Support HunyuanImage-image3.0 for t2i

configs/dist_infer/hunyuan_image3_t2i_dist_cfg.json

@@ -0,0 +1,32 @@
+{
+    "infer_steps": 50,
+    "sample_guide_scale": 5.0,
+    "flow_shift": 1.0,
+    "target_height": 1024,
+    "target_width": 1024,
+    "feature_caching": "NoCaching",
+    "enable_kv_cache": true,
+    "enable_text_kv_cache": true,
+    "use_taylor_cache": false,
+    "taylor_cache_interval": 5,
+    "taylor_cache_order": 2,
+    "taylor_cache_enable_first_enhance": false,
+    "taylor_cache_first_enhance_steps": 3,
+    "taylor_cache_enable_tailing_enhance": false,
+    "taylor_cache_tailing_enhance_steps": 1,
+    "taylor_cache_low_freqs_order": 2,
+    "taylor_cache_high_freqs_order": 2,
+    "pipeline_parallel": true,
+    "enable_cfg": true,
+    "hunyuan_cfg_mode": "parallel",
+    "parallel": {
+        "cfg_p_size": 2
+    },
+    "bot_task": "think_recaption",
+    "use_system_prompt": "en_unified",
+    "max_new_tokens": 2048,
+    "text_do_sample": true,
+    "text_top_k": 1024,
+    "text_top_p": 0.95,
+    "text_temperature": 0.6
+}

configs/dist_infer/hunyuan_image3_t2i_dist_cfg_kv_all_gather.json

@@ -0,0 +1,30 @@
+{
+    "infer_steps": 50,
+    "sample_guide_scale": 5.0,
+    "flow_shift": 1.0,
+    "target_height": 1024,
+    "target_width": 1024,
+    "feature_caching": "NoCaching",
+    "enable_kv_cache": true,
+    "enable_text_kv_cache": true,
+    "use_taylor_cache": false,
+    "pipeline_parallel": true,
+    "hunyuan_image3_pipeline_layout": "interleaved",
+    "enable_cfg": true,
+    "hunyuan_cfg_mode": "parallel",
+    "moe_impl": "flashinfer",
+    "attn_impl": "flash_attention_2",
+    "flashinfer_autotune_mode": "tune",
+    "parallel": {
+        "cfg_p_size": 2,
+        "seq_p_size": 2,
+        "seq_p_attn_type": "kv_all_gather"
+    },
+    "bot_task": "think_recaption",
+    "use_system_prompt": "en_unified",
+    "max_new_tokens": 2048,
+    "text_do_sample": true,
+    "text_top_k": 1024,
+    "text_top_p": 0.95,
+    "text_temperature": 0.6
+}

configs/dist_infer/hunyuan_image3_t2i_dist_cfg_ulysses.json

@@ -0,0 +1,30 @@
+{
+    "infer_steps": 50,
+    "sample_guide_scale": 5.0,
+    "flow_shift": 1.0,
+    "target_height": 1024,
+    "target_width": 1024,
+    "feature_caching": "NoCaching",
+    "enable_kv_cache": true,
+    "enable_text_kv_cache": true,
+    "use_taylor_cache": false,
+    "pipeline_parallel": true,
+    "hunyuan_image3_pipeline_layout": "interleaved",
+    "enable_cfg": true,
+    "hunyuan_cfg_mode": "parallel",
+    "moe_impl": "flashinfer",
+    "attn_impl": "sdpa",
+    "flashinfer_autotune_mode": "tune",
+    "parallel": {
+        "cfg_p_size": 2,
+        "seq_p_size": 2,
+        "seq_p_attn_type": "ulysses"
+    },
+    "bot_task": "think_recaption",
+    "use_system_prompt": "en_unified",
+    "max_new_tokens": 2048,
+    "text_do_sample": true,
+    "text_top_k": 1024,
+    "text_top_p": 0.95,
+    "text_temperature": 0.6
+}

configs/dist_infer/hunyuan_image3_t2i_dist_kv_all_gather.json

@@ -0,0 +1,29 @@
+{
+    "infer_steps": 50,
+    "sample_guide_scale": 5.0,
+    "flow_shift": 1.0,
+    "target_height": 1024,
+    "target_width": 1024,
+    "feature_caching": "NoCaching",
+    "enable_kv_cache": true,
+    "enable_text_kv_cache": true,
+    "use_taylor_cache": false,
+    "pipeline_parallel": true,
+    "enable_cfg": true,
+    "hunyuan_cfg_mode": "serial",
+    "moe_impl": "flashinfer",
+    "attn_impl": "sdpa",
+    "flashinfer_autotune_mode": "tune",
+    "parallel": {
+        "cfg_p_size": 1,
+        "seq_p_size": 2,
+        "seq_p_attn_type": "kv_all_gather"
+    },
+    "bot_task": "think_recaption",
+    "use_system_prompt": "en_unified",
+    "max_new_tokens": 2048,
+    "text_do_sample": true,
+    "text_top_k": 1024,
+    "text_top_p": 0.95,
+    "text_temperature": 0.6
+}

