The dataset viewer is not available for this split.
Error code: StreamingRowsError
Exception: TypeError
Message: Couldn't cast array of type
struct<pos: list<item: string>, kw: struct<has_mla: string, use_pdl: string, page_size: string, score_mod: string, aux_tensors: string>, out_slots: struct<3: list<item: string>, 4: list<item: string>>>
to
{'b1': Value('string'), 'b2': Value('string'), 'inplace': Value('string'), 'activation': Value('string'), 'is_gated': Value('string'), 'apply_router_weight_on_input': Value('string'), 'use_fp8_w8a8': Value('string'), 'use_int8_w8a8': Value('string'), 'use_int8_w8a16': Value('string'), 'use_int4_w4a16': Value('string'), 'per_channel_quant': Value('string'), 'w1_zp': Value('string'), 'w2_zp': Value('string'), 'a1_scale': Value('string'), 'a2_scale': Value('string'), 'block_shape': List(Value('string')), 'no_combine': Value('string'), 'routed_scaling_factor': Value('string'), 'gemm1_alpha': Value('string'), 'gemm1_limit': Value('string'), 'filter_expert': Value('string'), 'a1_q': Value('string'), 'fuse_swiglu_interleaved': Value('string'), 'gate_up_interleaved': Value('string'), 'swiglu_limit': Value('string'), 'hidden_states': Value('string'), 'w1': Value('string'), 'w2': Value('string'), 'topk_weights': Value('string'), 'topk_ids': Value('string'), 'w1_scale': Value('string'), 'w2_scale': Value('string')}
Traceback: Traceback (most recent call last):
File "/src/services/worker/src/worker/utils.py", line 149, in get_rows_or_raise
return get_rows(
dataset=dataset,
...<4 lines>...
column_names=column_names,
)
File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
return func(*args, **kwargs)
File "/src/services/worker/src/worker/utils.py", line 129, in get_rows
rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
File "/src/services/worker/src/worker/utils.py", line 544, in safe_iter
yield from ds.decode(False) if ds.features else ds
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 3052, in __iter__
for key, example in ex_iterable:
^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2559, in __iter__
for key, pa_table in self._iter_arrow():
~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2584, in _iter_arrow
for key, pa_table in self.ex_iterable._iter_arrow():
~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 558, in _iter_arrow
for key, pa_table in iterator:
^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 440, in _iter_arrow
for key, pa_table in self.generate_tables_fn(**gen_kwags):
~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 361, in _generate_tables
self._cast_table(pa_table, json_field_paths=json_field_paths),
~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
pa_table = table_cast(pa_table, self.info.features.arrow_schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2381, in table_cast
return cast_table_to_schema(table, schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2315, in cast_table_to_schema
cast_array_to_feature(
~~~~~~~~~~~~~~~~~~~~~^
table[name] if name in table_column_names else pa.array([None] * len(table), type=schema.field(name).type),
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
feature,
^^^^^^^^
)
^
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 1861, in wrapper
return pa.chunked_array([func(chunk, *args, **kwargs) for chunk in array.chunks])
~~~~^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2121, in cast_array_to_feature
casted_array_values = _c(array.values, feature.feature)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 1863, in wrapper
return func(array, *args, **kwargs)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2071, in cast_array_to_feature
_c(children[name] if name in array_fields else null_array, subfeature)
~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 1863, in wrapper
return func(array, *args, **kwargs)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2161, in cast_array_to_feature
raise TypeError(f"Couldn't cast array of type\n{_short_str(array.type)}\nto\n{_short_str(feature)}")
TypeError: Couldn't cast array of type
struct<pos: list<item: string>, kw: struct<has_mla: string, use_pdl: string, page_size: string, score_mod: string, aux_tensors: string>, out_slots: struct<3: list<item: string>, 4: list<item: string>>>
to
{'b1': Value('string'), 'b2': Value('string'), 'inplace': Value('string'), 'activation': Value('string'), 'is_gated': Value('string'), 'apply_router_weight_on_input': Value('string'), 'use_fp8_w8a8': Value('string'), 'use_int8_w8a8': Value('string'), 'use_int8_w8a16': Value('string'), 'use_int4_w4a16': Value('string'), 'per_channel_quant': Value('string'), 'w1_zp': Value('string'), 'w2_zp': Value('string'), 'a1_scale': Value('string'), 'a2_scale': Value('string'), 'block_shape': List(Value('string')), 'no_combine': Value('string'), 'routed_scaling_factor': Value('string'), 'gemm1_alpha': Value('string'), 'gemm1_limit': Value('string'), 'filter_expert': Value('string'), 'a1_q': Value('string'), 'fuse_swiglu_interleaved': Value('string'), 'gate_up_interleaved': Value('string'), 'swiglu_limit': Value('string'), 'hidden_states': Value('string'), 'w1': Value('string'), 'w2': Value('string'), 'topk_weights': Value('string'), 'topk_ids': Value('string'), 'w1_scale': Value('string'), 'w2_scale': Value('string')}Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
headkernel-mi308-bench
Five head kernels from AgentKernelArena (head_kernels_top5_20260920, commit 8fa0d73c81e5), packaged
to run on AMD Instinct MI308X (gfx942) with seeded synthetic inputs and independent PyTorch references.
