Merge branch 'main' into python_314

Author: CandiedCode

.github/workflows/cuda.yml

@@ -148,6 +148,10 @@ jobs:
         # Run Qwen 3.5 MoE tests (quantize roundtrip + TurboQuant KV cache + sampler)
         python -m pytest examples/models/qwen3_5_moe/test_quantize_roundtrip.py examples/models/qwen3_5_moe/test_turboquant.py examples/models/qwen3_5_moe/test_sampler.py -v -o "addopts="
 
+        # Run Gemma 4 31B tests (quant unit tests + pipeline integration tests)
+        pip install gguf
+        python -m pytest examples/models/gemma4_31b/quant/tests/ examples/models/gemma4_31b/tests/ -v -o "addopts="
+
   export-model-cuda-artifact:
     name: export-model-cuda-artifact
     # Skip this job if the pull request is from a fork (HuggingFace secrets are not available)

Makefile

@@ -91,7 +91,7 @@
 #
 # ==============================================================================
 
-.PHONY: voxtral-cuda voxtral-cpu voxtral-metal voxtral-mlx voxtral_realtime-cuda voxtral_realtime-cpu voxtral_realtime-metal voxtral_realtime-mlx voxtral_tts-cpu voxtral_tts-cuda whisper-cuda whisper-cuda-debug whisper-cpu whisper-metal parakeet-cuda parakeet-cuda-debug parakeet-cpu parakeet-metal parakeet-mlx parakeet-vulkan dinov2-cuda dinov2-cuda-debug sortformer-cuda sortformer-cpu silero-vad-cpu llama-cuda llama-cuda-debug llama-cpu llava-cpu gemma3-cuda gemma3-cpu qwen3_5_moe-cuda qwen3_5_moe-metal clean help
+.PHONY: voxtral-cuda voxtral-cpu voxtral-metal voxtral-mlx voxtral_realtime-cuda voxtral_realtime-cpu voxtral_realtime-metal voxtral_realtime-mlx voxtral_tts-cpu voxtral_tts-cuda whisper-cuda whisper-cuda-debug whisper-cpu whisper-metal parakeet-cuda parakeet-cuda-debug parakeet-cpu parakeet-metal parakeet-mlx parakeet-vulkan dinov2-cuda dinov2-cuda-debug sortformer-cuda sortformer-cpu silero-vad-cpu llama-cuda llama-cuda-debug llama-cpu llava-cpu gemma3-cuda gemma3-cpu gemma4_31b-cuda qwen3_5_moe-cuda qwen3_5_moe-metal clean help
 
 help:
 	@echo "This Makefile adds targets to build runners for various models on various backends. Run using \`make <target>\`. Available targets:"
@@ -126,6 +126,7 @@ help:
 	@echo "  llava-cpu           - Build Llava runner with CPU backend"
 	@echo "  gemma3-cuda         - Build Gemma3 runner with CUDA backend"
 	@echo "  gemma3-cpu          - Build Gemma3 runner with CPU backend"
+	@echo "  gemma4_31b-cuda     - Build Gemma 4 31B runner with CUDA backend"
 	@echo "  qwen3_5_moe-cuda    - Build Qwen3.5 MoE runner with CUDA backend"
 	@echo "  qwen3_5_moe-metal   - Build Qwen3.5 MoE runner with Metal backend"
 	@echo "  clean               - Clean build artifacts"
@@ -425,6 +426,15 @@ qwen3_5_moe-cuda:
 	@echo "✓ Build complete!"
 	@echo "  Binary: cmake-out/examples/models/qwen3_5_moe/qwen3_5_moe_runner"
 
+gemma4_31b-cuda:
+	@echo "==> Building and installing ExecuTorch with CUDA..."
+	cmake --workflow --preset llm-release-cuda
+	@echo "==> Building Gemma 4 31B runner with CUDA..."
+	cd examples/models/gemma4_31b && cmake --workflow --preset gemma4-31b-cuda
+	@echo ""
+	@echo "✓ Build complete!"
+	@echo "  Binary: cmake-out/examples/models/gemma4_31b/gemma4_31b_runner"
+
 qwen3_5_moe-metal:
 	@echo "==> Building and installing ExecuTorch with Metal..."
 	cmake --workflow --preset llm-release-metal

