dynamic distpatch of fp8 kernels (#14245)

Signed-off-by: Jeff Daily <jeff.daily@amd.com>
This commit is contained in:
Jeff Daily
2025-03-11 10:54:56 -04:00
committed by GitHub
parent 08a1a1121d
commit a1c8f3796c
25 changed files with 292 additions and 159 deletions
@@ -142,8 +142,7 @@ class QuarkW8A8Fp8MoEMethod(QuarkMoEMethod):
layer.w2_input_scale = torch.nn.Parameter(
layer.w2_input_scale.max(), requires_grad=False)
# If rocm, normalize the weights and scales to e4m3fnuz
if current_platform.is_rocm():
if current_platform.is_fp8_fnuz():
# Normalize the weights and scales
w13_weight, w13_weight_scale, w13_input_scale = \
normalize_e4m3fn_to_e4m3fnuz(
@@ -39,7 +39,7 @@ class QuarkW8A8Fp8(QuarkScheme):
logical_widths=layer.logical_widths,
)
if current_platform.is_rocm():
if current_platform.is_fp8_fnuz():
weight, max_w_scale, input_scale = normalize_e4m3fn_to_e4m3fnuz(
weight=weight,
weight_scale=max_w_scale,
@@ -55,7 +55,7 @@ class QuarkW8A8Fp8(QuarkScheme):
elif self.qscheme == "per_channel":
weight = layer.weight
if current_platform.is_rocm():
if current_platform.is_fp8_fnuz():
weight, weight_scale, input_scale = \
normalize_e4m3fn_to_e4m3fnuz(
weight=weight,