[AMD] Resolve Qwen3.5 MTP (speculative decoding) radix cache conflict. (#22908)

2026-04-20 18:17:11 -07:00
parent c7a4ebf3c8
commit ac08ebed65
1 changed files with 14 additions and 4 deletions
@@ -2328,6 +2328,16 @@ class ServerArgs:
                    self.disable_overlap_schedule = False
            else:
                if not self.disable_radix_cache:
+                    if is_hip():
+                        # On ROCm, extra_buffer is unsupported.
+                        # Automatically disable radix cache instead.
+                        logger.warning(
+                            f"Speculative decoding for {model_arch} is not compatible "
+                            "with radix cache on ROCm devices. "
+                            "Automatically disabling radix cache."
+                        )
+                        self.disable_radix_cache = True
+                    else:
                        raise ValueError(
                            f"Speculative decoding for {model_arch} is not compatible with radix cache when using --mamba-scheduler-strategy no_buffer."
                            "To use radix cache with speculative decoding, please use --mamba-scheduler-strategy extra_buffer and set SGLANG_ENABLE_SPEC_V2=1."