[NF4][FSDP] return contiguous quantization_factor (#298)

weifengpy · web-flow · commit 9dbc5542bc1a · 2024-05-31T11:37:05.000-07:00
Summary:

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:
diff --git a/torchao/dtypes/nf4tensor.py b/torchao/dtypes/nf4tensor.py
@@ -596,7 +596,7 @@ def double_quantize_scalers(
 
         return (
             quantized_scaler_blocks.flatten().to(torch.int8),
-            quantization_factor.view(n_scaler_blocks),
+            quantization_factor.view(n_scaler_blocks).contiguous(),
             scalers_1_mean,
         )
 

Original file line number	Diff line number	Diff line change
`@@ -596,7 +596,7 @@ def double_quantize_scalers(`
`596`	`596`
`597`	`597`	`return (`
`598`	`598`	`quantized_scaler_blocks.flatten().to(torch.int8),`
`599`		`- quantization_factor.view(n_scaler_blocks),`
	`599`	`+ quantization_factor.view(n_scaler_blocks).contiguous(),`
`600`	`600`	`scalers_1_mean,`
`601`	`601`	`)`
`602`	`602`