Skip to content
Open
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
60 changes: 60 additions & 0 deletions evidence/h3-ref2va-ascend-history-20261118/README.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,60 @@
# MiniMax-H3 Ref2VA historical Ascend 950PR record

This bundle is the compact, reviewable projection for the historical experiment
record tracked in Issue #520. It preserves searchable configuration, trial,
summary, and recomputation data while the privacy-reviewed raw material remains
in a versioned external archive.

The record covers MiniMax-H3 Ref2VA experiments collected between 2026-10-20
and 2026-11-18 on Atlas 150 / Ascend 950PR systems. The retained observations
include dense step sweeps, Cache-DiT, RainFusion, quantization, communication
probes, failures, retries, and partial VBench results.

## Verdict and interpretation

The overall performance and quality verdict is `INCONCLUSIVE`.

- Historical `PASS` strings are retained only as source observations; they are
not current InferRSI acceptance decisions.
- The 4.2242x 50-step engine-stack observation and the 7.3747x 10-step plus
Cache-DiT observation are separate comparisons and are not cumulative.
- Some VBench dimensions had one prompt per seed, source/runtime identities are
incomplete, no validator was frozen before search, and the original
containers are unavailable.
- The 448 index rows are source-derived observations, not 448 independent
trials.

## Retained core

- `plan/effective.json` records the bounded historical scope and environment.
- `results/trials.jsonl` indexes all 448 structured observations, including
failed and inconclusive records.
- `results/summary.json` preserves recorded ratios and limitations.
- `results/external_optimization_record.json` exposes the compact external
record without promoting it into gain ranking.
- `validator/recompute.py` checks only the arithmetic stated in the summary. It
does not rerun the model, VBench, or establish acceptance.
- `results/artifacts.json` is the authority for the external archive.

## Archive verification

Offline metadata and repository inventory checks do not download the archive:

```bash
python tools/verify_evidence_archive.py evidence/h3-ref2va-ascend-history-20261118/results/artifacts.json
python tools/validate_evidence.py
```

Explicit archive verification authenticates the compressed bytes and bounded
members before any replay comparison:

```bash
python tools/verify_evidence_archive.py \
evidence/h3-ref2va-ascend-history-20261118/results/artifacts.json \
--download
```

The public archive excludes media, model/cache payloads, the complete VBench
workspace, and unavailable live container state. Hash verification establishes
which redacted bytes were retained; it does not establish historical execution
correctness, model-quality equivalence, or a campaign `PASS`.
7 changes: 7 additions & 0 deletions evidence/h3-ref2va-ascend-history-20261118/patch/README.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,7 @@
# Patch status

