@@ -14,7 +14,6 @@ INSTALL_DIR="${WINDOWS_CUDA_INSTALL_DIR:-/opt/cuda-windows}"
 # Mapping of CUDA versions to their corresponding driver versions for Windows installers
 # Source: https://developer.nvidia.com/cuda-toolkit-archive
 declare -A CUDA_DRIVER_MAP=(
-    ["12.6"]="12.6.3:561.17"
     ["12.8"]="12.8.1:572.61"
     ["12.9"]="12.9.1:576.57"
     ["13.0"]="13.0.2:"
 
@@ -588,7 +588,7 @@ def to_v3_format(
             workflow_run_id: GitHub workflow run ID
             workflow_run_url: GitHub workflow run URL
             gpu_name: GPU device name (e.g., "Tesla V100", "A100")
-            cuda_driver_version: CUDA driver version (e.g., "12.6", "535.104.05")
+            cuda_driver_version: CUDA driver version (e.g., "13.0", "535.104.05")
 
         Returns:
             List of benchmark records in v3 format
@@ -873,7 +873,7 @@ def main():
         "--cuda_driver_version",
         type=str,
         default=None,
-        help="CUDA driver version (e.g., '12.6', '535.104.05') - optional for v3 format",
+        help="CUDA driver version (e.g., '13.0', '535.104.05') - optional for v3 format",
     )
     parser.add_argument("--verbose", action="store_true", help="Print verbose output")
 
 
@@ -46,13 +46,18 @@ import re
 import sys
 
 text = pathlib.Path(sys.argv[1]).read_text()
+# Under 8da4w, the base model's first token is a near-tie between ":\n" and ":"
+# that tips with small numeric changes, such as RMSNorm and QK-norm moving into
+# XNNPACK. Accept both observed continuations.
 pattern = re.compile(
-    r"^<\|im_start\|>user Calculate 15% of 80\?<\|im_end\|><\|im_start\|>assistant:\n"
-    r"(?:<think>\n)+"
+    r"^<\|im_start\|>user Calculate 15% of 80\?<\|im_end\|><\|im_start\|>assistant:"
+    r"(?:\n(?:<think>\n)+| me\n)"
     r"Okay, so I need to calculate 15% of 80\.",
     re.MULTILINE,
 )
-sys.exit(0 if pattern.match(text) else 1)
+if not pattern.match(text):
+    print(f"Expected a match for: {pattern.pattern}")
+    sys.exit(1)
 PY
 }
 
@@ -174,9 +179,6 @@ else
 fi
 
 ### QUANTIZATION & PROGRAM DATA SEPARATION ###
-EXPECTED_QUANT_PREFIX="<|im_start|>user Calculate 15% of 80?<|im_end|><|im_start|>assistant:
-<think>
-Okay, so I need to calculate 15% of 80."
 EXPECTED_QUANT_LORA_PREFIX="
 <|im_start|>user Calculate 15% of 80?<|im_end|><|im_start|>assistant
 To calculate 15% of 80, we can multiply 80 by 15/100.
@@ -238,11 +240,9 @@ NOW=$(date +"%H:%M:%S")
 echo "Finished at ${NOW}"
 RESULT=$(cat result.txt)
 if matches_base_response_prefix result.txt; then
-  echo "Expected result prefix: ${EXPECTED_QUANT_PREFIX}"
   echo "Actual result: ${RESULT}"
   echo "Test 3: Success"
 else
-  echo "Expected result prefix: ${EXPECTED_QUANT_PREFIX}"
   echo "Actual result: ${RESULT}"
   echo "Test 3: Failure; results not the same"
   cleanup_files
 
