Support mxfp8 KV cache in PD transfer (#35718)

python/sglang/srt/mem_cache/memory_pool.py

@@ -3403,6 +3403,9 @@ class MHATokenToKVPoolMXFP8(MHATokenToKVPool):
             ],
             device=self.device,
         )
+        # This override replaces the base allocation, so the PD-transfer
+        # descriptors for the packed data buffers are built here too.
+        self._kv_buffer_descs = self._build_kv_buffer_descs()
 
     def _clear_buffers(self):
         del self.k_buffer
@@ -3552,11 +3555,21 @@ class MHATokenToKVPoolMXFP8(MHATokenToKVPool):
     def load_cpu_copy(self, kv_cache_cpu, indices, mamba_indices=None):
         raise NotImplementedError("CPU offloading is unsupported for MXFP8 KV cache.")
 
-    def get_contiguous_buf_infos(self):
-        raise NotImplementedError(
-            "KV transfer / disaggregation is unsupported for MXFP8 KV cache "
-            "(scale buffers are not exposed)."
-        )
+    def get_kv_scale_buf_infos(self):
+        """(ptrs, lens, item_lens) for the UE8M0 scale buffers, k then v.
+
+        The interleaved layout puts pages on the leading axis, so a page's
+        scales are one contiguous row; the flat layout is per slot.
+        """
+        tensors = self.k_scale_buffer + self.v_scale_buffer
+        ptrs = [t.data_ptr() for t in tensors]
+        lens = [t.nbytes for t in tensors]
+        row_bytes = [t[0].nbytes for t in tensors]
+        if self.mxfp8_sf_interleaved:
+            item_lens = row_bytes
+        else:
+            item_lens = [rb * self.page_size for rb in row_bytes]
+        return ptrs, lens, item_lens
 
     def set_kv_buffer_prefix_valid(self, *args, **kwargs):
         raise NotImplementedError(