Add -igm / --ignore-recurrent-model flag to disable recurrent model checkpoint behavior

When a recurrent model (e.g. Qwen3.5-MoE) is used with the server,
checkpoints, banned strings disabling, and ctx_shift disabling are
automatically applied (from commit 'server: enable checkpoint for
recurrent models (ikawrakow#1310)'). The new -igm / --ignore-recurrent-model
flag skips all this, making recurrent models behave as non-recurrent.

- Added ignore_recurrent_model field to gpt_params (default: false)
- Added CLI parsing for -igm / --ignore-recurrent-model
- Wrapped the recurrent model block in server-context.cpp with the flag check

Author: Nexesenex

common/common.cpp

@@ -3054,6 +3054,10 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
         params.ctx_checkpoints_tolerance = std::stoi(argv[i]);
         return true;
     }
+    if (arg == "-igm" || arg == "--ignore-recurrent-model") {
+        params.ignore_recurrent_model = true;
+        return true;
+    }
     if (arg == "-cram" || arg == "--cache-ram") {
         CHECK_ARG
         params.cache_ram_mib = std::stoi(argv[i]);
@@ -3271,6 +3275,7 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param
     options.push_back({ "*",           "-ctx-ck, --ctx-checkpoints N",           "max number of context checkpoints to create per slot (default: %d)",params.ctx_checkpoints_n});
     options.push_back({ "*",           "-ctx-ck-i, --ctx-checkpoints-interval N",  "minimum number of tokens between each context checkpoint.  (default: %d, <=0 disable)",params.ctx_checkpoints_interval});
     options.push_back({ "*",           "-ctx-ck-t, --ctx-checkpoints-tolerance N", "the number of tokens before the full prompt to create the checkpoint.  (default: %d, <=0 disable)",params.ctx_checkpoints_tolerance});
+    options.push_back({ "*",           "-igm, --ignore-recurrent-model",       "disable recurrent model checkpoint behavior, treat as non-recurrent"});
     options.push_back({ "*",           "-cram, --cache-ram N",          "set the maximum cache size in MiB (default: %d, -1 - no limit, 0 - disable)",params.cache_ram_mib });
     options.push_back({ "*",           "-crs,  --cache-ram-similarity N",           "max of similarity of prompt tokens to cache tokens that triggers prompt cache (default: %.2f).",params.cache_ram_similarity });
     options.push_back({ "*",           "-cram-n-min --cache-ram-n-min N",           "minimum number of the cached tokens that triggers prompt cache (default: %d).", params.cache_ram_n_min });

common/common.h

@@ -572,6 +572,7 @@ struct gpt_params {
     float slot_prompt_similarity = 0.1f;
 
     bool do_checkpoint = false;               // do checkpoint for recurrent models only
+    bool ignore_recurrent_model = false;      // treat recurrent models as non-recurrent (disable checkpoints)
     int32_t ctx_checkpoints_n = 32;           // max number of context checkpoints per slot
     int32_t ctx_checkpoints_interval = n_batch * 2;   // minimum number of tokens between each context checkpoints
     int32_t ctx_checkpoints_tolerance = 5;    // the number of tokens before the full prompt to create the checkpoint

examples/server/server-context.cpp

@@ -1720,22 +1720,24 @@ bool server_context::launch_slot_with_task(server_slot& slot, server_task& task)
     slot.allow_ruless_prev = slot.allow_ruless;
 
     if (llama_model_has_recurrent(llama_get_model(slot.ctx))) {
-        params_base.can_ban_phrases = false;
-        bool do_checkpoint = params_base.ctx_checkpoints_n > 0;
-        // make checkpoints only for completion tasks
-        do_checkpoint = do_checkpoint && task.type == SERVER_TASK_TYPE_COMPLETION;
-        // make a checkpoint of the parts of the memory that cannot be rolled back.
-        // checkpoints are created only if:
-        // - the model architecture is marked as recurrent or hybrid
-        //
-        // TODO: try to make this conditional on the context or the memory module, instead of the model type
-        params_base.do_checkpoint = do_checkpoint;
-        if (slot.n_buffer != 0) {
-            LLAMA_LOG_WARN("banned strings is not supported by recurrent model, it will be disabled.\n");
-        }
-        if (params_base.ctx_shift) {
-            params_base.ctx_shift = false;
-            LOG_WARNING("%s\n", "ctx_shift is not supported by recurrent model, it will be disabled");
+        if (!params_base.ignore_recurrent_model) {
+            params_base.can_ban_phrases = false;
+            bool do_checkpoint = params_base.ctx_checkpoints_n > 0;
+            // make checkpoints only for completion tasks
+            do_checkpoint = do_checkpoint && task.type == SERVER_TASK_TYPE_COMPLETION;
+            // make a checkpoint of the parts of the memory that cannot be rolled back.
+            // checkpoints are created only if:
+            // - the model architecture is marked as recurrent or hybrid
+            //
+            // TODO: try to make this conditional on the context or the memory module, instead of the model type
+            params_base.do_checkpoint = do_checkpoint;
+            if (slot.n_buffer != 0) {
+                LLAMA_LOG_WARN("banned strings is not supported by recurrent model, it will be disabled.\n");
+            }
+            if (params_base.ctx_shift) {
+                params_base.ctx_shift = false;
+                LOG_WARNING("%s\n", "ctx_shift is not supported by recurrent model, it will be disabled");
+            }
         }
     }
     if (llama_model_is_split_mode_tensor_parallel(llama_get_model(slot.ctx))) {