[executorch][arm] Add Tosa to LLM extension

This diff is to facilitate the landing of #15556

Differential Revision: [D92280949](https://our.internmc.facebook.com/intern/diff/D92280949/)

[ghstack-poisoned]

Author: ssjia

backends/arm/quantizer/TARGETS

@@ -17,6 +17,11 @@ runtime.python_library(
     deps = [
         ":arm_quantizer_utils",
         ":quantization_annotator",
+        "//executorch/backends/arm:constants",
+        "//executorch/backends/arm:ethosu",
+        "//executorch/backends/arm:vgf",
+        "//executorch/backends/arm/tosa:specification",
+        "//executorch/backends/arm:arm_compile_spec",
         "//caffe2:torch",
         "//executorch/exir:lib",
         "//pytorch/ao:torchao",

examples/models/llama/export_llama_lib.py

@@ -1,6 +1,6 @@
 # Copyright (c) Meta Platforms, Inc. and affiliates.
 # All rights reserved.
-# Copyright 2025 Arm Limited and/or its affiliates.
+# Copyright 2025-2026 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.
@@ -37,6 +37,7 @@
     get_mps_partitioner,
     get_openvino_partitioner,
     get_qnn_partitioner,
+    get_tosa_partitioner,
     get_vulkan_partitioner,
     get_xnnpack_partitioner,
 )
@@ -46,6 +47,7 @@
     get_pt2e_quantization_params,
     get_pt2e_quantizers,
     get_qnn_quantizer,
+    get_tosa_quantizer,
     get_vulkan_quantizer,
 )
 from executorch.util.activation_memory_profiler import generate_memory_trace
@@ -210,6 +212,7 @@ def build_args_parser() -> argparse.ArgumentParser:
             "coreml_baseline_8a_c8w",
             "coreml_baseline_8a_c4w",
             "vulkan_8w",
+            "tosa_8a8w",
         ],
         help="Use PT2E quantization. Comma separated options. e.g. xnnpack_dynamic (for per channel 8 bit weight), xnnpack_dynamic_qc4 (for per channel 4 bit weight), embedding.",
     )
@@ -788,6 +791,11 @@ def get_quantizer_and_quant_params(llm_config):
             llm_config.quantization.pt2e_quantize.value
         )
         quantizers.append(coreml_quantizer)
+    if llm_config.backend.tosa.enabled and llm_config.quantization.pt2e_quantize:
+        tosa_quantizer = get_tosa_quantizer(
+            llm_config.backend.tosa.version, llm_config.quantization.pt2e_quantize.value
+        )
+        quantizers.append(tosa_quantizer)
     if llm_config.backend.vulkan.enabled and llm_config.quantization.pt2e_quantize:
         assert (
             len(quantizers) == 0
@@ -930,6 +938,32 @@ def _to_edge_and_lower_llama_openvino(
     return builder.to_executorch(passes=additional_passes)
 
 
+def _to_edge_and_lower_llama_tosa(
+    builder_exported,
+    modelname,
+    quantizers,
+    additional_passes,
+    tosa_spec,
+    verbose: bool = False,
+) -> LLMEdgeManager:
+
+    logging.info("Lowering model using TOSA partitioner")
+
+    partitioners = []
+    partitioners.append(get_tosa_partitioner(tosa_spec))
+
+    modelname = f"tosa_{modelname}"
+
+    builder = builder_exported.pt2e_quantize(quantizers).to_edge_transform_and_lower(
+        partitioners
+    )
+
+    if verbose:
+        print_delegation_info(builder.edge_manager.exported_program().graph_module)
+
+    return builder.to_executorch(passes=additional_passes)
+
+
 def _to_edge_and_lower_llama(  # noqa: C901
     builder_exported,
     modelname,
@@ -1119,7 +1153,10 @@ def _export_llama(llm_config: LlmConfig) -> LLMEdgeManager:  # noqa: C901
         additional_passes = [InitializedMutableBufferPass(["kv_cache_pos"])]
 
