[AMD] change fused rms quant interface for aiter upgrade (#14497)

2025-12-09 01:09:23 +08:00
parent 9a327bdfcf
commit 763888b5a8
2 changed files with 3 additions and 3 deletions
--- a/python/sglang/srt/models/deepseek_v2.py
+++ b/python/sglang/srt/models/deepseek_v2.py
@@ -1839,7 +1839,7 @@ class DeepseekV2AttentionMLA(nn.Module):
                current_stream.wait_stream(self.alt_stream)
            else:
                if _use_aiter_gfx95 and self.q_b_proj.weight.dtype == torch.uint8:
-                    q, k_nope, *_ = fused_rms_mxfp4_quant(
+                    q, _, k_nope, *_ = fused_rms_mxfp4_quant(
                        q,
                        self.q_a_layernorm.weight,
                        self.q_a_layernorm.variance_epsilon,