t5 : add example for text-to-text transfer transformer inference

examples/CMakeLists.txt

@@ -4,3 +4,4 @@ target_include_directories(ggml_utils PUBLIC ${CMAKE_CURRENT_SOURCE_DIR})
 add_subdirectory(gpt-2)
 add_subdirectory(gpt-j)
 add_subdirectory(whisper)
+add_subdirectory(t5)

examples/t5/CMakeLists.txt

@@ -0,0 +1,6 @@
+#
+# t5
+
+set(TEST_TARGET t5)
+add_executable(${TEST_TARGET} main.cpp)
+target_link_libraries(${TEST_TARGET} PRIVATE ggml ggml_utils)

examples/t5/README.md

@@ -0,0 +1,3 @@
+# t5
+
+ref: https://github.com/huggingface/transformers/blob/main/src/transformers/models/t5/modeling_t5.py

examples/t5/convert-flan-t5-pt-to-ggml.py

@@ -0,0 +1,25 @@
+import io
+import sys
+import torch
+
+import code
+
+from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
+
+if len(sys.argv) < 3:
+ print("Usage: convert-flan-t5-pt-to-ggml.py path-to-pt-model dir-output [use-f32]\n")
+ sys.exit(1)
+
+fname_inp=sys.argv[1] + "/pytorch_model.bin"
+
+try:
+ model_bytes = open(fname_inp, "rb").read()
+ with io.BytesIO(model_bytes) as fp:
+ checkpoint = torch.load(fp, map_location="cpu")
+except:
+ print("Error: failed to load PyTorch model file: %s" % fname_inp)
+ sys.exit(1)
+
+# list all keys
+for k in checkpoint.keys():
+ print(k)

examples/t5/main.cpp

@@ -0,0 +1,3 @@
+int main() {
+ return 0;
+}