LLM extension: add ethosu 8w16a and quantize scope plumbing (#19876)

- adds the `ethosu_8w16a` PT2E quantization mode
- introduces shared `quantization.quantize_scope` handling for Arm
backends
- wires the Arm quantize scope through the LLM export path
- passes Ethos-U system config and memory mode through the partitioner
setup



cc @digantdesai @freddan80 @per @zingo @oscarandersson8218 @mansnils
@Sebastian-Larsson @robell @rascani

---------

Signed-off-by: Xingguo Li <xingguo.li@arm.com>
Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>

Author: xingguo01

examples/models/llama/export_llama_lib.py

@@ -231,6 +231,7 @@ def build_args_parser() -> argparse.ArgumentParser:
             "vulkan_8w",
             "tosa_8a8w",
             "ethosu_8a8w",
+            "ethosu_16a8w",
             "vgf_8a8w",
             "vgf_16a8w",
         ],
@@ -845,17 +846,29 @@ def get_quantizer_and_quant_params(llm_config):
             llm_config.quantization.pt2e_quantize.value
         )
         quantizers.append(coreml_quantizer)
+    arm_quantize_scope = llm_config.quantization.quantize_scope.value
+    if (
+        arm_quantize_scope == "full"
+        and llm_config.backend.vgf.enabled
+        and llm_config.backend.vgf.quantize_scope.value != "full"
+    ):
+        arm_quantize_scope = llm_config.backend.vgf.quantize_scope.value
+
     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
+            llm_config.backend.tosa.version,
+            llm_config.quantization.pt2e_quantize.value,
+            arm_quantize_scope,
         )
         quantizers.append(tosa_quantizer)
     if llm_config.backend.ethosu.enabled and llm_config.quantization.pt2e_quantize:
         ethosu_quantizer = get_ethosu_quantizer(
             llm_config.backend.ethosu.target,
             llm_config.backend.ethosu.system_config,
             llm_config.backend.ethosu.memory_mode,
+            llm_config.backend.ethosu.extra_flags,
             llm_config.quantization.pt2e_quantize.value,
+            arm_quantize_scope,
         )
         quantizers.append(ethosu_quantizer)
     if llm_config.backend.vgf.enabled and llm_config.quantization.pt2e_quantize:
@@ -1054,6 +1067,7 @@ def _to_edge_and_lower_llama_arm(
                 llm_config.backend.ethosu.target,
                 llm_config.backend.ethosu.system_config,
                 llm_config.backend.ethosu.memory_mode,
+                llm_config.backend.ethosu.extra_flags,
             )
         )
         modelname = f"ethosu_{modelname}"

extension/llm/export/config/llm_config.py

@@ -377,6 +377,7 @@ class Pt2eQuantize(str, Enum):
     vulkan_8w = "vulkan_8w"
     tosa_8a8w = "tosa_8a8w"
     ethosu_8a8w = "ethosu_8a8w"
+    ethosu_16a8w = "ethosu_16a8w"
     vgf_8a8w = "vgf_8a8w"
     vgf_16a8w = "vgf_16a8w"
 
@@ -386,6 +387,11 @@ class SpinQuant(str, Enum):
     native = "native"
 
 
+class QuantizeScope(str, Enum):
+    full = "full"
+    linear = "linear"
+
+
 @dataclass
 class QuantizationConfig:
     """
@@ -403,6 +409,9 @@ class QuantizationConfig:
         use_spin_quant: Which spin quant mode to use. If unspecified, don't use
             spin quant.
         use_qat: Whether the checkpoint is quantization-awarely trained.
+        quantize_scope: Scope for Arm PT2E quantization. "full" quantizes the
+            full supported graph, while "linear" limits quantization to
+            torch.nn.Linear modules.
         calibration_tasks: Tasks for GPTQ calibration from lm_eval.
         calibration_limit: Number of samples used for calibration from lm_eval.
         calibration_seq_length: Sequence length for GPTQ calibration from lm_eval.
@@ -427,6 +436,7 @@ class QuantizationConfig:
     group_size: Optional[int] = None
     use_spin_quant: Optional[SpinQuant] = None
     use_qat: bool = False
+    quantize_scope: QuantizeScope = QuantizeScope.full
     calibration_tasks: Optional[List[str]] = None
     calibration_limit: Optional[int] = None
     calibration_seq_length: Optional[int] = None
@@ -587,6 +597,7 @@ class EthosUConfig:
     target: str = "ethos-u85-128"  # Default target, can be overridden.
     memory_mode: str = "default"
     system_config: str = "default"
+    extra_flags: List[str] = field(default_factory=list)
 
 
 class VgfQuantizeScope(str, Enum):
@@ -832,7 +843,9 @@ def from_args(cls, args: argparse.Namespace) -> "LlmConfig":  # noqa: C901
             llm_config.backend.vgf.quantize_scope = VgfQuantizeScope(
                 args.vgf_quantize_scope
             )
-
+            llm_config.quantization.quantize_scope = QuantizeScope(
+                args.vgf_quantize_scope
+            )
         # TorchAoKernels
         if any(
             hasattr(args, a)

