Fix merge conflicts

Author: pragupta

.ci/docker/ci_commit_pins/triton.txt

@@ -1,5 +1 @@
-<<<<<<< HEAD
-56765e8c1f6490e21312b46242ed78cb2dd46d35
-=======
 fccfc522864cf8bc172abe0cd58ae5581e2d44b9
->>>>>>> upstream/main

.ci/docker/requirements-ci.txt

@@ -111,15 +111,8 @@ ninja==1.11.1.3
 #Pinned versions: 1.11.1.3
 #test that import: run_test.py, test_cpp_extensions_aot.py,test_determination.py
 
-<<<<<<< HEAD
 numba==0.60.0 ; python_version == "3.9"
 numba==0.61.2 ; python_version > "3.9"
-=======
-numba==0.49.0 ; python_version < "3.9" and platform_machine != "s390x"
-numba==0.55.2 ; python_version == "3.9" and platform_machine != "s390x"
-numba==0.55.2 ; python_version == "3.10" and platform_machine != "s390x"
-numba==0.60.0 ; python_version == "3.12" and platform_machine != "s390x"
->>>>>>> upstream/main
 #Description: Just-In-Time Compiler for Numerical Functions
 #Pinned versions: 0.54.1, 0.49.0, <=0.49.1
 #test that import: test_numba_integration.py

aten/src/ATen/Context.cpp

@@ -615,34 +615,12 @@ at::ROCmFABackend Context::getROCmFAPreferredBackend() {
 
 void Context::setROCmFAPreferredBackend(at::ROCmFABackend b) {
 #ifdef USE_ROCM
-<<<<<<< HEAD
-  if(b == at::ROCmFABackend::Ck) {
-    static const bool ck_unsupported = []() {
-      static const std::vector<std::string> archs = {
-          "gfx90a",  "gfx942", "gfx950"
-      };
-      for (auto index: c10::irange(detail::getCUDAHooks().deviceCount())) {
-        if (!detail::getCUDAHooks().isGPUArch(archs, index)) {
-          TORCH_WARN_ONCE(
-            "Attempting to use CK on an unsupported architecture! Cannot set backend to CK");
-          return true;
-        }
-      }
-      return false;
-    }();
-    if(!ck_unsupported) rocm_fa_preferred_backend = b;
-  }
-  else {
-     rocm_fa_preferred_backend = b;
-  }
-=======
   static const bool hasCKSDPAFlag = hasCKSDPA();
   static const bool ckSupportedFlag = ckSupported();
   TORCH_CHECK((b != at::ROCmFABackend::Ck) || (hasCKSDPAFlag && ckSupportedFlag),
       "Cannot set preferred SDPA backend to CK since following conditions are not true: ",
       "architecture supported for CK: ", ckSupportedFlag,
       ", PyTorch built with CK SDPA support: ", hasCKSDPAFlag);
->>>>>>> upstream/main
 #endif
   rocm_fa_preferred_backend = b;
 }

aten/src/ATen/cuda/detail/OffsetCalculator.cuh

@@ -49,20 +49,12 @@ struct OffsetCalculator {
 #if defined(USE_ROCM)
     if ((dims > 0) && (dims <= 2)) {
       auto divmod = sizes_[0].divmod(linear_idx);
-<<<<<<< HEAD
-      #pragma unroll
-=======
 #pragma unroll
->>>>>>> upstream/main
       for (int arg = 0; arg < NARGS; arg++)
         offsets[arg] = divmod.mod * strides_[0][arg];
       if (dims >= 2) {
         divmod = sizes_[1].divmod(divmod.div);
-<<<<<<< HEAD
-        #pragma unroll
-=======
 #pragma unroll
->>>>>>> upstream/main
         for (int arg = 0; arg < NARGS; arg++)
           offsets[arg] += divmod.mod * strides_[1][arg];
       }

aten/src/ATen/cuda/tunable/GemmHipblaslt.h

@@ -91,10 +91,6 @@ constexpr hipDataType HipDataTypeFor<c10::Float4_e2m1fn_x2>() {
 #if ROCM_VERSION >= 70000
   return HIP_R_4F_E2M1;
 #else
-<<<<<<< HEAD
-  // Return HIP_R_4F_E2M1 enum value for earlier ROCm version.
-=======
->>>>>>> upstream/main
   return static_cast<hipDataType>(33);
 #endif
 }

aten/src/ATen/native/ConvUtils.h

@@ -362,18 +362,12 @@ inline at::MemoryFormat miopen_conv_suggest_memory_format(const at::Tensor& inpu
   }
 
