1.add scripts for ascend-npu, 2.bugfix for z-image in ascend-npu#1048
1.add scripts for ascend-npu, 2.bugfix for z-image in ascend-npu#1048helloyongyang merged 1 commit intomainfrom
Conversation
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Code Review
This pull request introduces support for the Ascend NPU platform across several models by adding configuration files and inference scripts, alongside a scheduler workaround for NPU-specific indexing limitations. Feedback focuses on optimizing attention kernels for the NPU, correcting a script reference to a missing configuration file, and improving script portability by replacing hardcoded or empty local paths with placeholders.
| "aspect_ratio": "16:9", | ||
| "prompt_template_encode": "<|im_start|>system\nDescribe the image by detailing the color, shape, size, texture, quantity, text, spatial relationships of the objects and background:<|im_end|>\n<|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n", | ||
| "prompt_template_encode_start_idx": 34, | ||
| "attn_type": "flash_attn3", |
There was a problem hiding this comment.
The attn_type is set to flash_attn3, which is typically optimized for NVIDIA Hopper GPUs. For Ascend NPU platforms, you should use npu_flash_attn to leverage the platform-specific optimized kernels.
| "attn_type": "flash_attn3", | |
| "attn_type": "npu_flash_attn", |
| --model_cls wan2.2_moe \ | ||
| --task t2v \ | ||
| --model_path $model_path \ | ||
| --config_json ${lightx2v_path}/configs/platforms/ascend_npu/wan_moe_t2v.json \ |
There was a problem hiding this comment.
The script references wan_moe_t2v.json, but this file is not included in the current pull request. Based on the added files, it seems you might have intended to use wan_ti2v_t2v.json or forgot to include the specific MoE configuration file.
| --config_json ${lightx2v_path}/configs/platforms/ascend_npu/wan_moe_t2v.json \ | |
| --config_json ${lightx2v_path}/configs/platforms/ascend_npu/wan_ti2v_t2v.json \ |
| lightx2v_path= | ||
| model_path= |
There was a problem hiding this comment.
The variables lightx2v_path and model_path are left empty. While these are often environment-specific, providing a placeholder or a default relative path (like in other scripts in this PR) would make the script more usable out-of-the-box.
| lightx2v_path=/data/nvme1/yongyang/ddc/yong/LightX2V | ||
| model_path=/data/nvme1/models/Qwen/Qwen-Image-2512 |
There was a problem hiding this comment.
These paths appear to be very specific to a local environment (/data/nvme1/yongyang/...). It is better to use generic placeholders or relative paths to ensure the script is portable across different systems.
2 participants
No description provided.