCALE_MODE_BILINEAR}) {
+        test_cases.emplace_back(new test_upscale(GGML_TYPE_F32, {512, 512, 3, 2}, 2, mode));
+        test_cases.emplace_back(new test_upscale(GGML_TYPE_F32, {512, 512, 3, 2}, 2, mode, true));
+        test_cases.emplace_back(new test_upscale_ext(GGML_TYPE_F32, {2, 5,  7, 11}, {5, 7, 11, 13}, mode));
+    }
+
     test_cases.emplace_back(new test_sum());
     test_cases.emplace_back(new test_sum_rows());
     test_cases.emplace_back(new test_mean());
-    test_cases.emplace_back(new test_upscale());
-    test_cases.emplace_back(new test_upscale(GGML_TYPE_F32, { 512, 512, 3, 1 }, 2, true));
-    test_cases.emplace_back(new test_upscale_ext());
     test_cases.emplace_back(new test_group_norm(GGML_TYPE_F32, {64, 64, 320, 1}));
     test_cases.emplace_back(new test_group_norm(GGML_TYPE_F32, {9, 9, 1280, 1}));
     test_cases.emplace_back(new test_acc());
@@ -4370,27 +4435,33 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     test_cases.emplace_back(new test_timestep_embedding());
     test_cases.emplace_back(new test_leaky_relu());
 
-    for (int hs : { 64, 80, 128, 256, }) {
-        for (bool mask : { true, false } ) {
-            for (float max_bias : { 0.0f, 8.0f }) {
-                if (!mask && max_bias > 0.0f) continue;
-                for (float logit_softcap : {0.0f, 10.0f}) {
-                    if (hs != 128 && logit_softcap != 0.0f) continue;
-                    for (int nh : { 4, }) {
-                        for (int nr : { 1, 4, 16 }) {
-                            if (nr == 16 && hs != 128) continue;
-                            for (int kv : { 512, 1024, }) {
-                                if (nr != 1 && kv != 512) continue;
-                                for (int nb : { 1, 3, 32, 35, }) {
-                                    for (ggml_prec prec : {GGML_PREC_F32, GGML_PREC_DEFAULT}) {
-                                        if (hs != 128 && prec == GGML_PREC_DEFAULT) continue;
-                                        for (ggml_type type_KV : {GGML_TYPE_F16, GGML_TYPE_BF16, GGML_TYPE_Q8_0, GGML_TYPE_Q4_0}) {
-                                            test_cases.emplace_back(new test_flash_attn_ext(
-                                                hs, nh, nr, kv, nb, mask, max_bias, logit_softcap, prec, type_KV));
-                                            // run fewer test cases permuted
-                                            if (mask == true && max_bias == 0.0f && logit_softcap == 0 && kv == 512) {
+    for (int hsk : { 64, 80, 128, 192, 256, 576 }) {
+        for (int hsv : { 64, 80, 128, 192, 256, 512 }) {
+            if (hsk != 192 && hsk != 576 && hsk != hsv) continue;
+            if (hsk == 192 && (hsv != 128 && hsv != 192)) continue;
+            if (hsk == 576 && hsv != 512) continue; // DeepSeek MLA
+
+            for (bool mask : { true, false } ) {
+                for (float max_bias : { 0.0f, 8.0f }) {
+                    if (!mask && max_bias > 0.0f) continue;
+                    for (float logit_softcap : {0.0f, 10.0f}) {
+                        if (hsk != 128 && logit_softcap != 0.0f) continue;
+                        for (int nh : { 4, }) {
+                            for (int nr : { 1, 4, 16 }) {
+                                if (nr == 16 && hsk != 128) continue;
+                                for (int kv : { 512, 1024, }) {
+                                    if (nr != 1 && kv != 512) continue;
+                                    for (int nb : { 1, 3, 32, 35, }) {
+                                        for (ggml_prec prec : {GGML_PREC_F32, GGML_PREC_DEFAULT}) {
+                                            if (hsk != 128 && prec == GGML_PREC_DEFAULT) continue;
+                                            for (ggml_type type_KV : {GGML_TYPE_F16, GGML_TYPE_BF16, GGML_TYPE_Q8_0, GGML_TYPE_Q4_0}) {
                                                 test_cases.emplace_back(new test_flash_attn_ext(
-                                                    hs, nh, nr, kv, nb, mask, max_bias, logit_softcap, prec, type_KV, {0, 2, 1, 3}));
+                                                    hsk, hsv, nh, nr, kv, nb, mask, max_bias, logit_softcap, prec, type_KV));
+                                                // run fewer test cases permuted
+                                                if (mask == true && max_bias == 0.0f && logit_softcap == 0 && kv == 512) {
+                                                    test_cases.emplace_back(new test_flash_attn_ext(
+                                                        hsk, hsv, nh, nr, kv, nb, mask, max_bias, logit_softcap, prec, type_KV, {0, 2, 1, 3}));
+                                                }
                                             }
                                         }
                                     }
@@ -4444,6 +4515,9 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_perf() {
     test_cases.emplace_back(new test_argmax(GGML_TYPE_F32, {1024, 10, 1, 1}));
     test_cases.emplace_back(new test_argmax(GGML_TYPE_F32, {32000, 512, 1, 1}));
 
