exp(acestep): int8 DiT fake-quant quality gate (OPT-0089)

packages/acestep/optimization/LEDGER.md

@@ -3,7 +3,7 @@
 Stage 2 was explicitly authorized on 2026-08-13. The approved baseline is
 frozen and measured optimization is active.
 
-Next available ID: `OPT-0089`.
+Next available ID: `OPT-0090`.
 
 | ID | Subsystem | Hypothesis | Evidence | Disposition | Result | Record | Implementation |
 | --- | --- | --- | --- | --- | --- | --- | --- |
@@ -95,6 +95,7 @@ Next available ID: `OPT-0089`.
 | OPT-0086 | Planner-enabled downstream scheduling | The already integrated exact OPT-0080 DiT/VAE depth-two policies depend on authenticated graph/package topology, not on whether conditioning was produced by the planner, so removing the planner-only selector exclusion can recover the proven downstream scheduling saving without changing planner or model math | pending | benchmark-only | Registered before implementation; require selector-negative tests plus planner-enabled control/candidate final-latent, waveform, WAV, cancellation, and topology identity before widening production | [record](experiments/OPT-0086-planner-downstream-depth2-scheduling.md) | allocation `1ddb65e751529936b3ef3cd48a6360386c7dd205`; no implementation yet |
 | OPT-0087 | Planner package-native dense GEMV | Route the already exact direct M1/M2 packed-BF16 kernel through all real planner decode layers and tied-head slices so the package can realize its useful bandwidth without OPT-0083's artificial submit-and-drain boundary around every isolated layer sample | positive | pending-integration | Actual Chrome preserved every full logit, appended K/V-cache evidence word, status, token, Philox cursor, topology, cancellation, and lifecycle result. Direct B won `16/16`, lowered every M1/M2 layer/head/model/complete median, reached `2.229363x` aggregate layer speedup, and reduced aggregate model-through-readback median `272.150 → 156.200 ms` (`115.950 ms`, projected `117.110 s/1,010`). Cleanup balanced `3,683/3,683` buffers and `133/133` maps with zero live resources. The fresh nominal launch passed; all later level-1/2 thermal observations are disclosed. Complete trajectory and planner-enabled product gates remain mandatory | [record](experiments/OPT-0087-planner-package-native-low-row-gemv.md), [result](results/OPT-0087/result.json) | allocation `84470d877ba69c6ef7821871f9bdb82d8c757747`; unchanged OPT-0083 direct kernel; implementation/harness `980623e0f9ec918f7a537b6a90e341adaad6d4b0`; pending integration gates |
 | OPT-0088 | Portable device support | Every subgroup-dependent production owner (OPT-0032/0037 dense K4, OPT-0051 K7 row-reuse, OPT-0048 ConvTranspose K4, attention query8/quad-query) can gain a workgroup-memory counterpart consuming the unchanged hosted packages, selected by the existing execution-profile machinery, so adapters without `subgroups` (Safari, Firefox, iOS) run the production graph instead of failing `FEATURE_UNAVAILABLE`; compatibility experiment, bounded slowdown expected and reported, `shader-f16` stays fail-closed | pending | pending-integration | Portable dense/K7/ConvTranspose owners landed with test-enforced bit-identical arithmetic (byte-equal WGSL arithmetic sections, re-exported rev7/rev8 index math); attention routes to the existing portable oracle (reordered-rounding vs subgroup reduction). End-to-end masked-subgroups waveform and timing gates pending | [record](experiments/OPT-0088-portable-no-subgroup-production-path.md) | kernels c272b2d/c48c050/373e90a; selection wiring pending |
+| OPT-0089 | DiT weight quantization | Weight-only symmetric int8 (per-32-K-block fp16 scales, round-to-nearest, clamp ±127) fake-quantization of all 264 rev7 DiT GEMM tensors, dequantized in place and run through the completely unchanged production graph, preserves end-to-end 30 s waveform quality within a small numerical envelope, so an int8-resident DiT (~1.51 GB + scales) is a credible answer to the observed iPhone 17 Safari OOM kill at `1,789,925,376 / 3,020,808,192` uploaded bytes (layer 14/24); pure quantization-damage gate, zero kernel changes, distinct mechanism from abandoned OPT-0058 activation-quantized DP4a | positive | benchmark-only | Per-tensor damage uniform and small: NRMSE `0.00515–0.00634` (median `0.00559`), min SNR `43.96 dB`, no outlier tensor/family, so no fp16-retention map needed. Determinism gate reproduced the pinned fp16 baseline WAV byte-exactly, then fake-quant vs fp16 on identical seeds gave lo-fi/12345 waveform NRMSE `0.0669` (Pearson `0.99777`, LSD ≈`3.5 dB`, RMS Δ `−0.050 dB`) and latin/424242 NRMSE `0.2268` (Pearson `0.97460`, LSD ≈`4.7 dB`, RMS Δ `+0.089 dB`; per-second max `5.19` is a near-silent-ending small-denominator artifact) — trajectory divergence of the 8-evaluation sampler, not noise-like corruption; zero non-finite samples and exact peak parity. Projected int8 DiT phase peak ≈`1.862 GB` tracked GPU (`1.51 GB` int8 + `94 MB` scales + fp16 norms/shared + measured `127 MB` overhead) versus the observed iPhone 17 kill at ≈`1.920 GB` — plausibly fits, marginal ≈`58 MB` margin; int8 kernel work justified, listening gate mandatory before any product claim | [record](experiments/OPT-0089-dit-int8-weight-fake-quant-gate.md), [quant result](results/OPT-0089/quant-error.json), [waveform result](results/OPT-0089/waveform-metrics.json) | `scripts/requantize-dit-int8.py` (repo root); fake-quant package `ef8355b9…` (models-local, not hosted); benchmark-only, no kernel or production change |
 
