Merge branch 'ggml-org:master' into glm45

Author: ddh0

common/chat.cpp

@@ -1646,7 +1646,7 @@ static void common_chat_parse_hermes_2_pro(common_chat_msg_parser & builder) {
         "|<function name=\"([^\"]+)\">"  // match 5 (function name again)
     );
 
-    if (auto res = builder.try_find_regex(open_regex)) {
+    while (auto res = builder.try_find_regex(open_regex)) {
         const auto & block_start = res->groups[1];
         std::string block_end = block_start.empty() ? "" : "```";
 
@@ -1668,7 +1668,6 @@ static void common_chat_parse_hermes_2_pro(common_chat_msg_parser & builder) {
                     builder.consume_literal(block_end);
                     builder.consume_spaces();
                 }
-                builder.add_content(builder.consume_rest());
             } else {
                 throw common_chat_msg_partial_exception("failed to parse tool call");
             }
@@ -1693,11 +1692,10 @@ static void common_chat_parse_hermes_2_pro(common_chat_msg_parser & builder) {
                     builder.consume_spaces();
                 }
             }
-            builder.add_content(builder.consume_rest());
         }
-    } else {
-        builder.add_content(builder.consume_rest());
     }
+
+    builder.add_content(builder.consume_rest());
 }
 
 static common_chat_params common_chat_params_init_without_tools(const common_chat_template & tmpl, const struct templates_params & inputs) {

ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -2106,12 +2106,12 @@ static void ggml_vk_load_shaders(vk_device& device) {
         s_mmq_wg_denoms = { 32,  64,  1 };
 
         // spec constants and tile sizes for quant matmul (Qi_K)
-        l_warptile_mmq_k = { 256, 64, 128, 64,  1 };
-        m_warptile_mmq_k = { 256, 32,  64, 64,  0 };
-        s_warptile_mmq_k = { 256, 32,  32, 128, 0 };
-        l_mmq_wg_denoms_k = { 64, 128, 1 };
-        m_mmq_wg_denoms_k = { 32,  64, 1 };
-        s_mmq_wg_denoms_k = { 32,  32, 1 };
+        l_warptile_mmq_k = { 256, 128, 256, 64, 1 };
+        m_warptile_mmq_k = { 256, 128, 128, 64, 1 };
+        s_warptile_mmq_k = { 256, 32,  64, 128, 0 };
+        l_mmq_wg_denoms_k = { 128, 256, 1 };
+        m_mmq_wg_denoms_k = { 128, 128, 1 };
+        s_mmq_wg_denoms_k = { 32,  64,  1 };
 
         // spec constants and tile sizes for quant matmul_id
         l_warptile_mmqid = { 256, 128, 128, 16, 0 };
@@ -5022,26 +5022,37 @@ static void ggml_vk_buffer_memset(vk_buffer& dst, size_t offset, uint32_t c, siz
     ggml_vk_queue_command_pools_cleanup(dst->device);
 }
 
-static uint32_t ggml_vk_guess_split_k(ggml_backend_vk_context * ctx, int m, int n, int k, const vk_pipeline& pipeline) {
+static uint32_t ggml_vk_guess_split_k(ggml_backend_vk_context * ctx, uint32_t m, uint32_t n, uint32_t k, const vk_pipeline& pipeline) {
     VK_LOG_DEBUG("ggml_vk_guess_split_k(" << m << ", " << n << ", " << k << ")");
 
     uint32_t split_k = 1;
-    if (ctx->device->shader_core_count != 0 && m >= (int)pipeline->wg_denoms[0] && n >= (int)pipeline->wg_denoms[1]) {
+    if (ctx->device->shader_core_count != 0 && m >= pipeline->wg_denoms[0] && n >= pipeline->wg_denoms[1]) {
         // If k is 'large' and the SMs will fill less than halfway, use split_k.
         uint32_t m_tiles = CEIL_DIV(m, pipeline->wg_denoms[0]);
         uint32_t n_tiles = CEIL_DIV(n, pipeline->wg_denoms[1]);
-        if (k >= 2048 && m_tiles * n_tiles < ctx->device->shader_core_count / 2) {
-            split_k = ctx->device->shader_core_count / (m_tiles * n_tiles);
-            // Clamp to 2 or 4
-            split_k = std::min(split_k, 4u);
-            if (split_k == 3) {
-                split_k = 2;
+
+        if (k >= 2048) {
+            if (m_tiles * n_tiles <= ctx->device->shader_core_count / 2) {
+                split_k = ctx->device->shader_core_count / (m_tiles * n_tiles);
+            } else if (m_tiles * n_tiles <= ctx->device->shader_core_count * 2 / 3) {
+                split_k = 3;
             }
-            if (ctx->device->coopmat2) {
-                // coopmat2 shader expects splits to be aligned to 256
-                while (split_k > 1 && ((k / split_k) % 256) != 0) {
-                    split_k /= 2;
+            // Cap the split at 8x. Unless k is huge this is a lot of overhead.
+            split_k = std::min(split_k, 8u);
+
+            // ggml_vk_matmul will align the splits to be a multiple of 256.
+            // If this rounded up size would cause the last split to be empty,
+            // then reduce the split count.
+            while (true) {
+                if (split_k == 1) {
+                    break;
+                }
+                uint32_t k_split = CEIL_DIV(k, split_k);
+                k_split = ROUNDUP_POW2(k_split, 256);
+                if (k_split * (split_k - 1) < k) {
+                    break;
                 }
+                split_k--;
             }
         }
     }
@@ -5053,9 +5064,22 @@ static vk_pipeline ggml_vk_guess_matmul_pipeline(ggml_backend_vk_context * ctx,
     VK_LOG_DEBUG("ggml_vk_guess_matmul_pipeline(" << m << ", " << n << ", " << aligned << ", " << ggml_type_name(src0_type) << ", " << ggml_type_name(src1_type) << ")");
 
