perf(kernel): optimize Blackwell grouped GEMM primitives

.agents/scripts/monitor_gpu.sh

@@ -23,18 +23,7 @@ while [[ $# -gt 0 ]]; do
     --interval) INTERVAL="$2"; shift 2 ;;
     --log) LOG="$2"; shift 2 ;;
     -h|--help)
-      cat <<'EOF'
-Watch a single GPU for foreign processes (anyone other than the current
-user) appearing during a long-running test. Intended companion to
-`/tir-test`: leave this running in a side terminal while pytest runs, and
-it will alert if someone else lands on the same GPU.
-
-Usage:
-  monitor_gpu.sh                       # uses $CUDA_VISIBLE_DEVICES, defaults to 0
-  monitor_gpu.sh --gpu 3               # watch GPU 3
-  monitor_gpu.sh --gpu 3 --interval 2  # poll every 2 seconds
-  monitor_gpu.sh --log /tmp/gpu.log    # also tee to a log file
-EOF
+      sed -n '2,12p' "$0" | sed 's/^# \{0,1\}//'
       exit 0 ;;
     *) echo "unknown arg: $1" >&2; exit 2 ;;
   esac

.agents/skills/tir-test/SKILL.md

@@ -2,9 +2,13 @@ Run the full TIRX test suite.
 
 ## Steps
 
-1. Select the least busy GPU to avoid conflicts:
+1. Install the kernel package and select the least busy GPU:
    ```bash
+   pip install -e /path/to/tirx-kernels-staging   # or sibling tirx-kernels checkout
    export CUDA_VISIBLE_DEVICES=$(nvidia-smi --query-gpu=index,memory.used --format=csv,noheader,nounits | sort -t',' -k2 -n | head -1 | cut -d',' -f1 | tr -d ' ')
+   export TVM_PATH=/path/to/tvm
+   export PYTHONPATH="${TVM_PATH}/python"
+   export TVM_LIBRARY_PATH="${TVM_PATH}/build/lib"
    ```
 
 2. Start the GPU monitor in the background so we can detect if anyone else lands on the same GPU mid-run:
@@ -15,18 +19,29 @@ Run the full TIRX test suite.
    trap 'kill $MON_PID 2>/dev/null' EXIT
    ```
 
-3. Run the full test suite with xdist parallelism:
+3. Import gate — bench workloads: fail fast if any kernel listed in `workloads.yaml` fails to import:
    ```bash
-   pytest tests/python/tirx/ -n auto
+   python -m tirx_kernels.bench_suite --check-imports
    ```
+   A non-zero exit means a pinned workload kernel failed to import — fix it before proceeding.
 
-4. Stop the monitor and check for foreign GPU usage during the run:
+4. Full kernel import gate (correctness test suite coverage):
+   ```bash
+   python -m tirx_kernels.registry --cc 10 --strict
+   ```
+
+5. Run the full test suite with xdist parallelism:
+   ```bash
+   pytest tests/python/tirx/ -n 16
+   ```
+
+6. Stop the monitor and check for foreign GPU usage during the run:
    ```bash
    kill $MON_PID 2>/dev/null; wait $MON_PID 2>/dev/null
    grep -E 'FOREIGN USER|\[FOREIGN\]' "$GPU_LOG" || echo "no foreign GPU usage observed"
    ```
 
-5. Report results: total passed, failed, skipped, errors. If any foreign-user events are present in step 4, mention them — flaky failures should be re-evaluated on a clean GPU before being attributed to code changes.
+7. Report results: total passed, failed, skipped, errors — and the import-gate results from steps 3–4. If any foreign-user events are present in step 6, mention them — flaky failures should be re-evaluated on a clean GPU before being attributed to code changes.
 
 ## Failure triage rules
 

.gitignore

@@ -183,7 +183,6 @@ cscope*
 perf
 .bash_history
 *.json
-!.claude/commands/tir-bench/baseline.json
 *.params
 *.ro
 *.onnx
@@ -295,7 +294,3 @@ python/bin/
 python/typing_extensions.py
 python/*.dist-info/
 pytest-of-bohanhou/
-
-# tir-bench run artifacts (regenerable; see .claude/commands/tir-bench.md)
-.tir-bench/
-.tir-bench-*/

AGENTS.md

@@ -85,8 +85,8 @@ the branch instead of the whole repository:
 pre-commit run --files <changed-file>...
 ```
 
-Use `.agents/skills/tir-build`, `.agents/skills/tir-test`, and
-`.agents/skills/tir-bench` when their workflows apply.
+Use `python -m tirx_kernels.bench_suite` in the **tirx-kernels** repo
+(`tirx_kernels/bench_suite/`) when that workflow applies.
 
 ## Coding Conventions