 Experiment IDs are allocated before code changes, never reused, and never
 removed from this table.

packages/acestep/optimization/experiments/OPT-0089-dit-int8-weight-fake-quant-gate.md

@@ -0,0 +1,193 @@
+# OPT-0089 — DiT int8 weight fake-quant quality gate (iPhone memory feasibility)
+
+## Status
+
+- Evidence: `positive` (numerical feasibility only; explicitly no listening
+  claim)
+- Disposition: `benchmark-only` (this experiment ships no kernel and
+  integrates nothing; it produces a quality verdict and a memory projection
+  that gate any future int8-weight DiT kernel work)
+- Date: 2026-08-24
+- Author/agent: Claude agent for Alex-Wengg
+- Risk class: `approximate` (weight values change; activations, kernels,
+  graph, sampler, and scheduling are untouched)
+
+## Motivation (measured breadcrumb)
+
+A real iPhone 17 Safari tab was OOM-killed during DiT dense-weight upload at
+`1,789,925,376 / 3,020,808,192` bytes (layer 14/24). The fp16 rev7 dense DiT
+package (`d3fc0020…`, 48 shards, `ACE_OPT_0009_DIT_MIXED_LAYER_BYTES =
+3,020,808,192`) cannot fit an iPhone Safari tab alongside the rest of the
+phase residency. An int8-weight path would roughly halve DiT resident weight
+bytes, but before any int8 kernel is designed the pure quantization damage
+must be measured with zero kernel changes.
+
+## Hypothesis
+
+Weight-only symmetric int8 quantization — one fp16 scale per 32-element block
+along the input (K) dimension, round-to-nearest, clamp ±127 — applied to all
+264 rank-two DiT GEMM/projection tensors in the rev7 package (216 fp16
+`dit-gemm-n256-k32-tile-major-v1` + 48 packed-bf16
+`dit-gemm-n128-k32-tile-major-v1` cross K/V projections; norms, biases,
+scale-shift tables, and constants untouched) preserves end-to-end 30-second
+generation quality within a small numerical envelope when the quantized
+values are dequantized back to the original storage dtype and run through the
+completely unchanged production graph (fake-quant). If true, an int8-resident
+DiT (~1.51 GB weights + ~94 MB fp16 scales) is a credible iPhone path and
+kernel work is justified; if quality collapses, int8-weight work stops here.
+
+## Relationship to OPT-0058 (int8 DP4a, abandoned)
+
+OPT-0058 quantized activations dynamically AND weights inside a DP4a compute
+kernel and was abandoned when adversarial finite-to-zero collapse exceeded
+its primitive gate before any timing. This experiment is a different
+mechanism, not a revisit of that abandoned kernel: no activation
+quantization, no DP4a, no kernel or arithmetic-order change of any kind —
+only weight values move, and the damage is measured end-to-end on the real
+product graph rather than on primitive adversarial fixtures. Any future int8
+kernel remains subject to OPT-0058's recorded lessons.
+
+## Method
+
+1. `scripts/requantize-dit-int8.py` (python3 + numpy, streams shard-by-shard)
+   reads the verified local mirror of hosted package
+   `v1/dit-revision7/d3fc0020…`, fake-quantizes exactly the 264 GEMM tensors
+   in their native tile layout (a 32-K block is physically contiguous as
+   `[n_tile, k_block, :, n_in_tile]`), and writes a complete
+   content-addressed package tree. The manifest is rewritten textually so
+   only the 48 shard SHA-256 values and its own SHA-256 change; the manifest
+   byte length (254,357) and every tensor record stay identical, so the
+   unchanged runtime loads the package once its two pinned identity constants
+   (`DIT_MANIFEST_SHA256`, `ACE_OPT_0009_DIT_DENSE_MANIFEST_SHA256`) are
+   pointed at the new manifest SHA-256 for the run (temporary local patch,
