Most of the bf16 inference error comes from the bf16 residual stream; an fp32 stream around the bf16 GEMMs costs ~1.5%
メンテナーはふだん 1 日以内に返信
まだ誰も着手していません。
評価
- 難易度
- 4/5
- 見積もり時間
- 3〜5日
- 初心者へのやさしさ
- 43/100
- issue の種類
- 機能追加
- 明瞭さ
- おおむね明確
- 活発さ
- 活発
調査の方向性
Start by reading MOSSImagePipeline.from_pretrained and the transformer’s _Block and _attention implementations, which the issue describes as handling dtype casts. Check how training sets activation precision and whether the pipeline can expose a stream dtype option. Done means the maintainers resolve the training precision question and decide whether to add an option for fp32 residual streams, with behavior and tests specified.
索引モデルが issue の本文から書いたものです。
説明
Hi, and thanks for releasing MOSS-Image. While building LoRA training and inference
tooling for it, we measured where the bf16 numerical error in the DiT comes from. The
result was larger than we expected, so we are sharing it here.
Setting
MOSSImagePipeline.from_pretrained loads the transformer with torch_dtype=torch.bfloat16
and nothing runs under autocast. So the residual stream x, the modulation, the gates and
the norm outputs are all bf16. Each block's x = x + gate * proj(...) rounds to an 8-bit
mantissa twice per block, over 42 blocks.
As a test we kept the 252 block linears (layers.N.proj.{q,k,v,o,ffn_up_gate,ffn_down})
in bf16 and cast every other parameter to fp32. The stream, modulation, gates, norms,
embedders and head then run in fp32 around the same bf16 block GEMMs. The model code
needs no change: _Block already casts its GEMM inputs to self.proj["q"].weight.dtype,
and _attention casts q / k to v.dtype, so attention stays bf16 too. On top of the
README example:
pipe.scheduler = type(pipe.scheduler).from_config(pipe.scheduler.config)
+for name, p in pipe.transformer.named_parameters():
+ if not (name.startswith("layers.") and ".proj." in name): # all but the 252 block linears
+ p.data = p.data.float()
pipe.to("cuda")
Prediction error
MOSS-Image-Base. Six images at ~4.1k image + 184–383 text tokens, σ ∈ {0.1, 0.5, 0.9},
one fixed ε per image. Each run is scored as ‖v − v_fp32‖ / ‖v_fp32‖ against an fp32
forward: every GEMM in fp32 on the same bf16 weights, fp32 norms, SDPA, TF32 off.
| σ | bf16 (as shipped) | fp32 stream | bf16, no reduced-precision reduction | loss bf16 / fp32 stream / fp32 |
|---|---|---|---|---|
| 0.1 | 4.81% | 1.64% | 4.80% | 0.5635 / 0.5612 / 0.5609 |
| 0.5 | 1.93% | 0.81% | 1.94% | 0.2253 / 0.2249 / 0.2249 |
| 0.9 | 2.08% | 0.59% | 2.12% | 0.3819 / 0.3814 / 0.3814 |
- The fp32 stream cuts the prediction error 2.4–3.5x.
- Its loss matches fp32 to the fourth decimal.
- Turning off cuBLAS reduced-precision reduction (
allow_bf16_reduced_precision_reduction = False)
does not move it. The error is in the stream, not the GEMM accumulation.
Effect on samples
Same prompts, seeds and initial noise. 1024², 50 steps, CFG 3 with block 22 skipped on
the uncond branch, shift 5. Mean |Δ| per image between bf16 and the fp32 stream, out of
255, over 3 prompts × 2 seeds:
| mean Δ | range | |
|---|---|---|
| Base | 9.5 | 5.0–16.4 |
| SFT | 7.3 | 3.5–14.0 |
Base usually keeps the composition and changes details: eyes, accessories, hands,
small marks. SFT can change the pose. In the third image below, the hands, legs and bag
move. To our eyes neither precision is consistently better; both give clean images.
