Voice et bot modif

venv/lib/python3.12/site-packages/onnxruntime/transformers/models/llama/benchmark.py

@@ -584,7 +584,7 @@ def get_args(rank=0):
         "--device",
         type=str,
         default="cuda" if torch.cuda.is_available() else "cpu",
-        choices=["cpu", "cuda", "rocm"],
+        choices=["cpu", "cuda"],
     )
     parser.add_argument("-id", "--device-id", type=int, default=0)
     parser.add_argument("-w", "--warmup-runs", type=int, default=5)
@@ -622,9 +622,6 @@ def get_args(rank=0):
         setattr(args, "execution_provider", f"{args.device.upper()}ExecutionProvider")  # noqa: B010
         if args.execution_provider == "CUDAExecutionProvider":
             args.execution_provider = (args.execution_provider, {"device_id": rank})
-        elif args.execution_provider == "ROCMExecutionProvider":
-            args.execution_provider = (args.execution_provider, {"device_id": rank})
-            args.device = "cuda"
 
     # Check that paths have been specified for any benchmarking with ORT
     if args.benchmark_type == "hf-ort":

venv/lib/python3.12/site-packages/onnxruntime/transformers/models/llama/benchmark_all.py

@@ -109,7 +109,7 @@ def get_args():
         "--device",
         type=str,
         required=True,
-        choices=["cpu", "cuda", "rocm"],
+        choices=["cpu", "cuda"],
         help="Device to benchmark models",
     )
 

venv/lib/python3.12/site-packages/onnxruntime/transformers/models/llama/convert_to_onnx.py

@@ -12,6 +12,7 @@ import shutil
 import subprocess
 import sys
 import tempfile
+import warnings
 from itertools import chain
 
 import onnx
@@ -234,6 +235,7 @@ def run_torchscript_separate_export(
         opset_version=torch_export_onnx_opset_version,
         do_constant_folding=True,
         verbose=args.verbose,
+        dynamo=False,
     )
 
     # Check decoder_model.onnx and save all external data to one file
@@ -293,6 +295,7 @@ def run_torchscript_separate_export(
         opset_version=torch_export_onnx_opset_version,
         do_constant_folding=True,
         verbose=args.verbose,
+        dynamo=False,
     )
 
     # Check decoder_with_past_model.onnx and save all external data to one file
@@ -631,7 +634,7 @@ def get_args():
         "--execution_provider",
         required=False,
         default="cpu",
-        choices=["cpu", "cuda", "rocm"],
+        choices=["cpu", "cuda"],
         help="Execution provider to verify parity with",
     )
 
@@ -670,6 +673,8 @@ def get_args():
 
     blockwise_group = parser.add_argument_group("blockwise (4-bit quantization)")
 
+    parser.add_argument("--bits", default=4, type=int, help="the target bits to represent weight")
+
     blockwise_group.add_argument(
         "--block_size",
         required=False,
@@ -799,6 +804,12 @@ def get_args():
 
 
 def main():
+    warnings.warn(
+        "This example is deprecated. Use the Olive recipe instead: "
+        "https://github.com/microsoft/olive-recipes/tree/main",
+        DeprecationWarning,
+        stacklevel=2,
+    )
     if version.parse(torch.__version__) < version.parse("2.2.0"):
         logger.error(f"Detected PyTorch version {torch.__version__}. Please upgrade and use v2.2.0 or newer.")
         return
@@ -988,6 +999,7 @@ def main():
                         model = onnx.load_model(fp_path, load_external_data=True)
                         quant = MatMulNBitsQuantizer(
                             model=model,
+                            bits=args.bits,
                             block_size=args.block_size,
                             is_symmetric=True,
                             accuracy_level=args.int4_accuracy_level,

venv/lib/python3.12/site-packages/onnxruntime/transformers/models/llama/llama_parity.py

@@ -228,7 +228,7 @@ def get_args(argv: list[str]):
         "--execution_provider",
         required=False,
         default="cpu",
-        choices=["cpu", "cuda", "rocm"],
+        choices=["cpu", "cuda"],
         help="Execution provider to verify parity with",
     )