fix windows build issue

Author: chinazhangchao

examples/models/llama/install_requirements.bat

@@ -0,0 +1,16 @@
+rem Install snakeviz for cProfile flamegraph
+rem Install sentencepiece for llama tokenizer
+pip install snakeviz sentencepiece
+
+rem Install torchao.
+pip install "%~dp0/../../../third-party/ao"
+
+rem Install lm-eval for Model Evaluation with lm-evalution-harness
+rem Install tiktoken for tokenizer
+pip install lm_eval==0.4.5
+pip install tiktoken blobfile
+rem Restore numpy if >= 2.0
+pip install "numpy<2.0"
+
+rem Call the install helper for further setup
+python examples/models/llama/install_requirement_helper.py

extension/llm/custom_ops/sdpa_with_kv_cache.py

@@ -10,6 +10,7 @@
 
 # pyre-unsafe
 
+import os
 import logging
 from pathlib import Path
 
@@ -24,7 +25,11 @@
     op2 = torch.ops.llama.fast_hadamard_transform.default
     assert op2 is not None
 except:
-    libs = list(Path(__file__).parent.resolve().glob("libcustom_ops_aot_lib.*"))
+    lib_name_pattern = "libcustom_ops_aot_lib.*"
+    if os.name == "nt":
+        lib_name_pattern = "custom_ops_aot_lib.*"
+
+    libs = list(Path(__file__).parent.resolve().glob(lib_name_pattern))
     assert len(libs) == 1, f"Expected 1 library but got {len(libs)}"
     logging.info(f"Loading custom ops library: {libs[0]}")
     torch.ops.load_library(libs[0])

extension/llm/runner/util.h

@@ -44,7 +44,7 @@ ET_EXPERIMENTAL void inline safe_printf(const char* piece) {
 ET_EXPERIMENTAL long inline time_in_ms() {
   // return time in milliseconds, for benchmarking the model speed
   struct timespec time;
-  clock_gettime(CLOCK_REALTIME, &time);
+  timespec_get(&time, TIME_UTC);
   return time.tv_sec * 1000 + time.tv_nsec / 1000000;
 }