backends/arm/README.md

@@ -249,10 +249,15 @@ Some tests, with `u55`, `u85` and `vgf` in the name require external dependencie
   ```
 
 In addition, some model tests in the Arm backend require third-party libraries or packages.
-To run these tests, you need to install the required dependencies by running the script `examples/arm/setup.sh` with the flag `--setup-test-dependency`.
+To run these tests, install the required dependencies directly:
 
-Please note that installing model test dependencies is a standalone process. When using the `--setup-test-dependency` flag,
-the script will install only the necessary dependencies for model tests, skipping all other setup procedures.
+```
+bash backends/arm/scripts/install_models_for_test.sh
+```
+
+Installing model test dependencies is a standalone process. The script installs
+only the dependencies needed for model tests, skipping all other setup
+procedures.
 
 ## Using git hooks
 

backends/arm/_passes/arm_pass_manager.py

@@ -150,7 +150,6 @@
 from executorch.backends.arm.common.arm_compile_spec import ArmCompileSpec
 from executorch.backends.arm.common.pipeline_config import (
     ArmPassPipelineConfig,
-    FuseDuplicateUsersConfig,
     SoftmaxDecompositionConfig,
 )
 from executorch.backends.arm.tosa.specification import (
@@ -238,9 +237,6 @@ def configure_skip_passes(
             case SoftmaxDecompositionConfig.STABLE:
                 skip_set.add(DecomposeMaskedFillPass)
 
-        if config.fuse_duplicate_users is FuseDuplicateUsersConfig.DISABLED:
-            skip_set.add(FuseDuplicateUsersPass)
-
         self._skip_pass_types = tuple(skip_set)
         skip_names = [skipped_pass.__name__ for skipped_pass in self._skip_pass_types]
         logger.debug(f"Passes in skip list: {skip_names}")
@@ -403,9 +399,6 @@ def _tosa_pipeline(
                 ConvertToClampPass(),
                 DecomposeTOSAUnsupportedClampPass(),
                 DecomposeGroupNormPass(),
-                DecomposeGruPass(),
-                DecomposeLstmPass(),
-                DecomposeRnnPass(),
                 DecomposeLayerNormPass(),
                 DecomposeVarPass(),
                 DecomposeMeanDimPass(exported_program.graph_module, self.tosa_spec),

backends/arm/_passes/decompose_gru_pass.py

@@ -1,5 +1,6 @@
 # Copyright (c) Meta Platforms, Inc. and affiliates.
 # All rights reserved.
+# Copyright 2026 Arm Limited and/or its affiliates.
 #
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
@@ -13,7 +14,6 @@
     create_node,
     get_getitem_users,
 )
-from executorch.backends.arm._passes.insert_table_ops import InsertTableOpsPass
 from executorch.exir.pass_base import ExportPass, PassResult
 
 
@@ -34,7 +34,7 @@ class DecomposeGruPass(ArmPass):
 
     """
 
-    _passes_required_after: Set[Type[ExportPass]] = {InsertTableOpsPass}
+    _passes_required_after: Set[Type[ExportPass]] = set()
 
     _TARGET = torch.ops.aten.gru.input
 

backends/arm/_passes/decompose_lstm_pass.py

@@ -1,5 +1,6 @@
 # Copyright (c) Meta Platforms, Inc. and affiliates.
 # All rights reserved.
+# Copyright 2026 Arm Limited and/or its affiliates.
 #
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
@@ -13,7 +14,6 @@
     create_node,
     get_getitem_users,
 )
-from executorch.backends.arm._passes.insert_table_ops import InsertTableOpsPass
 from executorch.exir.pass_base import ExportPass, PassResult
 
 
@@ -36,7 +36,7 @@ class DecomposeLstmPass(ArmPass):
 
     """
 
-    _passes_required_after: Set[Type[ExportPass]] = {InsertTableOpsPass}
+    _passes_required_after: Set[Type[ExportPass]] = set()
 
     _TARGET = torch.ops.aten.lstm.input
 

backends/arm/_passes/decompose_rnn_pass.py

@@ -1,5 +1,6 @@
 # Copyright (c) Meta Platforms, Inc. and affiliates.
 # All rights reserved.
+# Copyright 2026 Arm Limited and/or its affiliates.
 #
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
@@ -13,7 +14,6 @@
     create_node,
     get_getitem_users,
 )
-from executorch.backends.arm._passes.insert_table_ops import InsertTableOpsPass
 from executorch.exir.pass_base import ExportPass, PassResult
 
 
