[BugFix]Block count calculation error in MambaManager when hitting the external cache (vllm-project#11129)

### What this PR does / why we need it?
- When external cache hit, MambaManager will allocates 1 block for
external cache + 1 block for new scheduled tokens, But upstream
MambaManager#get_num_blocks_to_allocate says it only needs one extra
block besides the speculative blocks, which isn't right, this pr
overrides the get_num_blocks_to_allocate method to return one more block
for this condition

- vLLM version: v0.23.0
- vLLM main:
vllm-project/vllm@dc68bd8

Signed-off-by: Qingsong Zhang <1640410765@qq.com>

Author: zqs164

vllm_ascend/patch/__init__.py

@@ -951,12 +951,18 @@
 # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 #   1. `vllm.v1.core.single_type_kv_cache_manager.MambaManager`
 #    Why:
-#       Upstream hybrid prefix cache lookup does not support PCP/DCP.
+#       1. Upstream hybrid prefix cache lookup does not support PCP/DCP.
+#       2. Upstream MambaManager#get_num_blocks_to_allocate give the
+#          wrong number of blocks when an external cache hit occurred
 #    How:
-#       Replace MambaManager with AscendMambaManager for prefix cache hit lookup
-#       on hybrid Mamba paths (logical mamba block_size when caching is enabled).
+#       1. Replace MambaManager with AscendMambaManager for prefix cache hit lookup
+#          on hybrid Mamba paths (logical mamba block_size when caching is enabled).
+#       2. Override the get_num_blocks_to_allocate method to fix the number of blocks
+#          when hitting the external cache and loading synchronously
 #    Related PR (if no, explain why):
-#       https://github.com/vllm-project/vllm/pull/40996
+#       1. https://github.com/vllm-project/vllm/pull/40996
+#       2. https://github.com/vllm-project/vllm/pull/46892
 #    Future Plan:
-#       Upstream PR #40996 adds hybrid prefix cache lookup for DCP only; PCP is
-#       not supported yet. Remove this patch once upstream supports both PCP and DCP.
+#       1. Upstream PR #40996 adds hybrid prefix cache lookup for DCP only; PCP is
+#          not supported yet. Remove this patch once upstream supports both PCP and DCP.
+#       2. Remove this patch once upstream accept 46892 pr or fixed the bug by other pr.

vllm_ascend/patch/platform/patch_mamba_manager.py

@@ -4,6 +4,8 @@
 # Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
 # This file is a part of the vllm-ascend project.
 #
+from collections.abc import Sequence
+
 import vllm.v1.core.single_type_kv_cache_manager as single_type_kv_cache_manager
 from vllm.v1.core.single_type_kv_cache_manager import (
     BlockHashList,
@@ -47,5 +49,36 @@ def find_longest_cache_hit(
                 break
         return computed_blocks
 
+    def get_num_blocks_to_allocate(
+        self,
+        request_id: str,
+        num_tokens: int,
+        new_computed_blocks: Sequence[KVCacheBlock],
+        total_computed_tokens: int,
+        num_tokens_main_model: int,
+        apply_admission_cap: bool = False,
+    ) -> int:
+        num_new_blocks = super().get_num_blocks_to_allocate(
+            request_id,
+            num_tokens,
+            new_computed_blocks,
+            total_computed_tokens,
+            num_tokens_main_model,
+            apply_admission_cap,
+        )
+        # When external KV cache is loaded synchronously with new
+        # tokens, allocate_new_computed_blocks() allocates one
+        # extra block to hold the external cache content. Account
+        # for it here so the free-capacity check is accurate.
+        # (External tokens exist when total_computed_tokens exceeds
+        # what local prefix-cache hits cover; sync loading when
+        # num_tokens_main_model exceeds total_computed_tokens.)
+        has_external_tokens = total_computed_tokens > len(new_computed_blocks) * self.block_size
+        has_new_scheduled_tokens = num_tokens_main_model > total_computed_tokens
+        if has_external_tokens and has_new_scheduled_tokens:
+            # one more block for external computed tokens
+            num_new_blocks += 1
+        return num_new_blocks
+
 
 single_type_kv_cache_manager.MambaManager = AscendMambaManager