swa full used unless ctx shift and fast forward disabled

Author: LostRuins

common/arg.cpp

@@ -1446,6 +1446,14 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.n_keep = value;
         }
     ));
+    add_opt(common_arg(
+        {"--swa-full"},
+        string_format("use full-size SWA cache (default: %s)\n"
+            "[(more info)](https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)", params.swa_full ? "true" : "false"),
+        [](common_params & params) {
+            params.swa_full = true;
+        }
+    ));
     add_opt(common_arg(
         {"--no-context-shift"},
         string_format("disables context shift on infinite text generation (default: %s)", params.ctx_shift ? "disabled" : "enabled"),

common/common.cpp

@@ -1144,6 +1144,7 @@ struct llama_context_params common_context_params_to_llama(const common_params &
     cparams.flash_attn        = params.flash_attn;
     cparams.no_perf           = params.no_perf;
     cparams.op_offload        = !params.no_op_offload;
+    cparams.swa_full          = params.swa_full;
 
     if (params.reranking) {
         cparams.embeddings    = true;

common/common.h

@@ -319,6 +319,7 @@ struct common_params {
     bool flash_attn        = false; // flash attention
     bool no_perf           = false; // disable performance metrics
     bool ctx_shift         = true;  // context shift on inifinite text generation
+    bool swa_full          = false; // use full-size SWA cache (https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
 
     bool input_prefix_bos  = false; // prefix BOS to user inputs, preceding input_prefix
     bool use_mmap          = true;  // use mmap for faster loads

gpttype_adapter.cpp

@@ -583,6 +583,7 @@ static void speculative_decoding_setup(std::string spec_model_filename, const ll
     draft_ctx_params.flash_attn = base_ctx_params.flash_attn;
     draft_ctx_params.type_k = base_ctx_params.type_k;
     draft_ctx_params.type_v = base_ctx_params.type_v;
+    draft_ctx_params.swa_full = base_ctx_params.swa_full;
 
     llama_model * draftmodel = llama_model_load_from_file(spec_model_filename.c_str(), draft_model_params);
     draft_ctx = llama_init_from_model(draftmodel, draft_ctx_params);
@@ -1923,6 +1924,7 @@ ModelLoadResult gpttype_load_model(const load_model_inputs inputs, FileFormat in
     kcpp_data->use_smartcontext = inputs.use_smartcontext;
     kcpp_data->use_contextshift = inputs.use_contextshift;
     kcpp_data->use_fastforward = inputs.use_fastforward;
+    kcpp_data->swa_full = (inputs.use_fastforward || inputs.use_contextshift)?true:false;
     debugmode = inputs.debugmode;
     draft_ctx = nullptr;
     guidance_ctx = nullptr;
@@ -2318,6 +2320,7 @@ ModelLoadResult gpttype_load_model(const load_model_inputs inputs, FileFormat in
         }
 
         llama_ctx_params.flash_attn = kcpp_data->flash_attn;
+        llama_ctx_params.swa_full = kcpp_data->swa_full;
         llama_ctx_params.type_k = (inputs.quant_k>1?GGML_TYPE_Q4_0:(inputs.quant_k==1?GGML_TYPE_Q8_0:GGML_TYPE_F16));
         llama_ctx_params.type_v = (inputs.quant_v>1?GGML_TYPE_Q4_0:(inputs.quant_v==1?GGML_TYPE_Q8_0:GGML_TYPE_F16));
         llama_ctx_v4 = llama_init_from_model(llamamodel, llama_ctx_params);

include/llama.h

@@ -363,10 +363,11 @@ extern "C" {
 
         // Keep the booleans together and at the end of the struct to avoid misalignment during copy-by-value.
         bool embeddings;  // if true, extract embeddings (together with logits)
-        bool offload_kqv; // whether to offload the KQV ops (including the KV cache) to GPU
-        bool flash_attn;  // whether to use flash attention [EXPERIMENTAL]
-        bool no_perf;     // whether to measure performance timings
-        bool op_offload;  // whether to offload host tensor operations to device
+        bool offload_kqv; // offload the KQV ops (including the KV cache) to GPU
+        bool flash_attn;  // use flash attention [EXPERIMENTAL]
+        bool no_perf;     // measure performance timings
+        bool op_offload;  // offload host tensor operations to device
+        bool swa_full;    // use full-size SWA cache (https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
     };
 
     // model quantization parameters
@@ -732,10 +733,18 @@ extern "C" {
                        llama_pos   p1,
                              int   d);
 
+    // Returns the smallest position present in the KV cache for the specified sequence
+    // This is typically non-zero only for SWA caches
+    // Return -1 if the sequence is empty
+    LLAMA_API llama_pos llama_kv_self_seq_pos_min(
+            struct llama_context * ctx,
+                    llama_seq_id   seq_id);
+
     // Returns the largest position present in the KV cache for the specified sequence
+    // Return -1 if the sequence is empty
     LLAMA_API llama_pos llama_kv_self_seq_pos_max(
             struct llama_context * ctx,
-                     llama_seq_id   seq_id);
+                    llama_seq_id   seq_id);
 
