[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:
aoshen02
2026-08-05 17:24:47 -07:00
committed by GitHub
co-authored by Matej Sirovatka Rishi Puri Hsiao-Yuan Chen Claude Opus 5
parent 811622c410
commit 7c77868cdf
16 changed files with 104 additions and 80 deletions
@@ -592,7 +592,9 @@ class AutoAWQMoEMethod(FusedMoEMethodBase):
torch.empty(
num_experts,
hidden_size,
2 * intermediate_size_per_partition // self.quant_config.pack_factor,
self.moe.w13_num_shards
* intermediate_size_per_partition
// self.quant_config.pack_factor,
dtype=torch.int32,
),
requires_grad=False,
@@ -623,7 +625,7 @@ class AutoAWQMoEMethod(FusedMoEMethodBase):
torch.empty(
num_experts,
num_groups_w13,
intermediate_size_per_partition * 2,
intermediate_size_per_partition * self.moe.w13_num_shards,
dtype=params_dtype,
),
requires_grad=False,
@@ -644,7 +646,9 @@ class AutoAWQMoEMethod(FusedMoEMethodBase):
torch.empty(
num_experts,
num_groups_w13,
2 * intermediate_size_per_partition // self.quant_config.pack_factor,
self.moe.w13_num_shards
* intermediate_size_per_partition
// self.quant_config.pack_factor,
dtype=torch.int32,
),
requires_grad=False,