   // TODO: Remove PYTORCH_MIOPEN_SUGGEST_NHWC once ROCm officially supports NHWC in MIOpen
-<<<<<<< HEAD
-  // See #64427
-  static std::optional<bool> PYTORCH_MIOPEN_SUGGEST_NHWC = c10::utils::check_env("PYTORCH_MIOPEN_SUGGEST_NHWC");
-  static bool suggest_nhwc = PYTORCH_MIOPEN_SUGGEST_NHWC && *PYTORCH_MIOPEN_SUGGEST_NHWC;
-=======
   // See https://github.com/pytorch/pytorch/issues/64427.
   // non static variable is used to be able to change environment variable in runtime for testing
   // enabled by default for ROCm >= 7.0.0 with miopen 3.5
   int miopen_version = detail::getCUDAHooks().compiledWithMIOpen() ? detail::getCUDAHooks().versionMIOpen() : 0;
   bool is_miopen_3_5 = miopen_version >= 30500;  // ROCm 7.0
   bool suggest_nhwc = c10::utils::check_env("PYTORCH_MIOPEN_SUGGEST_NHWC").value_or(is_miopen_3_5);
->>>>>>> upstream/main
 
   auto input_memory_format = input.suggest_memory_format();
   auto weight_memory_format = weight.suggest_memory_format();
@@ -383,23 +377,17 @@ inline at::MemoryFormat miopen_conv_suggest_memory_format(const at::Tensor& inpu
     (input_memory_format  == at::MemoryFormat::ChannelsLast) ||
     (weight_memory_format == at::MemoryFormat::ChannelsLast)
   );
-<<<<<<< HEAD
-=======
   if (can_use_miopen_channels_last_2d) {
     return at::MemoryFormat::ChannelsLast;
   }
->>>>>>> upstream/main
 
   bool can_use_miopen_channels_last_3d = suggest_nhwc && (weight_ndim == 5) && (
     (input_memory_format  == at::MemoryFormat::ChannelsLast3d) ||
     (weight_memory_format == at::MemoryFormat::ChannelsLast3d)
   );
-<<<<<<< HEAD
-=======
   if (can_use_miopen_channels_last_3d) {
     return at::MemoryFormat::ChannelsLast3d;
   }
->>>>>>> upstream/main
 
   return at::MemoryFormat::Contiguous;
 }

aten/src/ATen/native/Convolution.cpp

@@ -1406,15 +1406,8 @@ static inline at::MemoryFormat determine_backend_memory_format(
     case ConvBackend::Miopen:
     case ConvBackend::MiopenDepthwise:
     case ConvBackend::MiopenTranspose:
-<<<<<<< HEAD
-      if (detail::getCUDAHooks().compiledWithMIOpen() && miopen_conv_use_channels_last(input, weight)) {
-        TORCH_INTERNAL_ASSERT((k == 4 || k == 5),
-            "Expected 4D or 5D input for miopen memory format selection in determine_backend_memory_format()");
-        backend_memory_format = (k == 5) ? at::MemoryFormat::ChannelsLast3d : at::MemoryFormat::ChannelsLast;
-=======
       if (detail::getCUDAHooks().compiledWithMIOpen()) {
         backend_memory_format = miopen_conv_suggest_memory_format(input, weight);
->>>>>>> upstream/main
       }
       break;
     case ConvBackend::Mkldnn:

aten/src/ATen/native/Normalization.cpp

@@ -537,20 +537,13 @@ BatchNormBackend _select_batch_norm_backend(
   }
 
   // TODO: Remove PYTORCH_MIOPEN_SUGGEST_NHWC_BATCHNORM once ROCm officially supports NHWC in MIOpen
-<<<<<<< HEAD
-  // See #64427
-  // non static variable is used to be able to change environment variable in runtime for testing
-  // enabled by default for ROCm >= 7.0.0
-  bool PYTORCH_MIOPEN_SUGGEST_NHWC_BATCHNORM = c10::utils::check_env("PYTORCH_MIOPEN_SUGGEST_NHWC_BATCHNORM").value_or(ROCM_VERSION >= 70000);
-=======
   // See https://github.com/pytorch/pytorch/issues/64427.
   // non static variable is used to be able to change environment variable in runtime for testing
   // enabled by default for ROCm >= 7.0.0 with miopen 3.5
   int miopen_version = detail::getCUDAHooks().compiledWithMIOpen() ? detail::getCUDAHooks().versionMIOpen() : 0;
   bool is_miopen_3_4 = miopen_version >= 30400;  // ROCm 6.4
   bool is_miopen_3_5 = miopen_version >= 30500;  // ROCm 7.0
   bool PYTORCH_MIOPEN_SUGGEST_NHWC_BATCHNORM = c10::utils::check_env("PYTORCH_MIOPEN_SUGGEST_NHWC_BATCHNORM").value_or(is_miopen_3_5);
->>>>>>> upstream/main
 
   if (
       detail::getCUDAHooks().compiledWithMIOpen()
@@ -565,17 +558,9 @@ BatchNormBackend _select_batch_norm_backend(
       && ((running_mean.defined() && running_var.defined())
         || (!running_mean.defined() && !running_var.defined() && training))
       && (input.suggest_memory_format() == MemoryFormat::Contiguous
-<<<<<<< HEAD
-#if (defined(USE_ROCM) && ROCM_VERSION >= 60500)
-        || (input.suggest_memory_format() == MemoryFormat::ChannelsLast && PYTORCH_MIOPEN_SUGGEST_NHWC_BATCHNORM)
-        || (input.suggest_memory_format() == MemoryFormat::ChannelsLast3d && PYTORCH_MIOPEN_SUGGEST_NHWC_BATCHNORM)
-#endif
-        )
-=======
           || (is_miopen_3_5 && PYTORCH_MIOPEN_SUGGEST_NHWC_BATCHNORM &&
               (input.suggest_memory_format() == MemoryFormat::ChannelsLast
                || input.suggest_memory_format() == MemoryFormat::ChannelsLast3d)))
->>>>>>> upstream/main
   ) {
     return BatchNormBackend::Miopen;
   }

aten/src/ATen/native/cuda/Blas.cpp

@@ -1300,17 +1300,10 @@ _scaled_mm_out_cuda(const Tensor& mat1, const Tensor& mat2,
     TORCH_CHECK(ROCM_VERSION >= 70000, "Float4_e2m1fn_x2 is only supported for ROCm 7.0 and above");
   }
   if (mat1.scalar_type() == ScalarType::Float8_e5m2 || mat2.scalar_type() == ScalarType::Float8_e5m2) {
-<<<<<<< HEAD
-    TORCH_CHECK(ROCM_VERSION >= 60000, "Float8_e5m2 is only supported for ROCm 6.0 and above");
-  }
-  if (mat1.scalar_type() == ScalarType::Float8_e4m3fn || mat2.scalar_type() == ScalarType::Float8_e4m3fn) {
-    TORCH_CHECK(ROCM_VERSION >= 60000, "Float8_e4m3fn is only supported for ROCm 6.0 and above");
-=======
     TORCH_CHECK(ROCM_VERSION >= 60500, "Float8_e5m2 is only supported for ROCm 6.5 and above");
   }
   if (mat1.scalar_type() == ScalarType::Float8_e4m3fn || mat2.scalar_type() == ScalarType::Float8_e4m3fn) {
     TORCH_CHECK(ROCM_VERSION >= 60500, "Float8_e4m3fn is only supported for ROCm 6.5 and above");
->>>>>>> upstream/main
   }
 #endif
   if (bias) {