Hailuo MiniMax H3 training support

README.es.md

@@ -110,6 +110,7 @@ SimpleTuner es compatible con las siguientes familias de modelos. El soporte det
 | **LTX Video 2** | 19B | Apache-2.0 | Sí |
 | **Lumina2** | 2B | Apache-2.0 | Sí |
 | **Mage-Flow** | 4B | MIT | Sí |
+| **MiniMax H3** | 33B | MiniMax H3 Community License | Aplican condiciones (exclusiones territoriales; autorización requerida en EE. UU./UE/Reino Unido/Corea del Sur) |
 | **OmniGen** | 3.8B | MIT | Sí |
 | **PixArt Sigma** | 0.6B-0.9B | OpenRAIL++ | Sí (restringido) |
 | **Qwen Image** | 20B | Apache-2.0 | Sí |

README.hi.md

@@ -110,6 +110,7 @@ SimpleTuner निम्नलिखित मॉडल families का सम
 | **LTX Video 2** | 19B | Apache-2.0 | हाँ |
 | **Lumina2** | 2B | Apache-2.0 | हाँ |
 | **Mage-Flow** | 4B | MIT | हाँ |
+| **MiniMax H3** | 33B | MiniMax H3 Community License | शर्तें लागू (territory exclusions; US/EU/UK/KR में authorization आवश्यक) |
 | **OmniGen** | 3.8B | MIT | हाँ |
 | **PixArt Sigma** | 0.6B-0.9B | OpenRAIL++ | हाँ (restricted) |
 | **Qwen Image** | 20B | Apache-2.0 | हाँ |

README.ja.md

@@ -110,6 +110,7 @@ SimpleTunerは以下のモデルファミリーをサポートしています。
 | **LTX Video 2** | 19B | Apache-2.0 | 可 |
 | **Lumina2** | 2B | Apache-2.0 | 可 |
 | **Mage-Flow** | 4B | MIT | 可 |
+| **MiniMax H3** | 33B | MiniMax H3 Community License | 条件付き（地域除外あり；米国/EU/英国/韓国は認可が必要） |
 | **OmniGen** | 3.8B | MIT | 可 |
 | **PixArt Sigma** | 0.6B-0.9B | OpenRAIL++ | 可（制限あり） |
 | **Qwen Image** | 20B | Apache-2.0 | 可 |

README.md

@@ -110,6 +110,7 @@ SimpleTuner supports the following model families. Detailed training feature sup
 | **LTX Video 2** | 19B | Apache-2.0 | Yes |
 | **Lumina2** | 2B | Apache-2.0 | Yes |
 | **Mage-Flow** | 4B | MIT | Yes |
+| **MiniMax H3** | 33B | MiniMax H3 Community License | Conditions apply (territory exclusions; authorization required in US/EU/UK/KR) |
 | **OmniGen** | 3.8B | MIT | Yes |
 | **PixArt Sigma** | 0.6B-0.9B | OpenRAIL++ | Yes (restricted) |
 | **Qwen Image** | 20B | Apache-2.0 | Yes |

README.pt-BR.md

@@ -110,6 +110,7 @@ SimpleTuner suporta as seguintes familias de modelos. O suporte detalhado a recu
 | **LTX Video 2** | 19B | Apache-2.0 | Sim |
 | **Lumina2** | 2B | Apache-2.0 | Sim |
 | **Mage-Flow** | 4B | MIT | Sim |
+| **MiniMax H3** | 33B | MiniMax H3 Community License | Condicoes aplicaveis (exclusoes territoriais; autorizacao exigida nos EUA/UE/Reino Unido/Coreia do Sul) |
 | **OmniGen** | 3.8B | MIT | Sim |
 | **PixArt Sigma** | 0.6B-0.9B | OpenRAIL++ | Sim (restrito) |
 | **Qwen Image** | 20B | Apache-2.0 | Sim |

README.zh.md

@@ -110,6 +110,7 @@ SimpleTuner 支持以下模型系列。详细的训练功能支持请参阅[快
 | **LTX Video 2** | 19B | Apache-2.0 | 是 |
 | **Lumina2** | 2B | Apache-2.0 | 是 |
 | **Mage-Flow** | 4B | MIT | 是 |
+| **MiniMax H3** | 33B | MiniMax H3 Community License | 有条件（含地区排除；美国/EU/英国/韩国需要授权） |
 | **OmniGen** | 3.8B | MIT | 是 |
 | **PixArt Sigma** | 0.6B-0.9B | OpenRAIL++ | 是（受限） |
 | **Qwen Image** | 20B | Apache-2.0 | 是 |

documentation/DATALOADER.md

@@ -682,8 +682,8 @@ For example, with 4 GPUs, `train_batch_size=4`, and `gradient_accumulation_steps
 To automatically adjust `repeats` when your dataset is smaller than the effective batch size, use the `--allow_dataset_oversubscription` flag (documented in [OPTIONS.md](OPTIONS.md#--allow_dataset_oversubscription)).
 
