Fix failed tests

Signed-off-by: Iman Tabrizian <[email protected]>

Author: Tabrizian

cpp/tensorrt_llm/batch_manager/cacheFormatter.cpp

@@ -47,6 +47,9 @@ BlockRange getBlockRangeForSending(BaseKVCacheManager* cacheManager, LlmRequest
 {
     auto poolNum = cacheManager->getBlockManager().getNumPools(
         /*includeBlockScalePools=*/false, /*includeIndexerKCachePools=*/false);
+
+    // Note: When recv side has CP, the requested seqLen is lesser than seqLen on the sender side as seqLen is
+    // distributed among CP ranks. So, we transfer all blocks from send side.
     if (poolNum > 1 || !cacheManager->isEnableBlockReuse() || lastBlockKey.uniqueTokens.size() == 0 || recvSideHasCP)
     {
         // disable reuse path, and vwsa don't support reuse.

cpp/tests/unit_tests/multi_gpu/cacheTransceiverTest.cpp

@@ -1022,7 +1022,8 @@ class AsymmetricalCacheTest : public ::testing::TestWithParam<AsymmetricTestPara
         }
         int kvFactor = mCacheState->getAttentionConfig().mKvFactor;
         int tokensPerBlock = mCacheState->getModelConfig().mTokensPerBlock;
-        int startTokenId = (blockId * mCpSize + mCpRank) * tokensPerBlock;
+        // We don't account for CP here because contextCP is always 1 currently.
+        int startTokenId = blockId * tokensPerBlock;
         int sizePerHead;
         if (isIndexerKCache)
         {
@@ -1117,7 +1118,8 @@ class AsymmetricalCacheTest : public ::testing::TestWithParam<AsymmetricTestPara
         }
         int kvFactor = mCacheState->getAttentionConfig().mKvFactor;
         int tokensPerBlock = mCacheState->getModelConfig().mTokensPerBlock;
-        int startTokenId = (blockId * mCpSize + mCpRank) * tokensPerBlock;
+        // We don't account for CP here because contextCP is always 1 currently.
+        int startTokenId = blockId * tokensPerBlock;
         int sizePerHead;
         if (isIndexerKCache)
         {

tests/integration/defs/accuracy/test_disaggregated_serving.py

@@ -45,7 +45,7 @@ def result(self):
 DuckLLM = namedtuple('DuckLLM', ['args', 'tokenizer', 'generate_async'])
 
 DEFAULT_TEST_TIMEOUT = 1800
-DEFAULT_SERVER_WAITING_TIMEOUT = 1200
+DEFAULT_SERVER_WAITING_TIMEOUT = 3600
 
 
 class MyThreadPoolExecutor(ThreadPoolExecutor):