default W4A16 alias to use group_size=128 (#94) (#95)

bfineran · web-flow · commit 42575fd2aaf1 · 2024-06-20T13:43:26.000-04:00
diff --git a/src/compressed_tensors/quantization/quant_scheme.py b/src/compressed_tensors/quantization/quant_scheme.py
@@ -111,7 +111,7 @@ def is_preset_scheme(name: str) -> bool:
     weights=QuantizationArgs(), input_activations=QuantizationArgs(symmetric=True)
 )
 
-W4A16 = dict(weights=QuantizationArgs(num_bits=4, symmetric=True))
+W4A16 = dict(weights=QuantizationArgs(num_bits=4, group_size=128))
 
 PRESET_SCHEMES = {
     "W8A8": W8A8,

Original file line number	Diff line number	Diff line change
`@@ -111,7 +111,7 @@ def is_preset_scheme(name: str) -> bool:`
`111`	`111`	`weights=QuantizationArgs(), input_activations=QuantizationArgs(symmetric=True)`
`112`	`112`	`)`
`113`	`113`
`114`		`-W4A16 = dict(weights=QuantizationArgs(num_bits=4, symmetric=True))`
	`114`	`+W4A16 = dict(weights=QuantizationArgs(num_bits=4, group_size=128))`
`115`	`115`
`116`	`116`	`PRESET_SCHEMES = {`
`117`	`117`	`"W8A8": W8A8,`