 When enabled, SimpleTuner will:
-- Calculate the minimum repeats needed for training
-- Automatically increase `repeats` to meet the requirement
+- Calculate the minimum repeats needed for each undersized aspect bucket
+- Pad only those buckets to meet the effective batch size
 - Log a warning showing the adjustment
 - **Respect manually-set repeats values** - if you explicitly configure `repeats` in your dataset config, the automatic adjustment will be skipped
 
@@ -1370,7 +1370,7 @@ Webshart datasets load WebDataset-style tar shards through the `webshart` packag
 - `metadata` is optional and points to a separate metadata location when captions or shard metadata are stored outside the shard source. For Hugging Face metadata repos such as `webshart/conceptual-captions-12m-webdataset-metadata`, pass the repo id; Webshart follows the source shard subfolder layout such as `data/`.
 - `metadata_backend` must be `webshart`; it reads dimensions and captions from Webshart metadata.
 - `caption_strategy` should be `webshart` to train from metadata captions, or `instanceprompt` to ignore stored captions.
-- `webshart.cache_dir` stores SimpleTuner metadata plus Webshart metadata and shard caches. `shard_cache_gb` and `parallel_downloads` are passed to Webshart's shard cache.
+- `webshart.cache_dir` stores SimpleTuner metadata plus Webshart metadata and shard caches. `shard_cache_gb` and `parallel_downloads` are passed to Webshart's shard cache; set `shard_cache_gb` to `0` to disable whole-shard caching and retain indexed range reads.
 
 This backend requires a Webshart build with `TarDataLoader.list_shard_sample_aspect_buckets()`.
 

documentation/OPTIONS.es.md

@@ -52,6 +52,15 @@ Donde `foo` es tu entorno de configuración; o simplemente usa `config/config.js
   - `diffusers` es el esquema estándar de PEFT/Diffusers.
   - `comfyui` convierte hacia/desde claves estilo ComfyUI (`diffusion_model.*` con tensores `lora_A/lora_B` y `.alpha`). Flux, Flux2, Lumina2 y Z-Image detectarán automáticamente entradas ComfyUI incluso si esto se deja en `diffusers`, pero cámbialo a `comfyui` para forzar salida ComfyUI al guardar.
 
+### `--minimax_h3_target_mode`
+
+- **Qué**: Controla si MiniMax-H3 incluye filas de audio objetivo.
+- **Opciones**: `auto`, `video`, `av`
+- **Predeterminado**: `auto`
+- **Notas**:
+  - `auto` se resuelve como solo video, omitiendo caché VAE de audio, colación y filas de audio objetivo para H3.
+  - Define `minimax_h3_target_mode` o `h3_target_mode` como `av` en una entrada de data backend para activar entrenamiento conjunto audio-video en un backend de audio auto-split o explícito.
+
 ### `--fuse_qkv_projections`
 