Three comparisons (bf16 | fp32 stream | |diff| ×4) follow in the first comment: Base
prompt 0 seed 1 (16.4/255), SFT prompt 2 seed 1 (14.0/255) and Base prompt 1 seed 1
(11.8/255).
These renders use our own sampler setup: an ER-SDE solver, Spectrum feature caching,
an INT8 text encoder and compiled blocks. The prediction-error table above is a single
forward and does not depend on it.
Cost. 20.8 s vs 20.5 s per 1024² image (+1.5%) and about 0.6 GB for the fp32
non-block parameters (RTX 5070 Ti, compiled blocks).
Questions
- What precision did the residual stream have in pre-training and SFT? The comment
"Match the reference activation dtype under fp32 weights + autocast" suggests fp32
master weights under bf16 autocast. If so, the linear outputs and the stream were
probably bf16 there too. The shipped bf16 inference would then match training, and the
fp32 stream would only be closer to exact arithmetic. If the stream was fp32 in
training, the fp32 stream would also match training better. - If that is useful, would you consider an option in the pipeline (e.g.
stream_dtype=torch.float32) for users who want the more exact forward at ~1.5% cost?
- 主要言語
- Python
- スター
- 2.2k
- フォーク
- 114
- 平均マージ
- 9日 1時間
- マージ済み PR（30日）
- 1
環境構築
このプロジェクトには開発コンテナ、Dockerfile、コントリビューションガイドがありません。まず README を読み、一般的な手順ははじめてのコントリビューションガイドを参照してください。
はじめの一歩
- issue を最後まで読み、次にプロジェクトのコントリビューションガイドを読みます。
- 着手することを issue にコメントします — 二人が同じ作業をするのを防げます。
- リポジトリをフォークし、ブランチを切って変更します。
- issue 番号を参照したプルリクエストを送ります。
OpenMOSS/MOSS-Image のほかの issue
-
Community project: hosted MOSS-Image studio with a visual JSON layout builder (no ComfyUI needed)オープン
難易度 1/5 1〜3時間 初心者へのやさしさ 65/100
OpenMOSS/MOSS-Image#27 · コメント 2 件 ·
メンテナーはふだん 1 日以内に返信
-
Apple Silicon (MPS): float64 timestep sinusoid fails — working 3-line fix対応中かも このイシューにリンクされたプルリクエストがオープン中、またはマージ済みです。 オープン
難易度 2/5 1〜3時間 初心者へのやさしさ 85/100
OpenMOSS/MOSS-Image#23 ·
メンテナーはふだん 1 日以内に返信
-
難易度 2/5 1〜3時間 初心者へのやさしさ 66/100
OpenMOSS/MOSS-Image#10 · コメント 1 件 ·
メンテナーはふだん 1 日以内に返信
-
難易度 1/5 1時間未満 初心者へのやさしさ 45/100
OpenMOSS/MOSS-Image#28 ·
メンテナーはふだん 1 日以内に返信
-
MOSS video?オープン
難易度 5/5 1週間以上 初心者へのやさしさ 20/100
OpenMOSS/MOSS-Image#24 ·
メンテナーはふだん 1 日以内に返信
OpenMOSS/MOSS-Image の issue をすべて見る
似ている issue
-
難易度 2/5 1〜3時間 初心者へのやさしさ 76/100
pytest-dev/pytest-html#1085 ·
-
難易度 2/5 1〜3時間 初心者へのやさしさ 90/100
pollen-robotics/reachy_mini#1527 ·
メンテナーはふだん 1 日以内に返信
-
難易度 1/5 1時間未満 初心者へのやさしさ 66/100
メンテナーはふだん 2 日以内に返信
-
難易度 2/5 1〜3時間 初心者へのやさしさ 72/100
メンテナーはふだん 1 日以内に返信
-
A seed and its negative give the same run, so realizations of a negative seed repeat each otherオープン
難易度 2/5 1〜3時間 初心者へのやさしさ 66/100
メンテナーはふだん 1 日以内に返信