configs/dist_infer/hunyuan_image3_t2i_dist_ulysses.json

@@ -0,0 +1,29 @@
+{
+    "infer_steps": 50,
+    "sample_guide_scale": 5.0,
+    "flow_shift": 1.0,
+    "target_height": 1024,
+    "target_width": 1024,
+    "feature_caching": "NoCaching",
+    "enable_kv_cache": true,
+    "enable_text_kv_cache": true,
+    "use_taylor_cache": false,
+    "pipeline_parallel": true,
+    "enable_cfg": true,
+    "hunyuan_cfg_mode": "serial",
+    "moe_impl": "flashinfer",
+    "attn_impl": "sdpa",
+    "flashinfer_autotune_mode": "tune",
+    "parallel": {
+        "cfg_p_size": 1,
+        "seq_p_size": 2,
+        "seq_p_attn_type": "ulysses"
+    },
+    "bot_task": "think_recaption",
+    "use_system_prompt": "en_unified",
+    "max_new_tokens": 2048,
+    "text_do_sample": true,
+    "text_top_k": 1024,
+    "text_top_p": 0.95,
+    "text_temperature": 0.6
+}

configs/hunyuan_image3/hunyuan_image3_i2i.json

@@ -0,0 +1,28 @@
+{
+    "infer_steps": 50,
+    "sample_guide_scale": 5.0,
+    "flow_shift": 1.0,
+    "image_size": "auto",
+    "feature_caching": "NoCaching",
+    "enable_kv_cache": true,
+    "enable_text_kv_cache": true,
+    "use_taylor_cache": false,
+    "taylor_cache_interval": 5,
+    "taylor_cache_order": 2,
+    "taylor_cache_enable_first_enhance": false,
+    "taylor_cache_first_enhance_steps": 3,
+    "taylor_cache_enable_tailing_enhance": false,
+    "taylor_cache_tailing_enhance_steps": 1,
+    "taylor_cache_low_freqs_order": 2,
+    "taylor_cache_high_freqs_order": 2,
+    "pipeline_parallel": true,
+    "enable_cfg": true,
+    "infer_align_image_size": true,
+    "bot_task": "think_recaption",
+    "use_system_prompt": "en_unified",
+    "max_new_tokens": 2048,
+    "text_do_sample": true,
+    "text_top_k": 1024,
+    "text_top_p": 0.95,
+    "text_temperature": 0.6
+}

configs/hunyuan_image3/hunyuan_image3_t2i.json

@@ -0,0 +1,28 @@
+{
+    "infer_steps": 50,
+    "sample_guide_scale": 5.0,
+    "flow_shift": 1.0,
+    "target_height": 1024,
+    "target_width": 1024,
+    "feature_caching": "NoCaching",
+    "enable_kv_cache": true,
+    "enable_text_kv_cache": true,
+    "use_taylor_cache": false,
+    "taylor_cache_interval": 5,
+    "taylor_cache_order": 2,
+    "taylor_cache_enable_first_enhance": false,
+    "taylor_cache_first_enhance_steps": 3,
+    "taylor_cache_enable_tailing_enhance": false,
+    "taylor_cache_tailing_enhance_steps": 1,
+    "taylor_cache_low_freqs_order": 2,
+    "taylor_cache_high_freqs_order": 2,
+    "pipeline_parallel": true,
+    "enable_cfg": true,
+    "bot_task": "think_recaption",
+    "use_system_prompt": "en_unified",
+    "max_new_tokens": 2048,
+    "text_do_sample": true,
+    "text_top_k": 1024,
+    "text_top_p": 0.95,
+    "text_temperature": 0.6
+}

lightx2v/common/ops/norm/rms_norm_weight.py

@@ -66,14 +66,17 @@ def _get_lora_attr_mapping(self):
         self.weight_diff = torch.tensor(0.0, dtype=GET_DTYPE(), device=AI_DEVICE)
 
     def _get_actual_weight(self):
-        if not hasattr(self, "weight_diff"):
+        if not hasattr(self, "weight_diff") or not self.has_diff:
             return self.weight
+        if self.weight_diff.device != self.weight.device or self.weight_diff.dtype != self.weight.dtype:
+            self.weight_diff = self.weight_diff.to(device=self.weight.device, dtype=self.weight.dtype)
         return self.weight + self.weight_diff
 
