init

embeds/SmolVLMVisionEmbeddings.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cab5d95b958d2cd6173c887bd98cab7accda22e8d321f7c348464e428b59f675
+size 2758065

infer_axmodel.py

@@ -12,10 +12,10 @@ from ml_dtypes import bfloat16
 
 
 device = "cuda" if torch.cuda.is_available() else "cpu"
-embeddings = torch.load("SmolVLMVisionEmbeddings.pkl", map_location=device, weights_only=False)
-embeds = np.load(os.path.join("./SmolVLM2-500M-Video-Instruct_1024_AXMODEL", "model.embed_tokens.weight.npy"))
+embeddings = torch.load("./embeds/SmolVLMVisionEmbeddings.pkl", map_location=device, weights_only=False)
+embeds = np.load(os.path.join("./smolvlm2_axmodel", "model.embed_tokens.weight.npy"))
 # connector = torch.load("SmolVLMConnector.pkl", map_location=device, weights_only=False)
-encoder = ort.InferenceSession(f'./export_onnx_model/vision_model.onnx', providers=["CPUExecutionProvider"])
+encoder = ort.InferenceSession(f'./vit_mdoel/vision_model.onnx', providers=["CPUExecutionProvider"])
 
 
 def run_vision_model(
@@ -164,8 +164,8 @@ def post_process(data, topk=1, topp=0.9, temperature=0.6):
 
 if __name__ == "__main__":
 
-    hf_model_path = "./SmolVLM2-500M-Video-Instruct/"
-    axmodel_path = "./SmolVLM2-500M-Video-Instruct_1024_AXMODEL"
+    hf_model_path = "./smolvlm2_tokenizer/"
+    axmodel_path = "./smolvlm2_axmodel"
     prompt = 'Can you describe this image?'
 
     processor = AutoProcessor.from_pretrained(hf_model_path)
@@ -176,7 +176,7 @@ if __name__ == "__main__":
         {
             "role": "user",
             "content": [
-                {"type": "image", "url": "./bee.jpg"},
+                {"type": "image", "url": "./assets/bee.jpg"},
                 {"type": "text", "text": prompt},
             ]
         },