Intended as the starting point for source-level kernel optimization (AVO).
| Task | Kernel | Shapes |
|---|---|---|
glm-5.3-flash__fused_moe_kernel |
Triton FP8 W8A8 block-scale fused MoE | 4 |
minimax-m3__gqa_share_sparse_fwd_kernel |
Triton sparse GQA prefill attention | 38 |
kimi-k3__fwd_grouped_kernel_stage1 |
Triton MLA decode, stage 1 | 4 |
qwen3.8-2.4t__fused_recurrent_gated_delta_rule_decode |
Triton gated delta-rule state update | 2 |
qwen3.8-2.4t__gemma_fused_add_rmsnorm |
Triton fused add + Gemma RMSNorm | 2 |
Source
- Tasks: AMD-AGI/AgentKernelArena, branch
head_kernels_top5_20260920, commit8fa0d73c81e55100d8e32d1963d6b247764fe56a(Apache-2.0). The upstream tasks target MI355X (gfx950), an internal MI355 image, and non-public captured inputs (reference_io.pt). - Selection: 16 upstream tasks were screened on MI308X; 10 run and pass, 6 need gfx950-only instructions
(FP4, scaled MFMA) or do not compile for gfx942. Five of the 10 are packaged here
(
docs/2026-10-26_headkernel-mi308-筛选报告.md). - Adaptation for MI308X: seeded synthetic inputs, independent PyTorch references, a negative control, a
public runtime image, and the changes listed in "Differences from the upstream tasks". The kernel sources
(
tasks/<task>/source/) are unchanged.
Requirements
- MI308X (gfx942) GPU, Docker with access to
/dev/kfdand/dev/dri. - Image
lmsysorg/sglang:v0.5.18-rocm720-mi30x(public, about 23 GB compressed). It supplies every dependency. Exact versions are inversions.lock.
docker pull lmsysorg/sglang@sha256:2a4b6896f2d69e5562545b5abaaf00e8d8dbb14da736fa2f5f5bd47ce00bfb0f
Validated platform
| Item | Value |
|---|---|
| GPU | AMD Instinct MI308X (gfx942), 80 CUs, max 1850 MHz, 192 GB HBM3 |
| Host | Linux 5.15, amdgpu driver 6.16.13, ROCm 7.2.0 |
| Image | lmsysorg/sglang@sha256:2a4b6896f2d69e5562545b5abaaf00e8d8dbb14da736fa2f5f5bd47ce00bfb0f (versions.lock) |
| GPU selection | ROCR_VISIBLE_DEVICES (HIP_VISIBLE_DEVICES does not filter HSA agents) |
Run
# correctness for all five tasks on GPU 7
GPU=7 bash scripts/run_in_container.sh python3 harness/run.py --mode correctness
# timing (runs the correctness check first for every shape)
GPU=7 bash scripts/run_in_container.sh python3 harness/run.py --mode performance
# one task, keep per-task JSON and logs
GPU=7 bash scripts/run_in_container.sh python3 harness/run.py --mode performance \
--task qwen3.8-2.4t__gemma_fused_add_rmsnorm --out runs/rmsnorm
scripts/run_in_container.sh starts a throwaway container (--rm), mounts only this directory at
/bench, selects the GPU with ROCR_VISIBLE_DEVICES, and returns file ownership to the caller.
harness/run.py exits nonzero if any task fails.