+    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 16416, 1, 128, {8,  1}, {4, 1}, {0, 2, 1, 3}));
+    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 128, 1, 16416, {8,  1}, {4, 1}, {0, 1, 2, 3}, true));
+
     for (int bs : {1, 2, 3, 4, 5, 8, 512}) {
         for (ggml_type type_a : all_types) {
             for (ggml_type type_b : {GGML_TYPE_F32}) {
@@ -4464,6 +4538,14 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_perf() {
         }
     }
 
+    for (int kv : { 4096, 8192, 16384, }) {
+        for (int hs : { 64, 128, }) {
+            for (int nr : { 1, 4, }) {
+                test_cases.emplace_back(new test_flash_attn_ext(hs, hs, 8, nr, kv, 1, true, 0, 0, GGML_PREC_F32, GGML_TYPE_F16));
+            }
+        }
+    }
+
     return test_cases;
 }
 

package/src/llama.cpp/tests/test-chat-template.cpp

@@ -19,6 +19,8 @@ static std::string normalize_newlines(const std::string & s) {
 #endif
 }
 
+#define U8C(x) (const char*)(u8##x)
+
 static common_chat_msg simple_msg(const std::string & role, const std::string & content) {
     common_chat_msg msg;
     msg.role = role;
@@ -35,6 +37,8 @@ int main(void) {
         {"assistant", "   I am an assistant   "},
         {"user", "Another question"},
     };
+
+    // std::string wrong = /* .template_str= */ u8"[gMASK]<sop>{% for item in messages %}{% if item['tools'] is defined %}<|system|>\n你是一个名为 ChatGLM 的人工智能助手。你是基于智谱AI训练的语言模型 GLM-4 模型开发的，你的任务是针对用户的问题和要求提供适当的答复和支持。\n\n# 可用工具{% set tools = item['tools'] %}{% for tool in tools %}{% if tool['type'] == 'function' %}\n\n## {{ tool['function']['name'] }}\n\n{{ tool['function'] | tojson(indent=4) }}\n......{% endif %}{% endfor %}{% endif %}{% if item['content'] %}<|{{ item['role'] }}|>{{ item['metadata'] }}\n{{ item['content'] }}{% endif %}{% endfor %}{% if add_generation_prompt %}<|assistant|>{% endif %}";
     struct TestCase {
         std::string name;
         std::string template_str;
@@ -177,24 +181,25 @@ int main(void) {
         },
         {
             /* .name= */ "ChatGLM4",
-            /* .template_str= */ u8"[gMASK]<sop>{% for item in messages %}{% if item['tools'] is defined %}<|system|>\n你是一个名为 ChatGLM 的人工智能助手。你是基于智谱AI训练的语言模型 GLM-4 模型开发的，你的任务是针对用户的问题和要求提供适当的答复和支持。\n\n# 可用工具{% set tools = item['tools'] %}{% for tool in tools %}{% if tool['type'] == 'function' %}\n\n## {{ tool['function']['name'] }}\n\n{{ tool['function'] | tojson(indent=4) }}\n......{% endif %}{% endfor %}{% endif %}{% if item['content'] %}<|{{ item['role'] }}|>{{ item['metadata'] }}\n{{ item['content'] }}{% endif %}{% endfor %}{% if add_generation_prompt %}<|assistant|>{% endif %}",
+            /* .template_str= */ U8C("[gMASK]<sop>{% for item in messages %}{% if item['tools'] is defined %}<|system|>\n你是一个名为 ChatGLM 的人工智能助手。你是基于智谱AI训练的语言模型 GLM-4 模型开发的，你的任务是针对用户的问题和要求提供适当的答复和支持。\n\n# 可用工具{% set tools = item['tools'] %}{% for tool in tools %}{% if tool['type'] == 'function' %}\n\n## {{ tool['function']['name'] }}\n\n{{ tool['function'] | tojson(indent=4) }}\n......{% endif %}{% endfor %}{% endif %}{% if item['content'] %}<|{{ item['role'] }}|>{{ item['metadata'] }}\n{{ item['content'] }}{% endif %}{% endfor %}{% if add_generation_prompt %}<|assistant|>{% endif %}"),
