Docs: give the local onboard sweep the same marker shape as CI

.github/workflows/ci.yml

@@ -595,8 +595,11 @@ jobs:
           # device fault/sync domain, so a fault on that device pays a multi-minute
           # teardown (issue #1425). Keep EVERY SDMA test out of this general sweep
           # — which also runs the aicore_op_timeout fault-injection test — and run
-          # them on their own dedicated device(s) in the SDMA step below, so an
-          # SDMA fault's slow teardown can never mix with ordinary fault recovery.
+          # them in the SDMA step below, after this sweep, so an SDMA fault's slow
+          # teardown can never mix with ordinary fault recovery. Device
+          # disjointness holds only on aarch64 (`task-submit --device auto`); the
+          # x86_64 branch has no task-submit and shares ${DEVICE_RANGE}, so there
+          # ordering is the separation.
           #
           # Selected by marker, not by path. `pytest --ignore=<moved path>` exits
           # 0 and says nothing, so a rename used to drop the quarantine silently
@@ -609,8 +612,8 @@ jobs:
               --run "python -m pytest examples tests/st -m 'not sdma' --platform a2a3 --device \$TASK_DEVICE -v --require-pto-isa --pto-session-timeout 600"
           fi
 
-      # SDMA pytest — every SDMA test runs here on its own dedicated device(s),
-      # isolated from the general sweep's aicore_op_timeout fault-injection test
+      # SDMA pytest — every SDMA test runs here, after the sweep above, isolated
+      # from the general sweep's aicore_op_timeout fault-injection test
       # (issue #1425). Both demos create their Worker with enable_sdma=True so the
       # runtime provisions the SDMA workspace: prefetch_async_demo on 1 device (L2),
       # sdma_async_completion_demo on 2 devices (L3, workspace provisioned per chip

README.md

@@ -71,7 +71,8 @@ See runtime docs per arch: [a2a3](src/a2a3/docs/runtimes.md), [a5](src/a5/docs/r
 pytest examples tests/st --platform a2a3sim
 
 # Hardware scene tests (requires Ascend device)
-pytest examples tests/st --platform a2a3 --device 4-7
+# SDMA cases are quarantined by marker, as in CI; run them separately with -m sdma
+pytest examples tests/st -m "not sdma" --platform a2a3 --device 4-7
 
 # Python unit tests
 pytest tests/ut -m "not requires_hardware" -v

docs/testing.md

@@ -46,8 +46,13 @@ ctest --test-dir tests/ut/cpp/build -L "^requires_hardware(_a2a3)?$" --output-on
 # Scene tests (pytest, @scene_test classes)
 pytest examples tests/st                          # all sim platforms (auto-parametrized)
 pytest examples tests/st --platform a2a3sim       # specific sim
-pytest examples tests/st --platform a2a3          # hardware
-pytest examples tests/st --platform a2a3 --device 4-7  # hardware with device pool
+pytest examples tests/st -m "not sdma" --platform a2a3          # hardware
+pytest examples tests/st -m "not sdma" --platform a2a3 --device 4-7  # hardware with device pool
+
+# SDMA cases run separately, as they do in CI: they are quarantined by
+# @pytest.mark.sdma so no fault-injection case shares a device with a
+# provisioned SDMA workspace (issue #1425)
+pytest examples tests/st -m sdma --platform a2a3 --device 4-5
 
 # Single scene test (standalone)
 python examples/a2a3/tensormap_and_ringbuffer/vector_example/test_vector_example.py -p a2a3sim
@@ -625,8 +630,8 @@ pytest examples tests/st --platform a2a3sim
 # Standalone (single case)
 python test_my_kernel.py -p a2a3sim
 
-# On hardware
-pytest examples tests/st --platform a2a3
+# On hardware (SDMA cases quarantined by marker; run them with -m sdma)
+pytest examples tests/st -m "not sdma" --platform a2a3
 ```
 
 Key fields:

docs/troubleshooting/a2a3-507899-aicpu-shared-so-fault.md

@@ -63,7 +63,7 @@ exception you must surface the CANN device slog, which is otherwise hidden:
 
    ```bash
    ASCEND_SLOG_PRINT_TO_STDOUT=1 ASCEND_GLOBAL_LOG_LEVEL=1 \
-   python -m pytest examples tests/st --platform a2a3 --device <range> -v \
+   python -m pytest examples tests/st -m "not sdma" --platform a2a3 --device <range> -v \
      --pto-session-timeout 600
    ```
 

docs/user/reference/cli.md

@@ -6,7 +6,7 @@ How you select what runs, turn diagnostics on, and read the artifacts back.
 
 ```bash
 pytest examples tests/st --platform a2a3sim            # simulation, no device
-pytest examples tests/st --platform a2a3 --device 4-7  # hardware
+pytest examples tests/st -m "not sdma" --platform a2a3 --device 4-7  # hardware (SDMA quarantined by marker)
 python examples/my_example/test_my_example.py -p a2a3sim   # standalone, no pytest
 ```
 

examples/README.md

@@ -17,7 +17,7 @@ on both simulators:
 
 ```bash
 pytest examples --platform a2a3sim
-pytest examples --platform a2a3 --device 0-1        # hardware
+pytest examples -m "not sdma" --platform a2a3 --device 0-1        # hardware (SDMA quarantined by marker)
 ```
 
 A single example:

src/a2a3/runtime/host_build_graph/docs/SUBMIT_BY_CLUSTER.md

@@ -210,7 +210,7 @@ python tests/st/a2a3/tensormap_and_ringbuffer/batch_paged_attention/test_batch_p
 Final validation:
 
 ```bash
-pytest examples tests/st --platform a2a3
+pytest examples tests/st -m "not sdma" --platform a2a3   # SDMA cases quarantined by marker; run them with -m sdma
 ```
 
 ## 14. Resolved Decisions

src/a2a3/runtime/tensormap_and_ringbuffer/docs/SUBMIT_BY_CLUSTER.md

@@ -208,7 +208,7 @@ python tests/st/a2a3/tensormap_and_ringbuffer/batch_paged_attention/test_batch_p
 Final validation:
 
 ```bash
-pytest examples tests/st --platform a2a3
+pytest examples tests/st -m "not sdma" --platform a2a3   # SDMA cases quarantined by marker; run them with -m sdma
 ```
 
 ## 14. Resolved Decisions

src/a5/runtime/tensormap_and_ringbuffer/docs/SUBMIT_BY_CLUSTER.md

@@ -208,7 +208,7 @@ python tests/st/a2a3/tensormap_and_ringbuffer/batch_paged_attention/test_batch_p
 Final validation:
 
 ```bash
-pytest examples tests/st --platform a2a3
+pytest examples tests/st -m "not sdma" --platform a2a3   # SDMA cases quarantined by marker; run them with -m sdma
 ```
 
 ## 14. Resolved Decisions