Merge branch 'main' into multimodal-eval-2

Author: Jack-Khuu

install/install_requirements.sh

@@ -81,7 +81,7 @@ then
   REQUIREMENTS_TO_INSTALL=(
     torch=="2.7.0.${PYTORCH_NIGHTLY_VERSION}"
     torchvision=="0.22.0.${VISION_NIGHTLY_VERSION}"
-    torchtune=="0.6.0"
+    #torchtune=="0.6.0" # no 0.6.0 on xpu nightly
   )
 else
   REQUIREMENTS_TO_INSTALL=(
@@ -115,11 +115,22 @@ fi
     "${REQUIREMENTS_TO_INSTALL[@]}"
 )
 
+# Temporatory instal torchtune nightly from cpu nightly link since no torchtune nightly for xpu now
+# TODO: Change to install torchtune from xpu nightly link, once torchtune xpu nightly is ready
+if [[ -x "$(command -v xpu-smi)" ]];
+then
+(
+  set -x
+  $PIP_EXECUTABLE install --extra-index-url  "https://download.pytorch.org/whl/nightly/cpu" \
+    torchtune=="0.6.0.${TUNE_NIGHTLY_VERSION}"
+)
+fi
+
 # For torchao need to install from github since nightly build doesn't have macos build.
 # TODO: Remove this and install nightly build, once it supports macos
 (
   set -x
-  $PIP_EXECUTABLE install git+https://github.com/pytorch/ao.git@2f97b0955953fa1a46594a27f0df2bc48d93e79d
+  $PIP_EXECUTABLE install git+https://github.com/pytorch/ao.git@7d8794622f3ac7ffa98761314019a20fba06edef
 )
 
 if [[ -x "$(command -v nvidia-smi)" ]]; then

torchchat/cli/builder.py

@@ -38,7 +38,7 @@
 from torchchat.utils.build_utils import (
     device_sync,
     is_cpu_device,
-    is_cuda_or_cpu_device,
+    is_cuda_or_cpu_or_xpu_device,
     name_to_dtype,
 )
 from torchchat.utils.measure_time import measure_time
@@ -581,7 +581,7 @@ def _initialize_model(
         _set_gguf_kwargs(builder_args, is_et=is_pte, context="generate")
 
     if builder_args.dso_path:
-        if not is_cuda_or_cpu_device(builder_args.device):
+        if not is_cuda_or_cpu_or_xpu_device(builder_args.device):
             print(
                 f"Cannot load specified DSO to {builder_args.device}. Attempting to load model to CPU instead"
             )
@@ -616,7 +616,7 @@ def do_nothing(max_batch_size, max_seq_length):
             raise RuntimeError(f"Failed to load AOTI compiled {builder_args.dso_path}")
 
     elif builder_args.aoti_package_path:
-        if not is_cuda_or_cpu_device(builder_args.device):
+        if not is_cuda_or_cpu_or_xpu_device(builder_args.device):
             print(
                 f"Cannot load specified PT2 to {builder_args.device}. Attempting to load model to CPU instead"
             )

torchchat/utils/build_utils.py

@@ -303,6 +303,5 @@ def get_device(device) -> str:
 def is_cpu_device(device) -> bool:
     return device == "" or str(device) == "cpu"
 
-
-def is_cuda_or_cpu_device(device) -> bool:
-    return is_cpu_device(device) or ("cuda" in str(device))
+def is_cuda_or_cpu_or_xpu_device(device) -> bool:
+    return is_cpu_device(device) or ("cuda" in str(device)) or ("xpu" in str(device))

torchchat/utils/quantize.py

@@ -747,7 +747,7 @@ def et_forward(self, indices: torch.Tensor) -> torch.Tensor:
             )
         else:
             return torch.ops.quantized_decomposed.embedding_4bit.dtype(
-                self.weight, self.scales, None, 0, 0, indices, dtype=self.dtype
+                self.weight, self.scales, None, -8, 7, indices, dtype=self.dtype
             )
 
     @torch.no_grad()