yangzhuxinyzx 60338fb87e [Kernel] Prefetch and gather disk-mapped PLE rows on CPU (#885) ...

* [Kernel] Prefetch and gather mapped PLE rows on CPU

Assisted-by: Codex
Signed-off-by: yangzhuxinyzx <153831768+yangzhuxinyzx@users.noreply.github.com>

* [Test] Capture and ablate mapped row reader admission

Assisted-by: Codex
Signed-off-by: yangzhuxinyzx <153831768+yangzhuxinyzx@users.noreply.github.com>

* [Bench] Fail fast and summarize the fixed C1 endpoint

Assisted-by: Codex
Signed-off-by: yangzhuxinyzx <153831768+yangzhuxinyzx@users.noreply.github.com>

* [Doc] Record installed PLE gather byte and cold-page checks

Assisted-by: Codex
Signed-off-by: yangzhuxinyzx <153831768+yangzhuxinyzx@users.noreply.github.com>

* [Bench] Reuse the endpoint for focused multi-seed anomaly triage

Assisted-by: Codex
Signed-off-by: yangzhuxinyzx <153831768+yangzhuxinyzx@users.noreply.github.com>

* [Bench] Preserve the GPU reservation across paired runs

Assisted-by: Codex
Signed-off-by: yangzhuxinyzx <153831768+yangzhuxinyzx@users.noreply.github.com>

* [Bench] Measure PLE cache faults and calibrate decode savings

Assisted-by: Codex
Signed-off-by: yangzhuxinyzx <153831768+yangzhuxinyzx@users.noreply.github.com>

* [Bench] Apply FP64 and three-seed admission to FP16 decode fusion

Assisted-by: Codex
Signed-off-by: yangzhuxinyzx <153831768+yangzhuxinyzx@users.noreply.github.com>

* [Bench] Record FP16 distribution differences without rejecting fusion

Assisted-by: Codex
Signed-off-by: yangzhuxinyzx <153831768+yangzhuxinyzx@users.noreply.github.com>

---------

Signed-off-by: yangzhuxinyzx <153831768+yangzhuxinyzx@users.noreply.github.com>

2026-10-05 12:18:47 +08:00

..

sweep

…

__init__.py

…

test_bench_startup.py

…

test_benchmark_sm70_decode.py

[Benchmark] Add QUASAR DFlash2 release gates

2026-09-02 07:32:19 +08:00

test_benchmark_sm70_dflash2_gsm8k.py

[Core] Preserve GLM DFlash target quality

2026-09-01 07:29:57 +08:00

test_benchmark_sm70_openai_stream.py

[Bugfix] Preserve chat templates in SM70 stream benchmark prompts

2026-10-02 19:13:59 +08:00

test_compare_sm70_dflash2_quality_matrix.py

…

test_compare_sm70_quality_trace_qsa.py

feat(qwen4-exp): add calibrated E4M3 QSA KV cache

2026-09-03 12:54:29 +08:00

test_custom_dataset_seed.py

…

test_custom_image_dataset.py

…

test_dflash2_concurrent_rounds.py

[Bugfix] Stop benchmark profiling after stream failure

2026-10-05 01:59:16 +08:00

test_dsv4_quality_tools.py

[Benchmark] Add QUASAR DFlash2 release gates

2026-09-02 07:32:19 +08:00

test_latency_cli.py

…

test_plot_filters.py

…

test_qwen38_dcp_datasets.py

[Test] Add paired DCP quality and numerical audit tools

2026-10-03 12:46:06 +08:00

test_qwen38_dcp_quality.py

[Test] Add paired DCP quality and numerical audit tools

2026-10-03 12:46:06 +08:00

test_qwen38_distribution_probe.py

[Kernel] Prefetch and gather disk-mapped PLE rows on CPU (#885)

2026-10-05 12:18:47 +08:00

test_qwen38_endpoint_preflight.py

[Kernel] Prefetch and gather disk-mapped PLE rows on CPU (#885)

2026-10-05 12:18:47 +08:00

test_qwen38_quality_health.py

[Kernel] Prefetch and gather disk-mapped PLE rows on CPU (#885)

2026-10-05 12:18:47 +08:00

test_random_dataset.py

…

test_random_multimodal_dataset_video.py

…

test_sampling_params.py

…

test_serve_cli.py

…

test_sm70_long_context_attention.py

[Test] Validate dynamic benchmark import specifications

2026-10-05 02:03:36 +08:00

test_sm70_paired_stats.py

[Build] Support SM72 targets and robust device diagnostics

2026-10-03 06:27:04 +08:00

test_sm70_quality_speed_matrix.py

[Benchmark] Admit healthy final answers without greedy identity

2026-09-05 13:44:14 +08:00

test_sm70_quasar_nvfp4_oracle.py

[Audit] Trace QUASAR TP partition drift through target sampling

2026-09-06 14:01:26 +08:00

test_sm70_qwen38_baseline.py

[Core] Select the qualified Qwen3.8 NVFP4 no-MTP defaults

2026-09-09 11:44:40 +08:00

test_sm70_qwen38_cohort.py

[Core] Stage PLE results in mapped host buffers

2026-10-03 14:04:54 +08:00

test_sm70_release_quality_gate.py

…

test_startup_cache.py

[Bugfix][SM70] Resolve channel-FP8 workspaces after AOT cache reload

2026-09-22 09:52:46 +08:00

test_throughput_cli.py

…

test_txt_slices_dataset.py

…