Merge branch 'main' into bugfix-fusedmoe

Author: wangxiyuan

.github/workflows/_e2e_test.yaml

@@ -103,12 +103,12 @@ jobs:
           pytest -sv tests/e2e/singlecard/test_sampler.py
           pytest -sv tests/e2e/singlecard/test_vlm.py
           pytest -sv tests/e2e/singlecard/multi-modal/test_internvl.py
+          pytest -sv tests/e2e/singlecard/test_xlite.py
 
           # ------------------------------------ v1 spec decode test ------------------------------------ #
           pytest -sv tests/e2e/singlecard/spec_decode_v1/test_v1_mtp_correctness.py
           pytest -sv tests/e2e/singlecard/spec_decode_v1/test_v1_mtp_torchair_correctness.py
-          # Fix me: test_eagle_correctness OOM error
-          #pytest -sv tests/e2e/singlecard/spec_decode_v1/test_v1_spec_decode.py
+          pytest -sv tests/e2e/singlecard/spec_decode_v1/test_v1_spec_decode.py
 
   e2e-2-cards:
     name: multicard-2

docs/source/tutorials/DeepSeek-V3.1.md

@@ -430,7 +430,6 @@ vllm serve /weights/DeepSeek-V3.1_w8a8mix_mtp \
   "engine_id": "0",
   "kv_connector_module_path": "vllm_ascend.distributed.mooncake_connector",
   "kv_connector_extra_config": {
-            "use_ascend_direct": true,
             "prefill": {
                     "dp_size": 2,
                     "tp_size": 8
@@ -510,7 +509,6 @@ vllm serve /weights/DeepSeek-V3.1_w8a8mix_mtp \
   "engine_id": "1",
   "kv_connector_module_path": "vllm_ascend.distributed.mooncake_connector",
   "kv_connector_extra_config": {
-            "use_ascend_direct": true,
             "prefill": {
                     "dp_size": 2,
                     "tp_size": 8
@@ -590,7 +588,6 @@ vllm serve /weights/DeepSeek-V3.1_w8a8mix_mtp \
   "engine_id": "2",
   "kv_connector_module_path": "vllm_ascend.distributed.mooncake_connector",
   "kv_connector_extra_config": {
-            "use_ascend_direct": true,
             "prefill": {
                     "dp_size": 2,
                     "tp_size": 8
@@ -670,7 +667,6 @@ vllm serve /weights/DeepSeek-V3.1_w8a8mix_mtp \
   "engine_id": "3",
   "kv_connector_module_path": "vllm_ascend.distributed.mooncake_connector",
   "kv_connector_extra_config": {
-            "use_ascend_direct": true,
             "prefill": {
                     "dp_size": 2,
                     "tp_size": 8

docs/source/user_guide/configuration/additional_config.md

@@ -26,6 +26,7 @@ The following table lists additional configuration options available in vLLM Asc
 
 | Name                                | Type | Default | Description                                                                                                                                   |
 |-------------------------------------|------|---------|-----------------------------------------------------------------------------------------------------------------------------------------------|
+| `xlite_graph_config`                | dict | `{}`    | Configuration options for xlite graph mode                                                                                                    |
 | `torchair_graph_config`             | dict | `{}`    | Configuration options for torchair graph mode                                                                                                    |
 | `weight_prefetch_config`            | dict | `{}`    | Configuration options for weight prefetch                                                                                                        |
 | `refresh`                           | bool | `false` | Whether to refresh global Ascend configuration content. This is usually used by rlhf or ut/e2e test case.                                      |
@@ -45,6 +46,12 @@ The following table lists additional configuration options available in vLLM Asc
 
 The details of each configuration option are as follows:
 
+**xlite_graph_config**
+| Name | Type | Default | Description |
+| ---- | ---- | ------- | ----------- |
+| `enabled` | bool | `False` | Whether to enable xlite graph mode. Currently only Llama or Qwen dense series models are supported. |
+| `full_mode` | bool | `False` | Whether to enable xlite for both the prefill and decode stages. By default, xlite is only enabled for the decode stage. |
+
 **torchair_graph_config**
 
 | Name | Type | Default | Description |

docs/source/user_guide/feature_guide/graph_mode.md

@@ -10,9 +10,10 @@ This guide provides instructions for using Ascend Graph Mode with vLLM Ascend. P
 
 From v0.9.1rc1 with V1 Engine, vLLM Ascend will run models in graph mode by default to keep the same behavior with vLLM. If you hit any issues, please feel free to open an issue on GitHub and fallback to the eager mode temporarily by setting `enforce_eager=True` when initializing the model.
 
-There are two kinds for graph mode supported by vLLM Ascend:
+There are three kinds for graph mode supported by vLLM Ascend:
 - **ACLGraph**: This is the default graph mode supported by vLLM Ascend. In v0.9.1rc1, Qwen and Deepseek series models are well tested.
 - **TorchAirGraph**: This is the GE graph mode. In v0.9.1rc1, only DeepSeek series models are supported.
+- **XliteGraph**: This is the euler xlite graph mode. In v0.11.0, only Llama and Qwen dense serise models are supported.
 
 ## Using ACLGraph
 ACLGraph is enabled by default. Take Qwen series models as an example, just set to use V1 Engine is enough.
@@ -57,9 +58,36 @@ vllm serve path/to/DeepSeek-R1-0528 --additional-config='{"torchair_graph_config
 
 You can find more details about additional configuration [here](../configuration/additional_config.md).
 