@@ -46,13 +46,18 @@ import re
 import sys
 
 text = pathlib.Path(sys.argv[1]).read_text()
+# Under 8da4w, the base model's first token is a near-tie between ":\n" and ":"
+# that tips with small numeric changes, such as RMSNorm and QK-norm moving into
+# XNNPACK. Accept both observed continuations.
 pattern = re.compile(
-    r"^<\|im_start\|>user Calculate 15% of 80\?<\|im_end\|><\|im_start\|>assistant:\n"
-    r"(?:<think>\n)+"
+    r"^<\|im_start\|>user Calculate 15% of 80\?<\|im_end\|><\|im_start\|>assistant:"
+    r"(?:\n(?:<think>\n)+| me\n)"
     r"Okay, so I need to calculate 15% of 80\.",
     re.MULTILINE,
 )
-sys.exit(0 if pattern.match(text) else 1)
+if not pattern.match(text):
+    print(f"Expected a match for: {pattern.pattern}")
+    sys.exit(1)
 PY
 }
 
@@ -107,10 +112,6 @@ EXPECTED_LORA_PREFIX="
 <|im_start|>user Calculate 15% of 80?<|im_end|><|im_start|>assistant
 To calculate 15% of 80"
 
-EXPECTED_BASE_PREFIX="<|im_start|>user Calculate 15% of 80?<|im_end|><|im_start|>assistant:
-<think>
-Okay, so I need to calculate 15% of 80."
-
 ### TEST 1: Run lora_forward method ###
 NOW=$(date +"%H:%M:%S")
 echo "Test 1: Multimethod lora_forward. Starting at ${NOW}"
@@ -151,7 +152,6 @@ if matches_base_response_prefix result_base.txt; then
   echo "Test 2 (base_forward): Success"
 else
   echo "Test 2 (base_forward): Failure"
-  echo "Expected result prefix: ${EXPECTED_BASE_PREFIX}"
   echo "Actual result: ${RESULT}"
   cleanup_files
   exit 1
 
@@ -103,7 +103,7 @@ def test_all_install_steps_preserve_exact_cu134_selection(self):
                         )
 
     def test_other_cuda_trains_keep_existing_pins(self):
-        for cuda in ((12, 6), (13, 0), (13, 2)):
+        for cuda in ((13, 0), (13, 2)):
             for machine in ("x86_64", "aarch64"):
                 with self.subTest(cuda=cuda, machine=machine):
                     core, local, domains, examples = self.install_commands(
@@ -159,7 +159,7 @@ def test_cpu_release_dependencies_are_opt_in(self):
                     )
 
     def test_cpu_release_option_keeps_cuda_pins(self):
-        for cuda in ((12, 6), (13, 0), (13, 2), (13, 4)):
+        for cuda in ((13, 0), (13, 2), (13, 4)):
             for machine in ("x86_64", "aarch64"):
                 with self.subTest(cuda=cuda, machine=machine):
                     self.assertEqual(
 
@@ -34,9 +34,7 @@ def _model_quant(entry):
 class CudaWorkflowTest(unittest.TestCase):
     def test_build_matrix_preserves_existing_cuda_versions(self):
         job = WORKFLOW["jobs"]["test-cuda-builds"]
-        self.assertEqual(
-            job["strategy"]["matrix"]["cuda-version"], ["12.6", "13.0", "13.4"]
-        )
+        self.assertEqual(job["strategy"]["matrix"]["cuda-version"], ["13.0", "13.4"])
         self.assertEqual(job["with"]["gpu-arch-version"], "${{ matrix.cuda-version }}")
 
     def test_cuda_builds_take_the_node_driver_on_an_unpinned_v3(self):
@@ -58,9 +56,7 @@ def test_cuda_runtime_update_precedes_build_and_propagates_failure(self):
 source() { printf 'BUILD %s\n' "$*"; printf 'LIBRARY PATH %s\n' "$LD_LIBRARY_PATH"; }
 """
         for version, conda_status in (
-            (version, status)
-            for version in ("12.6", "13.0", "13.4")
-            for status in (0, 1)
+            (version, status) for version in ("13.0", "13.4") for status in (0, 1)
         ):
             with self.subTest(version=version, conda_status=conda_status):
                 result = subprocess.run(
 
