[Speculative decoding] feat: add DFlash support

common/arg.cpp

@@ -3087,7 +3087,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params, bool value) {
             params.use_jinja = value;
         }
-    ).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_COMPLETION, LLAMA_EXAMPLE_CLI, LLAMA_EXAMPLE_MTMD}).set_env("LLAMA_ARG_JINJA"));
+    ).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_COMPLETION, LLAMA_EXAMPLE_CLI, LLAMA_EXAMPLE_MTMD, LLAMA_EXAMPLE_SPECULATIVE}).set_env("LLAMA_ARG_JINJA"));
     add_opt(common_arg(
         {"--reasoning-format"}, "FORMAT",
         "controls whether thought tags are allowed and/or extracted from the response, and in which format they're returned; one of:\n"
@@ -3143,7 +3143,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params, const std::string & value) {
             params.chat_template = value;
         }
-    ).set_examples({LLAMA_EXAMPLE_COMPLETION, LLAMA_EXAMPLE_CLI, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_MTMD}).set_env("LLAMA_ARG_CHAT_TEMPLATE"));
+    ).set_examples({LLAMA_EXAMPLE_COMPLETION, LLAMA_EXAMPLE_CLI, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_MTMD, LLAMA_EXAMPLE_SPECULATIVE}).set_env("LLAMA_ARG_CHAT_TEMPLATE"));
     add_opt(common_arg(
         {"--chat-template-file"}, "JINJA_TEMPLATE_FILE",
         string_format(
@@ -3467,6 +3467,20 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.speculative.p_min = std::stof(value);
         }
     ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}).set_env("LLAMA_ARG_DRAFT_P_MIN"));
+    add_opt(common_arg(
+        {"--eagle3"},
+        "use EAGLE3 speculative decoding with the draft model",
+        [](common_params & params) {
+            params.speculative.eagle3 = true;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_CLI}));
+    add_opt(common_arg(
+        {"--dflash"},
+        "use DFlash speculative decoding with the draft model",
+        [](common_params & params) {
+            params.speculative.dflash = true;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_CLI}));
     add_opt(common_arg(
         {"-cd", "--ctx-size-draft"}, "N",
         string_format("size of the prompt context for the draft model (default: %d, 0 = loaded from model)", params.speculative.n_ctx),

common/common.h

@@ -159,6 +159,7 @@ enum common_speculative_type {
     COMMON_SPECULATIVE_TYPE_NONE,          // no speculative decoding
     COMMON_SPECULATIVE_TYPE_DRAFT,         // draft model
     COMMON_SPECULATIVE_TYPE_EAGLE3,        // eagle draft model
+    COMMON_SPECULATIVE_TYPE_DFLASH,        // dflash draft model
     COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE,  // simple self-speculative decoding
     COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K,   // self-speculative decoding with n-gram keys only
     COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V, // self-speculative decoding with n-gram keys and 4 m-gram values
@@ -322,10 +323,14 @@ struct common_params_speculative {
 
     struct common_params_model mparams_dft;
 
+    llama_model * model_tgt = nullptr; // the target model
     llama_model * model_dft = nullptr; // a llama_model that can be shared by multiple speculative contexts
 
     llama_context_params cparams_dft; // these are the parameters for the draft llama_context
 
+    bool    eagle3       = false; // use EAGLE3 speculative decoding
+    bool    dflash       = false; // use DFlash speculative decoding
+
     int32_t n_ctx        = 0;  // draft context size
     int32_t n_gpu_layers = -1; // number of layers to store in VRAM for the draft model (-1 - use default)