after add cache, this version support flashinfer

[Chernobyllight]

No description provided.

[gemini-code-assist]

Code Review

This pull request introduces support for the HunyuanImage3 model, including its architecture, inference logic, and weight loading. It also adds configuration support for cache options and native implementations, along with updated scripts for running the model. The review feedback highlights potential runtime errors when accessing configuration attributes and suggests using safer getattr patterns, as well as replacing hardcoded absolute paths in shell scripts with dynamic path resolution to improve portability. One high-severity issue regarding the detection of runtime diff weights in the MoE implementation was also identified.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[gemini-code-assist]

The condition hasattr(linear, "weight_diff") always evaluates to True because weight_diff is initialized in __init__ as a dummy tensor. This prevents the use of flashinfer even when no runtime diff is active. Using getattr(linear, "has_diff", False) accurately detects active runtime diffs.

Suggested change

	if hasattr(linear, "weight_diff"):
	raise RuntimeError("HunyuanImage3 moe_impl='flashinfer' does not support runtime diff weights.")
	if getattr(linear, "has_diff", False):
	raise RuntimeError("HunyuanImage3 moe_impl='flashinfer' does not support runtime diff weights.")

[gemini-code-assist]

Directly accessing self.hunyuan_config.vae_dtype can raise an AttributeError if the attribute is missing from the configuration. Using getattr with a default value provides a safe fallback.

Suggested change

	vae_dtype = getattr(torch, self.hunyuan_config.vae_dtype, torch.float32)
	vae_dtype_str = getattr(self.hunyuan_config, "vae_dtype", "float32")
	vae_dtype = getattr(torch, vae_dtype_str, torch.float32)

[gemini-code-assist]

Directly accessing self.hunyuan_config.vae_autocast_dtype can raise an AttributeError if the attribute is missing from the configuration. Using getattr with a default value provides a safe fallback.

        vae_autocast_dtype_str = getattr(self.hunyuan_config, "vae_autocast_dtype", "float16")
        autocast_dtype = getattr(torch, vae_autocast_dtype_str, torch.float16)

[gemini-code-assist]

Directly accessing self.hunyuan_config.vae_autocast_dtype can raise an AttributeError if the attribute is missing from the configuration. Using getattr with a default value provides a safe fallback.

        vae_autocast_dtype_str = getattr(self.hunyuan_config, "vae_autocast_dtype", "float16")
        autocast_dtype = getattr(torch, vae_autocast_dtype_str, torch.float16)

[gemini-code-assist]

The script uses hardcoded absolute paths specific to a local environment. Replacing them with dynamic path resolution or placeholder paths makes the script portable across different environments.

Suggested change

	export lightx2v_path="/data/nvme0/lhd_codes/LightX2V"
	export model_path="/data/nvme0/lhd_codes/HunyuanImage-3.0-instruct/HunyuanImage-3-Instruct"
	export lightx2v_path=${lightx2v_path:-$(cd "$(dirname "${BASH_SOURCE[0]}")"/../.. && pwd)}
	export model_path=${model_path:-/path/to/HunyuanImage-3-Instruct}

[gemini-code-assist]

The script uses hardcoded absolute paths specific to a local environment. Replacing them with dynamic path resolution or placeholder paths makes the script portable across different environments.

Suggested change

	export lightx2v_path="${lightx2v_path:-/data/nvme0/lhd_codes/LightX2V}"
	export model_path="${model_path:-/data/nvme0/lhd_codes/HunyuanImage-3.0-instruct/HunyuanImage-3-Instruct}"
	export lightx2v_path=${lightx2v_path:-$(cd "$(dirname "${BASH_SOURCE[0]}")"/../.. && pwd)}
	export model_path=${model_path:-/path/to/HunyuanImage-3-Instruct}

[gemini-code-assist]

The script uses hardcoded absolute paths specific to a local environment. Replacing them with dynamic path resolution or placeholder paths makes the script portable across different environments.

Suggested change

	lightx2v_path="/data/nvme0/lhd_codes/LightX2V"
	model_path="/data/nvme0/models/Wan-AI/Wan2.1-I2V-14B-720P"
	lightx2v_path=${lightx2v_path:-$(cd "$(dirname "${BASH_SOURCE[0]}")"/../.. && pwd)}
	model_path=${model_path:-/path/to/Wan2.1-I2V-14B-720P}

[gemini-code-assist]

The script uses hardcoded absolute paths specific to a local environment. Replacing them with dynamic path resolution or placeholder paths makes the script portable across different environments.

Suggested change

	lightx2v_path="/data/nvme0/lhd_codes/LightX2V"
	model_path="/data/nvme0/models/Wan-AI/Wan2.1-T2V-1.3B"
	lightx2v_path=${lightx2v_path:-$(cd "$(dirname "${BASH_SOURCE[0]}")"/../.. && pwd)}
	model_path=${model_path:-/path/to/Wan2.1-T2V-1.3B}