chore(sa3): repin current upstream and version SAME-L engines - #319
Merged
Conversation
This was referenced Aug 3, 2026
Move the managed SA3 source from the old FP8 development lineage to current upstream plus the rebased TensorRT integer-attribute fix. The current upstream SAME-L plugin bakes its AOT/JIT implementation into the serialized engine. Include the plugin revision and implementation choice in the engine name and metadata so stale decoders are ignored instead of being paired with incompatible plugin code. Keep the plugin revision stable across unrelated vendor bumps so they do not force a 1.2 GB decoder rebuild. Existing engines remain beside the new artifact for rollback. Also accept git worktrees as DEMON_SA3_SRC overrides.
ryanontheinside
force-pushed
the
ryanontheinside/chore/sa3-current-upstream-pin
branch
from
August 17, 2026 19:16
3395004 to
b00a1f2
Compare
The repin to c076985 drops optimized/tensorRT/build/build_dit_fp8.py from the vendored tree — the ModelOpt PTQ builder lives in Stability PR #47 and was never merged upstream, so it exists at the old fork head and not at the new pin. Both places that told an operator to run it were therefore shipping an impossible remedy: the DIT_FP8_ONNX_FILES comment and, worse, the RuntimeError raised when --fp8 runs without --fp8-onnx, which is the one an operator hits at the moment they need it to be true. Point them at what is actually there instead. make_calib.py, build_dit_bf16.py and transplant_scales.py are all still vendored, but the transplant consumes a dit_fp8_calib.onnx only #47 produces, so the honest answer is that the graph cannot be produced from the pinned tree alone and has to arrive via --fp8-onnx. Every path the new text names was checked to exist at c076985. Comments and one error string; no logic change. DIT_FP8_ONNX_FILES is deliberately left alone — repointing the sidecar at dit_fp8lin.onnx.data is still gated on someone building and parity-checking an engine from it. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Two changes that travel together:
c076985, the head of Stability PR realtime_motion_graph_web: default to 60s low-fi fixture for first-time users #49(SWA plugin int-attribute fix for TensorRT 10.16 / NumPy 2.x). Without
it the same-L decoder engine cannot be built at all on current
TRT/NumPy.
(
{aot|jit}_{mma|ptx|jit}_v<plugin-revision>). Engines built against adifferent plugin revision or implementation are no longer silently
reused: discovery filters by the expected tag, metadata sidecars gate
rebuild/skip, and a missing engine falls back loudly to the eager
decoder (logged
mode=eager reason=no_trt_engine, realtime holds).Also includes a worktree-aware vendor
.gitcheck and unit tests for theengine-identity scheme.
Validation (2026-08-17, RTX 5090, TRT 10.16.1.11)
test_lora_facade_sessionfailuresreproduce identically on clean
main, pre-existing).8.47 ms vs 8.44 ms legacy per T=56 decode, VRAM delta nil.
selected, LoRA TRT refit, source swaps/uploads, denoise sweeps,
restart; 0 stale ticks throughout.
eager fallback, stream stays realtime.
Merge handoff
order. Branch history was rewritten onto current
main(d29d5f9), soreview the current diff, not older revisions.
(
ensure_sa3_vendormoves a clean checkout automatically; dirtycheckouts error loudly instead).
decode, still realtime on a 5090) until the builder is run once.