Update PyTorch pin

Signed-off-by: Whitney Tsang <[email protected]>

Author: whitneywhtsang

.github/pins/pytorch.txt

@@ -1 +1 @@
-3f1636ebef9b45e8a3cb0eb20d327ee6acb74be0
+54cc63b467f24242cf0d6538d3e1df39e553daf1

scripts/patch-pytorch.sh

@@ -38,5 +38,4 @@ echo "Applying PyTorch patches in $REPO_ROOT"
 apply_patch ./patch/flex_attn_143553.patch
 apply_patch ./patch/pytorch_fp64.patch
 apply_patch ./patch/pytorch_global_scratch.patch
-apply_patch ./patch/test_compile_subprocess.patch
 apply_patch ./patch/flex_decoding.patch

scripts/patch/flex_attn_143553.patch

@@ -1,8 +1,8 @@
 diff --git a/torch/_inductor/kernel/flex/flex_decoding.py b/torch/_inductor/kernel/flex/flex_decoding.py
-index 83c6b59cec96..e89981286ed8 100644
+index d3e6f14f44..cd2a52ac2c 100644
 --- a/torch/_inductor/kernel/flex/flex_decoding.py
 +++ b/torch/_inductor/kernel/flex/flex_decoding.py
-@@ -459,15 +459,12 @@ def create_flex_decoding_kernel(*args, **kwargs):
+@@ -241,15 +241,12 @@ def create_flex_decoding_kernel(*args, **kwargs):
              # m
              # if V.graph.sizevars.evaluate_expr(sympy.Lt(query.get_size()[-2], 0))
              # else  # Always use a BLOCK_M > 16 before Triton fix https://github.com/triton-lang/triton/pull/4061 is in pin

scripts/patch/flex_decoding.patch

@@ -1,8 +1,8 @@
 diff --git a/benchmarks/dynamo/common.py b/benchmarks/dynamo/common.py
-index 1088634ce911e8..4e70bf5bfce434 100644
+index 469ece2958..54ba0d7b3c 100644
 --- a/benchmarks/dynamo/common.py
 +++ b/benchmarks/dynamo/common.py
-@@ -1598,6 +1598,12 @@ def cast_to_fp32(model, inputs):
+@@ -1659,6 +1659,12 @@ def cast_to_fp32(model, inputs):
      return cast_to(torch.float32, model, inputs)
 
 
@@ -15,7 +15,7 @@ index 1088634ce911e8..4e70bf5bfce434 100644
  class DummyGradScaler:
      def scale(self, loss):
          return loss
-@@ -2097,10 +2103,24 @@ def record_status(accuracy_status, dynamo_start_stats):
+@@ -2162,10 +2168,24 @@ class BenchmarkRunner:
              model_fp64 = None
              inputs_fp64 = None
              try:
@@ -44,7 +44,7 @@ index 1088634ce911e8..4e70bf5bfce434 100644
                  self.init_optimizer(name, current_device, model_fp64.parameters())
                  fp64_outputs = self.run_n_iterations(
                      model_fp64, inputs_fp64, self.model_iter_fn
-@@ -2111,11 +2131,19 @@ def record_status(accuracy_status, dynamo_start_stats):
+@@ -2176,11 +2196,19 @@ class BenchmarkRunner:
                      else x,
                      fp64_outputs,
                  )

scripts/patch/pytorch_fp64.patch

@@ -1,10 +1,10 @@
 diff --git a/torch/_inductor/codegen/xpu/device_op_overrides.py b/torch/_inductor/codegen/xpu/device_op_overrides.py
-index 632cfd29f17..b08db340c12 100644
+index 99502ca2dd..5d538ec20c 100644
 --- a/torch/_inductor/codegen/xpu/device_op_overrides.py
 +++ b/torch/_inductor/codegen/xpu/device_op_overrides.py
 @@ -61,7 +61,7 @@ class XPUDeviceOpOverrides(DeviceOpOverrides):
-     def cpp_global_scratch(
-         self, idx: int, workspace: TritonScratchWorkspace
+     def cpp_scratch(
+         self, idx: int, workspace: TritonScratchWorkspace, prefix: Optional[str] = None
      ) -> Optional[tuple[list[str], str]]:
 -        return None
 +        return [f"void *global_scratch_{idx} = 0;"], f"global_scratch_{idx}"