@@ -41,29 +41,6 @@ _cuda_arch_aarch64_cu130="8.0 9.0 10.0 11.0 12.0"
 _cuda_arch_aarch64_cu132="${_cuda_arch_aarch64_cu130}"
 _cuda_arch_aarch64_cu134="${_cuda_arch_aarch64_cu130}"
 
-# The older CUDA train.
-#
-# The two architectures do not carry identical lists, because each covers what the published PyTorch
-# build for that architecture covers, and those differ. Matching them to each other instead would mean
-# advertising a GPU on one architecture that PyTorch cannot serve there.
-#
-# The smaller embedded modules are deliberately absent, with one exception. An embedded-only
-# architecture in a generic wheel would advertise a device the row cannot otherwise serve, since
-# those devices also need the CUDA, TensorRT and PyTorch pinned by their own software release
-# rather than the ones a generic wheel resolves.
-#
-# 8.7 is that exception. This is the only row whose CUDA major matches what that module's software
-# release ships, and the wheel declares no PyTorch, so the user supplies the build that carries
-# their architecture. Omitting it does not protect them from a bad pairing, it only removes the
-# device code they need.
-#
-# The floor is 8.0 rather than the oldest architecture PyTorch still carries. One of these sources compiles
-# an integer matrix-multiply path only at 8.0 and newer, so an older architecture would get a delegate that
-# loads, runs most models, and fails on one that needs that operator. Claiming hardware the delegate only
-# partly serves is the same problem the embedded modules have, so the row leaves it out for the same reason.
-_cuda_arch_x86_64_cu126="8.0 8.6 8.9 9.0"
-_cuda_arch_aarch64_cu126="8.0 8.7 9.0"
-
 # A CUDA train with no architecture list would leave the build detecting the builder's GPU, which is
 # the failure this file exists to prevent. Adding a train to the release matrix without adding its
 # architectures should fail loudly rather than silently produce a single-GPU wheel.
@@ -103,7 +80,6 @@ executorch_cuda_arch_list() {
   case "${machine}" in
     aarch64 | arm64)
       case "${train}" in
-        126) printf '%s' "${_cuda_arch_aarch64_cu126}" ;;
         130) printf '%s' "${_cuda_arch_aarch64_cu130}" ;;
         132) printf '%s' "${_cuda_arch_aarch64_cu132}" ;;
         134) printf '%s' "${_cuda_arch_aarch64_cu134}" ;;
@@ -112,7 +88,6 @@ executorch_cuda_arch_list() {
       ;;
     x86_64)
       case "${train}" in
-        126) printf '%s' "${_cuda_arch_x86_64_cu126}" ;;
         130) printf '%s' "${_cuda_arch_x86_64_cu130}" ;;
         132) printf '%s' "${_cuda_arch_x86_64_cu132}" ;;
         134) printf '%s' "${_cuda_arch_x86_64_cu134}" ;;
 
@@ -107,6 +107,7 @@
     "aoti_torch_index_put_out",
     "aoti_torch_cuda_scatter_reduce_two_out",
     "aoti_torch_cuda_scatter_src_out",
+    "aoti_torch_cuda_masked_select",
 )
 
 _QUANTIZED_KERNEL_SYMBOLS = (
@@ -1217,7 +1218,7 @@ def _wheel_cuda_train() -> str:
     """The CUDA train the installed wheel was built for, or "" for a CPU wheel.
 