             /* .expected_output= */ "[gMASK]<sop><|system|>\nYou are a helpful assistant<|user|>\nHello<|assistant|>\nHi there<|user|>\nWho are you<|assistant|>\n   I am an assistant   <|user|>\nAnother question<|assistant|>",
             /* .expected_output_jinja= */ "",
             /* .bos_token= */ "",
             /* .eos_token= */ "",
         },
-        {
-            /* .name= */ "GLMEdge",
-            /* .template_str= */ "{% for item in messages %}{% if item['role'] == 'system' %}<|system|>\n{{ item['content'] }}{% elif item['role'] == 'user' %}<|user|>\n{{ item['content'] }}{% elif item['role'] == 'assistant' %}<|assistant|>\n{{ item['content'] }}{% endif %}{% endfor %}<|assistant|>",
-            /* .expected_output= */ "<|system|>\nYou are a helpful assistant<|user|>\nHello<|assistant|>\nHi there<|user|>\nWho are you<|assistant|>\n   I am an assistant   <|user|>\nAnother question<|assistant|>",
-            /* .expected_output_jinja= */ "<|system|>\nYou are a helpful assistant<|user|>\nHello<|assistant|>\nHi there<|user|>\nWho are you<|assistant|>\n   I am an assistant   <|user|>\nAnother question<|assistant|>",
-            /* .bos_token= */ "",
-            /* .eos_token= */ "",
-        },
+        // TODO @ngxson : GLMEdge produces poor result without `[gMASK]<sop>`, so we're temporarily using GLM4 template for it. We should fix this in the future.
+        // {
+        //     /* .name= */ "GLMEdge",
+        //     /* .template_str= */ "{% for item in messages %}{% if item['role'] == 'system' %}<|system|>\n{{ item['content'] }}{% elif item['role'] == 'user' %}<|user|>\n{{ item['content'] }}{% elif item['role'] == 'assistant' %}<|assistant|>\n{{ item['content'] }}{% endif %}{% endfor %}<|assistant|>",
+        //     /* .expected_output= */ "<|system|>\nYou are a helpful assistant<|user|>\nHello<|assistant|>\nHi there<|user|>\nWho are you<|assistant|>\n   I am an assistant   <|user|>\nAnother question<|assistant|>",
+        //     /* .expected_output_jinja= */ "<|system|>\nYou are a helpful assistant<|user|>\nHello<|assistant|>\nHi there<|user|>\nWho are you<|assistant|>\n   I am an assistant   <|user|>\nAnother question<|assistant|>",
+        //     /* .bos_token= */ "",
+        //     /* .eos_token= */ "",
+        // },
         {
             /* .name= */ "MiniCPM-3B-OpenHermes-2.5-v2-GGUF",
-            /* .template_str= */ u8"{% for message in messages %}{% if message['role'] == 'user' %}{{'<用户>' + message['content'].strip() + '<AI>'}}{% else %}{{message['content'].strip()}}{% endif %}{% endfor %}",
-            /* .expected_output= */ u8"You are a helpful assistant<用户>Hello<AI>Hi there<用户>Who are you<AI>I am an assistant<用户>Another question<AI>",
+            /* .template_str= */ U8C("{% for message in messages %}{% if message['role'] == 'user' %}{{'<用户>' + message['content'].strip() + '<AI>'}}{% else %}{{message['content'].strip()}}{% endif %}{% endfor %}"),
+            /* .expected_output= */ U8C("You are a helpful assistant<用户>Hello<AI>Hi there<用户>Who are you<AI>I am an assistant<用户>Another question<AI>"),
             /* .expected_output_jinja= */ "",
