Merge remote-tracking branch 'refs/remotes/upstream/main' into pr2235-musa-python-overlay-20260819

Author: fay85

docs/src/content/docs/en/cli_reference.md

@@ -170,7 +170,6 @@ xLLM uses gflags to manage service startup parameters. `--model <PATH>` is the o
 | `enable_pd_ooc` | `bool` | `false` | Whether to enable online-offline co-location in disaggregated PD mode. |
 | `disagg_pd_port` | `int32` | `7777` | Listening port for the disaggregated PD bRPC server. |
 | `instance_role` | `string` | `"DEFAULT"` | Instance role, for example `DEFAULT`, `PREFILL`, `DECODE`, or `MIX`. |
-| `kv_cache_transfer_type` | `string` | `"Mooncake"` | KV Cache transfer type, for example `Mooncake`, `LlmDataDist`, or `HCCL`. |
 | `kv_cache_transfer_mode` | `string` | `"PUSH"` | KV Cache transfer mode, for example `PUSH` or `PULL`. |
 | `transfer_listen_port` | `int32` | `26000` | Listening port for KV Cache Transfer. |
 | `kv_push_dst_rotate` | `bool` | `false` | Rotate the destination-worker traversal order in `push_kv_blocks` per KV-split rank to spread incast across decode workers. |

docs/src/content/docs/zh/cli_reference.md

@@ -170,7 +170,6 @@ xLLM 使用 gflags 管理服务启动参数。`--model <PATH>` 是唯一必填
 | `enable_pd_ooc` | `bool` | `false` | 是否在 PD 分离模式下启用在线/离线混部。 |
 | `disagg_pd_port` | `int32` | `7777` | PD 分离 bRPC server 的监听端口。 |
 | `instance_role` | `string` | `"DEFAULT"` | 当前实例角色，例如 `DEFAULT`、`PREFILL`、`DECODE`、`MIX`。 |
-| `kv_cache_transfer_type` | `string` | `"Mooncake"` | KV Cache 传输类型，例如 `Mooncake`、`LlmDataDist`、`HCCL`。 |
 | `kv_cache_transfer_mode` | `string` | `"PUSH"` | KV Cache 传输模式，例如 `PUSH`、`PULL`。 |
 | `transfer_listen_port` | `int32` | `26000` | KV Cache Transfer 的监听端口。 |
 | `kv_push_dst_rotate` | `bool` | `false` | 在 `push_kv_blocks` 中按 KV-split rank 轮转遍历目标 worker，用于分散对 decode worker 的流量。 |

tests/core/framework/batch/batch_test.cpp

@@ -517,6 +517,25 @@ TEST(BatchInputBuilderTest, DSV4FirstChunkSlicesFullRemoteAllocation) {
             2u);
 }
 
+TEST(BatchInputBuilderTest, DSV4KvSplitMapsOneSourceBlockToTwoDecodeBlocks) {
+  BlockManager::Options options;
+  options.num_blocks(8).block_size(32);
+  BlockManagerImpl manager(options);
+  std::vector<Block> blocks = manager.allocate(2);
+  Sequence sequence = make_basic_sequence({1});
+  sequence.add_blocks(BlockType::C4, blocks);
+  const TransferKVInfo full_info =
+      make_info({100, 101, 102, 103}, BlockType::C4);
+
+  const TransferKVInfo info =
+      BatchInputBuilderTestPeer::build_step_transfer_info(
+          full_info, &sequence, /*seq_len=*/64, /*kv_split_size=*/2);
+
+  expect_mapping(info, BlockType::C4, block_ids(blocks), {100, 101, 102, 103});
+  EXPECT_EQ(sequence.kv_state().next_group_transfer_block_idx(BlockType::C4),
+            2u);
+}
+
 TEST(BatchInputBuilderTest, DSV4LaterChunkSkipsExpiredSWABlocks) {
   BlockManager::Options options;
   options.num_blocks(8).block_size(16);

tests/core/framework/config/config_json_test.cpp

@@ -40,6 +40,7 @@ inline constexpr std::string_view kInlineConfig = R"json({
   "max_tokens_per_batch": 8192,
   "max_seqs_per_batch": 64,
   "model_impl": "py",
+  "disable_graph_warmup": true,
   "python_graph_backend": "cudagraphs"
 })json";
 
@@ -112,6 +113,7 @@ class ConfigFlagGuard final {
         old_max_seqs_per_batch_(FLAGS_max_seqs_per_batch),
         old_model_impl_(FLAGS_model_impl),
         old_python_model_path_(FLAGS_python_model_path),
+        old_disable_graph_warmup_(FLAGS_disable_graph_warmup),
         old_python_graph_backend_(FLAGS_python_graph_backend) {}
 
   ~ConfigFlagGuard() {
@@ -122,6 +124,7 @@ class ConfigFlagGuard final {
     FLAGS_max_seqs_per_batch = old_max_seqs_per_batch_;
     FLAGS_model_impl = old_model_impl_;
     FLAGS_python_model_path = old_python_model_path_;
+    FLAGS_disable_graph_warmup = old_disable_graph_warmup_;
     FLAGS_python_graph_backend = old_python_graph_backend_;
   }
 
@@ -133,6 +136,7 @@ class ConfigFlagGuard final {
   int32_t old_max_seqs_per_batch_;
   std::string old_model_impl_;
   std::string old_python_model_path_;
+  bool old_disable_graph_warmup_;
   std::string old_python_graph_backend_;
 };
 