     def register_diff(self, weight_dict):
         if not self.lazy_load or self.create_cuda_buffer or self.create_cpu_buffer:
             if self.weight_diff_name in weight_dict:
                 self.weight_diff = weight_dict[self.weight_diff_name]
+                self.has_diff = True
                 logger.debug(f"Register Diff to {self.weight_name}")
 
     def load(self, weight_dict):

lightx2v/infer.py

@@ -12,6 +12,7 @@
 from lightx2v.models.runners.flux2.flux2_runner import Flux2DevRunner, Flux2KleinRunner  # noqa: F401
 from lightx2v.models.runners.hidream_o1_image.hidream_o1_image_runner import HidreamO1ImageRunner  # noqa: F401
 from lightx2v.models.runners.hunyuan3d.hunyuan3d_shape_runner import Hunyuan3DShapeRunner  # noqa: F401
+from lightx2v.models.runners.hunyuan_image3.hunyuan_image3_runner import HunyuanImage3Runner  # noqa: F401
 from lightx2v.models.runners.hunyuan_video.hunyuan_video_15_distill_runner import HunyuanVideo15DistillRunner  # noqa: F401
 from lightx2v.models.runners.hunyuan_video.hunyuan_video_15_runner import HunyuanVideo15Runner  # noqa: F401
 from lightx2v.models.runners.lingbot_video.lingbot_video_runner import LingBotVideoRunner  # noqa: F401
@@ -49,6 +50,17 @@
 from lightx2v_platform.registry_factory import PLATFORM_DEVICE_REGISTER
 
