Arm backend: reducing qtol to 1 after adjusting eps for llama (#18116)

- Reducing qtol from 75 to 1 after adjusting eps from 2^^-16 to 2^^-12

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

Signed-off-by: Xingguo Li <xingguo.li@arm.com>

Author: xingguo01

backends/arm/quantizer/arm_quantizer.py

@@ -76,6 +76,7 @@ def get_symmetric_quantization_config(
     act_qmax: int = 127,
     weight_qmin: int = -127,
     weight_qmax: int = 127,
+    eps: float = 2**-16,
 ) -> QuantizationConfig:
     """Create symmetric quantization config for activations and weights.
 
@@ -95,7 +96,7 @@ def get_symmetric_quantization_config(
         bias.
 
     """
-    extra_args: Dict[str, Any] = {"eps": 2**-16}
+    extra_args: Dict[str, Any] = {"eps": eps}
     if is_qat:
         if is_dynamic:
             act_observer_or_fake_quant_ctr = FakeQuantize

backends/arm/test/models/test_llama.py

@@ -102,11 +102,11 @@ def prepare_model(self):
         return llama_model, llama_inputs, llama_meta
 
 
-def _use_partial_quantizer(pipeline):
+def _use_partial_quantizer(pipeline, eps=2**-16):
     """Set the pipeline's quantizer to only include Linear layers."""
     pipeline.quantizer.set_global(None)
     pipeline.quantizer.set_module_type(
-        torch.nn.Linear, get_symmetric_quantization_config()
+        torch.nn.Linear, get_symmetric_quantization_config(eps=eps)
     )
 
 
@@ -209,12 +209,11 @@ def test_llama_tosa_INT_FP_partial_quant():
             tosa_extensions=["FP"],
             # Due to a few outliers, atol must be set high
             atol=1.1,
-            # TODO(MLETORCH-1875): reduce tolerance
-            qtol=75,
+            qtol=1,
             frobenius_threshold=None,
             cosine_threshold=None,
         )
-        _use_partial_quantizer(pipeline)
+        _use_partial_quantizer(pipeline, eps=2**-12)
         pipeline.run()
 
 
@@ -234,8 +233,7 @@ def test_llama_vgf_quant_partial_quant():
             quantize=True,
             # Due to a few outliers, atol must be set high
             atol=1.1,
-            # TODO(MLETORCH-1875): reduce tolerance
-            qtol=75,
+            qtol=1,
         )
-        _use_partial_quantizer(pipeline)
+        _use_partial_quantizer(pipeline, eps=2**-12)
         pipeline.run()