refactor compiler (#301)

scripts/qwen_image/qwen_image_i2i.sh

@@ -27,7 +27,6 @@ export TOKENIZERS_PARALLELISM=false
 export PYTHONPATH=${lightx2v_path}:$PYTHONPATH
 export DTYPE=BF16
 export PROFILING_DEBUG_LEVEL=2
-export ENABLE_GRAPH_MODE=false
 
 
 python -m lightx2v.infer \

scripts/qwen_image/qwen_image_t2i.sh

@@ -27,7 +27,6 @@ export TOKENIZERS_PARALLELISM=false
 export PYTHONPATH=${lightx2v_path}:$PYTHONPATH
 export DTYPE=BF16
 export PROFILING_DEBUG_LEVEL=2
-export ENABLE_GRAPH_MODE=false
 
 
 python -m lightx2v.infer \

scripts/qwen_image/qwen_image_t2i_offload.sh

@@ -27,7 +27,6 @@ export TOKENIZERS_PARALLELISM=false
 export PYTHONPATH=${lightx2v_path}:$PYTHONPATH
 export DTYPE=BF16
 export PROFILING_DEBUG_LEVEL=2
-export ENABLE_GRAPH_MODE=false
 
 
 python -m lightx2v.infer \

scripts/seko_talk/run_seko_talk_01_base.sh

@@ -10,7 +10,6 @@ source ${lightx2v_path}/scripts/base/base.sh
 
 export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
 
-export ENABLE_GRAPH_MODE=false
 export SENSITIVE_LAYER_DTYPE=None
 
 python -m lightx2v.infer \

scripts/seko_talk/run_seko_talk_02_fp8.sh

@@ -10,7 +10,6 @@ source ${lightx2v_path}/scripts/base/base.sh
 
 export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
 
-export ENABLE_GRAPH_MODE=false
 export SENSITIVE_LAYER_DTYPE=None
 
 python -m lightx2v.infer \

scripts/seko_talk/run_seko_talk_03_dist.sh

@@ -10,7 +10,6 @@ source ${lightx2v_path}/scripts/base/base.sh
 
 export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
 
-export ENABLE_GRAPH_MODE=false
 export SENSITIVE_LAYER_DTYPE=None
 
 torchrun --nproc-per-node 8 -m lightx2v.infer \

scripts/seko_talk/run_seko_talk_04_fp8_dist.sh

@@ -10,7 +10,6 @@ source ${lightx2v_path}/scripts/base/base.sh
 
 export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
 
-export ENABLE_GRAPH_MODE=false
 export SENSITIVE_LAYER_DTYPE=None
 
 torchrun --nproc-per-node 8 -m lightx2v.infer \

scripts/seko_talk/run_seko_talk_05_offload_fp8_4090.sh

@@ -10,7 +10,6 @@ source ${lightx2v_path}/scripts/base/base.sh
 
 export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
 
-export ENABLE_GRAPH_MODE=false
 export SENSITIVE_LAYER_DTYPE=None
 
 python -m lightx2v.infer \

scripts/seko_talk/run_seko_talk_06_offload_fp8_H100.sh

@@ -10,7 +10,6 @@ source ${lightx2v_path}/scripts/base/base.sh
 
 export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
 
-export ENABLE_GRAPH_MODE=false
 export SENSITIVE_LAYER_DTYPE=None
 
 python -m lightx2v.infer \

scripts/seko_talk/run_seko_talk_07_dist_offload.sh

@@ -10,7 +10,6 @@ source ${lightx2v_path}/scripts/base/base.sh
 
 export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
 
-export ENABLE_GRAPH_MODE=false
 export SENSITIVE_LAYER_DTYPE=None
 
 torchrun --nproc-per-node 4 -m lightx2v.infer \

scripts/seko_talk/run_seko_talk_08_5B_base.sh

@@ -10,7 +10,6 @@ source ${lightx2v_path}/scripts/base/base.sh
 
 export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
 
-export ENABLE_GRAPH_MODE=false
 export SENSITIVE_LAYER_DTYPE=None
 
 python -m lightx2v.infer \

scripts/seko_talk/run_seko_talk_09_base_fixed_min_area.sh

 #!/bin/bash
 
-lightx2v_path=
-model_path=
+lightx2v_path=/path/to/Lightx2v
+model_path=/path/to/SekoTalk-Distill
 
 export CUDA_VISIBLE_DEVICES=0
 
@@ -10,7 +10,6 @@ source ${lightx2v_path}/scripts/base/base.sh
 
 export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
 
-export ENABLE_GRAPH_MODE=false
 export SENSITIVE_LAYER_DTYPE=None
 
 python -m lightx2v.infer \

scripts/seko_talk/run_seko_talk_10_fp8_dist_fixed_min_area.sh

 #!/bin/bash
 
-lightx2v_path=
-model_path=
+lightx2v_path=/path/to/Lightx2v
+model_path=/path/to/SekoTalk-Distill-fp8
 
