fix mla swap_states

Signed-off-by: MengqingCao <cmq0113@163.com>

Author: MengqingCao

vllm_ascend/attention/attention_v1.py

@@ -27,11 +27,11 @@
 from vllm.forward_context import ForwardContext, get_forward_context
 from vllm.utils import direct_register_custom_op
 from vllm.v1.core.sched.output import SchedulerOutput
-from vllm.v1.worker.gpu_input_batch import InputBatch
 
 from vllm_ascend.ops.attention import vanilla_chunked_prefill
 from vllm_ascend.utils import (ACL_FORMAT_FRACTAL_NZ, aligned_16, is_310p,
                                nd_to_nz_2d, nd_to_nz_spec)
+from vllm_ascend.worker.npu_input_batch import InputBatch
 
 
 class AscendAttentionBackend(AttentionBackend):

vllm_ascend/attention/mla_v1.py

@@ -25,7 +25,8 @@
 
 if TYPE_CHECKING:
     from vllm.v1.core.sched.output import SchedulerOutput
-    from vllm.v1.worker.gpu_input_batch import InputBatch
+
+    from vllm_ascend.worker.npu_input_batch import InputBatch
 
 
 @dataclass

vllm_ascend/pool/__init__.py

@@ -26,6 +26,7 @@
 from vllm.multimodal.inputs import MultiModalKwargs, PlaceholderRange
 from vllm.pooling_params import PoolingParams
 from vllm.sampling_params import SamplingParams, SamplingType
+from vllm.utils import swap_dict_values
 from vllm.v1.outputs import LogprobsTensors
 from vllm.v1.sample.metadata import SamplingMetadata
 from vllm.v1.utils import copy_slice
@@ -423,6 +424,64 @@ def remove_request(self, req_id: str) -> Optional[int]:
         self.pooling_params.pop(req_id, None)
         return req_index
 
+    def swap_states(self, i1: int, i2: int) -> None:
+        old_id_i1 = self._req_ids[i1]
+        old_id_i2 = self._req_ids[i2]
+        self._req_ids[i1], self._req_ids[i2] =\
+            self._req_ids[i2], self._req_ids[i1] # noqa
+        self.req_output_token_ids[i1], self.req_output_token_ids[i2] =\
+            self.req_output_token_ids[i2], self.req_output_token_ids[i1]
+        assert old_id_i1 is not None and old_id_i2 is not None
+        self.req_id_to_index[old_id_i1], self.req_id_to_index[old_id_i2] =\
+            self.req_id_to_index[old_id_i2], self.req_id_to_index[old_id_i1]
+        self.num_tokens[i1], self.num_tokens[i2] =\
+            self.num_tokens[i2], self.num_tokens[i1]
+        self.num_tokens_no_spec[i1], self.num_tokens_no_spec[i2] =\
+            self.num_tokens_no_spec[i2], self.num_tokens_no_spec[i1]
+        self.num_prompt_tokens[i1], self.num_prompt_tokens[i2] =\
+            self.num_prompt_tokens[i2], self.num_prompt_tokens[i1]
+        self.num_computed_tokens_cpu[i1], self.num_computed_tokens_cpu[i2] =\
+            self.num_computed_tokens_cpu[i2], self.num_computed_tokens_cpu[i1]
+        self.temperature_cpu[i1], self.temperature_cpu[i2] =\
+            self.temperature_cpu[i2], self.temperature_cpu[i1]
+        self.top_p_cpu[i1], self.top_p_cpu[i2] =\
+            self.top_p_cpu[i2], self.top_p_cpu[i1]
+        self.top_k_cpu[i1], self.top_k_cpu[i2] =\
+            self.top_k_cpu[i2], self.top_k_cpu[i1]
+        self.frequency_penalties_cpu[i1], self.frequency_penalties_cpu[i2] =\
+            self.frequency_penalties_cpu[i2], self.frequency_penalties_cpu[i1]
+        self.presence_penalties_cpu[i1], self.presence_penalties_cpu[i2] =\
+            self.presence_penalties_cpu[i2], self.presence_penalties_cpu[i1]
+        self.repetition_penalties_cpu[i1], self.repetition_penalties_cpu[i2] =\
+            self.repetition_penalties_cpu[i2], self.repetition_penalties_cpu[i1]
+        self.min_p_cpu[i1], self.min_p_cpu[i2] =\
+            self.min_p_cpu[i2], self.min_p_cpu[i1]
+
+        # NOTE: the following is unsafe
+        # self.token_ids_cpu[i1, ...], self.token_ids_cpu[i2, ...], =\
+        #     self.token_ids_cpu[i2, ...], self.token_ids_cpu[i1, ...]
+        # instead, we need to temporiarily copy the data for one of the indices
+        # TODO(lucas): optimize this by only copying valid indices
+        tmp = self.token_ids_cpu[i1, ...].copy()
+        self.token_ids_cpu[i1, ...] = self.token_ids_cpu[i2, ...]
+        self.token_ids_cpu[i2, ...] = tmp
+
+        swap_dict_values(self.generators, i1, i2)
+        swap_dict_values(self.min_tokens, i1, i2)
+        swap_dict_values(self.bad_words_token_ids, i1, i2)
+
+        self.request_lora_mapping[i1], self.request_lora_mapping[i2] =\
+            self.request_lora_mapping[i2], self.request_lora_mapping[i1]
+        self.logit_bias[i1], self.logit_bias[i2] =\
+            self.logit_bias[i2], self.logit_bias[i1]
+
+        if self.allowed_token_ids_mask_cpu_tensor is not None:
+            self.allowed_token_ids_mask_cpu_tensor[i1], \
+                self.allowed_token_ids_mask_cpu_tensor[i2] =\
+                self.allowed_token_ids_mask_cpu_tensor[i2], \
+                    self.allowed_token_ids_mask_cpu_tensor[i1]
+        self.block_table.swap_row(i1, i2)
+
     def condense(self, empty_req_indices: list[int]) -> None:
         """Move non-empty requests down into lower, empty indices.