Add Jinja support for "indent" string filter (#19529)

* Add partial Jinja support for "indent" string filter

* Fully implement indent

* Add tests for all width variants.

* Update tests/test-jinja.cpp

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

* Fix getline ignoring trailing newlines

* Update common/jinja/value.cpp

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

* fix first indent condition

---------

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

common/jinja/value.cpp

@@ -4,6 +4,7 @@
 // for converting from JSON to jinja values
 #include <nlohmann/json.hpp>
 
+#include <sstream>
 #include <string>
 #include <cctype>
 #include <vector>
@@ -715,8 +716,46 @@ const func_builtins & value_string_t::get_builtins() const {
             return args.get_pos(0);
         }},
         {"tojson", tojson},
-        {"indent", [](const func_args &) -> value {
-            throw not_implemented_exception("String indent builtin not implemented");
+        {"indent", [](const func_args &args) -> value {
+            args.ensure_count(1, 4);
+            value val_input  = args.get_pos(0);
+            value val_width  = args.get_kwarg_or_pos("width", 1);
+            const bool first = args.get_kwarg_or_pos("first", 2)->as_bool(); // undefined == false
+            const bool blank = args.get_kwarg_or_pos("blank", 3)->as_bool(); // undefined == false
+            if (!is_val<value_string>(val_input)) {
+                throw raised_exception("indent() first argument must be a string");
+            }
+            std::string indent;
+            if (is_val<value_int>(val_width)) {
+                indent.assign(val_width->as_int(), ' ');
+            } else if (is_val<value_string>(val_width)) {
+                indent = val_width->as_string().str();
+            } else {
+                indent = "    ";
+            }
+            std::string indented;
+            std::string input = val_input->as_string().str();
+            std::istringstream iss = std::istringstream(input);
+            std::string line;
+            while (std::getline(iss, line)) {
+                if (!indented.empty()) {
+                    indented.push_back('\n');
+                }
+                if ((indented.empty() ? first : (!line.empty() || blank))) {
+                    indented += indent;
+                }
+                indented += line;
+            }
+            if (!input.empty() && input.back() == '\n') {
+                indented.push_back('\n');
+                if (blank) {
+                    indented += indent;
+                }
+            }
+
+            auto res = mk_val<value_string>(indented);
+            res->val_str.mark_input_based_on(val_input->as_string());
+            return res;
         }},
         {"join", [](const func_args &) -> value {
             throw not_implemented_exception("String join builtin not implemented");

ggml/src/ggml-webgpu/ggml-webgpu.cpp

@@ -1121,7 +1121,8 @@ static webgpu_command ggml_webgpu_mul_mat(webgpu_context & ctx,
         uint32_t batches       = dst->ne[2] * dst->ne[3];
         uint32_t output_groups = CEIL_DIV(dst->ne[0], decisions->outputs_per_wg);
         uint32_t total_wg      = output_groups * batches;
-        wg_x                   = total_wg % ctx->global_ctx->capabilities.limits.maxComputeWorkgroupsPerDimension;
+        // TODO: split large sizes into multiple batches to avoid way over-provisioning workgroups
+        wg_x = std::min(total_wg, ctx->global_ctx->capabilities.limits.maxComputeWorkgroupsPerDimension);
         wg_y = CEIL_DIV(total_wg, ctx->global_ctx->capabilities.limits.maxComputeWorkgroupsPerDimension);
     } else if (use_fast) {
         auto decisions = static_cast<ggml_webgpu_mul_mat_shader_decisions *>(pipeline.context.get());

tests/test-jinja.cpp

@@ -691,6 +691,48 @@ static void test_filters(testing & t) {
         "{\n  \"a\": 1,\n  \"b\": [\n    1,\n    2\n  ]\n}"
     );
 
+    test_template(t, "indent",
+        "{{ data|indent(2) }}",
+        {{ "data", "foo\nbar" }},
+        "foo\n  bar"
+    );
+
+    test_template(t, "indent first only",
+        "{{ data|indent(width=3,first=true) }}",
+        {{ "data", "foo\nbar" }},
+        "   foo\n   bar"
+    );
+
+    test_template(t, "indent blank lines and first line",
+        "{{ data|indent(width=5,blank=true,first=true) }}",
+        {{ "data", "foo\n\nbar" }},
+        "     foo\n     \n     bar"
+    );
+
+    test_template(t, "indent with default width",
+        "{{ data|indent() }}",
+        {{ "data", "foo\nbar" }},
+        "foo\n    bar"
+    );
+
+    test_template(t, "indent with no newline",
+        "{{ data|indent }}",
+        {{ "data", "foo" }},
+        "foo"
+    );
+
+    test_template(t, "indent with trailing newline",
+        "{{ data|indent(blank=true) }}",
+        {{ "data", "foo\n" }},
+        "foo\n    "
+    );
+
+    test_template(t, "indent with string",
+        "{{ data|indent(width='>>>>') }}",
+        {{ "data", "foo\nbar" }},
+        "foo\n>>>>bar"
+    );
+
     test_template(t, "chained filters",
         "{{ '  HELLO  '|trim|lower }}",
         json::object(),

tools/server/server-context.cpp

@@ -77,6 +77,7 @@ struct server_slot {
     size_t last_nl_pos = 0;
 
     std::string  generated_text;
+    std::string  debug_generated_text;
     llama_tokens generated_tokens;
 
     // idx of draft tokens in the main batch
@@ -425,7 +426,7 @@ struct server_slot {
 
             if (!only_metrics) {
                 res["prompt"] = ptask->tokens.detokenize(ctx, true);
-                res["generated"] = generated_text;
+                res["generated"] = generated_text.empty() ? debug_generated_text : generated_text;
             }
         }
 
@@ -1442,6 +1443,12 @@ private:
         res->id_slot = slot.id;
 
         res->index           = slot.task->index;
+
+        // keep copy of last generated text for debugging purposes
+        if (slots_debug) {
+            slot.debug_generated_text = slot.generated_text;
+        }
+
         // in stream mode, content and tokens are already in last partial chunk
         if (slot.task->params.stream) {
             res->content     = "";