server : fix img token logs (#16595)

* llama-quant: add support for mmproj

* Update src/llama.cpp

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

* check prefix instead

* small fix

---------

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

ggml/src/ggml-cuda/ggml-cuda.cu

@@ -273,6 +273,15 @@ static ggml_cuda_device_info ggml_cuda_init() {
         } else if (device_name.substr(0, 21) == "NVIDIA GeForce GTX 16") {
             turing_devices_without_mma.push_back({ id, device_name });
         }
+
+        // Temporary performance fix:
+        // Setting device scheduling strategy for iGPUs with cc121 to "spinning" to avoid delays in cuda synchronize calls.
+        // TODO: Check for future drivers the default scheduling strategy and
+        // remove this call again when cudaDeviceScheduleSpin is default.
+        if (prop.major == 12 && prop.minor == 1) {
+            CUDA_CHECK(cudaSetDeviceFlags(cudaDeviceScheduleSpin));
+        }
+
 #endif  // defined(GGML_USE_HIP)
     }
 

src/llama-arch.cpp

@@ -5,6 +5,7 @@
 #include <map>
 
 static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
+    { LLM_ARCH_CLIP,             "clip"             }, // dummy, only used by llama-quantize
     { LLM_ARCH_LLAMA,            "llama"            },
     { LLM_ARCH_LLAMA4,           "llama4"           },
     { LLM_ARCH_DECI,             "deci"             },
@@ -275,6 +276,10 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
 };
 
 static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_NAMES = {
+    {
+        LLM_ARCH_CLIP,
+        {},
+    },
     {
         LLM_ARCH_LLAMA,
         {

src/llama-arch.h

@@ -9,6 +9,7 @@
 //
 
 enum llm_arch {
+    LLM_ARCH_CLIP,
     LLM_ARCH_LLAMA,
     LLM_ARCH_LLAMA4,
     LLM_ARCH_DECI,

src/llama-model.cpp

@@ -478,7 +478,8 @@ void llama_model::load_hparams(llama_model_loader & ml) {
     ml.get_key(LLM_KV_GENERAL_NAME, name, false);
 
     // everything past this point is not vocab-related
-    if (hparams.vocab_only) {
+    // for CLIP models, we only need to load tensors, no hparams
+    if (hparams.vocab_only || ml.get_arch() == LLM_ARCH_CLIP) {
         return;
     }
 
@@ -20013,6 +20014,7 @@ int32_t llama_n_head(const llama_model * model) {
 llama_rope_type llama_model_rope_type(const llama_model * model) {
     switch (model->arch) {
         // these models do not use RoPE
+        case LLM_ARCH_CLIP:
         case LLM_ARCH_GPT2:
         case LLM_ARCH_GPTJ:
         case LLM_ARCH_MPT:

src/llama-quant.cpp

@@ -701,6 +701,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
         });
     }
 
+    bool is_clip_model = false;
     for (const auto * it : tensors) {
         const struct ggml_tensor * tensor = it->tensor;
 
@@ -714,12 +715,14 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
         } else if (name == LLM_TN(model.arch)(LLM_TENSOR_OUTPUT, "weight")) {
             qs.has_output = true;
         }
+
+        is_clip_model |= name.rfind("mm.", 0) == 0; // check the "mm." prefix
     }
 
     qs.n_ffn_down = qs.n_ffn_gate = qs.n_ffn_up = (int)model.hparams.n_layer;
 
     // sanity checks for models that have attention layers
-    if (qs.n_attention_wv != 0)
+    if (qs.n_attention_wv != 0 && !is_clip_model)
     {
         const auto & n_head_kv_iter = model.hparams.n_head_kv_arr.begin();
         // attention layers have a non-zero number of kv heads
@@ -881,6 +884,9 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
         // do not quantize relative position bias (T5)
         quantize &= name.find("attn_rel_b.weight") == std::string::npos;
 
+        // do not quantize specific multimodal tensors
+        quantize &= name.find(".position_embd.") == std::string::npos;
+
         ggml_type new_type;
         void * new_data;
         size_t new_size;

src/llama.cpp

@@ -124,6 +124,9 @@ static int llama_model_load(const std::string & fname, std::vector<std::string>
         } catch(const std::exception & e) {
             throw std::runtime_error("error loading model hyperparameters: " + std::string(e.what()));
         }
+        if (model.arch == LLM_ARCH_CLIP) {
+            throw std::runtime_error("CLIP cannot be used as main model, use it with --mmproj instead");
+        }
         try {
             model.load_vocab(ml);
         } catch(const std::exception & e) {

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");
                                 }
 
@@ -3839,14 +3839,14 @@ struct server_context {
 
                                         {
                                             const auto token = slot.prompt.tokens[i];
-                                            const auto piece = common_token_to_piece(ctx, token);
+                                            const auto piece = token != LLAMA_TOKEN_NULL ? common_token_to_piece(ctx, token) : "[mtmd]";
                                             ss0 << piece;
                                             st0 << std::setw(8) << token;
                                         }
 
                                         {
                                             const auto token = slot.task->tokens[i];
-                                            const auto piece = common_token_to_piece(ctx, token);
+                                            const auto piece = token != LLAMA_TOKEN_NULL ? common_token_to_piece(ctx, token) : "[mtmd]";
                                             ss1 << piece;
                                             st1 << std::setw(8) << token;
                                         }
@@ -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 @@ public:
             // 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
                 }
             }