Model ops add yaml qwen25

utils/model_ops/README.md

@@ -26,10 +26,11 @@ uv sync --group models-ops
 
 The driver scripts require an NVIDIA GPU, so install a CUDA-enabled build of
 PyTorch separately. The exact index URL depends on your CUDA version (replace
-`cu128` with the build that matches your driver, e.g. `cu121`, `cu124`):
+`cu130` with the build that matches your driver, e.g. `cu121`, `cu124`):
 
 ```
-uv pip install --upgrade torch --index-url https://download.pytorch.org/whl/cu128
+uv pip install --upgrade --force-reinstall "torch==2.13.0+cu130" --index-url https://download.pytorch.org/whl/cu130
+uv pip install mistral_common[opencv]
 ```
 
 Then change directory into `utils/models_ops/` to run the drivers (the absolute

utils/model_ops/models/ministral3_14b/run_huggingface.py

@@ -45,7 +45,7 @@ def main():
     model = Mistral3ForConditionalGeneration.from_pretrained(
         model_path,
         device_map=device,
-        torch_dtype=torch.float16,
+        torch_dtype=torch.bfloat16,
         trust_remote_code=False,
     )
     tokenizer = MistralCommonBackend.from_pretrained(model_path)

utils/model_ops/models/qwen25/run_huggingface.py

@@ -0,0 +1,78 @@
+# Copyright 2026 The Torch-Spyre Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import os
+
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer, StaticCache
+from utils.torchop_yaml import TorchOpCollector, require_cuda, setup_logging
+
+
+def main():
+    setup_logging()
+    require_cuda()
+
+    model_path = "Qwen/Qwen2.5-7B-Instruct"
+    prompt = "Give me a short introduction to large language model."
+    messages = [
+        {
+            "role": "system",
+            "content": "You are Qwen, created by Alibaba Cloud. You are a helpful assistant.",
+        },
+        {"role": "user", "content": prompt},
+    ]
+
+    device = "cuda"
+    model = AutoModelForCausalLM.from_pretrained(
+        model_path, device_map="auto", dtype=torch.bfloat16
+    ).to(device)
+    tokenizer = AutoTokenizer.from_pretrained(model_path)
+    text = tokenizer.apply_chat_template(
+        messages, tokenize=False, add_generation_prompt=True
+    )
+    encoded_input = tokenizer(text, return_tensors="pt").to(device)
+
+    past_key_values = StaticCache(config=model.config, max_cache_len=2048)
+
+    torch.backends.cuda.enable_flash_sdp(False)
+    torch.backends.cuda.enable_mem_efficient_sdp(False)
+    torch.backends.cuda.enable_math_sdp(True)
+
+    model.forward = torch.compile(model.forward)
+
+    with TorchOpCollector() as ctx:
+        with torch.no_grad():
+            model.generate(
+                **encoded_input,
+                past_key_values=past_key_values,
+                use_cache=True,
+                max_new_tokens=16,
+            )
+
+    # print traced torch op
+    for op in ctx.ops_list:
+        print(op)
+    print(f"Total ops traced: {len(ctx.ops_list)}")
+
+    # List of ops with generated test cases
+    print("List of ops with test cases generated")
+    for op in ctx.test_gen_ops:
+        print(op, ctx.test_case_count[op])
+    print(f"Total ops with test configs generated: {len(ctx.test_gen_ops)}")
+
+    ctx.write_yaml(os.path.basename(model_path))
+
+
+if __name__ == "__main__":
+    main()