[Feature] Unified memory: support decode context parallelism for Kimi-Linear (#36890)

test/registered/unit/model_executor/test_unified_out_cache_loc_rebind.py

@@ -70,6 +70,9 @@ def _armed_source(v2p, swa_map):
     )
     src.is_translating = True
     src._translate_full = lambda t, out=None: v2p[t.to(torch.int64)]
+    # The WRITE loc has its own translate because under DCP it arrives widened;
+    # at dcp_size == 1 it is the read translate, so arm it with the same fake.
+    src._translate_write_full = src._translate_full
     # Phase 2 derives from kernel-facing values through p2v + the swa v2p; arm
     # the inverse of the fake v2p (ps=1, both multipliers 1: kernel == physical,
     # and the expected swa loc for virtual t is swa_map[t]).