 
+def _str2bool(value):
+    if isinstance(value, bool):
+        return value
+    lowered = value.lower()
+    if lowered in ("1", "true", "yes", "y", "on"):
+        return True
+    if lowered in ("0", "false", "no", "n", "off"):
+        return False
+    raise argparse.ArgumentTypeError(f"Expected a boolean value, got {value!r}")
+
+
 def init_runner(config):
     torch.set_grad_enabled(False)
     runner = RUNNER_REGISTER[config["model_cls"]](config)
@@ -91,6 +103,7 @@ def main():
             "wan2.2_s2v",
             "hunyuan_video_1.5",
             "hunyuan_video_1.5_distill",
+            "hunyuan_image3",
             "hunyuan3d",
             "worldplay_distill",
             "worldplay_ar",
@@ -120,6 +133,110 @@ def main():
     parser.add_argument("--support_tasks", type=str, nargs="+", default=[], help="Set supported tasks for the model")
     parser.add_argument("--model_path", type=str, required=True)
     parser.add_argument("--config_json", type=str, required=True)
+    hunyuan_cache_args = parser.add_argument_group("HunyuanImage3 cache options")
+    hunyuan_cache_args.add_argument("--enable_kv_cache", "--enable-kv-cache", dest="enable_kv_cache", type=_str2bool, nargs="?", const=True, default=None)
+    hunyuan_cache_args.add_argument("--enable_text_kv_cache", "--enable-text-kv-cache", dest="enable_text_kv_cache", type=_str2bool, nargs="?", const=True, default=None)
+    hunyuan_cache_args.add_argument("--use_taylor_cache", "--use-taylor-cache", dest="use_taylor_cache", type=_str2bool, nargs="?", const=True, default=None)
+    hunyuan_cache_args.add_argument("--taylor_cache_interval", "--taylor-cache-interval", dest="taylor_cache_interval", type=int, default=None)
+    hunyuan_cache_args.add_argument("--taylor_cache_order", "--taylor-cache-order", dest="taylor_cache_order", type=int, default=None)
+    hunyuan_cache_args.add_argument(
+        "--taylor_cache_enable_first_enhance",
+        "--taylor-cache-enable-first-enhance",
+        dest="taylor_cache_enable_first_enhance",
+        type=_str2bool,
+        nargs="?",
+        const=True,
+        default=None,
+    )
+    hunyuan_cache_args.add_argument("--taylor_cache_first_enhance_steps", "--taylor-cache-first-enhance-steps", dest="taylor_cache_first_enhance_steps", type=int, default=None)
+    hunyuan_cache_args.add_argument(
+        "--taylor_cache_enable_tailing_enhance",
+        "--taylor-cache-enable-tailing-enhance",
+        dest="taylor_cache_enable_tailing_enhance",
+        type=_str2bool,
+        nargs="?",
+        const=True,
+        default=None,
+    )
+    hunyuan_cache_args.add_argument("--taylor_cache_tailing_enhance_steps", "--taylor-cache-tailing-enhance-steps", dest="taylor_cache_tailing_enhance_steps", type=int, default=None)
+    hunyuan_cache_args.add_argument("--taylor_cache_low_freqs_order", "--taylor-cache-low-freqs-order", dest="taylor_cache_low_freqs_order", type=int, default=None)
+    hunyuan_cache_args.add_argument("--taylor_cache_high_freqs_order", "--taylor-cache-high-freqs-order", dest="taylor_cache_high_freqs_order", type=int, default=None)
+    hunyuan_native_args = parser.add_argument_group("HunyuanImage3 native implementation options")
+    hunyuan_native_args.add_argument("--moe_impl", "--moe-impl", dest="moe_impl", type=str, choices=["eager", "flashinfer"], default=None)
+    hunyuan_native_args.add_argument(
+        "--hunyuan_cfg_mode",
+        "--hunyuan-cfg-mode",
+        dest="hunyuan_cfg_mode",
+        type=str,
+        choices=["batch", "serial", "parallel"],
+        default=None,
+        help="HunyuanImage3 CFG execution mode: batch packs cond/uncond, serial runs two batch=1 forwards, parallel splits cond/uncond by cfg rank.",
+    )
+    hunyuan_native_args.add_argument(
+        "--attn_impl",
+        "--attn-impl",
+        dest="attn_impl",
+        type=str,
+        choices=[
+            "eager",
+            "sdpa",
+            "torch_sdpa",
+            "flash_attention_2",
+            "flash_attn2",
+            "flash_attention_3",
+            "flash_attn3",
+            "sage_attn2",
+            "sage_attn3",
+        ],
+        default=None,
+    )
+    hunyuan_sp_args = parser.add_argument_group("HunyuanImage3 sequence parallel options")
+    hunyuan_sp_args.add_argument(
+        "--hunyuan_sp_size",
+        "--hunyuan-sp-size",
+        dest="hunyuan_sp_size",
+        type=int,
+        default=None,
+    )
+    hunyuan_sp_args.add_argument(
+        "--hunyuan_sp_attn_type",
+        "--hunyuan-sp-attn-type",
+        dest="hunyuan_sp_attn_type",
+        type=str,
+        choices=["kv_all_gather", "kv-all-gather", "ulysses"],
+        default=None,
+    )
+    hunyuan_sp_args.add_argument(
+        "--hunyuan_pipeline_layout",
+        "--hunyuan-pipeline-layout",
+        dest="hunyuan_image3_pipeline_layout",
+        type=str,
+        choices=["interleaved", "contiguous"],
+        default=None,
+    )
+    hunyuan_flashinfer_args = parser.add_argument_group("HunyuanImage3 FlashInfer autotune options")
+    hunyuan_flashinfer_args.add_argument("--flashinfer_autotune", "--flashinfer-autotune", dest="flashinfer_autotune", type=_str2bool, nargs="?", const=True, default=None)
+    hunyuan_flashinfer_args.add_argument(
+        "--flashinfer_autotune_mode",
+        "--flashinfer-autotune-mode",
+        dest="flashinfer_autotune_mode",
+        type=str,
+        choices=["off", "tune", "load"],
+        default=None,
+        help="'tune' profiles missing FlashInfer fused MoE tactics and saves cache; 'load' only loads cache; 'off' disables autotune.",
+    )
+    hunyuan_flashinfer_args.add_argument("--flashinfer_autotune_cache", "--flashinfer-autotune-cache", dest="flashinfer_autotune_cache", type=str, default=None)
+    hunyuan_flashinfer_args.add_argument("--flashinfer_tuning_buckets", "--flashinfer-tuning-buckets", dest="flashinfer_tuning_buckets", type=str, default=None)
+    hunyuan_flashinfer_args.add_argument(
+        "--flashinfer_autotune_round_up",
+        "--flashinfer-autotune-round-up",
+        dest="flashinfer_autotune_round_up",
+        type=_str2bool,
+        nargs="?",
+        const=True,
+        default=None,
+    )
+    hunyuan_flashinfer_args.add_argument("--flashinfer_tune_max_num_tokens", "--flashinfer-tune-max-num-tokens", dest="flashinfer_tune_max_num_tokens", type=int, default=None)
     parser.add_argument("--use_prompt_enhancer", action="store_true")
     parser.add_argument("--warmup", action="store_true", help="Warm up the model before inference. Disabled by default.")
     parser.add_argument("--prompt", type=str, default="", help="The input prompt for text-to-video generation")

lightx2v/models/networks/hunyuan_image3/__init__.py

@@ -0,0 +1,3 @@
+from .model import HunyuanImage3Model
+
+__all__ = ["HunyuanImage3Model"]