Update torch pin nightly to 20260310

.ci/docker/ci_commit_pins/pytorch.txt

@@ -1 +1 @@
-659af3c353e49b35c191cdd2dba3b3c79d0e6822
+a79095985ccd3a645bd2fc31c5ae4fcf215ef518

.ci/scripts/test_model_e2e.sh

@@ -248,7 +248,7 @@ if [ "$AUDIO_URL" != "" ]; then
 elif [[ "$MODEL_NAME" == *whisper* ]] || [ "$MODEL_NAME" = "voxtral_realtime" ]; then
   conda install -y -c conda-forge "ffmpeg<8"
   pip install datasets soundfile
-  pip install torchcodec==0.11.0.dev20260217 --extra-index-url https://download.pytorch.org/whl/nightly/cpu
+  pip install torchcodec==0.11.0.dev20260310 --extra-index-url https://download.pytorch.org/whl/nightly/cpu
   python -c "from datasets import load_dataset;import soundfile as sf;sample = load_dataset('distil-whisper/librispeech_long', 'clean', split='validation')[0]['audio'];sf.write('${MODEL_DIR}/$AUDIO_FILE', sample['array'][:sample['sampling_rate']*30], sample['sampling_rate'])"
 fi
 

examples/models/moshi/mimi/install_requirements.sh

@@ -8,7 +8,7 @@
 set -x
 
 sudo apt install ffmpeg -y
-pip install torchcodec==0.11.0.dev20260217 --extra-index-url https://download.pytorch.org/whl/nightly/cpu
+pip install torchcodec==0.11.0.dev20260310 --extra-index-url https://download.pytorch.org/whl/nightly/cpu
 pip install moshi==0.2.11
 pip install bitsandbytes soundfile einops
 # Run llama2/install requirements for torchao deps

runtime/core/portable_type/c10/c10/util/complex_math.h

@@ -86,6 +86,41 @@ C10_HOST_DEVICE inline c10::complex<T> pow(
 #endif
 }
 
+// Regression in ROCm 7.2. See https://github.com/ROCm/rocm-libraries/pull/3836.
+// Specialized version for complex<float> on AMD GPUs to use FMA-based
+// multiplication
+#if defined(__HIPCC__)
+namespace detail {
+// FMA-aware complex multiplication for float precision on AMD GPUs.
+// This prevents SLP vectorizer from breaking FMA formation, which causes
+// numerical precision loss in complex arithmetic.
+// The issue occurs when vectorizer packs scalar multiplies before backend
+// can form FMA instructions, resulting in double rounding instead of single.
+C10_HOST_DEVICE inline thrust::complex<float> complex_mul_fma(
+    thrust::complex<float> a,
+    thrust::complex<float> b) {
+  // Complex multiplication: (a.r + a.i*i) * (b.r + b.i*i)
+  // = (a.r*b.r - a.i*b.i) + (a.r*b.i + a.i*b.r)*i
+  // Using __builtin_fmaf ensures FMA at source level:
+  // real: a.r*b.r + (-(a.i*b.i)) = FMA(a.r, b.r, -(a.i*b.i))
+  // imag: a.i*b.r + a.r*b.i = FMA(a.r, b.i, a.i*b.r)
+  float real_part = __builtin_fmaf(a.real(), b.real(), -(a.imag() * b.imag()));
+  float imag_part = __builtin_fmaf(a.real(), b.imag(), a.imag() * b.real());
+  return thrust::complex<float>(real_part, imag_part);
+}
+} // namespace detail
+
+template <>
+C10_HOST_DEVICE inline c10::complex<float> pow(
+    const c10::complex<float>& x,
+    const c10::complex<float>& y) {
+  auto log_x = thrust::log(static_cast<thrust::complex<float>>(x));
+  auto y_log_x =
+      detail::complex_mul_fma(static_cast<thrust::complex<float>>(y), log_x);
+  return static_cast<c10::complex<float>>(thrust::exp(y_log_x));
+}
+#endif
+
 template <typename T>
 C10_HOST_DEVICE inline c10::complex<T> pow(
     const c10::complex<T>& x,

runtime/core/portable_type/c10/torch/headeronly/macros/Macros.h

@@ -629,7 +629,7 @@ __host__ __device__
 // This macro is used to find older C++ compilers
 // that don't support move optimization for return values.
 
-#if (defined(__GNUC__) && __GNUC__ < 13) || \
+#if (defined(__GNUC__) && __GNUC__ < 13 && __cplusplus < 202002L) || \
     (defined(__clang_major__) && __clang_major__ < 13)
 #define C10_RETURN_MOVE_IF_OLD_COMPILER 1
 #else

torch_pin.py

@@ -1,2 +1,2 @@
-TORCH_VERSION = "2.11.0"
-NIGHTLY_VERSION = "dev20260215"
+TORCH_VERSION = "2.12.0"
+NIGHTLY_VERSION = "dev20260310"