more cleanup

Signed-off-by: Lucas Wilkinson <lwilkinson@neuralmagic.com>

Author: LucasWilkinson

vllm/model_executor/layers/quantization/utils/quant_utils.py

@@ -53,7 +53,6 @@ def group_broadcast(t, shape):
 
 
 # Quantize assuming once scale per group of elements with shape group_shape,
-# Defaults to quantizing to correct platform specific float8 type
 # example group shapes:
 #  * (-1, -1)   for per-tensor quantization
 #  * (1, -1)    for per-row quantization
@@ -64,14 +63,14 @@ def group_broadcast(t, shape):
 def scaled_quantize(
     x: torch.Tensor,
     group_shape: Tuple[int, int],
-    dtype: torch.dtype,
+    tgt_dtype: torch.dtype,
 ) -> Tuple[torch.Tensor, torch.Tensor]:
     group_shape = _normalize_quant_group_shape(x, group_shape)
-    assert dtype.is_floating_point, \
+    assert tgt_dtype.is_floating_point, \
         "currently `scaled_quantize` only supports floating point dtypes " \
         "but could be extended to support other dtypes"
 
-    finfo = torch.finfo(dtype)
+    finfo = torch.finfo(tgt_dtype)
 
     # Reshape (M, N) into (BLK_M, BLOCK_SIZE_M, BLK_N, BLOCK_SIZE_N)
     assert x.ndim == 2
@@ -97,7 +96,7 @@ def scaled_quantize(
         .permute(0, 2, 1, 3)\
         .reshape(x.shape)
 
-    return x_scl_sat.to(dtype).contiguous(), scale.float().reciprocal()
+    return x_scl_sat.to(tgt_dtype).contiguous(), scale.float().reciprocal()
 
 
 # inverses `scaled_quantize`