added support for gpt4all original format

Makefile

@@ -238,8 +238,8 @@ ggml.o: ggml.c ggml.h
 ggml_blas.o: ggml.c ggml.h
 	$(CC)  $(CFLAGS) -DGGML_USE_OPENBLAS -c ggml.c -o ggml_blas.o
 
-ggml_old_v1.o: otherarch/ggml_old.c otherarch/ggml_old.h
-	$(CC)  $(CFLAGS)   -c otherarch/ggml_old.c -o ggml_old_v1.o
+ggml_v1.o: otherarch/ggml_v1.c otherarch/ggml_v1.h
+	$(CC)  $(CFLAGS)   -c otherarch/ggml_v1.c -o ggml_v1.o
 
 llama.o: llama.cpp llama.h
 	$(CXX) $(CXXFLAGS) -c llama.cpp -o llama.o
@@ -256,8 +256,8 @@ expose.o: expose.cpp expose.h
 llama_adapter.o: 
 	$(CXX) $(CXXFLAGS) -c llama_adapter.cpp -o llama_adapter.o
 	
-gptj_adapter.o: ggml_old_v1.o
-	$(CXX) $(CXXFLAGS) otherarch/gptj_old.cpp otherarch/utils.cpp ggml_old_v1.o gptj_adapter.cpp -o gptj_adapter.o
+gptj_adapter.o: ggml.o
+	$(CXX) $(CXXFLAGS) otherarch/gptj.cpp otherarch/utils.cpp ggml.o gptj_adapter.cpp -o gptj_adapter.o
 
 clean:
 	rm -vf *.o main quantize perplexity embedding main.exe quantize.exe llamacpp.dll llamacpp_blas.dll gpt2.exe gptj.exe
@@ -268,8 +268,11 @@ main: examples/main/main.cpp ggml.o llama.o common.o
 	@echo '====  Run ./main -h for help.  ===='
 	@echo
 
-gptj: ggml_old_v1.o
-	$(CXX) $(CXXFLAGS) otherarch/gptj_old.cpp otherarch/utils.cpp ggml_old_v1.o -o gptj $(LDFLAGS)
+gptj: ggml.o
+	$(CXX) $(CXXFLAGS) otherarch/gptj.cpp otherarch/utils.cpp ggml.o -o gptj $(LDFLAGS)
+
+gptjold: ggml_v1.o
+	$(CXX) $(CXXFLAGS) otherarch/gptj_old.cpp otherarch/utils.cpp ggml_v1.o -o gptj $(LDFLAGS)
 
 
 llamalib: ggml.o expose.o llama_adapter.o llamaextra.o common.o

gptj_adapter.cpp

@@ -35,7 +35,7 @@ bool gptj_load_model(const load_model_inputs inputs, FileFormat in_file_format)
     n_batch = params.n_batch = inputs.batch_size;
     modelname = params.model = inputs.model_filename;
 
-    if (!legacy_gptj_model_load(params.model, model, vocab)) {
+    if (!gptj_model_load(params.model, model, vocab)) {
         fprintf(stderr, "%s: failed to load model from '%s'\n", __func__, params.model.c_str());
         return false;
     }
@@ -151,7 +151,7 @@ generation_outputs gptj_generate(const generation_inputs inputs, generation_outp
                 printf("\rGenerating (%d / %d tokens)", (1 + params.n_predict - remaining_tokens), params.n_predict);
             }
            
-            if (!legacy_gptj_eval(model, params.n_threads, n_past, embd, logits, mem_per_token))
+            if (!gptj_eval(model, params.n_threads, n_past, embd, logits, mem_per_token))
             {
                 fprintf(stderr, "Failed to predict\n");
                 snprintf(output.text, sizeof(output.text), "%s", "");

llamaextra.cpp

@@ -272,7 +272,13 @@ void print_tok_vec(std::vector<int> &embd)
         vocab.id_to_token.resize(model.hparams.n_vocab);
         std::vector<char> tmp(64);
 
