Skip to content

fix: correct MXFP4 MoE weight shuffle for Quark models (MiniMax-M2.1) #632

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Draft
thpereir wants to merge 1 commit into
base: main
Choose a base branch
from
Draft

fix: correct MXFP4 MoE weight shuffle for Quark models (MiniMax-M2.1) #632

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
23 changes: 17 additions & 6 deletions atom/model_ops/moe.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -869,6 +869,7 @@ def process_weights_after_loading(self, layer):
True,
)
layer.w2_weight.data = shuffle_weight_a16w4(layer.w2_weight, 16, False)
layer.w2_weight.is_shuffled = True
shuffled_w2_scale = shuffle_scale_a16w4(
layer.w2_weight_scale.view(-1, layer.w2_weight_scale.shape[-1]),
self.num_experts,
Expand All @@ -883,16 +884,26 @@ def process_weights_after_loading(self, layer):
)
# quark method for moe, split it out?
elif self.quant_method == "quark":
shuffle_weights(layer.w13_weight, layer.w2_weight)
# W1/W13 (gate+up): both stage1 (ck_moe_stage1) and stage2
# (ck_moe_stage2) use DeviceMoeGemmMXBPreShuffle with preShuffleBuffer
# layout (gate_up=False). The a4w4 kernel expects gate_up=False even
# though w1 has 2*N rows.
layer.w13_weight.data = shuffle_weight_a16w4(layer.w13_weight, 16, False)
layer.w13_weight.is_shuffled = True
s0, s1, _ = layer.w13_weight_scale.shape
w13_weight_scale = layer.w13_weight_scale.view(s0 * s1, -1)
w13_weight_scale = fp4_utils.e8m0_shuffle(w13_weight_scale)
w13_weight_scale = shuffle_scale_a16w4(w13_weight_scale, s0, False)
layer.w13_weight_scale.data = w13_weight_scale.view(s0, s1, -1)

s0, s1, _ = layer.w2_weight_scale.shape
w2_weight_scale = layer.w2_weight_scale.view(s0 * s1, -1)
w2_weight_scale = fp4_utils.e8m0_shuffle(w2_weight_scale)
layer.w2_weight_scale.data = w2_weight_scale.view(s0, s1, -1)
# W2 (down-proj): same preShuffleBuffer layout with gate_up=False.
layer.w2_weight.data = shuffle_weight_a16w4(layer.w2_weight, 16, False)
layer.w2_weight.is_shuffled = True
shuffled_w2_scale = shuffle_scale_a16w4(
layer.w2_weight_scale.view(-1, layer.w2_weight_scale.shape[-1]),
self.num_experts,
False,
)
layer.w2_weight_scale = atom_parameter(shuffled_w2_scale)
return
else:
shuffle_weights(layer.w13_weight, layer.w2_weight)
Expand Down
Loading