+   never committed — the experiment is reproducible from the script plus the
+   documented invocation).
+2. Per-tensor quantization error (RMSE, NRMSE, max abs, max relative vs
+   tensor amax, SNR, zero-scale blocks) is emitted to a JSON report and
+   persisted under `optimization/results/OPT-0089/`.
+3. End-to-end: Chrome (puppeteer) against a dev server whose
+   `VITE_ACE_MODEL_ORIGIN` serves the local packages; fresh browser profile
+   per run; deterministic seeds.
+
+## Gates
+
+1. Environment determinism: a 30 s seed-12345 lo-fi fixture regenerated from
+   the UNMODIFIED local packages must reproduce the pinned known-good
+   baseline WAV SHA-256
+   `095267d7be0317ed9af10c64b8495d573b43207cd86615b6bbe66f27dc17895d`
+   before any fake-quant run is interpreted.
+2. Quality comparison, fake-quant vs fp16 baseline, on two content-distinct
+   prompt/seed pairs (lo-fi hip hop seed 12345; the default Latin-percussion
+   product prompt seed 424242), 30 s each: waveform NRMSE, max |diff|,
+   per-1-second-segment NRMSE profile, STFT log-magnitude spectral distance,
+   and peak/RMS deltas.
+3. Verdict recorded in the ledger: numerical envelope for both pairs,
+   per-tensor error outliers (with a mixed-precision fp16-retention map if
+   any tensor family is disproportionately damaged), and the projected
+   iPhone phase-peak residency versus the observed ~1.79 GB kill point.
+
+This experiment makes no listening, timing, or product claim. The standard
+listening-gate discipline applies before any int8 path could ever become a
+product profile; numerical closeness here authorizes only kernel-design
+work under a new ID.
+
+## Identity
+
+- Source package: local verified mirror of hosted `v1/dit-revision7/`
+  manifest `d3fc0020efcf60702db411da2fd4b93e9bb84f1437ed310aef01c892727e452f`
+  (source shard SHA-256s verified against the manifest before quantization)
+- Fake-quant package: manifest
+  `ef8355b9cffff466b018b51275923982b071234933fe8a32897915eeeb01fa36`,
+  manifest byte length preserved at `254,357`; 264 GEMM tensors quantized
+  (`3,019,898,880` bytes), 193 passthrough tensors byte-identical
+- Reference/VAE packages: unchanged hosted identities (`18f36c64…`,
+  `36a54d79…`) served from the same local origin
+- Runtime: completely unchanged production graph and profiles
+  (`opt-0009-fp16-fp32-dense-v1` dense DiT,
+  `opt-0070-fixed32-quad-query32-full-self-production-v1` attention,
+  production VAE tuple); temporary local patch of the three pinned
+  `d3fc0020…` identity constants only, reverted after the runs
+- Machine: Apple Silicon dev machine, stock Chrome via puppeteer, fresh
+  browser profile per run, dev server with
+  `VITE_ACE_MODEL_ORIGIN=/models/…` local origins
+- Output WAV SHA-256s: fp16 lo-fi `095267d7…` (pinned baseline reproduced),
+  fp16 latin `bad98ef7ffb0512b4138241c596669ae6f05ee0e7dc6790a9998a5aaf23ed043`,
+  int8 lo-fi `d46597d5641fd8cbe0b36ff0dc5ccc65bf5af5d7717a87b3f5cf3757ca2dcc25`,
+  int8 latin `221dce8174f58251b4a38bab770d5be39194ac3cda46fcf34be2bbe6650fb88f`
+
+## Results
+
+Authoritative artifacts:
+[quant-error.json](../results/OPT-0089/quant-error.json),
+[waveform-metrics.json](../results/OPT-0089/waveform-metrics.json).
+
+### Per-tensor quantization damage (weight level)
+
+Uniform and small across all 264 tensors: NRMSE `0.00515–0.00634`
+(median `0.00559`), minimum SNR `43.96 dB`, maximum per-element error
+`0.394%` of tensor amax, zero zero-scale blocks. Family medians are tightly
+clustered (o_proj `0.00535`, k/v/q_proj `0.00550–0.00558`, up/gate