 - **Qué**: Fusiona las proyecciones QKV en los bloques de atención del modelo para un uso más eficiente del hardware.
@@ -1880,6 +1889,7 @@ usage: train.py [-h] --model_family
                 [--flow_beta_schedule_beta FLOW_BETA_SCHEDULE_BETA]
                 [--flow_schedule_shift FLOW_SCHEDULE_SHIFT]
                 [--flow_schedule_auto_shift [FLOW_SCHEDULE_AUTO_SHIFT]]
+                [--audio_flow_schedule_shift AUDIO_FLOW_SCHEDULE_SHIFT]
                 [--flow_custom_timesteps FLOW_CUSTOM_TIMESTEPS]
                 [--flow_timesteps_mode {fixed-list,round-robin}]
                 [--flux_guidance_mode {constant,random-range}]
@@ -2000,7 +2010,7 @@ usage: train.py [-h] --model_family
                 [--rescale_betas_zero_snr [RESCALE_BETAS_ZERO_SNR]]
                 [--webhook_config WEBHOOK_CONFIG]
                 [--webhook_reporting_interval WEBHOOK_REPORTING_INTERVAL]
-                [--distillation_method {lcm,dcm,dmd,perflow,flow_dpo,anyflow}]
+                [--distillation_method {lcm,dcm,dmd,perflow,flow_dpo,anyflow,h3_drift}]
                 [--distillation_config DISTILLATION_CONFIG]
                 [--ema_validation {none,ema_only,comparison}]
                 [--local_rank LOCAL_RANK] [--ltx_train_mode {t2v,i2v}]
@@ -2351,6 +2361,9 @@ options:
                         Shift the noise schedule for flow-matching models
   --flow_schedule_auto_shift [FLOW_SCHEDULE_AUTO_SHIFT]
                         Auto-adjust schedule shift based on image resolution
+  --audio_flow_schedule_shift AUDIO_FLOW_SCHEDULE_SHIFT
+                        Shift the audio noise schedule for flow-matching
+                        models with audio latents
   --flow_custom_timesteps FLOW_CUSTOM_TIMESTEPS
                         Override flow-matching timestep sampling with a fixed
                         comma-separated list. The list is interpreted as
@@ -2733,7 +2746,7 @@ options:
                         Path to webhook configuration file
   --webhook_reporting_interval WEBHOOK_REPORTING_INTERVAL
                         Interval for webhook reports (seconds)
-  --distillation_method {lcm,dcm,dmd,perflow,flow_dpo,anyflow}
+  --distillation_method {lcm,dcm,dmd,perflow,flow_dpo,anyflow,h3_drift}
                         Method for model distillation
                         Distillation methods cannot be combined with
                         --train_text_encoder.

documentation/OPTIONS.hi.md

@@ -52,6 +52,15 @@ simpletuner configure config/foo/config.json
   - `diffusers` standard PEFT/Diffusers layout है।
   - `comfyui` keys को ComfyUI‑style में convert करता है (`diffusion_model.*` के साथ `lora_A/lora_B` और `.alpha` tensors)। Flux, Flux2, Lumina2, और Z‑Image ComfyUI inputs को auto‑detect करेंगे भले ही यह `diffusers` पर हो, लेकिन saving के लिए ComfyUI output force करने के लिए `comfyui` सेट करें।
 
+### `--minimax_h3_target_mode`
+
+- **What**: MiniMax-H3 target audio rows शामिल करे या नहीं, इसे नियंत्रित करता है।
+- **Choices**: `auto`, `video`, `av`
+- **Default**: `auto`
+- **Notes**:
+  - `auto` video-only में resolve होता है, जिससे H3 के लिए audio VAE cache, collate, और target audio rows skip होते हैं।
+  - auto-split या explicit audio backend को joint audio-video training में opt in करने के लिए data backend entry में `minimax_h3_target_mode` या `h3_target_mode` को `av` सेट करें।
+
 ### `--fuse_qkv_projections`
 
 - **What**: मॉडल के attention blocks में QKV projections को fuse करता है ताकि hardware का अधिक कुशल उपयोग हो।
@@ -1878,6 +1887,7 @@ usage: train.py [-h] --model_family
                 [--flow_beta_schedule_beta FLOW_BETA_SCHEDULE_BETA]
                 [--flow_schedule_shift FLOW_SCHEDULE_SHIFT]
                 [--flow_schedule_auto_shift [FLOW_SCHEDULE_AUTO_SHIFT]]
+                [--audio_flow_schedule_shift AUDIO_FLOW_SCHEDULE_SHIFT]
                 [--flow_custom_timesteps FLOW_CUSTOM_TIMESTEPS]
                 [--flow_timesteps_mode {fixed-list,round-robin}]
                 [--flux_guidance_mode {constant,random-range}]
@@ -1998,7 +2008,7 @@ usage: train.py [-h] --model_family
                 [--rescale_betas_zero_snr [RESCALE_BETAS_ZERO_SNR]]
                 [--webhook_config WEBHOOK_CONFIG]
                 [--webhook_reporting_interval WEBHOOK_REPORTING_INTERVAL]
-                [--distillation_method {lcm,dcm,dmd,perflow,flow_dpo,anyflow}]
+                [--distillation_method {lcm,dcm,dmd,perflow,flow_dpo,anyflow,h3_drift}]
                 [--distillation_config DISTILLATION_CONFIG]
                 [--ema_validation {none,ema_only,comparison}]
                 [--local_rank LOCAL_RANK] [--ltx_train_mode {t2v,i2v}]
@@ -2349,6 +2359,9 @@ options:
                         Shift the noise schedule for flow-matching models
   --flow_schedule_auto_shift [FLOW_SCHEDULE_AUTO_SHIFT]
                         Auto-adjust schedule shift based on image resolution
+  --audio_flow_schedule_shift AUDIO_FLOW_SCHEDULE_SHIFT
+                        Shift the audio noise schedule for flow-matching
+                        models with audio latents
   --flow_custom_timesteps FLOW_CUSTOM_TIMESTEPS
                         Override flow-matching timestep sampling with a fixed
                         comma-separated list. The list is interpreted as
@@ -2731,7 +2744,7 @@ options:
                         Path to webhook configuration file
   --webhook_reporting_interval WEBHOOK_REPORTING_INTERVAL
                         Interval for webhook reports (seconds)
-  --distillation_method {lcm,dcm,dmd,perflow,flow_dpo,anyflow}
+  --distillation_method {lcm,dcm,dmd,perflow,flow_dpo,anyflow,h3_drift}
                         Method for model distillation
                         Distillation methods cannot be combined with
                         --train_text_encoder.

