Update on "[ET-VK][int4] patch 4-bit source transformation quantizer to support linear modules with biases"

While LLaMa does not have biases, there are some models which will have biases in their linear modules.

Add support in the source transform quantizer for biases.

Differential Revision: [D69072087](https://our.internmc.facebook.com/intern/diff/D69072087/)

[ghstack-poisoned]

Author: Nathanael See

.buckconfig

@@ -33,3 +33,6 @@
       **/.git, \
       cmake-out, \
       pip-out
+
+[buck2]
+restarter=true

backends/arm/test/misc/test_tosa_spec.py

@@ -20,16 +20,22 @@
     "TOSA-0.80+MI+8k",
     "TOSA-0.80+BI+u55",
 ]
-test_valid_1_00_strings = [
-    "TOSA-1.00.0+INT+FP+fft",
-    "TOSA-1.00.0+FP+bf16+fft",
-    "TOSA-1.00.0+INT+int4+cf",
-    "TOSA-1.00.0+FP+cf+bf16+8k",
-    "TOSA-1.00.0+FP+INT+bf16+fft+int4+cf",
-    "TOSA-1.00.0+FP+INT+fft+int4+cf+8k",
+test_valid_1_0_strings = [
+    "TOSA-1.0.0+INT+FP+fft",
+    "TOSA-1.0.0+FP+bf16+fft",
+    "TOSA-1.0.0+INT+int4+cf",
+    "TOSA-1.0.0+FP+cf+bf16+8k",
+    "TOSA-1.0.0+FP+INT+bf16+fft+int4+cf",
+    "TOSA-1.0.0+FP+INT+fft+int4+cf+8k",
+    "TOSA-1.0+INT+FP+fft",
+    "TOSA-1.0+FP+bf16+fft",
+    "TOSA-1.0+INT+int4+cf",
+    "TOSA-1.0+FP+cf+bf16+8k",
+    "TOSA-1.0+FP+INT+bf16+fft+int4+cf",
+    "TOSA-1.0+FP+INT+fft+int4+cf+8k",
 ]
 
-test_valid_1_00_extensions = {
+test_valid_1_0_extensions = {
     "INT": ["int16", "int4", "var", "cf"],
     "FP": ["bf16", "fp8e4m3", "fp8e5m2", "fft", "var", "cf"],
 }
@@ -40,19 +46,19 @@
     "TOSA-0.80+8k",
     "TOSA-0.80+BI+MI",
     "TOSA-0.80+BI+U55",
-    "TOSA-1.00.0+fft",
-    "TOSA-1.00.0+fp+bf16+fft",
-    "TOSA-1.00.0+INT+INT4+cf",
-    "TOSA-1.00.0+BI",
-    "TOSA-1.00.0+FP+FP+INT",
-    "TOSA-1.00.0+FP+CF+bf16",
-    "TOSA-1.00.0+BF16+fft+int4+cf+INT",
+    "TOSA-1.0.0+fft",
+    "TOSA-1.0.0+fp+bf16+fft",
+    "TOSA-1.0.0+INT+INT4+cf",
+    "TOSA-1.0.0+BI",
+    "TOSA-1.0.0+FP+FP+INT",
+    "TOSA-1.0.0+FP+CF+bf16",
+    "TOSA-1.0.0+BF16+fft+int4+cf+INT",
 ]
 
 test_compile_specs = [
     ([CompileSpec("tosa_version", "TOSA-0.80+BI".encode())],),
     ([CompileSpec("tosa_version", "TOSA-0.80+BI+u55".encode())],),
-    ([CompileSpec("tosa_version", "TOSA-1.00.0+INT".encode())],),
+    ([CompileSpec("tosa_version", "TOSA-1.0.0+INT".encode())],),
 ]
 
 test_compile_specs_no_version = [
@@ -70,8 +76,8 @@ def test_version_string_0_80(self, version_string: str):
         assert isinstance(tosa_spec, Tosa_0_80)
         assert tosa_spec.profile in ["BI", "MI"]
 
-    @parameterized.expand(test_valid_1_00_strings)  # type: ignore[misc]
-    def test_version_string_1_00(self, version_string: str):
+    @parameterized.expand(test_valid_1_0_strings)  # type: ignore[misc]
+    def test_version_string_1_0(self, version_string: str):
         tosa_spec = TosaSpecification.create_from_string(version_string)
         assert isinstance(tosa_spec, Tosa_1_00)
         assert [profile in ["INT", "FP"] for profile in tosa_spec.profiles].count(
@@ -80,7 +86,7 @@ def test_version_string_1_00(self, version_string: str):
 
         for profile in tosa_spec.profiles:
             assert [
-                e in test_valid_1_00_extensions[profile] for e in tosa_spec.extensions
+                e in test_valid_1_0_extensions[profile] for e in tosa_spec.extensions
             ]
 
