[Silo] Bulk merge: kernel fixes and features (SplitK, MoE fixes, Qwen3-Next, pa_mqa OOB)

.github/scripts/build_aiter_triton.sh

@@ -9,7 +9,15 @@ dpkg -l | grep rocm || echo "No ROCm packages found."
 echo
 echo "==== Install dependencies and aiter ===="
 git config --global --add safe.directory /workspace
-pip install --upgrade pandas zmq einops numpy==1.26.2
+pip config set global.retries 15
+pip config set global.timeout 120
+pip install --upgrade pandas pyzmq einops numpy==1.26.2 || {
+    echo "WARNING: batch pip install failed, retrying packages individually..."
+    pip install --upgrade pandas || true
+    pip install --upgrade pyzmq || echo "WARNING: pyzmq unavailable (only needed by aiter.dist.shm_broadcast)"
+    pip install --upgrade einops
+    pip install --upgrade "numpy==1.26.2"
+}
 pip uninstall -y aiter || true
 pip install --upgrade "pybind11>=3.0.1"
 pip install --upgrade "ninja>=1.11.1"

.github/workflows/aiter-test.yaml

@@ -102,7 +102,7 @@ jobs:
               shopt -s nullglob &&
               rm -rf dist build aiter_meta ./*.egg-info &&
               pip install -r requirements.txt &&
-              pip install --upgrade pandas zmq einops numpy==1.26.2 &&
+              pip install --upgrade pandas pyzmq einops numpy==1.26.2 &&
               pip install --upgrade "pybind11>=3.0.1" &&
               pip install --upgrade "ninja>=1.11.1" &&
               pip install --upgrade setuptools_scm &&
@@ -372,7 +372,7 @@ jobs:
           bash -lc "
             pip uninstall -y amd-aiter aiter || true
             pip install -r requirements.txt
-            pip install --upgrade pandas zmq einops numpy==1.26.2
+            pip install --upgrade pandas pyzmq einops numpy==1.26.2
             pip install --upgrade 'pybind11>=3.0.1'
             pip install --upgrade 'ninja>=1.11.1'
             pip install tabulate
@@ -573,7 +573,7 @@ jobs:
           bash -lc "
             pip uninstall -y amd-aiter aiter || true
             pip install -r requirements.txt
-            pip install --upgrade pandas zmq einops numpy==1.26.2
+            pip install --upgrade pandas pyzmq einops numpy==1.26.2
             pip install --upgrade 'pybind11>=3.0.1'
             pip install --upgrade 'ninja>=1.11.1'
             pip install tabulate

.github/workflows/vllm_benchmark.yaml

@@ -99,7 +99,7 @@ jobs:
               pip config set global.retries 10
               pip config set global.index-url https://ausartifactory.amd.com/artifactory/api/pypi/hw-cpe-prod-remote/simple
               pip install -r requirements.txt
-              pip install --upgrade pandas zmq einops numpy==1.26.2
+              pip install --upgrade pandas pyzmq einops numpy==1.26.2
               pip install --upgrade "pybind11>=3.0.1"
               pip install --upgrade "ninja>=1.11.1"
               pip install --upgrade "setuptools_scm[toml]>=6.2" wheel packaging psutil