Merge branch 'vllm-project:main' into main

Author: zhaozx-cn

tests/e2e/common.sh

@@ -14,7 +14,7 @@ _err() { _red "Error: $*" && exit 1; }
 
 CURL_TIMEOUT=1
 CURL_COOLDOWN=5
-CURL_MAX_TRIES=180
+CURL_MAX_TRIES=300
 
 function wait_url_ready() {
   local serve_name="$1"

tests/e2e/doctests/001-quickstart-test.sh

@@ -57,8 +57,8 @@ function quickstart_online_test() {
 }
 
 _info "====> Start simple_test"
-simple_test
+time simple_test
 _info "====> Start quickstart_offline_test"
-quickstart_offline_test
+time quickstart_offline_test
 _info "====> Start quickstart_online_test"
-quickstart_online_test
+time quickstart_online_test

tests/e2e/doctests/002-pip-binary-installation-test.sh

@@ -59,4 +59,4 @@ function install_binary_test() {
 }
 
 _info "====> Start install_binary_test"
-install_binary_test
+time install_binary_test

tests/e2e/models/configs/DeepSeek-V2-Lite.yaml

@@ -3,11 +3,11 @@ tasks:
 - name: "gsm8k"
   metrics:
   - name: "exact_match,strict-match"
-    value: 0.375
+    value: 0.385
   - name: "exact_match,flexible-extract"
-    value: 0.375
+    value: 0.385
 tensor_parallel_size: 2
-batch_size: 8
+batch_size: 32
 gpu_memory_utilization: 0.7
 apply_chat_template: False
 fewshot_as_multiturn: False

tests/e2e/run_doctests.sh

@@ -22,7 +22,6 @@ set -eo errexit
 . $(dirname "$0")/common.sh
 
 export VLLM_USE_MODELSCOPE=true
-export VLLM_LOGGING_LEVEL=ERROR
 
 _info "====> Start Quickstart test"
 . "${SCRIPT_DIR}/doctests/001-quickstart-test.sh"

tests/ut/attention/test_attention_v1.py

@@ -435,6 +435,41 @@ def test_forward_decode_only_swa(self, mock_fused_infer_attention_score,
         mock_fused_infer_attention_score.assert_called_once()
         assert output.shape == (10, 8 * 64)
 
+    @patch('torch_npu._npu_reshape_and_cache')
+    @patch('torch_npu._npu_paged_attention')
+    @patch('torch_npu.npu_fused_infer_attention_score')
+    def test_forward_decode_only_swa_seq_len_mismatch(
+            self, mock_fused_infer_attention_score, mock_paged_attention,
+            mock_npu_reshape_and_cache):
+        """Test forward pass in DecodeOnly state when seq)len_mismatch"""
+        query = torch.randn(10, 8 * 64)
+        key = torch.randn(10, 8 * 64)
+        value = torch.randn(10, 8 * 64)
+        kv_cache = torch.empty(2, 5, 128, 8, 64)
+
+        metadata = self.attn_metadata
+        metadata.attn_state = AscendAttentionState.DecodeOnly
+        metadata.seq_lens = torch.tensor([10])  # len == 1 != query.size(0)==10
+        metadata.block_tables = torch.zeros(1, 5, dtype=torch.long)
+        metadata.num_actual_tokens = 10
+        metadata.slot_mapping = torch.zeros(10, dtype=torch.long)
+
+        mock_fused_infer_attention_score.return_value = (torch.ones(10, 8,
+                                                                    64), 1)
+
+        output = self.impl_swa.forward(self.layer_no_quant,
+                                       query,
+                                       key,
+                                       value,
+                                       kv_cache,
+                                       metadata,
+                                       trace_flag=False)
+
+        mock_paged_attention.assert_called_once()
+        mock_fused_infer_attention_score.assert_not_called()
+
+        assert output.shape == (10, 8 * 64)
+
     @patch('vllm_ascend.attention.attention_v1.is_310p', return_value=False)
     @patch('torch_npu._npu_reshape_and_cache')
     @patch('vllm_ascend.attention.attention_v1.vanilla_chunked_prefill')