Zhangdanyang/agent-phase2-held-out-aiter

Agent Phase 2 Held-Out Benchmark (AITER gfx950)

Held-out evaluation tasks for GPU kernel models targeting AMD MI350 (gfx950). Covers three kernel languages: Triton, OPUS (C++ HIP), and FlyDSL (Python MLIR). Zero overlap with training and dev data.

Dataset Structure

tasks/kernel.jsonl          # 150 task entries (geak_kernel_held_out_eval_v1)
artifacts/kernel/*/initial/ # Per-task kernel source + metadata + config + task_runner
receipts/                   # Benchmark result storage
manifest.json               # Dataset manifest
dataset_info.json           # Operator/language statistics
shape_summary.json          # Shape distribution
checksums.sha256            # Content integrity verification

Statistics

• Total tasks: 150 (50 per language)

By Language

Language	Count	File Type	Source
Triton	50	kernel.py	aiter/ops/triton/
OPUS	50	kernel.cu	aiter/csrc/
FlyDSL	50	kernel.py	aiter/ops/flydsl/

By Operator (27 types)

Operator	Count		Operator	Count
gemm	26		preshuffle_gemm	6
attention	18		mla	5
moe	14		mha	5
unknown	9		quantization	5
fmha	7		mxfp4_gemm	5
kv_cache	6		rmsnorm	4
conv	6		splitk_gemm	4
moe_sorting	6		all_reduce	3
block_scale_gemm	3		topk	3
normalization	3		deepgemm	2
batched_gemm	2		gemm_a16w16	2
gemm_a8w8	2		paged_attention	1
layernorm	1		sampling	1
delta_rule	1

Task Format

Each entry in tasks/kernel.jsonl:

{
  "task_id": "heldout-gfx950-gemm-opus-a1b2c3d4",
  "task_type": "direction_conditioned",
  "split": "held_out",
  "eval_only": true,
  "sft_enabled": false,
  "schema_version": "geak_kernel_held_out_eval_v1",
  "lane": "opus_gfx950",
  "contract": {
    "kernel_contract": {
      "operator": "gemm",
      "architecture": "gfx950",
      "language": "opus"
    }
  },
  "source_lineage_id": "aiter:csrc/gemm_kernel.cu",
  "initial_source": {
    "path": "artifacts/kernel/heldout-gfx950-gemm-opus-a1b2c3d4/initial",
    "sha256": "..."
  },
  "kernel_baseline": "aiter"
}

Per-Task Artifacts

Each task's artifacts/kernel/<task_id>/initial/ contains:

File	Description
kernel.py or kernel.cu	Complete kernel source (reference implementation)
metadata.json	Task metadata (operator, architecture, language, source)
config.yaml	Evaluation configuration (timeout, eval mode)
scripts/task_runner.py	Compile/correctness/performance runner

No Overlap

• Zero content overlap with agent-kernel-sft-gfx950-triton-HIP train.jsonl (3,670 samples)
• Zero content overlap with dev.jsonl (200 samples)
• Verified by SHA256 content hash comparison