[Bugfix] Size and iterate w13 by shard count for non-gated MoE (#51125)
Signed-off-by: aoshen02 <aoshen@inferact.ai> Co-authored-by: Matej Sirovatka <matej.sirovatka@gmail.com> Co-authored-by: Rishi Puri <riship@nvidia.com> Co-authored-by: Hsiao-Yuan Chen <littlecircle0730@users.noreply.github.com> Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
This commit is contained in:
co-authored by
Matej Sirovatka
Rishi Puri
Hsiao-Yuan Chen
Claude Opus 5
parent
811622c410
commit
7c77868cdf
@@ -728,10 +728,10 @@ class HummingMoEMethod(FusedMoEMethodBase):
|
||||
# The weight names of sublayer start with the prefix "{sublayer_name}_"
|
||||
layer.sublayer_configs = {
|
||||
"w13": {
|
||||
"shape_n": intermediate_size_per_partition * 2,
|
||||
"shape_n": intermediate_size_per_partition * self.moe.w13_num_shards,
|
||||
"shape_k": hidden_size,
|
||||
"tensors_attrs": self.weight_schema.get_padded_tensors_attrs(
|
||||
shape_n=intermediate_size_per_partition * 2,
|
||||
shape_n=intermediate_size_per_partition * self.moe.w13_num_shards,
|
||||
shape_k=hidden_size,
|
||||
num_experts=num_experts,
|
||||
param_dtype=params_dtype,
|
||||
|
||||
Reference in New Issue
Block a user