@@ -263,17 +267,25 @@ TEST(ConfigJsonTest, FromJsonUsesParsedOverrides) {
   // model and python_model_path are command-line-only: from_json neither reads
   // them nor touches their gflags, so both keep their pre-call values.
   EXPECT_EQ(model_config.python_model_path(), "");
+  EXPECT_TRUE(execution_config.disable_graph_warmup());
   EXPECT_EQ(execution_config.python_graph_backend(), "cudagraphs");
 
   EXPECT_EQ(FLAGS_model_impl, "py");
   EXPECT_EQ(FLAGS_python_model_path, old_python_model_path);
+  EXPECT_TRUE(FLAGS_disable_graph_warmup);
   EXPECT_EQ(FLAGS_python_graph_backend, "cudagraphs");
 
   EXPECT_EQ(kv_cache_config.kv_cache_dtype(), "auto");
   EXPECT_EQ(kv_cache_config.indexer_cache_dtype(), "auto");
   EXPECT_EQ(scheduler_config.max_decode_token_per_sequence(), 256);
 }
 
+TEST(ExecutionConfigTest, GraphWarmupIsEnabledByDefault) {
+  const ExecutionConfig execution_config;
+
+  EXPECT_FALSE(execution_config.disable_graph_warmup());
+}
+
 TEST(KVCacheConfigValidationTest, AcceptsSupportedIndexerCacheDtypes) {
   KVCacheConfig config;
 

tests/core/framework/config/disagg_pd_config_test.cpp

@@ -34,29 +34,20 @@ struct PrefixRoleCase {
 void set_values_requiring_mlu_normalization(DisaggPDConfig& disagg_pd_config,
                                             KVCacheConfig& kv_cache_config,
                                             SchedulerConfig& scheduler_config) {
-  disagg_pd_config.kv_cache_transfer_type("LlmDataDist")
-      .kv_cache_transfer_mode("PULL")
-      .enable_pd_ooc(true);
+  disagg_pd_config.kv_cache_transfer_mode("PULL").enable_pd_ooc(true);
   kv_cache_config.kv_cache_dtype("fp8").enable_prefix_cache(true);
   scheduler_config.enable_schedule_overlap(true);
 }
 
 void expect_normalized_values(const DisaggPDConfig& disagg_pd_config,
                               const KVCacheConfig& kv_cache_config,
                               const SchedulerConfig& scheduler_config) {
-  EXPECT_EQ(disagg_pd_config.kv_cache_transfer_type(), "Mooncake");
   EXPECT_EQ(disagg_pd_config.kv_cache_transfer_mode(), "PULL");
   EXPECT_FALSE(disagg_pd_config.enable_pd_ooc());
   EXPECT_EQ(kv_cache_config.kv_cache_dtype(), "auto");
   EXPECT_FALSE(scheduler_config.enable_schedule_overlap());
 }
 
-TEST(DisaggPDConfigTest, DefaultsToMooncakeTransfer) {
-  const DisaggPDConfig disagg_pd_config;
-
-  EXPECT_EQ(disagg_pd_config.kv_cache_transfer_type(), "Mooncake");
-}
-
 TEST(DisaggPDConfigTest, KeepsMluPrefixCacheForPrefillSideRoles) {
   const PrefixRoleCase cases[] = {
       {"PREFILL", true},
@@ -83,23 +74,14 @@ TEST(DisaggPDConfigTest, KeepsMluPrefixCacheForPrefillSideRoles) {
   }
 }
 
-TEST(DisaggPDConfigTest, ExposesParallelHeterogeneousShardPullOption) {
-  DisaggPDConfig disagg_pd_config;
-  EXPECT_FALSE(disagg_pd_config.enable_heterogeneous_pd());
-  EXPECT_TRUE(disagg_pd_config.enable_pd_parallel_shard_pull());
-
-  disagg_pd_config.enable_heterogeneous_pd(true);
-  disagg_pd_config.enable_pd_parallel_shard_pull(false);
-  EXPECT_TRUE(disagg_pd_config.enable_heterogeneous_pd());
-  EXPECT_FALSE(disagg_pd_config.enable_pd_parallel_shard_pull());
-
+TEST(DisaggPDConfigTest, OmitsRemovedHeterogeneousPullOptions) {
   const std::vector<std::string>& option_names =
       DisaggPDConfig::option_category().option_names;
-  EXPECT_NE(
+  EXPECT_EQ(
       std::find(
           option_names.begin(), option_names.end(), "enable_heterogeneous_pd"),
       option_names.end());
-  EXPECT_NE(std::find(option_names.begin(),
+  EXPECT_EQ(std::find(option_names.begin(),
                       option_names.end(),
                       "enable_pd_parallel_shard_pull"),
             option_names.end());

tests/core/framework/kv_cache/CMakeLists.txt

@@ -1,5 +1,22 @@
 include(cc_test)
 
+cc_test(
+  NAME
+    cache_layout_builder_test
+  SRCS
+    cache_layout_builder_test.cpp
+  DEPS
+    :kv_cache
+    GTest::gtest_main
+)
+target_link_libraries(cache_layout_builder_test
+                      PUBLIC
+                      Python::Python
+                      $<$<BOOL:${USE_NPU}>:ascendcl>
+                      $<$<BOOL:${USE_NPU}>:hccl>
+                      $<$<BOOL:${USE_NPU}>:c_sec>
+                      $<$<BOOL:${USE_NPU}>:nnopbase>)
+
 cc_test(
   NAME
     kv_cache_estimation_test