Bump both the ET (9-11) and PT(9-13) nightlies

install/.pins/et-pin.txt

@@ -1 +1 @@
-91298923a0076c1b41059efb6dad2876426e4b03
+6328d41eb87ee2fe8a8c0c5532e0b07ec4e90c73

install/install_requirements.sh

@@ -47,10 +47,11 @@ fi
 # NOTE: If a newly-fetched version of the executorch repo changes the value of
 # PYTORCH_NIGHTLY_VERSION, you should re-run this script to install the necessary
 # package versions.
-PYTORCH_NIGHTLY_VERSION=dev20240814
+
+PYTORCH_NIGHTLY_VERSION=dev20240913
 
 # Nightly version for torchvision
-VISION_NIGHTLY_VERSION=dev20240814
+VISION_NIGHTLY_VERSION=dev20240913
 
 # Nightly version for torchtune
 TUNE_NIGHTLY_VERSION=dev20240910
@@ -74,7 +75,7 @@ fi
 
 # pip packages needed by exir.
 REQUIREMENTS_TO_INSTALL=(
-  torch=="2.5.0.${PYTORCH_NIGHTLY_VERSION}"
+  torch=="2.6.0.${PYTORCH_NIGHTLY_VERSION}"
   torchvision=="0.20.0.${VISION_NIGHTLY_VERSION}"
   torchtune=="0.3.0.${TUNE_NIGHTLY_VERSION}"
 )

torchchat/export.py

@@ -194,7 +194,7 @@ def forward(self, x, freqs_cis, mask, input_pos=None):
             return self.wo(output)
 
     def replace_attention_with_custom_sdpa_attention(module: nn.Module):
-        from executorch.examples.models.llama2.custom_ops import (  # noqa
+        from executorch.extension.llm.custom_ops import (  # noqa
             sdpa_with_kv_cache,
         )
 
@@ -304,7 +304,6 @@ def export_for_et(model, device, output_path) -> str:
         edge_manager = edge_manager.to_backend(XnnpackDynamicallyQuantizedPartitioner())
         export_program = edge_manager.to_executorch(
             ExecutorchBackendConfig(
-                extract_constant_segment=True,
                 extract_delegate_segments=True,
                 passes=[
                     QuantFusionPass(),

torchchat/model.py

@@ -744,7 +744,7 @@ def apply_rotary_emb(x: Tensor, freqs_cis: Tensor) -> Tensor:
     from executorch.extension.pybindings import portable_lib as exec_lib
 
     # ET changed the way it's loading the custom ops so it's not included in portable_lib but has to be loaded separately.
-    from executorch.examples.models.llama2.custom_ops import sdpa_with_kv_cache  # no-qa
+    from executorch.extension.llm.custom_ops import sdpa_with_kv_cache  # no-qa
 
     class PTEModel(nn.Module):
         def __init__(self, config, path) -> None:

torchchat/utils/scripts/build_native.sh

@@ -53,12 +53,11 @@ if [ -z "${TORCHCHAT_ROOT}" ]; then
     # Get the absolute path of the current script
     SCRIPT_PATH="$( cd "$(dirname "$0")" >/dev/null 2>&1 ; pwd -P )"
     # Get the absolute path of the parent directory
-    TORCHCHAT_ROOT="$(dirname "$SCRIPT_PATH")"
-    source "$TORCHCHAT_ROOT/scripts/install_utils.sh"
-else
-    source "$TORCHCHAT_ROOT/torchchat/utils/scripts/install_utils.sh"
+    TORCHCHAT_ROOT="$(dirname "$(dirname "$(dirname "$SCRIPT_PATH")")")"
 fi
 
+source "$TORCHCHAT_ROOT/torchchat/utils/scripts/install_utils.sh"
+
 if [ -z "${ET_BUILD_DIR}" ]; then
     ET_BUILD_DIR="et-build"
 fi

torchchat/utils/scripts/install_utils.sh

@@ -13,7 +13,7 @@ if [ -z "$TORCHCHAT_ROOT" ]; then
 fi
 
 install_pip_dependencies() {
-  echo "Intalling common pip packages"
+  echo "Installing common pip packages"
   pip3 install wheel "cmake>=3.19" ninja zstd
   pushd ${TORCHCHAT_ROOT}
   pip3 install -r install/requirements.txt