+`0.00566–0.00571`, down_proj `0.00574`; worst single tensor
+`layers.23.self_attn.v_proj` at `0.00634`). No tensor or family is a
+disproportionate outlier, so no mixed-precision fp16-retention map is
+required at the weight level.
+
+### Gate 1 — environment determinism
+
+The fp16 run from the unmodified local packages reproduced the pinned
+baseline WAV byte-exactly (`095267d7…`). PASS.
+
+### Gate 2 — fake-quant vs fp16, two prompt/seed pairs (30 s each)
+
+- Lo-fi seed 12345: waveform NRMSE `0.0669` (SNR `23.50 dB`), Pearson
+  `0.99777`, max |diff| `0.3867`, per-second NRMSE median `0.0484` /
+  max `0.3264`, mean log-spectral distance `0.1772` log10 units
+  (≈ `3.5 dB`), p95 `0.3640`, peak Δ `+0.0000` (shared −1 dBFS peak
+  normalization), RMS Δ `−0.050 dB`.
+- Latin-percussion default prompt seed 424242: waveform NRMSE `0.2268`
+  (SNR `12.89 dB`), Pearson `0.97460`, max |diff| `0.9170`, per-second
+  NRMSE median `0.1173`, mean log-spectral distance `0.2343`
+  (≈ `4.7 dB`), p95 `0.4139`, peak Δ `+0.0000`, RMS Δ `+0.089 dB`. The
+  headline per-second maximum `5.19` is a small-denominator artifact of the
+  ending: at second 27 the baseline is essentially silent (segment RMS
+  `0.00061` vs `0.00322`) because the fake-quant realization sustains the
+  final decay slightly longer (second 26 RMS `0.0864` vs `0.0382`);
+  active-music seconds run `0.08–0.36`.
+
+Interpretation: the damage manifests as trajectory divergence of the
+8-evaluation sampler, not noise-like corruption — zero non-finite samples,
+exact peak parity, RMS within `0.09 dB`, high global correlation, flat
+per-second profiles except the latin ending decay. Waveform NRMSE therefore
+overstates perceptual change, but the outputs are not near-bit-identical
+realizations; the standard listening gate is mandatory before any product
+use of an int8 path.
+
+### Memory projection (iPhone)
+
+Quantized GEMM storage: `1,509,949,440` B int8 + `94,371,840` B fp16
+scales. DiT-phase resident weights become `1,735,340,288` B (including
+`909,312` B fp16 layer norms/tables and the `130,109,696` B fp16
+reference-shared tensors) versus `3,150,917,888` B today. Adding the
+measured desktop 30 s DiT-phase non-weight overhead
+(`3,277,864,192 − 3,150,917,888 = 126,946,304` B tracked arena/control)
+projects an int8 phase peak of ≈ `1.862 GB` tracked GPU bytes. The observed
+iPhone 17 kill point was ≈ `1.920 GB` tracked GPU (`1,789,925,376` dense
+bytes uploaded + `130,109,696` reference-shared already resident), and the
+pipeline destroys the conditioning phase before DiT upload begins, so that
+kill already reflects minimal-residency ordering. The projected int8 peak
+therefore sits ≈ `58 MB` under the observed kill boundary: plausibly fits,
+but marginal, and untracked tab overhead (JS staging, page, wasm) consumes
+an unknown share of the real Safari budget. Identified fallback levers:
+quantize the `130 MB` reference-shared DiT tensors, and/or int4 for the MLP
+family (`1.81 GB` of the fp16 bytes) which this gate's uniform error
+profile makes the natural next candidate.
+
+### Verdict
+
+Weight-only int8 per-32-K-block quantization does not collapse ACE-Step
+generation: the numerical feasibility gate passes and the projected DiT
+phase residency crosses under the observed iPhone OOM boundary. int8-weight
+kernel work (int8-resident storage with in-kernel dequant, new experiment
+ID) is justified. Per-denoise-step latent comparison and the standard
+listening gate are the mandatory next quality checkpoints; this record
+makes no listening, timing, or product claim.