A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 4, 11:15 UTC
[ROCm][Bugfix][GPTOSS]: fix input_ids and expert_map args for quark w4a8 gptoss (#41165)
Signed-off-by: Rohan138 <rohanpotdar138@gmail.com>
Rohan Potdar committed
Apr 29, 2026 at 18:39 UTC
3795d7acf431980e62e738493f437ae2a51549da
1 file changed
+5
-5
vllm/model_executor/layers/quantization/quark/quark_moe.py
+5
-5
@@ -1573,14 +1573,14 @@ class QuarkOCP_MX_MoEMethod_OSS(QuarkOCP_MX_MoEMethod):
1573
1574
def apply_monolithic(
1575
self,
1576
- layer: torch.nn.Module,
1576
+ layer: FusedMoE,
1577
x: torch.Tensor,
1578
router_logits: torch.Tensor,
1579
- expert_map: torch.Tensor | None = None,
1580
- ) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
1579
+ input_ids: torch.Tensor | None = None,
1580
+ ) -> torch.Tensor:
1581
if layer.enable_eplb:
1582
raise NotImplementedError(
1583
- "EPLB not supported for `QuarkW4MXFp4MoEMethod_OSS` yet."
1583
+ f"EPLB not supported for {self.__class__.__name__} yet."
1584
)
1585
1586
from vllm.model_executor.layers.fused_moe.experts.gpt_oss_triton_kernels_moe import ( # noqa: E501
@@ -1597,7 +1597,7 @@ class QuarkOCP_MX_MoEMethod_OSS(QuarkOCP_MX_MoEMethod):
1597
topk=layer.top_k,
1598
renormalize=layer.renormalize,
1599
global_num_experts=layer.global_num_experts,
1600
- expert_map=expert_map,
1600
+ expert_map=layer.expert_map,
1601
quant_config=self.moe_quant_config,
1602
apply_router_weight_on_input=layer.apply_router_weight_on_input,
1603
unpadded_N_w1=self.moe.intermediate_size_per_partition_unpadded * 2,