[Bugfix] Fix GraniteMoeShared weight loading broken by #41184 (#47031)
Signed-off-by: <Michal Ganczarenko> <michal.ganczarenko@intel.com> Co-authored-by: Kunshang Ji <kunshang.ji@intel.com>
Michał Ganczarenko committed
Jul 1, 2026 at 17:39 UTC
63fcce4de1563309ea5195ba98d0a2e1ba4f5831
1 file changed
+3
-3
vllm/model_executor/models/granitemoeshared.py
+3
-3
@@ -214,11 +214,11 @@ class GraniteMoeSharedModel(nn.Module):
214
for e in range(p.size(0)):
215
w1_name = n.replace(
216
".block_sparse_moe.input_linear.weight",
217
- f".block_sparse_moe.experts.routed_experts.{e}.w1.weight",
217
+ f".block_sparse_moe.experts.{e}.w1.weight",
218
)
219
w3_name = n.replace(
220
".block_sparse_moe.input_linear.weight",
221
- f".block_sparse_moe.experts.routed_experts.{e}.w3.weight",
221
+ f".block_sparse_moe.experts.{e}.w3.weight",
222
)
223
w1_param, w3_param = p[e].chunk(2, dim=0)
224
assert w1_name not in new_weights
@@ -229,7 +229,7 @@ class GraniteMoeSharedModel(nn.Module):
229
for e in range(p.size(0)):
230
w2_name = n.replace(
231
".block_sparse_moe.output_linear.weight",
232
- f".block_sparse_moe.experts.routed_experts.{e}.w2.weight",
232
+ f".block_sparse_moe.experts.{e}.w2.weight",
233
)
234
w2_param = p[e]
235
assert w2_name not in new_weights