rocprofv3 --kernel-trace --output-format csv -d benchmarks/results/2026-06-03-hipengine-qwen35-native-c8-grouped-projection-profile/profile-c8-grouped-projection -- env HIP_VISIBLE_DEVICES=1 python3 scripts/qwen35_batch_retained_bench.py --model /models/hipengine/Qwen3.6-35B-A3B-PARO-full4096-e5-packed-MTP-BF16 --fixture /tmp/hipengine-prebench/fixtures/qwen36_paro_8x512_prompt_ids.json --prompt-length 512 --batch-size 8 --decode-tokens 128 --warmup-decode-tokens 8 --max-layers 40 --compiler-version-file benchmarks/results/2026-06-03-hipengine-qwen35-native-c8-grouped-projection-profile/hipcc-version.txt --require-cached-build --batch-decode-moe-path grouped_compact --batch-sample-mode batched_lm_head --batch-sample-eq-ok --batch-sample-eq-artifact benchmarks/results/2026-06-02-hipengine-qwen35-c8-native-batch-sampler-equality.json --batch-sample-eq-rows 8 --c1-baseline-json benchmarks/results/2026-06-02-hipengine-qwen35-native-c8-profiler-preflight/native-baseline-c1.json --serial-bridge-json benchmarks/results/2026-06-02-hipengine-qwen35-native-c8-profiler-preflight/serial-bridge-c8.json --primitive-correctness-json benchmarks/results/2026-06-03-hipengine-qwen35-native-c248-primitive-attached/primitive-c8.json --projection-dispatch-artifact benchmarks/results/2026-06-03-hipengine-qwen35-native-c248-projection-dispatch-catalog/summary.json --profiler-json benchmarks/results/2026-06-03-hipengine-qwen35-native-c8-grouped-projection-profile/profiler-c8-grouped-projection.json --json benchmarks/results/2026-06-03-hipengine-qwen35-native-c8-grouped-projection-profile/native-profiled-c8-grouped-projection.json 
