From be8b427dffdb3111403ae6e8e3c25cfb61b01964 Mon Sep 17 00:00:00 2001 From: plotarmordev <299844489+plotarmordev@users.noreply.github.com> Date: Sat, 31 Oct 2026 03:17:56 +0800 Subject: [PATCH] docs: I record the v1.1.9 frame-cap and GPU A/B results The PR #33 GPU A/B passed (G1 corpus captures byte-identical, G2, G3), and the issue #32 diagnostic gave the same frame-cap outcome on v1.1.9 as on the previous pin (tagged 3/5 capped, untagged 0/5, per-reply bytes, durations and RMS identical). I replace the untested/pending wording in README-dialogue, README and CHANGELOG. The cause stays open in #32; the Mac line is unchanged. --- CHANGELOG.md | 5 +++-- README.md | 6 +++--- recipe/README-dialogue.md | 13 ++++++++----- 3 files changed, 14 insertions(+), 10 deletions(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index 79d1901..4b2b8b6 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -5,7 +5,8 @@ ### Changed - The full-model engine pins TensorFold v1.1.9 with an empty patch series: every behaviour of the - old patches 0001–0015 is upstream (per upstream source; GPU A/B pending). Two differences remain: with the mid-silence rule on, + old patches 0001–0015 is upstream. A CUDA GPU A/B against the previous pin passed: corpus captures + byte-identical, request sampling and speaker tags pass. Two differences remain: with the mid-silence rule on, `X-Silence-Mid-Trips` reports only whether the returned take tripped (0 or 1), and a temperature of `-0` or `-1e-400` reaches the sampler as `+0`. `scripts/build-engine.sh` accepts an empty full-model series only through an empty `SHA256SUMS`; an empty NVFP4 manifest or a missing @@ -37,7 +38,7 @@ including native/Python catalog writers. Legacy CUDA long-form continuations retain the active speaker across splits and redraws; reference-count and malformed-tag validation remain enforced. No voice-quality claim is made. Tagged full-model requests - could run to the frame cap when measured on the previous pin; see the + can run to the frame cap, with the same outcome on the previous pin and on v1.1.9; see the [known stopping limitation](recipe/README-dialogue.md#known-full-model-stopping-limitation) and [issue #32](https://github.com/plotarmordev/drama-tensorfold/issues/32). - Per-request HTTP speech `temperature`, `top_p`, `top_k` and exact unsigned 64-bit diff --git a/README.md b/README.md index 856a708..fd83ec0 100644 --- a/README.md +++ b/README.md @@ -134,8 +134,8 @@ Guidance-enabled replies carry the configured γ in `X-Cfg-Gamma`; builtin voice Request text and reference transcripts accept canonical speaker tags 0 through 15. See [speaker-tagged dialogue](recipe/README-dialogue.md) for catalog examples, profile-specific reference limits and legacy long-form speaker continuation, and its -[known full-model stopping limitation](recipe/README-dialogue.md#known-full-model-stopping-limitation) -([issue #32](https://github.com/plotarmordev/drama-tensorfold/issues/32)). +[known full-model stopping limitation](recipe/README-dialogue.md#known-full-model-stopping-limitation), +unchanged on v1.1.9 ([issue #32](https://github.com/plotarmordev/drama-tensorfold/issues/32)). - **Whole reply:** `Content-Length` framing and an `X-Finish-Reason` header: `stop`, `length` or, with the silence guard on, `silence`. @@ -190,7 +190,7 @@ feature. Each feature's notes are in the engine tree's `CHANGELOG.md` or `docs/` | Request-local speech sampling and exact integer seeds, with parsed-JSON boundary docs | HTTP `temperature`, `top_p`, `top_k`, `seed` | [#1277](https://github.com/ashhart/TensorFold/pull/1277) | | Canonical speaker tags 0..15 in speech and catalog transcripts | speaker tags in text | [#1278](https://github.com/ashhart/TensorFold/pull/1278) | -Every old patch's behaviour is in v1.1.9 (per upstream source; GPU A/B pending) with two differences. With the mid-silence rule on, `X-Silence-Mid-Trips` +Every old patch's behaviour is in v1.1.9, and a CUDA GPU A/B against the previous pin passed (corpus captures byte-identical; request sampling and speaker tags pass), with two differences. With the mid-silence rule on, `X-Silence-Mid-Trips` reports only whether the returned take tripped (0 or 1); takes discarded by silence/short/register redraws are no longer counted. A temperature of `-0` or `-1e-400` reaches the sampler as `+0` rather than `-0`; both were already greedy. `drama3-voice`, the upstream tool that writes attested voice files for `--voices`, is in TensorFold; the diff --git a/recipe/README-dialogue.md b/recipe/README-dialogue.md index 6a8bd7d..7a3210a 100644 --- a/recipe/README-dialogue.md +++ b/recipe/README-dialogue.md @@ -45,14 +45,14 @@ full-model and Metal profiles do not gain a long-form splitter from this feature ## Known full-model stopping limitation -These results were measured on the previous full-model pin, TensorFold `d1018b31` -plus patches 0001–0015, not on v1.1.9. On one GB10 with that CUDA full-model build, this page's two-speaker +These results were first measured on the previous full-model pin, TensorFold `d1018b31` +plus patches 0001–0015. On one GB10 with that CUDA full-model build, this page's two-speaker example reached the 1024-frame cap (47.55 seconds, `X-Finish-Reason: length`) in three of five seeds with a matching reference catalog. The same text without speaker tags stopped normally in all five. Relabeling both references and input to speakers 0/1 still reached the cap in two of five seeds; using the tagged input without a reference catalog did so in one of five. Neither variation is a -reliable workaround. +reliable workaround; both were measured on the previous pin only. The CUDA NVFP4 profile, with its default guidance, stopped normally on all five tagged and all five untagged requests in this small check. The engine pins, @@ -60,8 +60,11 @@ weights and guidance differ, so these observations do not establish the cause or a guarantee for NVFP4. Metal was not tested in this follow-up. See the [settings and per-seed results](../results/dialogue-stopping-20261030.json). -The current full-model pin, TensorFold v1.1.9, has not been tested for this -behaviour; that check and the investigation are tracked in +The current full-model pin, TensorFold v1.1.9, gave the same outcome on the same +GPU: with seeds 1–5, the tagged dialogue reached the frame cap in three of five +and the untagged text in none, with per-reply bytes, durations and RMS identical +to the previous pin. The limitation is unchanged on v1.1.9, not fixed; the +investigation is tracked in [issue #32](https://github.com/plotarmordev/drama-tensorfold/issues/32). Treat `X-Finish-Reason: length` as an incomplete generation. The cause remains unresolved; accepting a speaker tag does not guarantee natural termination.