Under rocprofv3 (detected via LD_PRELOAD / ROCP_TOOL_LIBRARIES), harness/run.py replaces itself with
the child instead of spawning it: a GPU child process deadlocks at start-up under the profiler, which then waits
forever. Profile one task at a time (--task <name>).
Inside an existing container of the same image, run python3 harness/run.py ... from the repository
root. Each task also has its own entry, python3 tasks/<task>/ut/unittest.py [--mode correctness]
(default: performance), which is the same run as harness/run.py --task <task>.
What is checked
For each task and each shape:
- The code under test is
tasks/<task>/source/. The harness restores theut/kernel_src/*symlinks tosource/(a Hugging Face download turns them into plain copies), copies the source intout/_cand_overlay/, and verifies that the target function resolves to that copy, not to the library in the image. - Inputs are generated from
torch.Generator(device="cuda").manual_seed(0)at the task's real shapes (ut/meta.json, never shrunk). Index inputs (page tables, top-k blocks, expert ids) follow the kernel's contract; rules are documented inadapters/<task>.py. RMSNorm and the gated delta rule use the task's ownut/cases.pybuilders. - Outputs must be finite and match
refs/<task>.py, a plain-PyTorch reference, under the task rule|out-ref| <= tol*RMS(ref) + tol*|ref|withtolfrommeta.json. In-place state updates are compared too. - Negative control: corrupting one element of each output must be rejected.
Performance mode then times each shape with the task's own ut/harness_lib.py:time_op (CUDA events,
10 warmup, 50 samples, median, cache flushed before each sample, graph replay when the deployment
regime uses graphs) and prints
GEAK_RESULT_LATENCY_MS=<ms> (<shape id>)
Timing uses the same inputs with verify_inputs=False where the task supports it (RMSNorm): the
input-integrity probe clones and compares every input, which would add device work to the timed window
and block graph capture. If a task requests graph timing and capture fails, the harness prints
HK_WARNING graph capture failed ... and the per-shape JSON records timer=cuda_event instead of
cuda_event_graph.
Baseline (original kernels)
Medians of three performance runs on one idle MI308X (results/baseline_mi308x.csv, one row per shape,
with the run-to-run spread and the source run). GLM uses the FNUZ-weight run of 2026-11-03; the other
tasks use the fixed-timing run of 2026-10-29.
| Task | Shapes | Timer | Fastest shape (ms) | Slowest shape (ms) | Geomean (ms) |
|---|---|---|---|---|---|
glm-5.3-flash__fused_moe_kernel |
4 | eager | 0.537 | 6.310 | 1.085 |
minimax-m3__gqa_share_sparse_fwd_kernel |
38 | graph | 0.025 | 26.25 | 15.80 |
kimi-k3__fwd_grouped_kernel_stage1 |
4 | graph | 0.037 | 2.523 | 0.351 |
qwen3.8-2.4t__fused_recurrent_gated_delta_rule_decode |
2 | graph | 0.258 | 0.310 | 0.283 |
qwen3.8-2.4t__gemma_fused_add_rmsnorm |
2 | graph | 0.017 | 0.262 | 0.067 |
How to compare:
- Absolute latencies depend on the node and on load from neighbouring GPUs (differences of tens of percent have been seen). Use the table to check that a setup is sane (within about 10%), not as a target.
- To measure a speedup, time the original and the candidate on the same GPU in the same session, interleaved (A B B A), and report per-shape ratios and their geomean.
- Shapes of a few tens of microseconds vary by up to about 8% between runs (Kimi
oracle_bs1_ctx395); GLM decode under eager timing had one 39% outlier run (rand_m19_decode), so take the median of at least three runs.
AVO reference results
The five tasks were optimized at source level with AVO. Optimized kernels are not distributed: rerun AVO with the version and configuration below, and verify the result as described here.
