Upload the original exports

.gitignore

@@ -1 +1,2 @@
 .DS_STORE
+/onnx-dest

Makefile

@@ -1,4 +1,3 @@
-.SHELLFLAGS := -e -c
 SHELL := /bin/bash
 
 # Configuration variables
@@ -18,24 +17,63 @@ TRANSFORMERS_PYTHON = $(TRANSFORMERS_JS_PATH)/.venv/bin/python3
 PARTS = A B C D E
 
 define progress_bar
-total=$$(echo $(1) | wc -w | tr -d ' '); \
-current=0; \
-for item in $(1); do \
-    current=$$((current + 1)); \
-    printf "\r   Progress: \033[1;32m["; \
-    for ((i=0; i<current*20/total; i++)); do printf "="; done; \
-    printf "\033[0m"; \
-    for ((i=current*20/total; i<20; i++)); do printf " "; done; \
-    printf "\033[1;32m]\033[0m $$current/$$total "; \
-    printf "\033[1;34m$$item\033[K\033[0m\n"; \
-    cmd="$(2)"; \
-    cmd=$$(echo "$$cmd" | sed "s|{}|$$item|g"); \
-    $$cmd; \
-done; \
-printf "\n"
+	printf "\r   Progress: \033[1;32m["; \
+	_done=$$(($1 * 20 / $2)); \
+	for ((i=0; i<_done; i++)); do printf "="; done; \
+	printf "\033[0m"; \
+	_left=$$((20 - _done)); \
+	for ((i=0; i<_left; i++)); do printf " "; done; \
+	printf "\033[1;32m]\033[0m $1/$2  Processing: \033[1;34m%s\033[K\033[0m\r" "$3"
 endef
 
-.PHONY: all all-in-one clean clean-large-files export fix-gpu-buffers quantize quantize-% slim
+# See https://github.com/pytorch/pytorch/issues/94280#issuecomment-2089196400
+# Original export scripts export a bunch of tensor files, so we merge into one / two files instead.
+export-merged-source-models: export-merged-source-models-first-pass export-merged-source-models-second-pass
+	@echo "✅ Exporting merged source models complete"
+
+export-merged-source-models-first-pass:
+	@echo "💾 First pass: Export all models with merged tensors..."
+	@mkdir -p $(ONNX_DEST_DIR)
+	@files=`find $(ONNX_SRC_DIR) -name "*.onnx"`; \
+	total=`echo "$$files" | wc -w | tr -d ' '`; \
+	echo "Files found (first pass): $$total"; \
+	current=0; \
+	for item in $$files; do \
+	current=$$((current + 1)); \
+	$(call progress_bar,$$current,$$total,$$item); \
+	$(NATIVE_PYTHON) -u -c "import onnx, os, sys; src='$$item'; dest_dir='$(ONNX_DEST_DIR)'; \
+		m = onnx.load(src); \
+		d = os.path.join(dest_dir, os.path.basename(src)); \
+		onnx.save_model(m, d, all_tensors_to_one_file=True, save_as_external_data=True, location=os.path.basename(d)+'.data')" || exit 1; \
+	done; \
+	echo "✅ Done first pass"
+
+export-merged-source-models-second-pass:
+	@echo "💾 Second pass: Converting large models to external data format..."
+	@files=`find $(ONNX_DEST_DIR) -name "*.onnx"`; \
+	total=`echo "$$files" | wc -w | tr -d ' '`; \
+	echo "Files found (second pass): $$total"; \
+	current=0; \
+	for item in $$files; do \
+		current=$$((current + 1)); \
+		$(call progress_bar,$$current,$$total,$$item); \
+		$(NATIVE_PYTHON) -c 'import onnx, os, sys; \
+			src = """'"$$item"'"""; \
+			total_size = os.path.getsize(src); \
+			total_size += os.path.getsize(src + ".data") if os.path.exists(src + ".data") else 0; \
+			needs_external = total_size > 2e9; \
+			onnx.save_model( \
+				onnx.load(src), \
+				src, \
+				save_as_external_data=needs_external, \
+				all_tensors_to_one_file=True, \
+				location=(os.path.basename(src) + ".data") if needs_external else None \
+			); \
+			not needs_external and os.path.exists(src + ".data") and os.remove(src + ".data") \
+			' || exit 1; \
+	done; \
+	echo "✅ Done second models"
+
 
 all-in-one: export quantize clean-large-files slim fix-gpu-buffers
 	@echo "✨ All done! ONNX models exported, slimmed, quantized and fixed"
@@ -46,12 +84,12 @@ export: export-abcd export-e
 export-abcd:
 	@echo "🚀 Exporting parts A, B, C, D..."
 	cd ../Native-LLM-for-Android/Export_ONNX/QwenVL && \
-	../../.venv/bin/python3 QwenVL_Export_ABCD.py "Qwen/Qwen2-VL-2B-Instruct"
+	$(NATIVE_PYTHON) QwenVL_Export_ABCD.py "Qwen/Qwen2-VL-2B-Instruct"
 
 export-e:
 	@echo "🚀 Exporting part E..."
 	cd ../Native-LLM-for-Android/Export_ONNX/QwenVL && \
-	../../.venv/bin/python3 QwenVL_Export_E.py "Qwen/Qwen2-VL-2B-Instruct"
+	$(NATIVE_PYTHON) QwenVL_Export_E.py "Qwen/Qwen2-VL-2B-Instruct"
 
 slim:
 	@echo "🗜️  Slimming ONNX models..."

onnx/QwenVL_A.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7668776b6d8a7dbbd5344de5948f9e7040cce04ac4fafff9155204dd2e0ef561
+size 341395

onnx/QwenVL_A.onnx.data

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a1bdde323eb76c15f6eab14966d5b802c51a8d9559d5260ad3cf9e868ef160bf
+size 5322682368

onnx/QwenVL_B.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7b2a741d6586465346e5c552c1d375da0b8321dd76a4d5498c0dd267ccd523b6
+size 233983352

onnx/QwenVL_C.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a239bb5f47b6589f4db8d9a3b57ada13cabee3508851769d473f3bd2338da732
+size 6384

onnx/QwenVL_D.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1d70b7429fc137486f82683d68953dd8a60d72466071fd22104bf5ff77e4460e
+size 25215

onnx/QwenVL_E.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0b602930000f109874f028142d62fc488908d65e30be235565efa310d3d32c89
+size 1505816