server : fix mtmd checkpoints (ggml-org#16591)

Author: ggerganov

tools/server/server.cpp

@@ -3812,7 +3812,7 @@ struct server_context {
                             if (slot.n_past > 0 && slot.n_past < (int) slot.prompt.tokens.size()) {
                                 const auto pos_min = llama_memory_seq_pos_min(llama_get_memory(ctx), slot.id);
                                 if (pos_min == -1) {
-                                    SLT_ERR(slot, "n_past = %d, cache_tokens.size() = %d, seq_id = %d, pos_min = %d\n", slot.n_past, (int) slot.prompt.tokens.size(), slot.id, pos_min);
+                                    SLT_ERR(slot, "n_past = %d, slot.prompt.tokens.size() = %d, seq_id = %d, pos_min = %d\n", slot.n_past, (int) slot.prompt.tokens.size(), slot.id, pos_min);
                                     GGML_ABORT("pos_min == -1, but n_past > 0 - should not happen: https://github.com/ggml-org/llama.cpp/pull/13833#discussion_r2116181237");
                                 }
 
@@ -3860,7 +3860,7 @@ struct server_context {
                                 }
 
                                 if (pos_min > pos_min_thold) {
-                                    SLT_WRN(slot, "n_past = %d, cache_tokens.size() = %d, seq_id = %d, pos_min = %d, n_swa = %d\n", slot.n_past, (int) slot.prompt.tokens.size(), slot.id, pos_min, n_swa);
+                                    SLT_WRN(slot, "n_past = %d, slot.prompt.tokens.size() = %d, seq_id = %d, pos_min = %d, n_swa = %d\n", slot.n_past, (int) slot.prompt.tokens.size(), slot.id, pos_min, n_swa);
 
                                     // search for a context checkpoint
                                     const auto it = std::find_if(
@@ -4028,7 +4028,7 @@ struct server_context {
                         }
                     }
 
-                    // SLT_INF(slot, "new cache_tokens: %s\n", slot.cache_tokens.str().c_str());
+                    // SLT_INF(slot, "new slot.prompt.tokens: %s\n", slot.slot.prompt.tokens.str().c_str());
 
                     SLT_INF(slot, "prompt processing progress, n_past = %d, n_tokens = %d, progress = %f\n", slot.n_past, batch.n_tokens, (float) slot.n_past / slot.n_prompt_tokens());
 

tools/server/utils.hpp

@@ -1237,9 +1237,10 @@ struct server_tokens {
             // allowed to resize      ^                    ^
             // disallowed to resize          ^      ^             ^
             if (n > 0) {
-                llama_token last_token = tokens[n - 1];
                 // make sure we never remove tokens in the middle of an image
-                if (last_token == LLAMA_TOKEN_NULL) {
+                // note that the case where we keep a full image at the end is allowed:
+                //   tokens[n - 1] == LLAMA_TOKEN_NULL && tokens[n] != LLAMA_TOKEN_NULL
+                if (tokens[n - 1] == LLAMA_TOKEN_NULL && tokens[n] == LLAMA_TOKEN_NULL) {
                     find_chunk(n - 1); // will throw an error if the token is not begin-of-chunk
                 }
             }