     if (ctx->device->coopmat2) {
+        const uint32_t shader_core_count = ctx->device->shader_core_count;
+        const uint32_t tiles_l = CEIL_DIV(m, mmp->a_l->wg_denoms[0]) * CEIL_DIV(n, mmp->a_l->wg_denoms[1]);
+        const uint32_t tiles_m = CEIL_DIV(m, mmp->a_m->wg_denoms[0]) * CEIL_DIV(n, mmp->a_m->wg_denoms[1]);
+
         // Use large shader when the N dimension is greater than the medium shader's tile size
         uint32_t crossover_large = mmp->m->wg_denoms[1];
-        if ((ctx->device->mul_mat_l[src0_type] && (n > crossover_large)) || (!ctx->device->mul_mat_m[src0_type] && !ctx->device->mul_mat_s[src0_type])) {
+
+        // Prefer large over medium if either:
+        // - medium or large tiles would overfill the GPU
+        // - large tiles with a split_k==3 fits in the GPU and medium tiles with split_k==2 does not
+        //   (medium with split_k==2 is probably better if it fits - more workgroups running and less split_k overhead)
+        bool prefer_large = tiles_m > shader_core_count || tiles_l > shader_core_count ||
+                            // split_k==3 with large tiles likely better than medium tiles with no split_k.
+                            (tiles_l <= shader_core_count / 3 && tiles_m > shader_core_count / 2);
+
+        if ((ctx->device->mul_mat_l[src0_type] && (n > crossover_large && prefer_large)) || (!ctx->device->mul_mat_m[src0_type] && !ctx->device->mul_mat_s[src0_type])) {
             return aligned ? mmp->a_l : mmp->l;
         }
         // Use medium shader when the N dimension is greater than the small shader's tile size
@@ -5099,7 +5123,11 @@ static void ggml_vk_matmul(
 
     GGML_ASSERT(batch_stride_d == m * n);
 
-    const vk_mat_mat_push_constants pc1 = { m, n, k, stride_a, stride_b, stride_d, batch_stride_a, batch_stride_b, batch_stride_d, CEIL_DIV(k, split_k), ne02, ne12, broadcast2, broadcast3, padded_n };
+    // Round the split size up to a multiple of 256 (k-quant alignment)
+    uint32_t k_split = CEIL_DIV(k, split_k);
+    k_split = ROUNDUP_POW2(k_split, 256);
+
+    const vk_mat_mat_push_constants pc1 = { m, n, k, stride_a, stride_b, stride_d, batch_stride_a, batch_stride_b, batch_stride_d, k_split, ne02, ne12, broadcast2, broadcast3, padded_n };
     // Make sure enough workgroups get assigned for split k to work
     ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { a, b, split_k_buffer }, pc1, { (CEIL_DIV(m, pipeline->wg_denoms[0]) * pipeline->wg_denoms[0]) * split_k, n, batch });
     ggml_vk_sync_buffers(subctx);

scripts/compare-llama-bench.py

@@ -326,7 +326,7 @@ def __init__(self, tool: str = "llama-bench"):
 
         # Set table name and schema based on tool
         if self.tool == "llama-bench":
-            self.table_name = "test"
+            self.table_name = "llama_bench"
             db_fields = LLAMA_BENCH_DB_FIELDS
             db_types = LLAMA_BENCH_DB_TYPES
         elif self.tool == "test-backend-ops":
@@ -409,17 +409,17 @@ def __init__(self, data_file: str, tool: Any):
 