extension/llm/export/partitioner_lib.py

@@ -252,6 +252,7 @@ def get_ethosu_partitioner(
     target: str,
     system_config: Optional[str] = None,
     memory_mode: Optional[str] = None,
+    extra_flags: Optional[List[str]] = None,
 ):
     from executorch.backends.arm.ethosu.compile_spec import EthosUCompileSpec
     from executorch.backends.arm.ethosu.partitioner import EthosUPartitioner
@@ -260,6 +261,7 @@ def get_ethosu_partitioner(
         target,
         system_config=None if system_config == "default" else system_config,
         memory_mode=None if memory_mode == "default" else memory_mode,
+        extra_flags=extra_flags,
     )
 
     return EthosUPartitioner(compile_spec)

extension/llm/export/quantizer_lib.py

@@ -323,7 +323,7 @@ def get_vulkan_quantizer(pt2e_quantize: str):
     return quantizer
 
 
-def get_tosa_quantizer(version: str, pt2e_quantize: str):
+def get_tosa_quantizer(version: str, pt2e_quantize: str, quantize_scope: str):
     from executorch.backends.arm.quantizer.arm_quantizer import (
         get_symmetric_quantization_config,
         TOSAQuantizer,
@@ -335,34 +335,76 @@ def get_tosa_quantizer(version: str, pt2e_quantize: str):
     quantizer = TOSAQuantizer(compile_spec)
 
     if pt2e_quantize == "tosa_8a8w":
-        quantizer.set_global(get_symmetric_quantization_config())
+        quantization_config = get_symmetric_quantization_config()
     else:
         raise ValueError(f"Unsupported quantizer specification {pt2e_quantize}")
 
+    _apply_arm_quantize_scope(
+        quantizer,
+        quantization_config=quantization_config,
+        quantize_scope=quantize_scope,
+        backend_name="TOSA",
+    )
     return quantizer
 
 
 def get_ethosu_quantizer(
-    target: str, system_config: str, memory_mode: str, pt2e_quantize: str
+    target: str,
+    system_config: str,
+    memory_mode: str,
+    extra_flags: Optional[List[str]],
+    pt2e_quantize: str,
+    quantize_scope: str,
 ):
     from executorch.backends.arm.ethosu.compile_spec import EthosUCompileSpec
     from executorch.backends.arm.quantizer.arm_quantizer import (
         EthosUQuantizer,
+        get_symmetric_a16w8_quantization_config,
         get_symmetric_quantization_config,
     )
 
-    compile_spec = EthosUCompileSpec(target, system_config, memory_mode)
+    compile_spec = EthosUCompileSpec(
+        target=target,
+        system_config=None if system_config == "default" else system_config,
+        memory_mode=None if memory_mode == "default" else memory_mode,
+        extra_flags=extra_flags,
+    )
 
     quantizer = EthosUQuantizer(compile_spec)
 
     if pt2e_quantize == "ethosu_8a8w":
-        quantizer.set_global(get_symmetric_quantization_config())
+        quantization_config = get_symmetric_quantization_config()
+    elif pt2e_quantize == "ethosu_16a8w":
+        quantization_config = get_symmetric_a16w8_quantization_config()
     else:
         raise ValueError(f"Unsupported quantizer specification {pt2e_quantize}")
 
+    _apply_arm_quantize_scope(
+        quantizer,
+        quantization_config=quantization_config,
+        quantize_scope=quantize_scope,
+        backend_name="Ethos-U",
+    )
     return quantizer
 
 
+def _apply_arm_quantize_scope(
+    quantizer,
+    *,
+    quantization_config,
+    quantize_scope: str,
+    backend_name: str,
+):
+    if quantize_scope == "full":
+        quantizer.set_global(quantization_config)
+    elif quantize_scope == "linear":
+        quantizer.set_module_type(torch.nn.Linear, quantization_config)
+    else:
+        raise ValueError(
+            f"Unsupported {backend_name} quantization scope {quantize_scope}"
+        )
+
+
 def get_vgf_quantizer(
     compile_spec: Optional[str],
     compiler_flags: Optional[List[str]],
@@ -392,11 +434,10 @@ def get_vgf_quantizer(
     else:
         raise ValueError(f"Unsupported quantizer specification {pt2e_quantize}")
 
-    if quantize_scope == "full":
-        quantizer.set_global(quantization_config)
-    elif quantize_scope == "linear":
-        quantizer.set_module_type(torch.nn.Linear, quantization_config)
-    else:
-        raise ValueError(f"Unsupported VGF quantization scope {quantize_scope}")
-
+    _apply_arm_quantize_scope(
+        quantizer,
+        quantization_config=quantization_config,
+        quantize_scope=quantize_scope,
+        backend_name="VGF",
+    )
     return quantizer