build: move the MUSA stack to torch 2.11.0.post2, torchada 0.1.90 and deep_ep post2

README.md

@@ -35,8 +35,8 @@ vLLM-MUSA is the Moore Threads backend for [vLLM](https://github.com/vllm-projec
 providing an OpenAI-compatible inference and serving engine for MUSA GPUs.
 
 - v0.28.0-dev development line with the vLLM V1 engine.
-- Uses the pinned MUSA stack: PyTorch/torch_musa 2.11.0.post1 (MUSA 5.2.0),
-  MATE 0.2.6, and torchada 0.1.83; this branch remains an upgrade candidate.
+- Uses the pinned MUSA stack: PyTorch/torch_musa 2.11.0.post2 (MUSA 5.2.0),
+  MATE 0.2.7, and torchada 0.1.90; this branch remains an upgrade candidate.
 - MUSA-native attention, communication, custom ops, and compilation support.
 - Per-checkpoint recipes for Qwen and DeepSeek-V4-Flash on S5000.
 

README_CN.md

@@ -34,8 +34,8 @@ vLLM-MUSA 是摩尔线程为 [vLLM](https://github.com/vllm-project/vllm) 提供
 为 MUSA GPU 提供 OpenAI 兼容接口的推理服务引擎。
 
 - v0.28.0-dev 系列开发版本，采用 vLLM V1 引擎。
-- 使用固定的 MUSA 软件栈：PyTorch/torch_musa 2.11.0.post1（MUSA 5.2.0）、
-  MATE 0.2.6 和 torchada 0.1.83；当前分支仍属于升级候选版本。
+- 使用固定的 MUSA 软件栈：PyTorch/torch_musa 2.11.0.post2（MUSA 5.2.0）、
+  MATE 0.2.7 和 torchada 0.1.90；当前分支仍属于升级候选版本。
 - 提供 MUSA 原生注意力机制、通信、自定义算子和编译支持。
 - 针对 S5000 上 Qwen 和 DeepSeek-V4-Flash 的各个模型检查点提供配置示例。
 

docker/README.md

@@ -59,7 +59,7 @@ bash docker/build_image.sh
 With the defaults this produces:
 
 ```
-vllm-musa:ubuntu22.04_py3.10_musa_runtime_5.2_pytorch_release_2.11.0.post1_musa5.2.0
+vllm-musa:ubuntu22.04_py3.10_musa_runtime_5.2_pytorch_release_2.11.0.post2_musa5.2.0
 ```
 
 The release image accepts the model and engine arguments directly. This block
@@ -228,7 +228,7 @@ Two build-time details make this work on such a host:
 ```bash
 docker run --rm <MUSA GPU flags> \
   --entrypoint python \
-  vllm-musa:ubuntu22.04_py3.10_musa_runtime_5.2_pytorch_release_2.11.0.post1_musa5.2.0 \
+  vllm-musa:ubuntu22.04_py3.10_musa_runtime_5.2_pytorch_release_2.11.0.post2_musa5.2.0 \
   -c "import torch, torch_musa; print('musa available:', torch.musa.is_available())"
 ```
 

docs/installation-cn.md

@@ -35,7 +35,7 @@ docker pull "${VLLM_MUSA_IMAGE}"
 该 registry tag 独立于当前分支发布，可能使用不同的依赖版本。请将镜像内已安装的
 软件包版本与分支固定版本核对；需要完全一致时，请使用下文构建的本地镜像。
 在本次 backport 时，已发布标签中的 MATE/Flash 为 0.2.4、tilelang_musa 为 0.1.8，
-而当前分支分别固定为 0.2.6 和 0.1.12。
+而当前分支分别固定为 0.2.7 和 0.1.12。
 
 镜像以 `vllm serve` 为入口点；在模型路径后追加
 [服务配置示例（英文）](cookbook/README.md)中的引擎参数，并为 `docker run`

docs/installation.md

@@ -40,7 +40,7 @@ different dependency revision. Verify the installed package versions against
 the branch pins; build and use the local image below when exact parity is
 required.
 At the time of this backport, the published tag reports MATE/Flash 0.2.4 and
-tilelang_musa 0.1.8, while this branch pins 0.2.6 and 0.1.12 respectively.
+tilelang_musa 0.1.8, while this branch pins 0.2.7 and 0.1.12 respectively.
 