 export CUDA_VISIBLE_DEVICES=0,1,2,3
 
@@ -10,7 +10,6 @@ source ${lightx2v_path}/scripts/base/base.sh
 
 export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
 
-export ENABLE_GRAPH_MODE=false
 export SENSITIVE_LAYER_DTYPE=None
 
 torchrun --nproc-per-node 4 -m lightx2v.infer \

scripts/seko_talk/run_seko_talk_11_fp8_dist_fixed_shape.sh

 #!/bin/bash
 
-lightx2v_path=
-model_path=
+lightx2v_path=/path/to/Lightx2v
+model_path=/path/to/SekoTalk-Distill-fp8
 
 export CUDA_VISIBLE_DEVICES=0,1,2,3
 
@@ -10,7 +10,6 @@ source ${lightx2v_path}/scripts/base/base.sh
 
 export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
 
-export ENABLE_GRAPH_MODE=false
 export SENSITIVE_LAYER_DTYPE=None
 
 torchrun --nproc-per-node 4 -m lightx2v.infer \

scripts/seko_talk/run_seko_talk_12_fp8_dist_fixed_shape_8gpus_1s.sh

@@ -10,7 +10,6 @@ source ${lightx2v_path}/scripts/base/base.sh
 
 export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
 
-export ENABLE_GRAPH_MODE=false
 export SENSITIVE_LAYER_DTYPE=None
 
 torchrun --nproc-per-node 8 -m lightx2v.infer \

scripts/seko_talk/run_seko_talk_13_fp8_dist_bucket_shape_8gpus_5s_realtime.sh

@@ -10,7 +10,6 @@ source ${lightx2v_path}/scripts/base/base.sh
 
 export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
 
-export ENABLE_GRAPH_MODE=false
 export SENSITIVE_LAYER_DTYPE=None
 
 torchrun --nproc-per-node 8 -m lightx2v.infer \

scripts/seko_talk/run_seko_talk_14_fp8_dist_bucket_shape_8gpus_1s_realtime.sh

@@ -10,7 +10,6 @@ source ${lightx2v_path}/scripts/base/base.sh
 
 export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
 
-export ENABLE_GRAPH_MODE=false
 export SENSITIVE_LAYER_DTYPE=None
 
 torchrun --nproc-per-node 8 -m lightx2v.infer \

scripts/seko_talk/run_seko_talk_15_base_compile.sh

+#!/bin/bash
+
+lightx2v_path=/path/to/Lightx2v
+model_path=/path/to/SekoTalk-Distill
+
+export CUDA_VISIBLE_DEVICES=0
+
+# set environment variables
+source ${lightx2v_path}/scripts/base/base.sh
+
+export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
+
+export SENSITIVE_LAYER_DTYPE=None
+
+python -m lightx2v.infer \
+--model_cls seko_talk \
+--task i2v \
+--model_path $model_path \
+--config_json ${lightx2v_path}/configs/seko_talk/seko_talk_15_base_compile.json \
+--prompt  "The video features a male speaking to the camera with arms spread out, a slightly furrowed brow, and a focused gaze." \
+--negative_prompt 色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走 \
+--image_path ${lightx2v_path}/assets/inputs/audio/seko_input.png \
+--audio_path ${lightx2v_path}/assets/inputs/audio/seko_input.wav \
+--save_video_path ${lightx2v_path}/save_results/output_lightx2v_seko_talk.mp4

scripts/seko_talk/run_seko_talk_16_fp8_dist_compile.sh

+#!/bin/bash
+
+lightx2v_path=/path/to/Lightx2v
+model_path=/path/to/SekoTalk-Distill-fp8
+
+export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+
+# set environment variables
+source ${lightx2v_path}/scripts/base/base.sh
+
+export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
+
+export SENSITIVE_LAYER_DTYPE=None
+
+torchrun --nproc-per-node 8 -m lightx2v.infer \
+--model_cls seko_talk \
+--task i2v \
+--model_path $model_path \
+--config_json ${lightx2v_path}/configs/seko_talk/seko_talk_16_fp8_dist_compile.json \
+--prompt  "The video features a male speaking to the camera with arms spread out, a slightly furrowed brow, and a focused gaze." \
+--negative_prompt 色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走 \
+--image_path ${lightx2v_path}/assets/inputs/audio/seko_input.png \
+--audio_path ${lightx2v_path}/assets/inputs/audio/seko_input.wav \
+--save_video_path ${lightx2v_path}/save_results/output_lightx2v_seko_talk.mp4

scripts/server/start_server.sh

@@ -9,7 +9,6 @@ export CUDA_VISIBLE_DEVICES=0
 # set environment variables
 source ${lightx2v_path}/scripts/base/base.sh
 
-export ENABLE_GRAPH_MODE=false
 
 # Start API server with distributed inference service
 python -m lightx2v.api_server \