common/parser: add proper reasoning tag prefill reading (#20424)

* Implement proper prefill extraction

* Refactor cli parameters, update docs, move reasoning budget sampler part to common/reasoning-budget.cpp

* Update tools/server/server-task.cpp

* refactor: move grammars to variant, remove grammar_external, handle exception internally

* Make code less C++y

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

common/common.h

@@ -3,12 +3,14 @@
 #pragma once
 
 #include "ggml-opt.h"
+#include "ggml.h"
 #include "llama-cpp.h"
 
 #include <set>
 #include <sstream>
 #include <string>
 #include <string_view>
+#include <variant>
 #include <vector>
 #include <map>
 
@@ -178,6 +180,43 @@ enum common_speculative_type {
     COMMON_SPECULATIVE_TYPE_COUNT          // number of types, unknown type
 };
 
+// Grammar type enumeration
+enum common_grammar_type {
+    COMMON_GRAMMAR_TYPE_NONE,           // no grammar set
+    COMMON_GRAMMAR_TYPE_USER,           // user-provided GBNF (--grammar / "grammar" API field)
+    COMMON_GRAMMAR_TYPE_OUTPUT_FORMAT,  // auto-generated from JSON schema (--json-schema / "json_schema" API field)
+    COMMON_GRAMMAR_TYPE_TOOL_CALLS,     // auto-generated by chat template parser for function calling
+};
+
+// Grammar variant struct with type and grammar string
+struct common_grammar {
+    common_grammar_type type = COMMON_GRAMMAR_TYPE_NONE;
+    std::string grammar;
+
+    // Default constructor - no grammar
+    common_grammar() = default;
+
+    // Constructor with type and grammar string
+    common_grammar(common_grammar_type t, std::string g) : type(t), grammar(std::move(g)) {
+        GGML_ASSERT(type != COMMON_GRAMMAR_TYPE_NONE || !grammar.empty());
+    }
+
+    // Check if a grammar is set
+    bool empty() const { return type == COMMON_GRAMMAR_TYPE_NONE || grammar.empty(); }
+};
+
+// Returns the raw grammar string, or empty string if no grammar is set.
+inline const std::string & common_grammar_value(const common_grammar & g) {
+    return g.grammar;
+}
+
+// Returns true when the generation_prompt should be prefilled into the grammar sampler.
+// Only output-format and tool-call grammars need prefill; user-supplied grammars must not be prefilled.
+inline bool common_grammar_needs_prefill(const common_grammar & g) {
+    return g.type == COMMON_GRAMMAR_TYPE_OUTPUT_FORMAT
+        || g.type == COMMON_GRAMMAR_TYPE_TOOL_CALLS;
+}
+
 // sampling parameters
 struct common_params_sampling {
     uint32_t seed = LLAMA_DEFAULT_SEED; // the seed used to initialize llama_sampler
@@ -228,7 +267,7 @@ struct common_params_sampling {
         COMMON_SAMPLER_TYPE_TEMPERATURE,
     };
 
-    std::string                         grammar; // optional BNF-like grammar to constrain sampling
+    common_grammar              grammar;      // optional grammar constraint (user / output-format / tool-calls)
     bool                                grammar_lazy = false;
     std::vector<common_grammar_trigger> grammar_triggers; // optional triggers (for lazy grammars)
     std::set<llama_token>               preserved_tokens;
@@ -236,10 +275,15 @@ struct common_params_sampling {
     std::vector<llama_logit_bias> logit_bias;     // logit biases to apply
     std::vector<llama_logit_bias> logit_bias_eog; // pre-calculated logit biases for EOG tokens
 
+    // The assistant generation prompt already prefilled into the prompt.
+    // Fed to the grammar sampler (to advance past pre-existing tokens) and used
+    // to determine the reasoning budget sampler's initial state.
+    // Only applied when the grammar is of output-format or tool-calls type.
+    std::string generation_prompt;
+
     // reasoning budget sampler parameters
     // these are populated by the server/CLI based on chat template params
     int32_t                  reasoning_budget_tokens   = -1;   // -1 = disabled, >= 0 = token budget
-    bool                     reasoning_budget_activate_immediately = false;
     std::vector<llama_token> reasoning_budget_start;           // start tag token sequence
     std::vector<llama_token> reasoning_budget_end;             // end tag token sequence
     std::vector<llama_token> reasoning_budget_forced;          // forced sequence (message + end tag)