This historical record does not claim a complete, source-pinned candidate patch.
Patch fragments and source-side diagnostics that survived collection remain in
the external archive. Incomplete source identity is one reason the campaign
verdict remains `INCONCLUSIVE`; the archive must not be used as an upstream
implementation patch without a separately accepted reproduction.
14 changes: 14 additions & 0 deletions evidence/h3-ref2va-ascend-history-20261118/plan/effective.json
Original file line number Diff line number Diff line change
@@ -0,0 +1,14 @@
{
"campaign_id": "h3-ref2va-ascend-history-20261118",
"kind": "external_experiment_record",
"backend": "ascend",
"hardware": "4-8 logical Ascend 950PR devices on Atlas 150/A5",
"model": "MiniMax-H3 Ref2VA",
"scenario": "diffusion.video_generation",
"collection_window": [
"2026-10-20",
"2026-11-18"
],
"source_revision": "663f3b327998526a13aa31e72ad729025c18983c",
"verdict": "INCONCLUSIVE"
}
58 changes: 58 additions & 0 deletions evidence/h3-ref2va-ascend-history-20261118/results/artifacts.json
Original file line number Diff line number Diff line change
@@ -0,0 +1,58 @@
{
"schema_version": 2,
"archive": {
"id": "h3-ref2va-ascend-history-20261118-v2",
"file": "h3-ref2va-ascend-history-20261118-v2.tar.gz",
"url": "https://github.com/InferMatrixLab/InferRSI/releases/download/evidence-h3-ref2va-ascend-history-20261118-v2/h3-ref2va-ascend-history-20261118-v2.tar.gz",
"sha256": "778f5c7c73cea6ee0f545a59fa91f9a40c79818545f86d49f001d2b36585a0cd",
"bytes": 5454990,
"format": "tar+gzip",
"source_revision": "663f3b327998526a13aa31e72ad729025c18983c",
"root": "h3-ref2va-ascend-history-20261118",
"files": 2684,
"uncompressed_bytes": 40740727,
"key_files": {
"configuration": "h3-ref2va-ascend-history-20261118/plan/effective.json",
"trials": "h3-ref2va-ascend-history-20261118/results/trials.jsonl",
"summary": "h3-ref2va-ascend-history-20261118/results/summary.json",
"record": "h3-ref2va-ascend-history-20261118/records/external_optimization_record.json",
"validator": "h3-ref2va-ascend-history-20261118/validator/recompute.py",
"redaction_inventory": "h3-ref2va-ascend-history-20261118/results/redactions.json"
},
"retention_owner": "JingjingWu6"
},
"verification": {
"inputs": [
{
"path": "h3-ref2va-ascend-history-20261118/results/summary.json",
"sha256": "40186cf06b23026b9e8ff5fd102d5e40a60c78c81d601002651d6b1ed844dcab",
"role": "control"
},
{
"path": "h3-ref2va-ascend-history-20261118/plan/effective.json",
"sha256": "50165f47ae6e2ca959c5480047aad5de4a554df9c68a4996ddf4d9110dadb35c",
"role": "configuration"
},
{
"path": "h3-ref2va-ascend-history-20261118/validator/recompute.py",
"sha256": "eacd63f7c0f37a06c11e3d9a6b1020159a5a558c52b5236d811fc230678a82f0",
"role": "validator"
}
],
"command": ["python", "validator/recompute.py"],
"trials": "results/trials.jsonl",
"summary": "results/summary.json",
"absolute_tolerance": 1e-12,
"relative_tolerance": 0.0
},
"availability": {
"verdict": "INCONCLUSIVE",
"included": "Redacted text logs, JSON/JSONL results, configurations, selected profiler outputs, source references, trial index, summary, and arithmetic recomputation.",
"excluded": [
"Generated media and controlled original inputs",
"Model weights, caches, and the complete VBench workspace",
"Deleted container state and incompletely retained runtime/source identities"
],
"limit": "Archive integrity and arithmetic replay do not establish model quality, service gains, or campaign acceptance."
}
}
Original file line number Diff line number Diff line change
@@ -0,0 +1,42 @@
{
"schema_version": 1,
"kind": "external_optimization_record",
"candidate_id": "minimax-h3-historical-ascend",
"backend": "ascend",
"protocol": "historical-mixed-ref2va-protocols",
"state": "completed",
"source_refs": [
{
"path": "results/trials.jsonl",
"sha256": "c62886a8b9d0cbd75e9e21f419ae912d47504f20dc4862d8b01f3d61d00a413f"
}
],
"performance_verdict": "INCONCLUSIVE",
"quality": {
"verdict": "INCONCLUSIVE"
},
"metrics": [
{
"name": "50-step-engine-stack-recorded-speedup",
"unit": "x",
"baseline": 701.954369,
"samples": [
166.172687
],
"aggregate": 4.224246364867411,
"aggregation": "recorded_ratio"
},
{
"name": "10-step-cachedit-three-seed-recorded-speedup",

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Please label this as diffusion-stage speedup in the standalone record. The indexed timings observation contains both diffuse and e2e columns, and 7.3747× is the mean of the three diffusion ratios, not the end-to-end request ratios. The unit x and recorded-speedup name alone don't preserve that distinction for downstream readers.

"unit": "x",
"samples": [
7.71175565987307,
7.2540855539449725,
7.158391805120384
],
"aggregate": 7.374744339646142,
"aggregation": "three_seed_arithmetic_mean"
}
],
"note": "Recorded historical measurements, not final acceptance. Performance and quality conclusions are separate."
}
42 changes: 42 additions & 0 deletions evidence/h3-ref2va-ascend-history-20261118/results/summary.json
Original file line number Diff line number Diff line change
@@ -0,0 +1,42 @@
{
"schema_version": 1,
"campaign_id": "h3-ref2va-ascend-history-20261118",
"verdict": "INCONCLUSIVE",
"trial_rows": 448,
"status_counts": {
"FAILED": 7,
"RECORDED": 404,
"RECORDED_VALUE": 4,
"COMPLETED": 27,
"INCONCLUSIVE": 6
},
"failed_or_regressing": 7,
"interrupted_or_incomplete": 0,
"duplicates": 0,
"performance": {
"opus_50_step_recorded_speedup": 4.224246364867411,
"opus_derivation": "701.954369 / 166.172687 from the retained InferRSI evaluation and trial result JSONs.",
"deepseek_10step_cachedit_recorded_speedup": 7.374744339646142,
"deepseek_seed_group_speedups": [
7.71175565987307,
7.2540855539449725,
7.158391805120384
],
"deepseek_range": [
7.158391805120384,
7.71175565987307
],
"deepseek_note": "Arithmetic mean of the three retained per-seed paired speedups; each seed has 11 paired requests."
},
"quality": {
"verdict": "INCONCLUSIVE",
"note": "Performance and quality are kept separate. Historical thresholds were not frozen in an acceptance issue and some VBench dimensions have n=1 per seed."
},
"limitations": [
"No pre-existing acceptance issue or frozen current-main validators.",
"Historical PASS labels are retained only as source observations.",
"Containers were deleted; only retained files were collected.",
"<VBENCH_ROOT> was sampled for structured/text evidence, not mirrored in full.",
"Media files and large caches are not included in this archive."
]
}
Loading
Loading