     # export_to_edge
-    builder_exported = _prepare_for_llama_export(llm_config).export()
+    builder_manager = _prepare_for_llama_export(llm_config)
+    if llm_config.backend.tosa.enabled:
+        builder_manager.skip_dim_order = False
+    builder_exported = builder_manager.export()
     builder_exported.run_canonical_optimizations()
     modelname = builder_exported.modelname
 
@@ -1162,6 +1199,15 @@ def _export_llama(llm_config: LlmConfig) -> LLMEdgeManager:  # noqa: C901
             openvino_device=llm_config.backend.openvino.device,
             verbose=llm_config.debug.verbose,
         )
+    elif llm_config.backend.tosa.enabled:
+        builder = _to_edge_and_lower_llama_tosa(
+            builder_exported,
+            modelname,
+            quantizers,
+            additional_passes,
+            llm_config.backend.tosa.version,
+            verbose=llm_config.debug.verbose,
+        )
     else:
         builder = _to_edge_and_lower_llama(
             builder_exported,

examples/models/llama/tests/BUCK

@@ -22,6 +22,7 @@ fbcode_target(_kind = python_unittest,
     ],
     deps = [
         "//caffe2:torch",
+        "//executorch/backends/arm/quantizer:lib",
         "//executorch/examples/models/llama:export_library",
         "//executorch/examples/models/llama:llama_transformer",
         "//pytorch/ao:torchao",
@@ -97,6 +98,7 @@ fbcode_target(_kind = python_unittest,
     ],
     deps = [
         "//caffe2:torch",
+        "//executorch/backends/arm/quantizer:lib",
         "//executorch/examples/models/llama:export_library",
         "//executorch/examples/models/llama:llama_transformer",
         "//executorch/extension/pybindings:portable_lib",

examples/models/llama/tests/test_export_llama_lib.py

@@ -1,17 +1,21 @@
 # Copyright (c) Meta Platforms, Inc. and affiliates.
 # All rights reserved.
+# Copyright 2025-2026 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.
 
 import unittest
 
+from executorch.backends.arm.quantizer.arm_quantizer import TOSAQuantizer
+
 from executorch.devtools.backend_debug import get_delegation_info
 from executorch.examples.models.llama.export_llama_lib import (
     _export_llama,
     build_args_parser,
+    get_quantizer_and_quant_params,
 )
-from executorch.extension.llm.export.config.llm_config import LlmConfig
+from executorch.extension.llm.export.config.llm_config import LlmConfig, Pt2eQuantize
 
 UNWANTED_OPS = [
     "aten_permute_copy_default",
@@ -48,3 +52,17 @@ def test_has_expected_ops_and_op_counts(self):
 
         for op, _op_info in delegation_info.delegation_by_operator.items():
             self.assertTrue(op not in UNWANTED_OPS)
+
+    def test_get_quantizer_and_quant_params_returns_tosa_quantizer(self):
+        llm_config = LlmConfig()
+        llm_config.backend.tosa.enabled = True
+        llm_config.quantization.pt2e_quantize = Pt2eQuantize.tosa_8a8w
+
+        pt2e_quant_params, quantizers, quant_dtype = get_quantizer_and_quant_params(
+            llm_config
+        )
+
+        self.assertIsNone(pt2e_quant_params)
+        self.assertIsNone(quant_dtype)
+        self.assertEqual(len(quantizers), 1)
+        self.assertIsInstance(quantizers[0], TOSAQuantizer)

extension/llm/export/builder.py

@@ -1,5 +1,6 @@
 # Copyright (c) Meta Platforms, Inc. and affiliates.
 # All rights reserved.
+# Copyright 2025-2026 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.
@@ -96,6 +97,7 @@ def __init__(
         dynamic_shapes: Optional[Any] = None,
         save_exported_program: bool = False,
         generate_etrecord: bool = False,
+        skip_dim_order: bool = True,
     ):
         # Store necessary constructor arguments.
         self.model = model
@@ -118,6 +120,7 @@ def __init__(
         self.dynamic_shapes = dynamic_shapes
         self.save_exported_program = save_exported_program
         self.generate_etrecord = generate_etrecord
+        self.skip_dim_order = skip_dim_order
 