documentation/OPTIONS.ja.md

@@ -52,6 +52,15 @@ simpletuner configure config/foo/config.json
   - `diffusers` は標準の PEFT/Diffusers 形式です。
   - `comfyui` は ComfyUI 形式（`diffusion_model.*` と `lora_A/lora_B` + `.alpha`）に変換します。Flux、Flux2、Lumina2、Z-Image は `diffusers` のままでも ComfyUI 入力を自動検出しますが、保存時に ComfyUI 出力を強制したい場合は `comfyui` を指定してください。
 
+### `--minimax_h3_target_mode`
+
+- **内容**: MiniMax-H3 がターゲット音声行を含めるかを制御します。
+- **選択肢**: `auto`, `video`, `av`
+- **既定**: `auto`
+- **注記**:
+  - `auto` は video-only として扱われ、H3 の audio VAE cache、collate、ターゲット音声行を省略します。
+  - auto-split または明示的な audio backend で joint audio-video training を使う場合は、data backend entry に `minimax_h3_target_mode` または `h3_target_mode` を `av` として設定します。
+
 ### `--fuse_qkv_projections`
 
 - **内容**: モデルのアテンションブロック内 QKV 投影を融合し、ハードウェア効率を高めます。
@@ -1881,6 +1890,7 @@ usage: train.py [-h] --model_family
                 [--flow_beta_schedule_beta FLOW_BETA_SCHEDULE_BETA]
                 [--flow_schedule_shift FLOW_SCHEDULE_SHIFT]
                 [--flow_schedule_auto_shift [FLOW_SCHEDULE_AUTO_SHIFT]]
+                [--audio_flow_schedule_shift AUDIO_FLOW_SCHEDULE_SHIFT]
                 [--flow_custom_timesteps FLOW_CUSTOM_TIMESTEPS]
                 [--flow_timesteps_mode {fixed-list,round-robin}]
                 [--flux_guidance_mode {constant,random-range}]
@@ -2001,7 +2011,7 @@ usage: train.py [-h] --model_family
                 [--rescale_betas_zero_snr [RESCALE_BETAS_ZERO_SNR]]
                 [--webhook_config WEBHOOK_CONFIG]
                 [--webhook_reporting_interval WEBHOOK_REPORTING_INTERVAL]
-                [--distillation_method {lcm,dcm,dmd,perflow,flow_dpo,anyflow}]
+                [--distillation_method {lcm,dcm,dmd,perflow,flow_dpo,anyflow,h3_drift}]
                 [--distillation_config DISTILLATION_CONFIG]
                 [--ema_validation {none,ema_only,comparison}]
                 [--local_rank LOCAL_RANK] [--ltx_train_mode {t2v,i2v}]
@@ -2351,6 +2361,9 @@ options:
                         Shift the noise schedule for flow-matching models
   --flow_schedule_auto_shift [FLOW_SCHEDULE_AUTO_SHIFT]
                         Auto-adjust schedule shift based on image resolution
+  --audio_flow_schedule_shift AUDIO_FLOW_SCHEDULE_SHIFT
+                        Shift the audio noise schedule for flow-matching
+                        models with audio latents
   --flow_custom_timesteps FLOW_CUSTOM_TIMESTEPS
                         Override flow-matching timestep sampling with a fixed
                         comma-separated list. The list is interpreted as
@@ -2733,7 +2746,7 @@ options:
                         Path to webhook configuration file
   --webhook_reporting_interval WEBHOOK_REPORTING_INTERVAL
                         Interval for webhook reports (seconds)
-  --distillation_method {lcm,dcm,dmd,perflow,flow_dpo,anyflow}
+  --distillation_method {lcm,dcm,dmd,perflow,flow_dpo,anyflow,h3_drift}
                         Method for model distillation
                         Distillation methods cannot be combined with
                         --train_text_encoder.