     Read from the local version segment, which is the only place the wheel states what
-    it was built for. `1.5.0+cu126` gives "126".
+    it was built for. `1.5.0+cu132` gives "132".
     """
     local = importlib.metadata.version("executorch").partition("+")[2]
     return local[2:] if local.startswith("cu") else ""
@@ -1360,7 +1361,7 @@ def _resolve_required(required):
     ),
     # The CUDA components. Required exactly when the wheel says it is a CUDA wheel,
     # which is decided at check time rather than here: a fixed False meant a wheel
-    # tagged +cu126 carrying no CUDA library at all passed every check in this file,
+    # tagged +cu132 carrying no CUDA library at all passed every check in this file,
     # while a fixed True would fail every CPU wheel. The marker is the string these
     # rows are keyed on below.
     (
@@ -1457,7 +1458,7 @@ def test_each_component_has_one_owner() -> None:
     shipped = {path.name for path in _shipped_shared_objects(_installed_package_dir())}
     for what, symbols, owner, required in _OWNED_COMPONENTS:
         # Resolved here rather than in the table, because it depends on the installed
-        # wheel. A fixed False let a wheel tagged +cu126 ship with no CUDA library at
+        # wheel. A fixed False let a wheel tagged +cu132 ship with no CUDA library at
         # all and still pass, which is the whole point of the conditional rows.
         required = _resolve_required(required)
         present = any(name.startswith(owner) for name in shipped)
 
@@ -1,6 +1,6 @@
 ---
 name: arm
-description: Develop, debug, or explain the Arm backend in backends/arm/ and its community-facing examples in examples/arm/. Covers partitioning, TOSA lowering, VGF, Ethos-U, quantization, graph passes, memory layout, and node visitors. Not for development of other backends.
+description: Develop, debug, or explain the Arm backend in backends/arm/ and its community-facing examples in examples/arm/. Covers partitioning, TOSA lowering, VGF, Ethos-U, quantization, graph passes, memory layout, node visitors, performance optimizations. Not for development of other backends.
 ---
 
 # Arm Backend Development
@@ -200,3 +200,49 @@ change. If execution requires approval or is blocked, report that rather than
 claiming it passed.
 For documentation-only changes, use focused documentation or skill validation
 and whitespace checks instead; explanations without file changes need no hook.
+
+
+# Optimising for Performance for the Arm backends.
+
+## Arm VGF backend
+The Arm VGF backend targets Arm Mali GPU with Neural Engine technology(NX). The
+Arm Neural Engine technology provides silicon acceleration for the most intense
+mathematical computations in a neural network. Here are a few principles how to
+optimise for performance for Arm Mali GPUs with NX.
+
+1. **Check that the model is quantized.**
+   Inspect the TOSA graph for unintended float32 computation. For performance, prefer full
+   static quantization of the model to 8x8 (8-bit activations and weights), or 16x8 (16-bit
+   activations and 8-bit weights).
+   Quantized inputs and outputs alone do not mean the internal model is
+   quantized.
+
+2. **Treat every NX/shader boundary as a memory cost.**
+   When execution switches from NX to a shader and back, intermediate tensors
+   must be materialized in shared memory. Budget for NX writing its output to external memory,
+   the shader reading it and writing its result, and the following NX graph reading
+   that result. These transfers will generate substantial DRAM traffic, especially
+   if the tensor sizes are significant. Do enough useful work in the shader to
+   justify the transfers and dispatches.
+
+   For example, if a bilinear RESIZE upsamples a tensor immediately before a
+   shader, consider moving the RESIZE into that shader. The shader can often
+   perform bilinear sampling efficiently. NX then writes the smaller tensor and
+   the shader reads it, avoiding the expanded intermediate's write and read.
+   Doubling both spatial dimensions makes that intermediate four times larger.
+   Preserve the original resize coordinates and quantized rounding behavior.
+
+3. **Avoid small NX graphs between shaders.**
+   Model Converter can split NX work according to the source order around
+   CUSTOM shader operations. A SLICE or a few elementwise operations between
+   independent shaders can therefore create a separate DataGraph. Group
+   independent shaders together and put their TOSA postprocessing afterwards
+   when dependencies allow. Consecutive shaders do not need a DataGraph between
+   them. Check the compiled VGF to confirm the small graphs disappeared.
+
+4. **Fuse work to avoid intermediate tensors.**
+   Consider combining related shaders and nearby operations such as
+   PixelShuffle, resizing or blending. For example, a shader that warps two
+   images can also blend them and write only the final image, avoiding two
+   temporary images and a separate NX blend. Check the serializer's supported
+   number of CUSTOM outputs before designing the fusion.