[ROCm] Upgrade MoRI version on rocm dockers required for WideEP DP16 DI CI enablement and fixes for combine API (#56073)

Signed-off-by: Shiksha Patel <shiksha.patel@amd.com>
Signed-off-by: lcskrishna <lollachaitanya@gmail.com>
Co-authored-by: shikamd123 <shikpate@amd.com>
Co-authored-by: Andreas Karatzas <akaratza@amd.com>
This commit is contained in:
Chaitanya Sri Krishna Lolla
2026-09-29 18:07:31 +00:00
committed by GitHub
co-authored by shikamd123 Andreas Karatzas
parent afac509a33
commit 467d81d9a0
4 changed files with 142 additions and 3 deletions
@@ -27,6 +27,7 @@ class MoriPrepareAndFinalize(mk.FusedMoEPrepareAndFinalizeModular):
self.num_dispatchers_ = num_dispatchers
self.max_tokens_per_rank = max_tokens_per_rank
self.use_fp8_dispatch = use_fp8_dispatch
self._dispatch_topk_ids: torch.Tensor | None = None
@property
def activation_format(self) -> mk.FusedMoEActivationFormat:
@@ -92,6 +93,12 @@ class MoriPrepareAndFinalize(mk.FusedMoEPrepareAndFinalizeModular):
# mori expects one scale slot per token; broadcast.
scale = scale.expand(a1.shape[0], 1).contiguous()
# mori's combine() reduces over this rank's own [num_tokens, topk]
# routing. The modular kernel rebinds topk_ids to the dispatched ids
# returned below before it calls finalize(), so hold a reference to
# the pre-dispatch tensor (ROCm/mori#475).
self._dispatch_topk_ids = topk_ids
(
dispatch_a1,
dispatch_weights,
@@ -122,9 +129,13 @@ class MoriPrepareAndFinalize(mk.FusedMoEPrepareAndFinalizeModular):
weight_and_reduce_impl: mk.TopKWeightAndReduce,
) -> None:
num_token = output.shape[0]
assert self._dispatch_topk_ids is not None, (
"finalize() called before prepare(); mori combine() needs the "
"routing tensor that dispatch() was given"
)
result = self.mori_op.combine(
fused_expert_output,
None,
topk_ids,
self._dispatch_topk_ids,
)[0]
output.copy_(result[:num_token])