biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Quark] Fix MoE fp8 activation scale handling on mi300 (#34386)

Browse Source 
Signed-off-by: Bowen Bao <bowenbao@amd.com>
This commit is contained in:
Bowen Bao
2026-02-19 21:27:14 -08:00
committed by GitHub
parent a1a2d79442
commit d9e62c03eb
1 changed files with 3 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
vllm/model_executor/layers/quantization/quark/quark_moe.py
View File

@@ -858,7 +858,7 @@ class QuarkOCP_MX_MoEMethod(QuarkMoEMethod):
layer.w2_input_scale = None
def process_weights_after_loading(self, layer):
if self.static_input_scales:
if self.static_input_scales and self.input_dtype == "fp8":
# firstly, process activations if fp8 static input
if layer.w13_input_scale is None or layer.w2_input_scale is None:
raise ValueError(
@@ -883,14 +883,14 @@ class QuarkOCP_MX_MoEMethod(QuarkMoEMethod):
if current_platform.is_fp8_fnuz():
# Normalize the weights and scales
_, _, w13_input_scale = normalize_e4m3fn_to_e4m3fnuz(
torch.empty_like(layer.w13_weight, dtype=torch.float8_e4m3fnuz),
torch.empty_like(layer.w13_weight, dtype=torch.float8_e4m3fn),
torch.empty_like(
layer.w13_weight_scale, dtype=layer.w13_weight_scale.dtype
),
layer.w13_input_scale,
)
_, _, w2_input_scale = normalize_e4m3fn_to_e4m3fnuz(
torch.empty_like(layer.w2_weight, dtype=torch.float8_e4m3fnuz),
torch.empty_like(layer.w2_weight, dtype=torch.float8_e4m3fn),
torch.empty_like(
layer.w2_weight_scale, dtype=layer.w13_weight_scale.dtype
),