[Bugfix][Quantization] Fix MXFP8 startup crash on layers below mm_mxfp8 shape limits (#54223)
Signed-off-by: samuelkim7 <samuelmwkim@gmail.com> Signed-off-by: mgoin <mgoin64@gmail.com> Co-authored-by: mgoin <mgoin64@gmail.com>
This commit is contained in:
@@ -431,7 +431,7 @@ def test_mxfp8_linear_emulation_bf16_at_load(
|
||||
layer.weight = torch.nn.Parameter(w_fp8.clone(), requires_grad=False)
|
||||
layer.weight_scale = torch.nn.Parameter(w_scale.clone(), requires_grad=False)
|
||||
|
||||
kernel = EmulationMxfp8LinearKernel(Mxfp8LinearLayerConfig())
|
||||
kernel = EmulationMxfp8LinearKernel(Mxfp8LinearLayerConfig(weight_shape=(N, K)))
|
||||
kernel.process_weights_after_loading(layer)
|
||||
|
||||
if dequant_at_load:
|
||||
@@ -482,7 +482,9 @@ def test_mxfp8_rocm_native_unaligned_k_dequantizes_at_load(shape):
|
||||
layer.weight_scale = torch.nn.Parameter(w_scale.clone(), requires_grad=False)
|
||||
return layer
|
||||
|
||||
kernel = rocm_native.RocmDotScaledMxfp8LinearKernel(Mxfp8LinearLayerConfig())
|
||||
kernel = rocm_native.RocmDotScaledMxfp8LinearKernel(
|
||||
Mxfp8LinearLayerConfig(weight_shape=(N, K))
|
||||
)
|
||||
layer = make_layer()
|
||||
kernel.process_weights_after_loading(layer)
|
||||
assert (layer.weight.element_size() >= 2) is not aligned
|
||||
|
||||
Reference in New Issue
Block a user