AVO version:
| Item | Value |
|---|---|
| Repository | AMD-BRAIN-Internal/PerfSkills-avo (AMD internal) |
| Branch | GEAK-v4-integrated |
| Commit | 01b60f937a24b3235f2ea897fda9bdc868f0d9bf |
| Entry | python3 avo-workflow/run_geak_avo.py handoff.json result.json (validate first with --dry-run) |
| Agent | Claude Code 2.1.308 (Claude Opus 5.5), Node 22 |
Configuration (one AVO run per task, two tasks in parallel on one 8-GPU node):
| Setting | Value |
|---|---|
operator |
team (default start_policy=diverse) |
| GPUs per task | 4 (gpu_ids 0,1,2,3 or 4,5,6,7) |
budget_s |
43200 (12 h, counted from workflow start; 10% is reserved for finalization) |
kernel_path |
the repository root (AVO copies it; the harness needs harness/, refs/, adapters/) |
hw |
gfx942 |
apply_to_original |
false |
task |
see docs/avo_task_template.md |
Results (speedup = original latency / optimized latency; per-shape median of five interleaved rounds on one GPU, then the geomean over all shapes; every shape passed the correctness check and the negative control):
| Task | Speedup |
|---|---|
minimax-m3__gqa_share_sparse_fwd_kernel |
4.39x |
glm-5.3-flash__fused_moe_kernel |
3.19x |
kimi-k3__fwd_grouped_kernel_stage1 |
2.64x |
qwen3.8-2.4t__gemma_fused_add_rmsnorm |
1.26x |
qwen3.8-2.4t__fused_recurrent_gated_delta_rule_decode |
1.02x |
Notes for rerunning AVO on MI308X with this commit:
- Tell AVO the GPU has 80 CUs. Its knowledge file reads the CU count with
rocminfo | grep -m1, which returns the CPU core count on this node, and its roofline table forgfx942is MI300X. - Tasks with many shapes (MiniMax, 38) can stop with
TypeError: (perCase || []).map is not a functionafter a commit. Committed versions stay inavo_repo(tagsavo-vN) and can be verified. - Review every patch: no output caching, no work skipped on repeated inputs or the fixed seed, no shape-specialised branches, no lowered precision.
Differences from the upstream tasks
- No
reference_io.pt: the upstream captured tensors are in non-public OCI storage. Synthetic inputs change timing (sparse attention selections are more spread out; MoE routing is uniform or synthetic), not whether a kernel runs and is correct. Speedups are valid for these inputs only. - Hardware is MI308X, not MI355X. Shapes are kept as captured on MI355X (Kimi split counts assume 256 CUs).
- GLM fused MoE: activations are quantized to
e4m3fnuzon gfx942 (MI355X usese4m3fn). Since 2026-11-03 the expert weights are also given to the kernel ase4m3fnuzon gfx942, converted with sglang'snormalize_e4m3fn_to_e4m3fnuz, as sglang does at model load on MI300-series GPUs. Before that date they weree4m3fn, and the original ran about 30x slower than in deployment; GLM speedups measured before that date are not comparable (results/validation_20261103_glm_fnuz). The model path points tomodel_cfg/GLM-5.2-FP8-config(configuration only).SGLANG_USE_AITER=0is forced byharness/run.py: the image defaults to 1, and that path dropsrouted_scaling_factor. kimi-k3__fwd_grouped_kernel_stage1starts from a prior optimized candidate, not upstream code.qwen3.8-2.4t__gemma_fused_add_rmsnorm: the upstream contract text rounds the residual sum to BF16 before the norm, while the source computes the norm from the unrounded FP32 sum. The reference follows the source. Optimized versions must keep the source behavior.- Each task's
ut/unittest.pyis the benchmark entry described in "Run". The upstream scripts that need the non-public captured data (ut/unittest.py,ut/leg_runner.py,scripts/task_runner.py) were removed; the Kimi seam helpers they held are inut/cases.py. config.yamlcommands callharness/run.py.
Layout
tasks/ five task directories (source/ is the editable kernel; ut/unittest.py is the entry)
adapters/ synthetic input builders (GLM, Kimi, MiniMax)
refs/ PyTorch references
harness/ run.py (entry) and child.py (one task per fresh process)
scripts/ run_in_container.sh, compare_runs.py (timing spread across runs)
model_cfg/ GLM configuration files used by the GLM task
results/ screening_20261026 (16 tasks, rocm/atom image), validation runs, baseline_mi308x.csv
docs/ screening plan/report (Chinese), avo_task_template.md (AVO task prompt)
- Downloads last month
- 193