@@ -30,7 +30,7 @@ class DecomposeRnnPass(ArmPass):
 
     """
 
-    _passes_required_after: Set[Type[ExportPass]] = {InsertTableOpsPass}
+    _passes_required_after: Set[Type[ExportPass]] = set()
 
     _TARGETS = {
         torch.ops.aten.rnn_tanh.input,

backends/arm/common/pipeline_config.py

@@ -14,24 +14,12 @@ class SoftmaxDecompositionConfig(Enum):
     STABLE = auto()  # Stable softmax, no masked fill decomposition
 
 
-class FuseDuplicateUsersConfig(Enum):
-    ENABLED = auto()
-    DISABLED = auto()
-
-
 @dataclass
 class ArmPassPipelineConfig:
     softmax: SoftmaxDecompositionConfig = SoftmaxDecompositionConfig.MASKED
-    fuse_duplicate_users: FuseDuplicateUsersConfig = FuseDuplicateUsersConfig.ENABLED
-
-    def disable_fuse_duplicate_users(self) -> None:
-        self.fuse_duplicate_users = FuseDuplicateUsersConfig.DISABLED
 
     def is_default(self) -> bool:
-        return (
-            self.softmax is SoftmaxDecompositionConfig.MASKED
-            and self.fuse_duplicate_users is FuseDuplicateUsersConfig.ENABLED
-        )
+        return self.softmax is SoftmaxDecompositionConfig.MASKED
 
     def to_dict(self) -> dict[str, str]:
         return {f.name: getattr(self, f.name).name for f in fields(self)}

backends/arm/runtime/VGFBackend.cpp

@@ -5,10 +5,10 @@
  * LICENSE file in the root directory of this source tree.
  */
 
-#include <list>
-#include <numeric>
+#include <cinttypes>
 using namespace std;
 
+#include <c10/util/safe_numerics.h>
 #include <executorch/runtime/backend/interface.h>
 #include <executorch/runtime/core/error.h>
 #include <executorch/runtime/core/evalue.h>
@@ -191,8 +191,18 @@ class VGFBackend final : public ::executorch::runtime::BackendInterface {
       if (!io->is_input)
         continue;
 
-      size_t io_size = accumulate(
-          io->size.begin(), io->size.end(), io->elt_size, std::multiplies<>());
+      size_t io_size = io->elt_size;
+      for (int64_t dim : io->size) {
+        ET_CHECK_OR_RETURN_ERROR(
+            dim >= 0,
+            InvalidArgument,
+            "Negative dimension in IO size: %" PRId64,
+            dim);
+        ET_CHECK_OR_RETURN_ERROR(
+            !c10::mul_overflows(io_size, static_cast<size_t>(dim), &io_size),
+            InvalidArgument,
+            "Overflow computing IO buffer size");
+      }
 
       void* data;
       if (!repr->map_io(io, &data)) {
@@ -226,8 +236,18 @@ class VGFBackend final : public ::executorch::runtime::BackendInterface {
       if (io->is_input)
         continue;
 
-      size_t io_size = accumulate(
-          io->size.begin(), io->size.end(), io->elt_size, std::multiplies<>());
+      size_t io_size = io->elt_size;
+      for (int64_t dim : io->size) {
+        ET_CHECK_OR_RETURN_ERROR(
+            dim >= 0,
+            InvalidArgument,
+            "Negative dimension in IO size: %" PRId64,
+            dim);
+        ET_CHECK_OR_RETURN_ERROR(
+            !c10::mul_overflows(io_size, static_cast<size_t>(dim), &io_size),
+            InvalidArgument,
+            "Overflow computing IO buffer size");
+      }
 
       void* data;
       if (!repr->map_io(io, &data)) {

backends/arm/test/misc/test_compile_spec.py

@@ -5,10 +5,7 @@
 
 import warnings
 
-from executorch.backends.arm.common.pipeline_config import (
-    FuseDuplicateUsersConfig,
-    SoftmaxDecompositionConfig,
-)
+from executorch.backends.arm.common.pipeline_config import SoftmaxDecompositionConfig
 from executorch.backends.arm.ethosu import EthosUCompileSpec
 from executorch.backends.arm.tosa.compile_spec import TosaCompileSpec
 from executorch.backends.arm.vgf import VgfCompileSpec
@@ -66,11 +63,11 @@ def test_compile_spec_vgf_no_quant():
         EthosUCompileSpec._from_list(spec_list)
 
 
-def test_compile_spec_vgf_defaults_to_enabled_fuse_duplicate_users():
+def test_compile_spec_vgf_uses_default_pipeline_config():
     compile_spec = VgfCompileSpec()
     pipeline_config = compile_spec._get_pass_pipeline_config()
 
-    assert pipeline_config.fuse_duplicate_users == FuseDuplicateUsersConfig.ENABLED
+    assert pipeline_config.is_default()
 
 
 def test_compile_spec_tosa_INT():