+## Using XliteGraph
+
+If you want to run Llama or Qwen dense series models with xlite graph mode, please install xlite, and set xlite_graph_config.
+
+```bash
+pip install xlite
+```
+
+Offline example:
+
+```python
+import os
+from vllm import LLM
+
+# xlite supports the decode-only mode by default, and the full mode can be enabled by setting: "full_mode": True
+model = LLM(model="path/to/Qwen3-32B", tensor_parallel_size=8, additional_config={"xlite_graph_config": {"enabled": True, "full_mode": True}})
+outputs = model.generate("Hello, how are you?")
+```
+
+Online example:
+
+```shell
+vllm serve path/to/Qwen3-32B --tensor-parallel-size 8 --additional-config='{"xlite_graph_config": {"enabled": true, "full_mode": true}}'
+```
+
+You can find more details abort xlite [here](https://gitee.com/openeuler/GVirt/blob/master/xlite/README.md)
+
 ## Fallback to the Eager Mode
 
-If both `ACLGraph` and `TorchAirGraph` fail to run, you should fallback to the eager mode.
+If `ACLGraph`, `TorchAirGraph` and `XliteGraph` all fail to run, you should fallback to the eager mode.
 
 Offline example:
 

docs/source/user_guide/feature_guide/kv_pool.md

@@ -41,7 +41,6 @@ The environment variable **MOONCAKE_CONFIG_PATH** is configured to the full path
     "metadata_server": "P2PHANDSHAKE",
     "protocol": "ascend",
     "device_name": "",
-    "use_ascend_direct": true,
     "alloc_in_same_node": true,
     "master_server_address": "xx.xx.xx.xx:50088",
     "global_segment_size": "1GB" (1024MB/1048576KB/1073741824B/1073741824)
@@ -52,7 +51,6 @@ The environment variable **MOONCAKE_CONFIG_PATH** is configured to the full path
 **metadata_server**: Configured as **P2PHANDSHAKE**.  
 **protocol:** Configured for Ascend to use Mooncake's HCCL communication.  
 **device_name**: ""  
-**use_ascend_direct**: Indicator for using ADXL engine.  
 **alloc_in_same_node**: Indicator for preferring local buffer allocation strategy.  
 **master_server_address**: Configured with the IP and port of the master service.  
 **global_segment_size**: Expands the kvcache size registered by the PD node to the master.
@@ -133,7 +131,7 @@ python3 -m vllm.entrypoints.openai.api_server \
                     }  
         ]
     }
-    }' > p.log 2>&1
+    }'
 ```
 
 `decode` Node：
@@ -177,7 +175,6 @@ python3 -m vllm.entrypoints.openai.api_server \
                 "kv_role": "kv_consumer",
                 "kv_port": "20002",
                 "kv_connector_extra_config": {
-                    "use_ascend_direct": true,
                     "prefill": {
                         "dp_size": 1,
                         "tp_size": 1
@@ -196,7 +193,7 @@ python3 -m vllm.entrypoints.openai.api_server \
             }
         ]
     }
-    }' > d.log 2>&1
+    }'
 ```
 
 #### 2、Start proxy_server.

examples/external_online_dp/run_dp_template.sh

@@ -29,4 +29,4 @@ vllm serve model_path \
     --trust-remote-code \
     --gpu-memory-utilization 0.9 \
     --quantization ascend \
-    --speculative-config '{"num_speculative_tokens": 1, "method":"deepseek_mtp"}' \
+    --speculative-config '{"num_speculative_tokens": 1, "method":"mtp"}'

mypy.ini

@@ -27,3 +27,6 @@ ignore_missing_imports = True
 [mypy-msprobe.*]
 ignore_missing_imports = True
 allow_untyped_imports = True
+
+[mypy-xlite.*]
+ignore_missing_imports = True

requirements-dev.txt

@@ -20,4 +20,5 @@ soundfile
 pytest_mock
 msserviceprofiler>=1.2.2
 mindstudio-probe>=8.3.0
-arctic-inference==0.1.1
+arctic-inference==0.1.1
+xlite

tests/e2e/nightly/features/test_mtpx_deepseek_r1_0528_w8a8.py

@@ -74,10 +74,7 @@ async def test_models(model: str, mode: str) -> None:
         "VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS": "3600000"
     }
     additional_config: dict[str, Any] = {}
-    speculative_config = {
-        "num_speculative_tokens": 2,
-        "method": "deepseek_mtp"
-    }
+    speculative_config = {"num_speculative_tokens": 2, "method": "mtp"}
     compilation_config = {
         "cudagraph_capture_sizes": [56],
         "cudagraph_mode": "FULL_DECODE_ONLY"

tests/e2e/nightly/features/test_prefix_cache_deepseek_r1_0528_w8a8.py

@@ -84,10 +84,7 @@ async def test_models(model: str) -> None:
         "chunked_prefill_for_mla": True,
         "enable_weight_nz_layout": True
     }
-    speculative_config = {
-        "num_speculative_tokens": 1,
-        "method": "deepseek_mtp"
-    }
+    speculative_config = {"num_speculative_tokens": 1, "method": "mtp"}
     server_args = [
         "--quantization", "ascend", "--data-parallel-size", "2",
         "--tensor-parallel-size", "8", "--enable-expert-parallel", "--port",