             /* .bos_token= */ "",
             /* .eos_token= */ "",
@@ -202,7 +207,7 @@ int main(void) {
         {
             /* .name= */ "DeepSeek-V2",
             /* .template_str= */ "{% if not add_generation_prompt is defined %}{% set add_generation_prompt = false %}{% endif %}{{ bos_token }}{% for message in messages %}{% if message['role'] == 'user' %}{{ 'User: ' + message['content'] + '\n\n' }}{% elif message['role'] == 'assistant' %}{{ 'Assistant: ' + message['content'] + eos_token }}{% elif message['role'] == 'system' %}{{ message['content'] + '\n\n' }}{% endif %}{% endfor %}{% if add_generation_prompt %}{{ 'Assistant:' }}{% endif %}",
-            /* .expected_output= */ u8"You are a helpful assistant\n\nUser: Hello\n\nAssistant: Hi there<｜end▁of▁sentence｜>User: Who are you\n\nAssistant:    I am an assistant   <｜end▁of▁sentence｜>User: Another question\n\nAssistant:",
+            /* .expected_output= */ U8C("You are a helpful assistant\n\nUser: Hello\n\nAssistant: Hi there<｜end▁of▁sentence｜>User: Who are you\n\nAssistant:    I am an assistant   <｜end▁of▁sentence｜>User: Another question\n\nAssistant:"),
             /* .expected_output_jinja= */ "",
             /* .bos_token= */ "",
             /* .eos_token= */ "<｜end▁of▁sentence｜>",
@@ -256,7 +261,7 @@ int main(void) {
         },
         {
             /* .name= */ "Infinigence/Megrez-3B-Instruct",
-            /* .template_str= */ u8"{% for message in messages %}{% if loop.first and messages[0]['role'] != 'system' %}{{ '<|role_start|>system<|role_end|>你是Megrez-3B-Instruct，将针对用户的问题给出详细的、积极的回答。<|turn_end|>' }}{% endif %}{{ '<|role_start|>' + message['role'] + '<|role_end|>' + message['content'] + '<|turn_end|>' }}{% endfor %}{% if add_generation_prompt %}{{ '<|role_start|>assistant<|role_end|>' }}{% endif %}",
+            /* .template_str= */ U8C("{% for message in messages %}{% if loop.first and messages[0]['role'] != 'system' %}{{ '<|role_start|>system<|role_end|>你是Megrez-3B-Instruct，将针对用户的问题给出详细的、积极的回答。<|turn_end|>' }}{% endif %}{{ '<|role_start|>' + message['role'] + '<|role_end|>' + message['content'] + '<|turn_end|>' }}{% endfor %}{% if add_generation_prompt %}{{ '<|role_start|>assistant<|role_end|>' }}{% endif %}"),
             /* .expected_output= */ "<|role_start|>system<|role_end|>You are a helpful assistant<|turn_end|><|role_start|>user<|role_end|>Hello<|turn_end|><|role_start|>assistant<|role_end|>Hi there<|turn_end|><|role_start|>user<|role_end|>Who are you<|turn_end|><|role_start|>assistant<|role_end|>   I am an assistant   <|turn_end|><|role_start|>user<|role_end|>Another question<|turn_end|><|role_start|>assistant<|role_end|>",
             /* .expected_output_jinja= */ "",
             /* .bos_token= */ "",
@@ -270,6 +275,22 @@ int main(void) {
             /* .bos_token= */ "",
             /* .eos_token= */ "",
         },
+        {
+            /* .name= */ "yandex/YandexGPT-5-Lite-8B-instruct",
