llama : add Qwen2VL support + multimodal RoPE (ggerganov#10361)

* Barebone Qwen2VL LLM convertor

* Add Qwen2VL cli entrypoint

* [WIP] add qwen2vl arch

* Verify m-rope output

* Add vl-rope/2d-rope support for qwen2vl ViT

* update qwen2vl cli tool

* update 5D tensor op workaround

* [WIP] qwen2vl vision model

* make batch and clip utils compatible with qwen2vl

* [WIP] create inference workflow, gguf convert script but fix

* correcting vision-rope behavior, add the missing last layer back to ViT

* add arg parser to qwen2vl_surgery

* replace variable size array with vector

* cuda-gdb cmake preset

* add fp32 mrope, vision rope kernel

* add fp16 support for qwen2vl and m-rope

* add `GGML_ROPE_TYPE_MROPE`, `GGML_ROPE_TYPE_VISION`

* fix rope op mode switching, out dated func args

* update `llama_hparams`

* update to keep up stream changes

* resolve linter, test errors

* add makefile entry, update speical image padding token

* add mrope unit test, fix few compiler warnings

* rename `mrope` related function, params

* minor updates on debug util, bug fixs

* add `m-rope` testcase to `test-backend-ops`

* Apply suggestions from code review

Co-authored-by: Georgi Gerganov <[email protected]>

* fix traililng whitespce

* store `llama_hparams.rope_sections` with fixed size array

* update position id tensor size check in GGML_OP_ROPE

* minor updates

* update `ggml_backend_*_supports_op` of unsupported backends

* remote old `rope_section` compare operator

---------

Co-authored-by: Georgi Gerganov <[email protected]>

Makefile

@@ -22,6 +22,7 @@ BUILD_TARGETS = \
 	llama-infill \
 	llama-llava-cli \
 	llama-minicpmv-cli\
+	llama-qwen2vl-cli\
 	llama-lookahead \
 	llama-lookup \
 	llama-lookup-create \
@@ -1404,6 +1405,14 @@ llama-minicpmv-cli: examples/llava/minicpmv-cli.cpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) $< $(filter-out %.h $<,$^) -o $@ $(LDFLAGS) -Wno-cast-qual
 
+llama-qwen2vl-cli: examples/llava/qwen2vl-cli.cpp \
+	examples/llava/llava.cpp \
+	examples/llava/llava.h \
+	examples/llava/clip.cpp \
+	examples/llava/clip.h \
+	$(OBJ_ALL)
+	$(CXX) $(CXXFLAGS) $< $(filter-out %.h $<,$^) -o $@ $(LDFLAGS) -Wno-cast-qual
+
 ifeq ($(UNAME_S),Darwin)
 swift: examples/batched.swift
 	(cd examples/batched.swift; make build)

README.md

@@ -110,6 +110,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 - [x] [Mini CPM](https://huggingface.co/models?search=MiniCPM)
 - [x] [Moondream](https://huggingface.co/vikhyatk/moondream2)
 - [x] [Bunny](https://github.com/BAAI-DCAI/Bunny)
+- [x] [Qwen2-VL](https://huggingface.co/collections/Qwen/qwen2-vl-66cee7455501d7126940800d)
 
 </details>
 

convert_hf_to_gguf.py

@@ -2001,6 +2001,29 @@ def set_gguf_parameters(self):
                 self.gguf_writer.add_rope_scaling_orig_ctx_len(self.hparams["rope_scaling"]["original_max_position_embeddings"])
 
 
+@Model.register("Qwen2VLForConditionalGeneration")
+class Qwen2VLModel(Model):
+    model_arch = gguf.MODEL_ARCH.QWEN2VL
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        mrope_section = self.hparams["rope_scaling"]["mrope_section"]
+        mrope_section += [0] * max(0, 4 - len(mrope_section))
+        self.gguf_writer.add_rope_dimension_sections(mrope_section)
+
+    def set_vocab(self):
+        try:
+            self._set_vocab_sentencepiece()
+        except FileNotFoundError:
+            self._set_vocab_gpt2()
+
+    def get_tensors(self) -> Iterator[tuple[str, Tensor]]:
+        for name, data in super().get_tensors():
+            if name.startswith("visual."):
+                continue
+            yield name, data
+
+
 @Model.register("Qwen2MoeForCausalLM")
 class Qwen2MoeModel(Model):
     model_arch = gguf.MODEL_ARCH.QWEN2MOE

examples/llava/CMakeLists.txt

@@ -43,3 +43,10 @@ set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME llama-minicpmv-cli)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_17)
+
+set(TARGET llama-qwen2vl-cli)
+add_executable(${TARGET} qwen2vl-cli.cpp)
+set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME llama-qwen2vl-cli)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_17)