-        for (int i = 0; i < model.hparams.n_vocab; i++) {
+        int32_t vocabloops = model.hparams.n_vocab;
+        if(vocabloops==32001 && legacy_file_format)
+        {
+            printf("---\n!! WARNING: Model appears to be GPT4ALL v1 model, triggering compatibility fix !!\n---\n");
+            vocabloops -= 1;
+        }
+        for (int i = 0; i < vocabloops; i++) {
             uint32_t len;
             fin.read((char *) &len, sizeof(len));
 

otherarch/ggml_v1.c

@@ -1,4 +1,4 @@
-#include "ggml_old.h"
+#include "ggml_v1.h"
 
 #if defined(_MSC_VER) || defined(__MINGW32__)
 #include <malloc.h> // using malloc.h with MSC/MINGW

otherarch/gptj.cpp

@@ -14,7 +14,19 @@
 #include <iostream>
 #include <unistd.h>
 
-
+bool should_transpose_layer(std::string name)
+{
+       
+    if(name.find(".mlp.fc_in.weight")!=std::string::npos || 
+    name.find(".attn.out_proj.weight")!=std::string::npos || 
+    name.find(".attn.q_proj.weight")!=std::string::npos || 
+    name.find(".attn.k_proj.weight")!=std::string::npos || 
+    name.find(".attn.v_proj.weight")!=std::string::npos)
+    {
+        return true;
+    }
+    return false;
+}
 
 // load the model's weights from a file
 bool gptj_model_load(const std::string & fname, gptj_model & model, gpt_vocab & vocab) {
@@ -139,6 +151,7 @@ bool gptj_model_load(const std::string & fname, gptj_model & model, gpt_vocab &
 
         ctx_size += (5 + 10*n_layer)*256; // object overhead
 
+        ctx_size = ctx_size * 3 / 2;
         printf("%s: ggml ctx size = %6.2f MB\n", __func__, ctx_size/(1024.0*1024.0));
     }
 
@@ -279,6 +292,7 @@ bool gptj_model_load(const std::string & fname, gptj_model & model, gpt_vocab &
                 fprintf(stderr, "%s: tensor '%s' has wrong size in model file\n", __func__, name.data());
                 return false;
             }
+          
 
             if (tensor->ne[0] != ne[0] || tensor->ne[1] != ne[1]) {
                 fprintf(stderr, "%s: tensor '%s' has wrong shape in model file: got [%d, %d], expected [%d, %d]\n",
@@ -312,7 +326,7 @@ bool gptj_model_load(const std::string & fname, gptj_model & model, gpt_vocab &
             }
 
             fin.read(reinterpret_cast<char *>(tensor->data), ggml_nbytes(tensor));
-
+          
             //printf("%42s - [%5d, %5d], type = %6s, %6.2f MB\n", name.data(), ne[0], ne[1], ftype == 0 ? "float" : "f16", ggml_nbytes(tensor)/1024.0/1024.0);
             total_size += ggml_nbytes(tensor);
             if (++n_tensors % 8 == 0) {

otherarch/gptj_old.cpp

@@ -1,4 +1,4 @@
-#include "ggml_old.h"
+#include "ggml_v1.h"
 #include "otherarch.h"
 
 #include "utils.h"

otherarch/otherarch.h

@@ -72,3 +72,5 @@ struct gptj_model {
 
 bool legacy_gptj_model_load(const std::string &fname, gptj_model &model, gpt_vocab &vocab);
 bool legacy_gptj_eval(const gptj_model &model, const int n_threads, const int n_past, const std::vector<gpt_vocab::id> &embd_inp, std::vector<float> &embd_w, size_t &mem_per_token);
+bool gptj_model_load(const std::string &fname, gptj_model &model, gpt_vocab &vocab);
+bool gptj_eval(const gptj_model &model, const int n_threads, const int n_past, const std::vector<gpt_vocab::id> &embd_inp, std::vector<float> &embd_w, size_t &mem_per_token);