talk-llama : sync llama.cpp

examples/talk-llama/llama-quant.cpp

@@ -7,14 +7,12 @@
 #include <algorithm>
 #include <cmath>
 #include <cstring>
+#include <cinttypes>
 #include <fstream>
 #include <mutex>
 #include <thread>
 #include <unordered_map>
 
-// TODO: replace with ggml API call
-#define QK_K 256
-
 static void zeros(std::ofstream & file, size_t n) {
     char zero = 0;
     for (size_t i = 0; i < n; ++i) {
@@ -154,8 +152,10 @@ static ggml_type llama_tensor_get_type(quantize_state_impl & qs, ggml_type new_t
         if (qs.params->output_tensor_type < GGML_TYPE_COUNT) {
             new_type = qs.params->output_tensor_type;
         } else {
-            int nx = tensor->ne[0];
-            if (arch == LLM_ARCH_FALCON || nx % QK_K != 0) {
+            const int64_t nx = tensor->ne[0];
+            const int64_t qk_k = ggml_blck_size(new_type);
+
+            if (arch == LLM_ARCH_FALCON || nx % qk_k != 0) {
                 new_type = GGML_TYPE_Q8_0;
             }
             else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS ||
@@ -235,7 +235,7 @@ static ggml_type llama_tensor_get_type(quantize_state_impl & qs, ggml_type new_t
         else if ((ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) &&
                 use_more_bits(qs.i_attention_wv, qs.n_attention_wv)) new_type = GGML_TYPE_Q6_K;
         else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && qs.i_attention_wv < 4) new_type = GGML_TYPE_Q5_K;
-        if (qs.model.type == MODEL_70B) {
+        if (qs.model.type == LLM_TYPE_70B) {
             // In the 70B model we have 8 heads sharing the same attn_v weights. As a result, the attn_v.weight tensor is
             // 8x smaller compared to attn_q.weight. Hence, we can get a nice boost in quantization accuracy with
             // nearly negligible increase in model size by quantizing this tensor with more bits:
@@ -367,20 +367,19 @@ static ggml_type llama_tensor_get_type(quantize_state_impl & qs, ggml_type new_t
     //    if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_S) new_type = GGML_TYPE_Q4_K;
     //}
     bool convert_incompatible_tensor = false;
-    if (new_type == GGML_TYPE_Q2_K    || new_type == GGML_TYPE_Q3_K    || new_type == GGML_TYPE_Q4_K   ||
-        new_type == GGML_TYPE_Q5_K    || new_type == GGML_TYPE_Q6_K    || new_type == GGML_TYPE_IQ4_XS ||
-        new_type == GGML_TYPE_IQ2_XS  || new_type == GGML_TYPE_IQ2_XXS || new_type == GGML_TYPE_IQ2_S  ||
-        new_type == GGML_TYPE_IQ3_XXS || new_type == GGML_TYPE_IQ1_S   || new_type == GGML_TYPE_IQ3_S  ||
-        new_type == GGML_TYPE_IQ1_M) {
-        int nx = tensor->ne[0];
-        int ny = tensor->ne[1];
-        if (nx % QK_K != 0) {
-            LLAMA_LOG_WARN("\n\n%s : tensor cols %d x %d are not divisible by %d, required for %s", __func__, nx, ny, QK_K, ggml_type_name(new_type));
+    {
+        const int64_t nx = tensor->ne[0];
+        const int64_t ny = tensor->ne[1];
+        const int64_t qk_k = ggml_blck_size(new_type);
+
+        if (nx % qk_k != 0) {
+            LLAMA_LOG_WARN("\n\n%s : tensor cols %" PRId64 " x %" PRId64 " are not divisible by %" PRId64 ", required for %s", __func__, nx, ny, qk_k, ggml_type_name(new_type));
             convert_incompatible_tensor = true;
         } else {
             ++qs.n_k_quantized;
         }
     }
+
     if (convert_incompatible_tensor) {
         switch (new_type) {
             case GGML_TYPE_TQ1_0:
@@ -526,18 +525,20 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
         auto v = (std::vector<llama_model_kv_override>*)params->kv_overrides;
         kv_overrides = v->data();
     }
+
     llama_model_loader ml(fname_inp, use_mmap, /*check_tensors*/ true, kv_overrides);
     ml.init_mappings(false); // no prefetching
 
-    llama_model model;
-    llm_load_arch   (ml, model);
-    llm_load_hparams(ml, model);
-    llm_load_stats  (ml, model);
+    llama_model model(llama_model_default_params());
+
+    model.load_arch   (ml);
+    model.load_hparams(ml);
+    model.load_stats  (ml);
 
     struct quantize_state_impl qs(model, params);
 
     if (params->only_copy) {
-        ftype = model.ftype;
+        ftype = ml.ftype;
     }
     const std::unordered_map<std::string, std::vector<float>> * imatrix_data = nullptr;
     if (params->imatrix) {
@@ -621,7 +622,8 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
 
     qs.n_ffn_down = qs.n_ffn_gate = qs.n_ffn_up = (int)model.hparams.n_layer;
 
-    // sanity checks
+    // sanity checks for models that have attention layers
+    if (qs.n_attention_wv != 0)
     {
         const auto & n_head_kv_iter = model.hparams.n_head_kv_arr.begin();
         // attention layers have a non-zero number of kv heads
@@ -759,6 +761,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
         quantize &= name.find("time_mix_w2.weight") == std::string::npos;
         quantize &= name.find("time_mix_decay_w1.weight") == std::string::npos;
         quantize &= name.find("time_mix_decay_w2.weight") == std::string::npos;
+        quantize &= name.find("time_mix_lerp_fused.weight") == std::string::npos;
 
         // do not quantize relative position bias (T5)
         quantize &= name.find("attn_rel_b.weight") == std::string::npos;
@@ -875,7 +878,8 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
 
         // update the gguf meta data as we go
         gguf_set_tensor_type(ctx_outs[cur_split].get(), name.c_str(), new_type);
-        gguf_set_tensor_data(ctx_outs[cur_split].get(), name.c_str(), new_data, new_size);
+        GGML_ASSERT(gguf_get_tensor_size(ctx_outs[cur_split].get(), gguf_find_tensor(ctx_outs[cur_split].get(), name.c_str())) == new_size);
+        gguf_set_tensor_data(ctx_outs[cur_split].get(), name.c_str(), new_data);
 
         // write tensor data + padding
         fout.write((const char *) new_data, new_size);