[ROCm][Bugfix] Fix GPT-OSS Quark MXFP4 MoE loading - emulation buffer not block-aligned (#43979)

Signed-off-by: xuebwang-amd <xuebwang@amd.com>
This commit is contained in:
xuebwang-amd
2026-07-18 03:49:41 +00:00
committed by GitHub
parent da64db78b9
commit f12b80c6ef
3 changed files with 59 additions and 7 deletions
@@ -1063,12 +1063,9 @@ class QuarkOCP_MX_MoEMethod(QuarkMoEMethod):
act_dtype=act_dtype,
moe_parallel_config=moe_parallel_config,
)
# In case quantization emulation backend is used, there is no need to apply
# MXFP4-specific padding logic as the compute happens in higher precision.
if (
self.mxfp4_backend is not None
and self.mxfp4_backend != Mxfp4MoeBackend.EMULATION
):
# Round per-partition sizes up to each backend's requirement. Emulation is
# handled inside the helper too (OCP MX block alignment), so no special-case.
if self.mxfp4_backend is not None:
hidden_size, intermediate_size_per_partition = (
mxfp4_round_up_hidden_size_and_intermediate_size(
self.mxfp4_backend, hidden_size, intermediate_size_per_partition