 The image uses `vllm serve` as its entrypoint, so pass the model path followed
 by the engine arguments from the [serving cookbook](cookbook/README.md). Apply

requirements/common.txt

@@ -1,5 +1,5 @@
 # Runtime Python dependencies that are portable across MUSA images.
-torchada==0.1.83
+torchada==0.1.90
 mthreads-ml-py>=2.4.2
 numpy==1.26
 openai>=2.24.0

requirements/musa_private.txt

@@ -12,18 +12,18 @@
 # projects on public PyPI, so they must resolve from the Moore Threads index
 # only. Versions below track the MUSA release_5.2.0 Python Wheel guide.
 
-torch==2.11.0.post1+musa5.2.0
-torch_musa==2.11.0.post1+musa5.2.0
-torchvision==0.26.0.post1+musa5.2.0
-torchaudio==2.11.0+musa5.2.0
+torch==2.11.0.post2+musa5.2.0
+torch_musa==2.11.0.post2+musa5.2.0
+torchvision==0.26.0.post2+musa5.2.0
+torchaudio==2.11.0.post2+musa5.2.0
 mate==0.2.7
 mate-mubin==0.2.7
 flash_attn_3==0.2.7+musa
 flash_mla==0.2.7+musa
 deep-gemm==0.2.7+musa
 flashinfer-python==0.2.7+musa
 sageattention==0.2.7+musa
-deep_ep==1.1.0+musa5.2.0torch2.11.0.post1
+deep_ep==1.1.0+musa5.2.0torch2.11.0.post2
 tilelang_musa==0.1.12+musa.2
 # The MUSA 5.2.0 Triton backend remains pinned independently of the PyTorch
 # release; verify this pin in the 2.11 image before changing it.

tests/test_inplace_porting_contract.py

@@ -16,14 +16,14 @@
         ),
         "torchada": "torchada==0.1.77",
     },
-    "torch==2.11.0.post1+musa5.2.0": {
+    "torch==2.11.0.post2+musa5.2.0": {
         "private": (
-            "torch_musa==2.11.0.post1+musa5.2.0",
-            "torchvision==0.26.0.post1+musa5.2.0",
-            "torchaudio==2.11.0+musa5.2.0",
-            "deep_ep==1.1.0+musa5.2.0torch2.11.0.post1",
+            "torch_musa==2.11.0.post2+musa5.2.0",
+            "torchvision==0.26.0.post2+musa5.2.0",
+            "torchaudio==2.11.0.post2+musa5.2.0",
+            "deep_ep==1.1.0+musa5.2.0torch2.11.0.post2",
         ),
-        "torchada": "torchada==0.1.83",
+        "torchada": "torchada==0.1.90",
     },
 }
 
@@ -55,23 +55,23 @@ def _declared_musa_stack():
 def test_supported_musa_stack_contract_cases_are_explicit():
     assert set(SUPPORTED_MUSA_STACKS) == {
         "torch==2.9.1.post1+musa5.2.0",
-        "torch==2.11.0.post1+musa5.2.0",
+        "torch==2.11.0.post2+musa5.2.0",
     }
     assert (
         SUPPORTED_MUSA_STACKS["torch==2.9.1.post1+musa5.2.0"]["torchada"]
         == "torchada==0.1.77"
     )
     assert (
-        SUPPORTED_MUSA_STACKS["torch==2.11.0.post1+musa5.2.0"]["torchada"]
-        == "torchada==0.1.83"
+        SUPPORTED_MUSA_STACKS["torch==2.11.0.post2+musa5.2.0"]["torchada"]
+        == "torchada==0.1.90"
     )
     assert (
         "torchvision==0.24.1.post1+musa5.2.0"
         in SUPPORTED_MUSA_STACKS["torch==2.9.1.post1+musa5.2.0"]["private"]
     )
     assert (
-        "torchvision==0.26.0.post1+musa5.2.0"
-        in SUPPORTED_MUSA_STACKS["torch==2.11.0.post1+musa5.2.0"]["private"]
+        "torchvision==0.26.0.post2+musa5.2.0"
+        in SUPPORTED_MUSA_STACKS["torch==2.11.0.post2+musa5.2.0"]["private"]
     )
     for torch_pin, expected in SUPPORTED_MUSA_STACKS.items():
         assert _select_musa_stack({torch_pin}) is expected