+            /* .template_str= */ "<s>{%- set names = {'assistant': ' Ассистент:', 'user': ' Пользователь:'} %}\n{%- set tools_prefix = 'Тебе доступны следующие функции:' %}\n{%- macro __render_tool(tool) %}\n    {%- set name = tool.function.name %}\n    {%- set description = tool.function.description|default('') %}\n    {%- set parameters = tool.function.parameters|tojson %}\n    {{- '\\n' }}function {{ '{' }}'name':'{{ name }}',\n    {%- if tool.function.description %}'description':'{{ description }}',{% endif %}\n'parameters':{{ parameters }}\n    {{- '}' }}\n{%- endmacro %}\n{%- macro __render_tools(tools) %}\n    {{- tools_prefix }}\n    {%- for tool in tools %}\n        {{- __render_tool(tool) }}\n    {%- endfor %}\n    {{- '\\n\\n' }}\n{%- endmacro %}\n{%- macro __render_tool_message(message) %}\n    {{- '\\n\\nРезультат вызова' }} {{ message.name }}: {{ message.content }} {{ '\\n\\n' }}\n{%- endmacro %}\n{%- if tools -%}\n    {{- __render_tools(tools) }}\n{%- endif -%}\n{%- macro __render_user_message(message) %}\n{{ names.user }} {{ message.content + '\\n\\n' }}\n{%- endmacro %}\n{%- macro __render_assistant_message(message) %}\n    {{- names.assistant }}\n    {%- set call = message['function_call'] %}\n    {%- if call %}\n        {{- '\\n[TOOL_CALL_START]' }}{{ call.name }}{{ '\\n' }}{{ call.arguments|tojson }}\n    {%- else %}\n        {{- ' ' + message.content + '\\n\\n' }}\n    {%- endif %}\n{%- endmacro %}\n{%- if not add_generation_prompt is defined %}\n{%- set add_generation_prompt = false %}\n{%- endif %}\n{%- for message in messages %}\n    {%- if message['role'] == 'user' %}\n        {{- __render_user_message(message) }}\n    {%- endif %}\n    {%- if message.role == 'assistant' and not loop.last %}\n        {{- __render_assistant_message(message) }}\n    {%- endif %}\n    {%- if message.role == 'tool' %}\n        {{- __render_tool_message(message) }}\n    {%- endif %}\n    {%- if loop.last %}\n        {{- ' Ассистент:[SEP]' }}\n    {%- endif %}\n{%- endfor %}\n",
+            /* .expected_output= */ "<s> Пользователь: Hello\n\n Ассистент: Hi there\n\n Пользователь: Who are you\n\n Ассистент:    I am an assistant   \n\n Пользователь: Another question\n\n Ассистент:[SEP]",
+            /* .expected_output_jinja= */ "<s> Пользователь: You are a helpful assistant\nHello\n\n Ассистент: Hi there\n\n Пользователь: Who are you\n\n Ассистент:    I am an assistant   \n\n Пользователь: Another question\n\n Ассистент:[SEP]",
+            /* .bos_token= */ "",
+            /* .eos_token= */ "",
+        },
+        {
+            /* .name= */ "inclusionAI/Ling-lite",
+            /* .template_str */ "{% for message in messages %}{% set role = message['role'] | lower %}{% if role == 'user' %}{% set role = 'HUMAN' %}{% endif %}{% set role = role | upper %}{{ '<role>' + role + '</role>' + message['content'] }}{% endfor %}{% if add_generation_prompt %}{{ '<role>ASSISTANT</role>' }}{% endif %}",
+            /* .expected_output= */ "<role>SYSTEM</role>You are a helpful assistant<role>HUMAN</role>Hello<role>ASSISTANT</role>Hi there<role>HUMAN</role>Who are you<role>ASSISTANT</role>   I am an assistant   <role>HUMAN</role>Another question<role>ASSISTANT</role>",
+            /* .expected_output_jinja= */ "",
+            /* .bos_token= */ "",
+            /* .eos_token= */ "",
+        },
     };
     std::vector<char> formatted_chat(1024);
     int32_t res;

