From 392d1b4d2eaced844026196b2eea3b284a0e2554 Mon Sep 17 00:00:00 2001 From: Ameen Patel Date: Tue, 7 Jul 2026 08:53:55 -0700 Subject: [PATCH] [BugFix][LoRA] Refresh punica metadata when LoRA slots are reassigned under an unchanged mapping (#47725) Signed-off-by: AmeenP Co-authored-by: Claude Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com> --- tests/lora/test_lora_manager.py | 49 +++++++++++++++++++++++++++++++++ vllm/lora/model_manager.py | 8 +++++- 2 files changed, 56 insertions(+), 1 deletion(-) diff --git a/tests/lora/test_lora_manager.py b/tests/lora/test_lora_manager.py index 80a3b6dd9c..1db05b4b09 100644 --- a/tests/lora/test_lora_manager.py +++ b/tests/lora/test_lora_manager.py @@ -651,6 +651,55 @@ def test_lru_lora_model_manager(default_vllm_config, dist_init, dummy_model, dev assert manager.device == device +@pytest.mark.parametrize("device", DEVICES) +def test_set_adapter_mapping_refreshes_after_slot_reassignment( + default_vllm_config, dist_init, dummy_model, device +): + # An out-of-band add_lora() can LRU-evict and reassign GPU slots while the + # running batch -- and therefore its LoRAMapping -- is unchanged. The + # punica metadata must still be re-derived, otherwise in-flight requests + # are routed to the evicted layout and decode with the wrong adapter. + model = dummy_model + model_lora1 = create_lora(1, model, ["dense1", "dense2", "lm_head"], device=device) + model_lora2 = create_lora(2, model, ["dense1", "dense2", "lm_head"], device=device) + model_lora3 = create_lora(3, model, ["dense1", "dense2", "lm_head"], device=device) + manager = LRUCacheLoRAModelManager( + model, + 2, + 2, + 2, + LoRAConfig( + max_lora_rank=8, max_cpu_loras=3, max_loras=2, lora_dtype=DEFAULT_DTYPE + ), + device=device, + vllm_config=default_vllm_config, + ) + punica_wrapper = manager.punica_wrapper_mapping[DEFAULT_LANGUAGE_WRAPPER_KEY] + + assert manager.add_adapter(model_lora1) + assert manager.activate_adapter(1) + assert manager.add_adapter(model_lora2) + assert manager.activate_adapter(2) + assert manager.lora_index_to_id == [1, 2] + + # Two in-flight requests, one token each, on adapters 1 and 2. + manager.set_adapter_mapping(LoRAMapping((1, 2), (1, 2))) + assert punica_wrapper.token_lora_indices.tolist() == [0, 1] + + # Out-of-band add_lora() with both slots held by the running batch: + # activating 3 evicts 1; re-activating the batch's adapters lands them + # in swapped slots while the batch itself is unchanged. + assert manager.add_adapter(model_lora3) + assert manager.activate_adapter(3) + assert manager.activate_adapter(1) + assert manager.activate_adapter(2) + assert manager.lora_index_to_id == [2, 1] + + # Identical mapping, but the metadata must follow the new slot layout. + manager.set_adapter_mapping(LoRAMapping((1, 2), (1, 2))) + assert punica_wrapper.token_lora_indices.tolist() == [1, 0] + + @pytest.mark.parametrize("device", DEVICES) def test_lru_cache_worker_adapter_manager(dist_init, dummy_model, device, tmp_path): lora_config = LoRAConfig( diff --git a/vllm/lora/model_manager.py b/vllm/lora/model_manager.py index bc3d278af4..8b5d97c1e2 100644 --- a/vllm/lora/model_manager.py +++ b/vllm/lora/model_manager.py @@ -117,6 +117,7 @@ class LoRAModelManager: self.packed_modules: dict[str, list[str]] = {} self.modules: dict[str, BaseLayerWithLoRA] = {} self._last_mapping: LoRAMapping | None = None + self._last_slot_layout: tuple[int | None, ...] | None = None is_moe = is_moe_model(self.model) self._is_moe = is_moe @@ -1147,9 +1148,14 @@ class LoRAModelManager: return True def set_adapter_mapping(self, mapping: LoRAMapping) -> None: - if self._last_mapping != mapping: + # The punica metadata derives from the slot layout as well as the + # mapping: an out-of-band add_lora() can LRU-evict and reassign slots + # while the running batch, and thus the mapping, is unchanged. + slot_layout = tuple(self.lora_index_to_id) + if self._last_mapping != mapping or self._last_slot_layout != slot_layout: self._set_adapter_mapping(mapping) self._last_mapping = mapping + self._last_slot_layout = slot_layout def remove_adapter(self, adapter_id: int) -> bool: self.deactivate_adapter(adapter_id)