[Core][Model] Terratorch backend integration (#23513)

Signed-off-by: Michele Gazzetti <michele.gazzetti1@ibm.com>
Signed-off-by: Christian Pinto <christian.pinto@ibm.com>
Co-authored-by: Christian Pinto <christian.pinto@ibm.com>
Co-authored-by: Cyrus Leung <tlleungac@connect.ust.hk>

examples/offline_inference/prithvi_geospatial_mae.py

@@ -45,7 +45,11 @@ datamodule_config = {
 class PrithviMAE:
     def __init__(self, model):
         self.model = LLM(
-            model=model, skip_tokenizer_init=True, dtype="float16", enforce_eager=True
+            model=model,
+            skip_tokenizer_init=True,
+            dtype="float16",
+            enforce_eager=True,
+            model_impl="terratorch",
         )
 
     def run(self, input_data, location_coords):

examples/offline_inference/prithvi_geospatial_mae_io_processor.py

@@ -37,6 +37,7 @@ def main():
         # The maximum number depends on the available GPU memory
         max_num_seqs=32,
         io_processor_plugin="prithvi_to_tiff_india",
+        model_impl="terratorch",
     )
 
     pooling_params = PoolingParams(task="encode", softmax=False)

examples/online_serving/prithvi_geospatial_mae.py

@@ -15,6 +15,7 @@ import requests
 #   https://github.com/christian-pinto/prithvi_io_processor_plugin
 # - start vllm in serving mode with the below args
 #   --model='christian-pinto/Prithvi-EO-2.0-300M-TL-VLLM'
+#   --model-impl terratorch
 #   --task embed --trust-remote-code
 #   --skip-tokenizer-init --enforce-eager
 #   --io-processor-plugin prithvi_to_tiff_india