fix prefix cache and ep

Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>

Author: wangxiyuan

.github/workflows/vllm_ascend_test_full.yaml

@@ -129,12 +129,12 @@ jobs:
           # We found that if running aclgraph tests in batch, it will cause AclmdlRICaptureBegin error. So we run
           # the test separately.
 
-          #pytest -sv tests/e2e/singlecard/test_aclgraph.py
-          #pytest -sv tests/e2e/singlecard/test_ascend_scheduler.py
-          #pytest -sv tests/e2e/singlecard/test_camem.py
-          #pytest -sv tests/e2e/singlecard/test_chunked.py
-          #pytest -sv tests/e2e/singlecard/test_embedding.py
-          #pytest -sv tests/e2e/singlecard/test_guided_decoding.py
+          pytest -sv tests/e2e/singlecard/test_aclgraph.py
+          pytest -sv tests/e2e/singlecard/test_ascend_scheduler.py
+          pytest -sv tests/e2e/singlecard/test_camem.py
+          pytest -sv tests/e2e/singlecard/test_chunked.py
+          pytest -sv tests/e2e/singlecard/test_embedding.py
+          pytest -sv tests/e2e/singlecard/test_guided_decoding.py
           #pytest -sv tests/e2e/singlecard/test_ilama_lora.py
           pytest -sv tests/e2e/singlecard/test_profile_execute_duration.py
           pytest -sv tests/e2e/singlecard/test_quantization.py
@@ -211,21 +211,21 @@ jobs:
           VLLM_USE_MODELSCOPE: True
         run: |
           #pytest -sv tests/e2e/multicard/test_data_parallel.py
-          #pytest -sv tests/e2e/multicard/test_expert_parallel.py
+          pytest -sv tests/e2e/multicard/test_expert_parallel.py
           # external_launcher test is not stable enough. Fix it later
           # pytest -sv tests/e2e/multicard/test_external_launcher.py
           #pytest -sv tests/e2e/multicard/test_fused_moe_allgather_ep.py
           #pytest -sv tests/e2e/multicard/test_ilama_lora_tp2.py
 
           # To avoid oom, we need to run the test in a single process.
-          pytest -sv tests/e2e/multicard/test_offline_inference_distributed.py::test_models_distributed_QwQ
-          pytest -sv tests/e2e/multicard/test_offline_inference_distributed.py::test_models_distributed_DeepSeek_multistream_moe
-          pytest -sv tests/e2e/multicard/test_offline_inference_distributed.py::test_models_distributed_Qwen3_W8A8
-          pytest -sv tests/e2e/multicard/test_offline_inference_distributed.py::test_models_distributed_Qwen3_W4A8DYNAMIC
-          pytest -sv tests/e2e/multicard/test_offline_inference_distributed.py::test_models_distributed_DeepSeek_W4A8DYNAMIC
-          pytest -sv tests/e2e/multicard/test_offline_inference_distributed.py::test_sp_for_qwen3_moe
-          pytest -sv tests/e2e/multicard/test_offline_inference_distributed.py::test_models_distributed_Qwen_Dense_with_flashcomm_v1
-          pytest -sv tests/e2e/multicard/test_offline_inference_distributed.py::test_models_distributed_Qwen_Dense_with_prefetch_mlp_weight
+          #pytest -sv tests/e2e/multicard/test_offline_inference_distributed.py::test_models_distributed_QwQ
+          #pytest -sv tests/e2e/multicard/test_offline_inference_distributed.py::test_models_distributed_DeepSeek_multistream_moe
+          #pytest -sv tests/e2e/multicard/test_offline_inference_distributed.py::test_models_distributed_Qwen3_W8A8
+          #pytest -sv tests/e2e/multicard/test_offline_inference_distributed.py::test_models_distributed_Qwen3_W4A8DYNAMIC
+          #pytest -sv tests/e2e/multicard/test_offline_inference_distributed.py::test_models_distributed_DeepSeek_W4A8DYNAMIC
+          #pytest -sv tests/e2e/multicard/test_offline_inference_distributed.py::test_sp_for_qwen3_moe
+          #pytest -sv tests/e2e/multicard/test_offline_inference_distributed.py::test_models_distributed_Qwen_Dense_with_flashcomm_v1
+          #pytest -sv tests/e2e/multicard/test_offline_inference_distributed.py::test_models_distributed_Qwen_Dense_with_prefetch_mlp_weight
 
           #pytest -sv tests/e2e/multicard/test_pipeline_parallel.py
           pytest -sv tests/e2e/multicard/test_prefix_caching.py

vllm_ascend/platform.py

@@ -25,6 +25,7 @@
 from torch.distributed.distributed_c10d import PrefixStore
 from vllm.logger import logger
 from vllm.platforms import Platform, PlatformEnum
+from vllm.utils import cdiv
 
 from vllm_ascend.ascend_config import (check_ascend_config, get_ascend_config,
                                        init_ascend_config)
@@ -245,6 +246,12 @@ def check_and_update_config(cls, vllm_config: VllmConfig) -> None:
         if cache_config:
             if cache_config.block_size is None:
                 cache_config.block_size = 128
+            else:
+                target_block_size = cdiv(cache_config.block_size, 64) * 64
+                if target_block_size > 256:
+                    target_block_size = 256
+                cache_config.block_size = target_block_size
+
             if cache_config.enable_prefix_caching and cache_config.block_size != 128:
                 logger.warning(
                     "If prefix caching is enabled, block size must be set to 128."

vllm_ascend/worker/block_table.py

@@ -166,7 +166,9 @@ def compute_slot_mapping(self, req_indices: np.ndarray,
             self.slot_mapping_np[:req_indices.shape[0]] = np.where(
                 mask, slot_mapping, -1)
         else:
-            if self.block_size == self.kernel_sizes[0]:
+            assert self.kernel_sizes is not None
+            if self.block_size == self.kernel_sizes[0] or self.kernel_sizes[
+                    0] == 0:
                 # IMPORTANT: In hybrid mode, positions are in logical block space,
                 # but we need to map them to the correct logical block table indices
                 logical_block_idx = positions // self.block_size