package/src/llama.cpp/tests/test-chat.cpp

@@ -11,8 +11,9 @@
 #include <string>
 
 #include "chat.h"
-#include "llama-grammar.h"
-#include "unicode.h"
+
+#include "../src/unicode.h"
+#include "../src/llama-grammar.h"
 
 using json = nlohmann::ordered_json;
 
@@ -569,6 +570,7 @@ static void test_template_output_parsers() {
     {
         // Not supported yet
         auto tmpls = read_templates("models/templates/CohereForAI-c4ai-command-r-plus-tool_use.jinja");
+        assert_equals(COMMON_CHAT_FORMAT_CONTENT_ONLY, common_chat_templates_apply(tmpls.get(), inputs_no_tools).format);
         assert_equals(COMMON_CHAT_FORMAT_GENERIC, common_chat_templates_apply(tmpls.get(), inputs_tools).format);
     }
     {
@@ -665,6 +667,7 @@ static void test_template_output_parsers() {
         auto tmpls = read_templates("models/templates/NousResearch-Hermes-2-Pro-Llama-3-8B-tool_use.jinja");
         std::vector<std::string> end_tokens{ "<|im_end|>" };
 
+        assert_equals(COMMON_CHAT_FORMAT_CONTENT_ONLY, common_chat_templates_apply(tmpls.get(), inputs_no_tools).format);
         assert_equals(COMMON_CHAT_FORMAT_HERMES_2_PRO, common_chat_templates_apply(tmpls.get(), inputs_tools).format);
         assert_equals(
             COMMON_CHAT_FORMAT_HERMES_2_PRO,
@@ -793,6 +796,7 @@ static void test_template_output_parsers() {
         auto tmpls = read_templates("models/templates/meta-llama-Llama-3.1-8B-Instruct.jinja");
         std::vector<std::string>   end_tokens{ "<|eom_id|>", "<|eot_id|>" };
 
+        assert_equals(COMMON_CHAT_FORMAT_CONTENT_ONLY, common_chat_templates_apply(tmpls.get(), inputs_no_tools).format);
         assert_equals(COMMON_CHAT_FORMAT_LLAMA_3_X, common_chat_templates_apply(tmpls.get(), inputs_tools).format);
         assert_equals(COMMON_CHAT_FORMAT_LLAMA_3_X_WITH_BUILTIN_TOOLS,
                       common_chat_templates_apply(tmpls.get(), inputs_tools_builtin).format);
@@ -815,6 +819,7 @@ static void test_template_output_parsers() {
         std::vector<std::string>   end_tokens{ "<|eom_id|>", "<|eot_id|>" };
 
         assert_equals(COMMON_CHAT_FORMAT_LLAMA_3_X, common_chat_templates_apply(tmpls.get(), inputs_tools).format);
+        assert_equals(COMMON_CHAT_FORMAT_CONTENT_ONLY, common_chat_templates_apply(tmpls.get(), inputs_no_tools).format);
 
         test_templates(tmpls.get(), end_tokens, message_assist, tools, "Hello, world!\nWhat's up?", /* expect_grammar_triggered= */ false);
         test_templates(tmpls.get(), end_tokens, message_assist_call, tools,
@@ -824,6 +829,8 @@ static void test_template_output_parsers() {
         auto tmpls = read_templates("models/templates/meetkai-functionary-medium-v3.1.jinja");
         std::vector<std::string>   end_tokens{ "<|eom_id|>", "<|eot_id|>" };
 