     @parameterized.expand(test_invalid_strings)  # type: ignore[misc]
@@ -103,3 +109,15 @@ def test_create_from_invalid_compilespec(self, compile_specs: list[CompileSpec])
             tosa_spec = TosaSpecification.create_from_compilespecs(compile_specs)
 
         assert tosa_spec is None
+
+    @parameterized.expand(test_valid_0_80_strings)
+    def test_correct_string_representation_0_80(self, version_string: str):
+        tosa_spec = TosaSpecification.create_from_string(version_string)
+        assert isinstance(tosa_spec, Tosa_0_80)
+        assert f"{tosa_spec}" == version_string
+
+    @parameterized.expand(test_valid_1_0_strings)
+    def test_correct_string_representation_1_0(self, version_string: str):
+        tosa_spec = TosaSpecification.create_from_string(version_string)
+        assert isinstance(tosa_spec, Tosa_1_00)
+        assert f"{tosa_spec}" == version_string

backends/arm/tosa_specification.py

@@ -1,4 +1,4 @@
-# Copyright 2024 Arm Limited and/or its affiliates.
+# Copyright 2024-2025 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.
@@ -14,7 +14,9 @@
 import re
 from typing import List
 
-from executorch.exir.backend.compile_spec_schema import CompileSpec
+from executorch.exir.backend.compile_spec_schema import (  # type: ignore[import-untyped]
+    CompileSpec,
+)
 from packaging.version import Version
 
 
@@ -131,7 +133,7 @@ def __init__(self, version: Version, extras: List[str]):
     def __repr__(self):
         extensions = ""
         if self.level_8k:
-            extensions += "+8K"
+            extensions += "+8k"
         if self.is_U55_subset:
             extensions += "+u55"
         return f"TOSA-{str(self.version)}+{self.profile}{extensions}"
@@ -207,7 +209,10 @@ def _get_extensions_string(self) -> str:
         return "".join(["+" + e for e in self.extensions])
 
     def __repr__(self):
-        return f"TOSA-{self.version}{self._get_profiles_string()}{self._get_profiles_string()}"
+        extensions = self._get_extensions_string()
+        if self.level_8k:
+            extensions += "+8k"
+        return f"TOSA-{self.version}{self._get_profiles_string()}{extensions}"
 
     def __hash__(self) -> int:
         return hash(str(self.version) + self._get_profiles_string())

backends/cadence/aot/functions_hifi.yaml

@@ -71,11 +71,11 @@
   kernels:
     - arg_meta: null
       kernel_name: cadence::impl::HiFi::full_out
-      
+
 - op: gt.Scalar_out
   kernels:
     - arg_meta: null
-      kernel_name: torch::executor::gt_scalar_out      
+      kernel_name: torch::executor::gt_scalar_out
 
 - op: gelu.out
   kernels:
@@ -100,7 +100,7 @@
 - op: mean.out
   kernels:
     - arg_meta: null
-      kernel_name: cadence::impl::HiFi::mean_dim_out   
+      kernel_name: cadence::impl::HiFi::mean_dim_out
 
 - op: minimum.out
   kernels:
@@ -213,3 +213,13 @@
   kernels:
     - arg_meta: null
       kernel_name: cadence::impl::HiFi::quantized_linear_per_tensor_out
+
+- func: cadence::quantized_fully_connected.out(Tensor src, Tensor weight, Tensor bias, int src_zero_point, Tensor weight_zero_point, Tensor out_multiplier, Tensor out_shift, int out_zero_point, Tensor? offset, *, Tensor(a!) out) -> Tensor(a!)
+  kernels:
+    - arg_meta: null
+      kernel_name: cadence::impl::HiFi::quantized_fully_connected_out
+
+- func: cadence::quantized_fully_connected.per_tensor_out(Tensor src, Tensor weight, Tensor bias, int src_zero_point, int weight_zero_point, int out_multiplier, int out_shift, int out_zero_point, Tensor? offset, *, Tensor(a!) out) -> Tensor(a!)
+  kernels:
+    - arg_meta: null
+      kernel_name: cadence::impl::HiFi::quantized_fully_connected_per_tensor_out

backends/cadence/hifi/operators/op_clamp.cpp

@@ -48,7 +48,7 @@ namespace impl {
 namespace HiFi {
 namespace native {
 
-Tensor& clamp_tensor_out(
+Tensor& clamp_Tensor_out(
     RuntimeContext& ctx,
     const Tensor& in,
     const executorch::aten::optional<Tensor>& min_opt,