[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
@@ -523,7 +523,7 @@ class BitsAndBytesMoEMethod(FusedMoEMethodBase):
quant_ratio = calculate_quant_ratio(params_dtype)
# Fused gate_up_proj (column parallel)
w13_total_size = (
hidden_size * 2 * intermediate_size_per_partition
hidden_size * self.moe.w13_num_shards * intermediate_size_per_partition
) // quant_ratio
w13_qweight = torch.nn.Parameter(
torch.empty(
@@ -541,10 +541,10 @@ class BitsAndBytesMoEMethod(FusedMoEMethodBase):
{
"num_experts": num_experts,
"input_dim": hidden_size,
"output_dim": 2 * intermediate_size_per_partition,
"output_dim": self.moe.w13_num_shards * intermediate_size_per_partition,
"experts_shape": (
num_experts,
intermediate_size_per_partition * 2,
intermediate_size_per_partition * self.moe.w13_num_shards,
hidden_size,
),
"pack_factor": quant_ratio,