Merge pull request #126 from BrandonGroth/tiny_models

test: Add tests for save_for_aiu functionality w/ tiny models

Author: BrandonGroth

.gitignore

@@ -14,8 +14,8 @@ htmlcov/
 durations/*
 coverage*.xml
 qcfg.json
-models
 configs
+pytest.out
 
 # IDEs
 .vscode/

fms_mo/quant/quantizers.py

@@ -3029,7 +3029,7 @@ def __init__(
             self.register_buffer("clip_valn", torch.zeros(perGp[0]))
         else:
             self.register_buffer(
-                "clip_val", torch.zeros(perCh) if perCh else torch.Tensor([1.0])
+                "clip_val", torch.zeros(perCh) if perCh else torch.Tensor([0.0])
             )
             self.register_buffer(
                 "clip_valn", torch.zeros(perCh) if perCh else torch.Tensor([0.0])

fms_mo/utils/aiu_utils.py

@@ -232,6 +232,14 @@ def process_weight(
     is_w_recomputed = False
     if layer_name + ".quantize_weight.clip_val" in model.state_dict():
         w_cv = model.state_dict()[layer_name + ".quantize_weight.clip_val"]
+
+        # Check that clip values are initialized
+        if torch.any(w_cv.isclose(torch.tensor(0.0))):
+            raise ValueError(
+                f"Quantization clip values for {layer_name=} have near-zero values and "
+                "are likely uninitialized."
+            )
+
         if w_cv.numel() > 1:
             w_cv = w_cv.unsqueeze(dim=1)
         weight_int_as_fp = torch.clamp(127 / w_cv * weight_pre_quant, -127, 127).round()

fms_mo/utils/qconfig_utils.py

@@ -816,8 +816,7 @@ def qconfig_save(
 
     # Save config as json
     if os.path.isfile(fname):
-        message = f"{fname} already exist, will overwrite."
-        warnings.warn(message, UserWarning)
+        logger.info(f"{fname} already exist, will overwrite.")
     with open(fname, "w", encoding="utf-8") as outfile:
         json.dump(temp_qcfg, outfile, indent=4)