+        assert_equals(COMMON_CHAT_FORMAT_CONTENT_ONLY,
+                      common_chat_templates_apply(tmpls.get(), inputs_no_tools).format);
         assert_equals(COMMON_CHAT_FORMAT_FUNCTIONARY_V3_1_LLAMA_3_1,
                       common_chat_templates_apply(tmpls.get(), inputs_tools).format);
 
@@ -851,6 +858,7 @@ static void test_template_output_parsers() {
         auto tmpls = read_templates("models/templates/fireworks-ai-llama-3-firefunction-v2.jinja");
         std::vector<std::string>   end_tokens{ "<|eot_id|>" };
 
+        assert_equals(COMMON_CHAT_FORMAT_CONTENT_ONLY, common_chat_templates_apply(tmpls.get(), inputs_no_tools).format);
         assert_equals(COMMON_CHAT_FORMAT_FIREFUNCTION_V2, common_chat_templates_apply(tmpls.get(), inputs_tools).format);
 
         test_templates(tmpls.get(), end_tokens, message_assist, tools, "Hello, world!\nWhat's up?", /* expect_grammar_triggered= */ false);
@@ -862,6 +870,7 @@ static void test_template_output_parsers() {
         auto tmpls = read_templates("models/templates/deepseek-ai-DeepSeek-R1-Distill-Llama-8B.jinja");
         std::vector<std::string>   end_tokens{ "<｜end▁of▁sentence｜>" };
 
+        assert_equals(COMMON_CHAT_FORMAT_DEEPSEEK_R1,                   common_chat_templates_apply(tmpls.get(), inputs_no_tools).format);
         assert_equals(COMMON_CHAT_FORMAT_DEEPSEEK_R1,                   common_chat_templates_apply(tmpls.get(), inputs_tools).format);
         assert_equals(COMMON_CHAT_FORMAT_DEEPSEEK_R1_EXTRACT_REASONING, common_chat_templates_apply(tmpls.get(), inputs_tools_think).format);
 
@@ -891,6 +900,7 @@ static void test_template_output_parsers() {
         auto tmpls = read_templates("models/templates/llama-cpp-deepseek-r1.jinja");
         std::vector<std::string>   end_tokens{ "<｜end▁of▁sentence｜>" };
 
+        assert_equals(COMMON_CHAT_FORMAT_DEEPSEEK_R1,                   common_chat_templates_apply(tmpls.get(), inputs_no_tools).format);
         assert_equals(COMMON_CHAT_FORMAT_DEEPSEEK_R1,                   common_chat_templates_apply(tmpls.get(), inputs_tools).format);
         assert_equals(COMMON_CHAT_FORMAT_DEEPSEEK_R1_EXTRACT_REASONING, common_chat_templates_apply(tmpls.get(), inputs_tools_think).format);
 

package/src/llama.cpp/{examples/gbnf-validator/gbnf-validator.cpp → tests/test-gbnf-validator.cpp}

@@ -1,5 +1,5 @@
-#include "unicode.h"
-#include "llama-grammar.h"
+#include "../src/unicode.h"
+#include "../src/llama-grammar.h"
 
 #include <cstdio>
 #include <cstdlib>

package/src/llama.cpp/tests/test-grammar-integration.cpp

@@ -2,10 +2,11 @@
 #undef NDEBUG
 #endif
 
-#include "unicode.h"
-#include "llama-grammar.h"
 #include "json-schema-to-grammar.h"
 
+#include "../src/unicode.h"
+#include "../src/llama-grammar.h"
+
 #include <cassert>
 #include <string>
 #include <vector>