     // Defragment the KV cache
     // This will be applied:
@@ -945,9 +954,12 @@ extern "C" {
     // Requires KV cache.
     // For encode-decoder contexts, processes the batch using the decoder.
     // Positive return values does not mean a fatal error, but rather a warning.
-    //   0 - success
-    //   1 - could not find a KV slot for the batch (try reducing the size of the batch or increase the context)
-    // < 0 - error. the KV cache state is restored to the state before this call
+    // Upon non-zero return values, the KV cache state is restored to the state before this call
+    //    0 - success
+    //    1 - could not find a KV slot for the batch (try reducing the size of the batch or increase the context)
+    //    2 - aborted
+    //   -1 - invalid input batch
+    // < -1 - error
     LLAMA_API int32_t llama_decode(
             struct llama_context * ctx,
               struct llama_batch   batch);

otherarch/otherarch.h

@@ -56,6 +56,7 @@ struct kcpp_params {
     bool use_smartcontext            = false;
     bool use_contextshift            = false;
     bool use_fastforward             = false;
+    bool swa_full                    = true;
 };
 
 // default hparams (GPT-J 6B)

src/llama-context.cpp

@@ -93,6 +93,7 @@ llama_context::llama_context(
     }
 
     cparams.n_ubatch = std::min(cparams.n_batch, params.n_ubatch == 0 ? params.n_batch : params.n_ubatch);
+
     cparams.op_offload = params.op_offload;
 
     const uint32_t n_ctx_per_seq = cparams.n_ctx / cparams.n_seq_max;
@@ -176,8 +177,9 @@ llama_context::llama_context(
     // init the memory module
     if (!hparams.vocab_only) {
         llama_memory_params params_mem = {
-            /*.type_k =*/ params.type_k,
-            /*.type_v =*/ params.type_v,
+            /*.type_k   =*/ params.type_k,
+            /*.type_v   =*/ params.type_v,
+            /*.swa_full =*/ params.swa_full,
         };
 
         memory.reset(model.create_memory(params_mem, cparams));
@@ -947,8 +949,6 @@ int llama_context::decode(llama_batch & inp_batch) {
 
         // find KV slot
         if (!kv_self->find_slot(ubatch)) {
-            LLAMA_LOG_WARN("%s: failed to find KV cache slot for ubatch of size %d\n", __func__, ubatch.n_tokens);
-
             return 1;
         }
 
@@ -2093,6 +2093,7 @@ llama_context_params llama_context_default_params() {
         /*.flash_attn                  =*/ false,
         /*.no_perf                     =*/ true,
         /*.op_offload                  =*/ true,
+        /*.swa_full                    =*/ true,
     };
 
     return result;
@@ -2467,6 +2468,15 @@ void llama_kv_self_seq_div(
     kv->seq_div(seq_id, p0, p1, d);
 }
 
+llama_pos llama_kv_self_seq_pos_min(llama_context * ctx, llama_seq_id seq_id) {
+    const auto * kv = ctx->get_kv_self();
+    if (!kv) {
+        return -1;
+    }
+
+    return kv->seq_pos_min(seq_id);
+}
+
 // deprecated
 llama_pos llama_kv_cache_seq_pos_max(llama_context * ctx, llama_seq_id seq_id) {
     return llama_kv_self_seq_pos_max(ctx, seq_id);
@@ -2475,7 +2485,7 @@ llama_pos llama_kv_cache_seq_pos_max(llama_context * ctx, llama_seq_id seq_id) {
 llama_pos llama_kv_self_seq_pos_max(llama_context * ctx, llama_seq_id seq_id) {
     const auto * kv = ctx->get_kv_self();
     if (!kv) {
-        return 0;
+        return -1;
     }
 
     return kv->seq_pos_max(seq_id);
@@ -2637,7 +2647,21 @@ int32_t llama_encode(
 int32_t llama_decode(
         llama_context * ctx,
           llama_batch   batch) {
-    const int ret = ctx->decode(batch);
+    int ret = ctx->decode(batch);
+
+    // defrag and try again
+    // TODO: distinguish return code when we are sure that even after defrag there is no space available
+    if (ret == 1) {
+        llama_kv_self_defrag(ctx);
+        ret = ctx->decode(batch);
+
+        if (ret == 1) {
+            LLAMA_LOG_WARN("%s: failed to find KV cache slot for batch of size %d\n", __func__, batch.n_tokens);
+
+            return ret;
+        }
+    }
+
     if (ret != 0) {
         LLAMA_LOG_ERROR("%s: failed to decode, ret = %d\n", __func__, ret);
     }