         # Tool selection logic
         if tool is None:
-            if "test" in table_names:
-                self.table_name = "test"
+            if "llama_bench" in table_names:
+                self.table_name = "llama_bench"
                 self.tool = "llama-bench"
             elif "test_backend_ops" in table_names:
                 self.table_name = "test_backend_ops"
                 self.tool = "test-backend-ops"
             else:
                 raise RuntimeError(f"No suitable table found in database. Available tables: {table_names}")
         elif tool == "llama-bench":
-            if "test" in table_names:
-                self.table_name = "test"
+            if "llama_bench" in table_names:
+                self.table_name = "llama_bench"
                 self.tool = "llama-bench"
             else:
                 raise RuntimeError(f"Table 'test' not found for tool 'llama-bench'. Available tables: {table_names}")

tests/test-chat.cpp

@@ -953,6 +953,33 @@ static void test_template_output_parsers() {
                 /* is_partial= */ false,
                 {COMMON_CHAT_FORMAT_HERMES_2_PRO}));
 
+        // Test multiple tool calls
+        common_chat_msg message_assist_multiple_calls;
+        message_assist_multiple_calls.role = "assistant";
+        message_assist_multiple_calls.content = "";
+        message_assist_multiple_calls.tool_calls.push_back({"special_function", "{\"arg1\": 1}", ""});
+        message_assist_multiple_calls.tool_calls.push_back({"python", "{\"code\":\"print('hello')\"}", ""});
+
+        assert_msg_equals(
+            message_assist_multiple_calls,
+            common_chat_parse(
+                "<tool_call>\n"
+                "{\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}\n"
+                "</tool_call>\n"
+                "<tool_call>\n"
+                "{\"name\": \"python\", \"arguments\": {\"code\":\"print('hello')\"}}\n"
+                "</tool_call>",
+                /* is_partial= */ false,
+                {COMMON_CHAT_FORMAT_HERMES_2_PRO}));
+
+        assert_msg_equals(
+            message_assist_multiple_calls,
+            common_chat_parse(
+                "<function=special_function>{\"arg1\": 1}</function>\n"
+                "<function=python>{\"code\":\"print('hello')\"}</function>",
+                /* is_partial= */ false,
+                {COMMON_CHAT_FORMAT_HERMES_2_PRO}));
+
         assert_msg_equals(
             simple_assist_msg(
                 "This is not a tool call:",
@@ -1039,6 +1066,22 @@ static void test_template_output_parsers() {
                       "<tool_call>\n"
                       "{\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}\n"
                       "</tool_call>");
+
+        // Test multiple tool calls with template
+        common_chat_msg message_assist_multiple_calls_template;
+        message_assist_multiple_calls_template.role = "assistant";
+        message_assist_multiple_calls_template.content = "";
+        message_assist_multiple_calls_template.tool_calls.push_back({"special_function", "{\"arg1\": 1}", ""});
+        message_assist_multiple_calls_template.tool_calls.push_back({"python", "{\"code\":\"print('test')\"}", ""});
+
+        test_templates(tmpls.get(), end_tokens, message_assist_multiple_calls_template, tools,
+                      "<tool_call>\n"
+                      "{\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}\n"
+                      "</tool_call>\n"
+                      "<tool_call>\n"
+                      "{\"name\": \"python\", \"arguments\": {\"code\":\"print('test')\"}}\n"
+                      "</tool_call>");
+
         test_templates(tmpls.get(), end_tokens, message_assist_call_python_lines, tools,
                       "<tool_call>\n"
                       "{\"name\": \"python\", \"arguments\": {\"code\":\"# This is a program:\\nprint('hey')\"}}\n"

tools/llama-bench/llama-bench.cpp

@@ -1738,7 +1738,7 @@ struct sql_printer : public printer {
 
     void print_header(const cmd_params & params) override {
         std::vector<std::string> fields = test::get_fields();
-        fprintf(fout, "CREATE TABLE IF NOT EXISTS test (\n");
+        fprintf(fout, "CREATE TABLE IF NOT EXISTS llama_bench (\n");
         for (size_t i = 0; i < fields.size(); i++) {
             fprintf(fout, "  %s %s%s\n", fields.at(i).c_str(), get_sql_field_type(fields.at(i)).c_str(),
                     i < fields.size() - 1 ? "," : "");
@@ -1749,7 +1749,7 @@ struct sql_printer : public printer {
     }
 
     void print_test(const test & t) override {
-        fprintf(fout, "INSERT INTO test (%s) ", join(test::get_fields(), ", ").c_str());
+        fprintf(fout, "INSERT INTO llama_bench (%s) ", join(test::get_fields(), ", ").c_str());
         fprintf(fout, "VALUES (");
         std::vector<std::string> values = t.get_values();
         for (size_t i = 0; i < values.size(); i++) {