Merge branch 'v0.6.2-dev_wm' into 'v0.6.2-dev'

修复单测中test_long_context、test_attention部分错误

See merge request dcutoolkit/deeplearing/vllm!33

tests/kernels/test_attention.py

@@ -206,7 +206,7 @@ def test_paged_attention(
         opcheck(torch.ops._C.paged_attention_v1,
                 (output, query, key_cache, value_cache, num_kv_heads, scale,
                  block_tables, seq_lens, block_size, max_seq_len, alibi_slopes,
-                 kv_cache_dtype, k_scale, v_scale, 0, 0, 0, 64, 0),
+                 kv_cache_dtype, k_scale, v_scale, 0, 0, 0, 64, 0, None, 0),
                 cond=(head_size == HEAD_SIZES[0]
                       and block_size == BLOCK_SIZES[0]))
 
@@ -248,7 +248,7 @@ def test_paged_attention(
                     (output, exp_sums, max_logits, tmp_output, query,
                      key_cache, value_cache, num_kv_heads, scale, block_tables,
                      seq_lens, block_size, max_seq_len, alibi_slopes,
-                     kv_cache_dtype, k_scale, v_scale, 0, 0, 0, 64, 0),
+                     kv_cache_dtype, k_scale, v_scale, 0, 0, 0, 64, 0, None, 0),
                     cond=(head_size == HEAD_SIZES[0]
                           and block_size == BLOCK_SIZES[0]))
 

tests/lora/test_long_context.py

@@ -111,7 +111,7 @@ def lora_llm(long_context_infos):
     llm = vllm.LLM("meta-llama/Llama-2-13b-chat-hf",
                    enable_lora=True,
                    max_num_seqs=16,
-                   max_loras=2,
+                   max_loras=8,
                    long_lora_scaling_factors=tuple(scaling_factors),
                    max_num_batched_tokens=4096 * 8,
                    tensor_parallel_size=4,

tests/lora/test_punica_variation.py

@@ -20,7 +20,7 @@ from vllm.utils import seed_everything
 from .utils import (generate_data, generate_data_for_expand_nslices,
                     ref_torch_groupgemm)
 
-HIDDEN_SIZES = [4097]
+HIDDEN_SIZES = [1024]
 
 BATCHES = [1, 4, 16, 32]
 NUM_LORA = [1, 8, 32, 128]

vllm/lora/request.py

@@ -83,7 +83,9 @@ class LoRARequest(
         and comparison lora adapter across engines.
         """
         return isinstance(value,
-                          self.__class__) and self.lora_name == value.lora_name
+                          self.__class__) and self.lora_name == value.lora_name and \
+                            self.lora_int_id == value.lora_int_id and \
+                            self.lora_path == value.lora_path
 
     def __hash__(self) -> int:
         """