fix condition

Author: dsikka

src/compressed_tensors/compressors/model_compressors/model_compressor.py

@@ -377,9 +377,9 @@ def compress(
 
         compressed_state_dict = state_dict
 
-        quantized_modules_to_args: Dict[str, QuantizationArgs] = (
-            map_modules_to_quant_args(model)
-        )
+        quantized_modules_to_args: Dict[
+            str, QuantizationArgs
+        ] = map_modules_to_quant_args(model)
 
         if self.quantization_compressor is not None:
             compressed_state_dict = self.quantization_compressor.compress(

src/compressed_tensors/quantization/lifecycle/initialize.py

@@ -193,17 +193,16 @@ def _initialize_scale_zero_point(
             module, f"{base_name}_global_scale", init_global_scale
         )
 
-        
     # TODO: consider erroring out in the future as if the dtype if not one fo these,
     # there is likely bug
-   
+
     if scale_dtype not in [
         torch.float16,
         torch.bfloat16,
         torch.float32,
         FP8_E4M3_DATA.dtype,
     ]:
-      scale_dtype = torch.float16
+        scale_dtype = torch.float16
 
     # initializes empty scale, zero point, and g_idx parameters for the module
     init_scale = Parameter(

src/compressed_tensors/quantization/utils/helpers.py

@@ -82,7 +82,6 @@ def calculate_qparams(
     zp_dtype = FP8_E4M3_DATA.dtype
 
     if quantization_args.symmetric:
-        # TODO: update for NVFP4 when applying observers
         max_val_pos = torch.max(torch.abs(min_vals), torch.abs(max_vals))
 
         if (
@@ -96,10 +95,15 @@ def calculate_qparams(
             # Divide over bit range over max value?
             scales = max_val_pos / (float(bit_range) / 2)
 
-        # TODO: clamp not implemented for FP8 - we shouldn't need to clamp this anyway as we're
-        # casting to FP8 on line 92?
-        if scales.dtype != FP8_E4M3_DATA.dtype:
+        if scales.dtype == FP8_E4M3_DATA.dtype:
+            # use the next largest fp8 value from 0
+            # Optionally, we swap to use the reciporcal
+            scales = torch.where(
+                scales == 0, torch.tensor(0.125, dtype=FP8_E4M3_DATA.dtype), scales
+            )
+        else:
             scales = torch.clamp(scales, min=torch.finfo(torch.float32).eps)
+
         zero_points = torch.zeros(scales.shape, device=device, dtype=min_vals.dtype)
     else:
         scales = (max_vals - min_vals) / float(bit_range)