From df9135e3a9a6708bb62e6484d239e2b4ea212ed7 Mon Sep 17 00:00:00 2001
From: Concedo <39025047+LostRuins@users.noreply.github.com>
Date: Fri, 23 Jun 2023 18:41:23 +0800
Subject: [PATCH] fixing memory bugs

---
 gpttype_adapter.cpp    | 8 ++++++--
 koboldcpp.py           | 2 +-
 llama.cpp              | 4 ++--
 model_adapter.cpp      | 2 +-
 otherarch/llama_v2.cpp | 4 ++--
 5 files changed, 12 insertions(+), 8 deletions(-)

diff --git a/gpttype_adapter.cpp b/gpttype_adapter.cpp
index d0ddaf99b..4e087bd65 100644
--- a/gpttype_adapter.cpp
+++ b/gpttype_adapter.cpp
@@ -308,8 +308,12 @@ ModelLoadResult gpttype_load_model(const load_model_inputs inputs, FileFormat in
     params.memory_f16 = inputs.f16_kv;
     params.n_ctx = inputs.max_context_length;
 
-    neox_ctx_v2.hparams.n_ctx = gptj_ctx_v1.hparams.n_ctx = gptj_ctx_v2.hparams.n_ctx = gpt2_ctx_v1.hparams.n_ctx = gpt2_ctx_v2.hparams.n_ctx
-    = neox_ctx_v3.hparams.n_ctx = gptj_ctx_v3.hparams.n_ctx = gptj_ctx_v3.hparams.n_ctx = mpt_ctx_v3.hparams.n_ctx = params.n_ctx;
+    neox_ctx_v2.hparams.n_ctx  = neox_ctx_v3.hparams.n_ctx
+    = gptj_ctx_v1.hparams.n_ctx = gptj_ctx_v2.hparams.n_ctx = gptj_ctx_v3.hparams.n_ctx
+    = gpt2_ctx_v1.hparams.n_ctx = gpt2_ctx_v2.hparams.n_ctx = gpt2_ctx_v3.hparams.n_ctx
+    = mpt_ctx_v3.hparams.n_ctx = params.n_ctx;
+
+    bool calc_mem_with_scratch = ggml_cpu_has_gpublas();
 
     printf("System Info: %s\n", llama_print_system_info());
     SetQuantsUnshuffled(false);
diff --git a/koboldcpp.py b/koboldcpp.py
index aa5426262..76e94b84a 100644
--- a/koboldcpp.py
+++ b/koboldcpp.py
@@ -225,7 +225,7 @@ maxhordectx = 1024
 maxhordelen = 256
 modelbusy = False
 defaultport = 5001
-KcppVersion = "1.32"
+KcppVersion = "1.32.1"
 showdebug = True
 
 class ServerRequestHandler(http.server.SimpleHTTPRequestHandler):
diff --git a/llama.cpp b/llama.cpp
index 27d3d4a0a..aa67038e0 100644
--- a/llama.cpp
+++ b/llama.cpp
@@ -80,7 +80,7 @@ static const std::map<e_model, size_t> & MEM_REQ_SCRATCH0()
         { MODEL_3B,    256ull * MB },
         { MODEL_7B,    512ull * MB },
         { MODEL_13B,   512ull * MB },
-        { MODEL_30B,   512ull * MB },
+        { MODEL_30B,   640ull * MB },
         { MODEL_65B,  1024ull * MB },
     };
     return k_sizes;
@@ -92,7 +92,7 @@ static const std::map<e_model, size_t> & MEM_REQ_SCRATCH1()
         { MODEL_3B,    256ull * MB },
         { MODEL_7B,    512ull * MB },
         { MODEL_13B,   512ull * MB },
-        { MODEL_30B,   512ull * MB },
+        { MODEL_30B,   640ull * MB },
         { MODEL_65B,  1024ull * MB },
     };
     return k_sizes;
diff --git a/model_adapter.cpp b/model_adapter.cpp
index 547a8a1ef..da9fa193e 100644
--- a/model_adapter.cpp
+++ b/model_adapter.cpp
@@ -98,7 +98,7 @@ void print_tok_vec(std::vector<float> &embd)
        //we need to read more to determine
        int32_t vocabsiz = 0;
        fin.read((char *) &vocabsiz, sizeof(int32_t));
-       if(vocabsiz==4096) //actually the d_model for mpt
+       if(vocabsiz==4096 || vocabsiz==7168) //actually the d_model for mpt
        {
            fileformat = FileFormat::MPT_1;
        }
diff --git a/otherarch/llama_v2.cpp b/otherarch/llama_v2.cpp
index 167f3e9c3..2f8e168ca 100644
--- a/otherarch/llama_v2.cpp
+++ b/otherarch/llama_v2.cpp
@@ -59,7 +59,7 @@ static const std::map<e_model2, size_t> & MEM_REQ_SCRATCH0_2()
         { MODEL_UNKNOWN_2, 512ull * MB_2 },
         { MODEL_7B_2,    512ull * MB_2 },
         { MODEL_13B_2,   512ull * MB_2 },
-        { MODEL_30B_2,   512ull * MB_2 },
+        { MODEL_30B_2,   640ull * MB_2 },
         { MODEL_65B_2,  1024ull * MB_2 },
     };
     return k_sizes;
@@ -71,7 +71,7 @@ static const std::map<e_model2, size_t> & MEM_REQ_SCRATCH1_2()
         { MODEL_UNKNOWN_2, 512ull * MB_2 },
         { MODEL_7B_2,    512ull * MB_2 },
         { MODEL_13B_2,   512ull * MB_2 },
-        { MODEL_30B_2,   512ull * MB_2 },
+        { MODEL_30B_2,   640ull * MB_2 },
         { MODEL_65B_2,  1024ull * MB_2 },
     };
     return k_sizes;