documentation/OPTIONS.md

@@ -52,6 +52,15 @@ Where `foo` is your config environment - or just use `config/config.json` if you
   - `diffusers` is the standard PEFT/Diffusers layout.
   - `comfyui` converts to/from ComfyUI-style keys (`diffusion_model.*` with `lora_A/lora_B` and `.alpha` tensors). Flux, Flux2, Lumina2, and Z-Image will auto-detect ComfyUI inputs even if this is left at `diffusers`, but set it to `comfyui` to force ComfyUI output when saving.
 
+### `--minimax_h3_target_mode`
+
+- **What**: Controls whether MiniMax-H3 includes target audio rows.
+- **Choices**: `auto`, `video`, `av`
+- **Default**: `auto`
+- **Notes**:
+  - `auto` resolves to video-only, skipping audio VAE caching, collation, and target audio rows for H3.
+  - Set `minimax_h3_target_mode` or `h3_target_mode` to `av` in a data backend entry to opt an auto-split or explicit audio backend into joint audio-video training.
+
 ### `--fuse_qkv_projections`
 
 - **What**: Fuses the QKV projections in the model's attention blocks to make more efficient use of hardware.
@@ -1884,6 +1893,7 @@ usage: train.py [-h] --model_family
                 [--flow_beta_schedule_beta FLOW_BETA_SCHEDULE_BETA]
                 [--flow_schedule_shift FLOW_SCHEDULE_SHIFT]
                 [--flow_schedule_auto_shift [FLOW_SCHEDULE_AUTO_SHIFT]]
+                [--audio_flow_schedule_shift AUDIO_FLOW_SCHEDULE_SHIFT]
                 [--flow_custom_timesteps FLOW_CUSTOM_TIMESTEPS]
                 [--flow_timesteps_mode {fixed-list,round-robin}]
                 [--flux_guidance_mode {constant,random-range}]
@@ -2004,7 +2014,7 @@ usage: train.py [-h] --model_family
                 [--rescale_betas_zero_snr [RESCALE_BETAS_ZERO_SNR]]
                 [--webhook_config WEBHOOK_CONFIG]
                 [--webhook_reporting_interval WEBHOOK_REPORTING_INTERVAL]
-                [--distillation_method {lcm,dcm,dmd,perflow,flow_dpo,anyflow}]
+                [--distillation_method {lcm,dcm,dmd,perflow,flow_dpo,anyflow,h3_drift}]
                 [--distillation_config DISTILLATION_CONFIG]
                 [--ema_validation {none,ema_only,comparison}]
                 [--local_rank LOCAL_RANK] [--ltx_train_mode {t2v,i2v}]
@@ -2355,6 +2365,9 @@ options:
                         Shift the noise schedule for flow-matching models
   --flow_schedule_auto_shift [FLOW_SCHEDULE_AUTO_SHIFT]
                         Auto-adjust schedule shift based on image resolution
+  --audio_flow_schedule_shift AUDIO_FLOW_SCHEDULE_SHIFT
+                        Shift the audio noise schedule for flow-matching
+                        models with audio latents
   --flow_custom_timesteps FLOW_CUSTOM_TIMESTEPS
                         Override flow-matching timestep sampling with a fixed
                         comma-separated list. The list is interpreted as
@@ -2739,7 +2752,7 @@ options:
                         Path to webhook configuration file
   --webhook_reporting_interval WEBHOOK_REPORTING_INTERVAL
                         Interval for webhook reports (seconds)
-  --distillation_method {lcm,dcm,dmd,perflow,flow_dpo,anyflow}
+  --distillation_method {lcm,dcm,dmd,perflow,flow_dpo,anyflow,h3_drift}
                         Method for model distillation
                         Distillation methods cannot be combined with
                         --train_text_encoder.