add fp8 autotune ut and fix bug in autotune (#1638)

Signed-off-by: xin3he <[email protected]>

neural_compressor/torch/quantization/__init__.py

@@ -31,6 +31,7 @@
     get_default_hqq_config,
     FP8Config,
     get_default_fp8_config,
+    get_default_fp8_config_set,
 )
 
 from neural_compressor.torch.quantization.autotune import (

neural_compressor/torch/quantization/autotune.py

@@ -72,10 +72,13 @@ def autotune(
         tuning_logger.trial_end(trial_index)
         if tuning_monitor.need_stop():
             logger.info("Stopped tuning.")
+            del q_model  # maybe gc.collect() is needed for memory release
             best_quant_config: BaseConfig = tuning_monitor.get_best_quant_config()
             # !!! Make sure to use deepcopy only when inplace is set to `True`.
-            quantize(deepcopy(model), quant_config=best_quant_config, run_fn=run_fn, run_args=run_args, inplace=True)
-            best_quant_model = model  # quantize model inplace
+            q_model = quantize(
+                deepcopy(model), quant_config=best_quant_config, run_fn=run_fn, run_args=run_args, inplace=True
+            )
+            best_quant_model = q_model  # quantize model inplace
             break
     tuning_logger.tuning_end()
     return best_quant_model

neural_compressor/torch/quantization/config.py

@@ -941,14 +941,23 @@ def get_config_set_for_tuning(cls) -> Union[None, "FP8Config", List["FP8Config"]
 
 
 def get_default_fp8_config() -> FP8Config:
-    """Generate the default gptq config.
+    """Generate the default fp8 config.
 
     Returns:
-        the default gptq config.
+        the default fp8 config.
     """
     return FP8Config()
 
 
+def get_default_fp8_config_set() -> FP8Config:
+    """Generate the default fp8 config set.
+
+    Returns:
+        the default fp8 config.
+    """
+    return FP8Config.get_config_set_for_tuning()
+
+
 ##################### Algo Configs End ###################################
 
 

test/3x/torch/quantization/habana_fp8/test_fp8.py

@@ -18,8 +18,14 @@
         FP8Matmul,
         Matmul,
     )
-    from neural_compressor.torch.quantization import quantize
-    from neural_compressor.torch.quantization.config import FP8Config, get_default_fp8_config
+    from neural_compressor.torch.quantization import (
+        FP8Config,
+        TuningConfig,
+        autotune,
+        get_default_fp8_config,
+        get_default_fp8_config_set,
+        quantize,
+    )
 
     torch.set_grad_enabled(False)
 
@@ -164,3 +170,31 @@ def calib_func(model):
         assert isinstance(m.fc1, FP8Linear), "Unexpected result. Please double check."
         assert isinstance(m.mm, FP8Matmul), "Unexpected result. Please double check."
         assert isinstance(m.bmm, FP8BatchMatmul), "Unexpected result. Please double check."
+
+    def test_autotune(self):
+        m = copy.deepcopy(self.model)
+        inp = self.inp
+        fp32_out = m(inp)
+
+        def calib_func(model):
+            model(inp)
+
+        accu_list = [1.0, 0.9, 0.99]
+
+        def eval_func(model):
+            nonlocal accu_list
+            return accu_list.pop()
+
+        tune_config = TuningConfig(
+            config_set=get_default_fp8_config_set(),
+            tolerable_loss=0.01,
+        )
+        best_model = autotune(
+            model=m,
+            tune_config=tune_config,
+            run_fn=calib_func,
+            eval_fns=eval_func,
+        )
+        assert isinstance(best_model.fc1, FP8Linear), "Unexpected result. Please double check."
+        assert isinstance(best_model.mm, FP8Matmul), "Unexpected result. Please double check."
+        assert isinstance(best_model.bmm, FP8BatchMatmul), "Unexpected result. Please double check."