[bugfix][LoRA] Fix the lora accuracy issue introduced by the upstream vLLM changed.

tests/e2e/singlecard/test_llama32_lora.py

@@ -3,7 +3,6 @@
 
 from unittest.mock import patch
 
-import pytest
 import vllm
 import vllm.config
 from vllm.lora.request import LoRARequest
@@ -126,7 +125,6 @@ def generate_and_test(llm, llama32_lora_files, tensorizer_config_dict: dict | No
     print("removing lora")
 
 
-@pytest.mark.skip(reason="fix me")
 @patch.dict("os.environ", {"VLLM_USE_MODELSCOPE": "False"})
 def test_llama_lora(llama32_lora_files):
     vllm_model = VllmRunner(

vllm_ascend/worker/model_runner_v1.py

@@ -1848,6 +1848,7 @@ def _determine_batch_execution_and_padding(
         # be improved in model runner v2)
         force_uniform_decode: bool | None = None,
         force_has_lora: bool | None = None,
+        force_num_active_loras: int | None = None,
         num_encoder_reqs: int = 0,
     ) -> tuple[CUDAGraphMode, BatchDescriptor, bool, torch.Tensor | None, CUDAGraphStat | None]:
         num_tokens_padded = self._pad_for_sequence_parallelism(num_tokens)
@@ -1864,7 +1865,12 @@ def _determine_batch_execution_and_padding(
         # Encoder-decoder models only support CG for decoder_step > 0 (no enc_output
         # is present). Also, chunked-prefill is disabled, so batch are uniform.
         has_encoder_output = self.model_config.is_encoder_decoder and num_encoder_reqs > 0
-        has_lora = len(self.input_batch.lora_id_to_lora_request) > 0 if force_has_lora is None else force_has_lora
+        num_active_loras = (
+            force_num_active_loras
+            if force_num_active_loras is not None
+            else len(self.input_batch.lora_id_to_lora_request)
+        )
+        has_lora = num_active_loras > 0 if force_has_lora is None else force_has_lora
 
         # ruff: noqa: E731
         def dispatch_cudagraph(num_tokens, disable_full=False, valid_modes=None):
@@ -1877,6 +1883,7 @@ def dispatch_cudagraph(num_tokens, disable_full=False, valid_modes=None):
                     has_lora=has_lora,
                     uniform_decode=uniform_decode,
                     disable_full=disable_full,
+                    num_active_loras=num_active_loras,
                 )
             else:
                 return self.cudagraph_dispatcher.dispatch(
@@ -1885,6 +1892,7 @@ def dispatch_cudagraph(num_tokens, disable_full=False, valid_modes=None):
                     uniform_decode=uniform_decode,
                     valid_modes=valid_modes,
                     invalid_modes={CUDAGraphMode.FULL} if disable_full else None,
+                    num_active_loras=num_active_loras,
                 )
 
         cudagraph_mode, batch_descriptor = dispatch_cudagraph(num_tokens_padded, use_cascade_attn or has_encoder_output)
@@ -2182,7 +2190,6 @@ def _dummy_run(
         allow_microbatching: bool = True,
         skip_eplb: bool = False,
         remove_lora: bool = True,
-        activate_lora: bool = False,
         is_graph_capturing: bool = False,
         num_active_loras: int = 0,
     ) -> tuple[torch.Tensor, torch.Tensor]:
@@ -2246,7 +2253,8 @@ def _dummy_run(
             # `force_has_lora` is used for cudagraph capture; because LoRA is
             # activated later in the context manager, but we need to know the
             # LoRA state when determining the batch descriptor for capture
-            force_has_lora=activate_lora,
+            force_has_lora=num_active_loras > 0,
+            force_num_active_loras=num_active_loras,
         )
         if self.use_cp:
             self.pcp_manager.init_batch_info(
@@ -2321,6 +2329,11 @@ def _dummy_run(
             self.lora_config,
             num_scheduled_tokens,
             num_sampled_tokens,
+            remove_lora,
+            # TODO: The next line is a temporary workaround
+            # to fix the accuracy issue of test_llama32_lora.py,
+            # which is introduced by vllm-project/vllm#32005
+            num_active_loras=(self.lora_config.max_loras if self.lora_config is not None else num_active_loras),
         ):
             # Make sure padding doesn't exceed max_num_tokens
             assert num_tokens_padded <= self.max_num_tokens