fix(llama-cpp): adapt grpc-server.cpp to upstream API changes

[walcz-de]

Summary

The llama.cpp pin at 187a45637054 (2026-04-23) tightened the upstream server API in two ways that break backend/cpp/llama-cpp/grpc-server.cpp. This PR adapts the call sites so the rocm-7.x / hipblas build (and presumably any other build that pulls this pin) compiles again.

What broke

grpc-server.cpp:1648:47: error: no member named 'logit_bias_eog' in 'server_context_meta'
grpc-server.cpp:2432:47: error: no member named 'logit_bias_eog' in 'server_context_meta'
grpc-server.cpp:2838:40: error: use of undeclared identifier 'get_media_marker'

Root cause

1. server_task::params_from_json_cmpl was reduced from five parameters to four. The logit_bias_eog vector argument was removed and server_context_meta no longer carries the field — it now lives only on common_params / sampling. The upstream server-context.cpp was updated accordingly (task.params = server_task::params_from_json_cmpl(ctx_server.vocab, params, meta->slot_n_ctx, data);) but grpc-server.cpp was not.

2. get_media_marker() is no longer a member of the server context accessor surface. The MTMD module exposes mtmd_default_marker() (declared in tools/mtmd/mtmd.h, line 104) which returns the same default marker string the upstream server uses internally. mtmd.h is already included transitively by server-context.cpp, so no new include is required.

Fix

diff --git a/backend/cpp/llama-cpp/grpc-server.cpp b/backend/cpp/llama-cpp/grpc-server.cpp
@@ -1645,7 +1645,6 @@
                         ctx_server.impl->vocab,
                         params_base,
                         ctx_server.get_meta().slot_n_ctx,
-                        ctx_server.get_meta().logit_bias_eog,
                         data);
                 task.id_slot = json_value(data, "id_slot", -1);
@@ -2429,7 +2428,6 @@
                         ctx_server.impl->vocab,
                         params_base,
                         ctx_server.get_meta().slot_n_ctx,
-                        ctx_server.get_meta().logit_bias_eog,
                         data);
                 task.id_slot = json_value(data, "id_slot", -1);
@@ -2835,7 +2833,7 @@
         if (ctx_server.impl->mctx != nullptr) {
-            response->set_media_marker(get_media_marker());
+            response->set_media_marker(mtmd_default_marker());
         }

Test plan

• Rebuilt the rocm7-llama-cpp backend image against this patch on AMD gfx1151 / Strix Halo. Earlier build with the same llama.cpp pin failed at grpc-server.cpp:1648 with the error above (9m56s wall, exit 2). After the patch the same build completes cleanly: Backends fertig: 1 OK, 0 FEHLER (3m10s).
• No behavioural change intended: the upstream API drop simply means logit_bias_eog is sourced from params_base.sampling.logit_bias_eog inside params_from_json_cmpl (where the upstream server-context.cpp reads it from), and mtmd_default_marker() returns the same string get_media_marker() previously returned.
• CI matrix on other backends (cuda, vulkan, metal) — same call sites, same fix, expected to pass.