         # Note: treat this as the source of truth for the result of
         # torch.export'ing a model. If the overall ExportedProgram is needed,
@@ -197,7 +200,7 @@ def _get_dynamic_shape(self) -> Any:
     def _get_edge_config(self) -> EdgeCompileConfig:
         edge_config = EdgeCompileConfig(
             _check_ir_validity=False,
-            _skip_dim_order=True,
+            _skip_dim_order=self.skip_dim_order,
         )
         return edge_config
 

extension/llm/export/config/llm_config.py

@@ -1,5 +1,6 @@
 # Copyright (c) Meta Platforms, Inc. and affiliates.
 # All rights reserved.
+# Copyright 2025-2026 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.
@@ -288,6 +289,7 @@ class Pt2eQuantize(str, Enum):
     coreml_baseline_8a_c8w = "coreml_baseline_8a_c8w"
     coreml_baseline_8a_c4w = "coreml_baseline_8a_c4w"
     vulkan_8w = "vulkan_8w"
+    tosa_8a8w = "tosa_8a8w"
 
 
 class SpinQuant(str, Enum):
@@ -474,6 +476,16 @@ class TorchAOKernelsConfig:
     use_torchao_kernels_tied_embedding: bool = False
 
 
+@dataclass
+class TosaConfig:
+    """
+    Configures the TOSA backend.
+    """
+
+    enabled: bool = False
+    version: str = "TOSA-1.0+INT"
+
+
 @dataclass
 class BackendConfig:
     """
@@ -488,6 +500,7 @@ class BackendConfig:
     mps: MPSConfig = field(default_factory=MPSConfig)
     openvino: OpenvinoConfig = field(default_factory=OpenvinoConfig)
     torchao: TorchAOKernelsConfig = field(default_factory=TorchAOKernelsConfig)
+    tosa: TosaConfig = field(default_factory=TosaConfig)
 
 
 ################################################################################

extension/llm/export/partitioner_lib.py

@@ -1,5 +1,6 @@
 # Copyright (c) Meta Platforms, Inc. and affiliates.
 # All rights reserved.
+# Copyright 2025-2026 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.
@@ -236,3 +237,12 @@ def get_qnn_partitioner(
         # TODO: if deprecated legacy export, skip_mutable_buffer can be set False
         skip_mutable_buffer=True,
     )
+
+
+def get_tosa_partitioner(version: str):
+    from executorch.backends.arm.tosa.compile_spec import TosaCompileSpec
+    from executorch.backends.arm.tosa.partitioner import TOSAPartitioner
+
+    compile_spec = TosaCompileSpec(version)
+
+    return TOSAPartitioner(compile_spec)

extension/llm/export/quantizer_lib.py

@@ -1,5 +1,6 @@
 # Copyright (c) Meta Platforms, Inc. and affiliates.
 # All rights reserved.
+# Copyright 2025-2026 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.
@@ -320,3 +321,22 @@ def get_vulkan_quantizer(pt2e_quantize: str):
 
     quantizer = VulkanQuantizer().set_global(config)
     return quantizer
+
+
+def get_tosa_quantizer(version: str, pt2e_quantize: str):
+    from executorch.backends.arm.quantizer.arm_quantizer import (
+        get_symmetric_quantization_config,
+        TOSAQuantizer,
+    )
+    from executorch.backends.arm.tosa.compile_spec import TosaCompileSpec
+
+    compile_spec = TosaCompileSpec(version)
+
+    quantizer = TOSAQuantizer(compile_spec)
+
+    if pt2e_quantize == "tosa_8a8w":
+        quantizer.set_global(get_symmetric_quantization_config())
+    else:
+        raise ValueError(f"Unsupported quantizer specification {pt2e_quantize}")
+
+    return quantizer